SYSTEM IDENTIFICATION with MATLAB.

Linear Models
Marvin L.
What Are Model Objects?
In this section...
Model Objects Represent Linear Systems on page 1-3 About Model Data on page 1-3

Model Objects Represent Linear Systems

In Control System Toolbox, System Identification Toolbox, and Robust Control Toolbox software,
you represent linear systems as model objects. Model objects are specialized data containers that
encapsulate model data and other attributes in a structured way. Model objects allow you to manipulate
linear systems as single entities rather than keeping track of multiple data vectors, matrices, or cell
arrays.
Model objects can represent single-input, single-output (SISO) systems or multiple-input, multipleoutput (MIMO) systems. You can represent both continuous- and discrete-time linear systems.
The main families of model objects are:
Numeric Models Basic representation of linear systems with fixed numerical coefficients. This
family also includes identified models that have coefficients estimated with System Identification
Toolbox software.
Generalized Models Representations that combine numeric coefficients with tunable or uncertain
coefficients. Generalized models support tasks such as parameter studies or compensator tuning using
Robust Control Toolbox tuning commands.

About Model Data

The data encapsulated in your model object depends on the model type you use. For example:
Transfer functions store the numerator and denominator coefficients
State-space models store the A, B, C,and D matrices that describe the dynamics of the system
PID controller models store the proportional, integral, and derivative gains
Other model attributes stored as model data include time units, names for the model inputs or outputs,
and time delays.
Note All model objects are MATLABobjects, but working with them does not require a background in
object-oriented programming. To learn more about objects and object syntax, see Classes in the
MATLAB Language in the MATLAB documentation.

Concepts Types of Model Objects on page 1-5

Types of Model Objects
The following diagram illustrates the relationships between the types of model objects in Control
System Toolbox, Robust Control Toolbox, and System Identification Toolbox software. Model types that
begin with idrequire System Identification Toolbox software. Model types that begin with urequire
Robust Control Toolbox software. All other model types are available with Control System Toolbox
software.
GENERALIZED MODELS Control Design Blocks
NUMERIC MODELS Numeric
LTI Models
tf
zpk
ss
frd
pid
pidstd

Identified LTI Models

idgrey
idtf
idss
idfrd
idpoly
idproc

Identified
Nonlinear Models
idnlarx
idnlgrey
idnlhw

Dynamic
Control Design Blocks
Tunable
ltiblock.tf
ltiblock.ss
ltiblock.gain
ltiblock.pid
ltiblock.pid2
Uncertain
ultidyn
udyn

Switch Block
loopswitch

Static
Control Design Blocks Tunable
realp
Uncertain
ureal
ucomplex
ucomplexm

Generalized LTI Models

Tunable and/or Uncertain genss
genfrd
Uncertain Only
uss
ufrd

Generalized Matrices Tunable

genmat
Uncertain umat

DYNAMIC SYSTEM MODELS STATIC MODELS

The diagram illustrates the following two overlapping broad classifications of model object types:
Dynamic System Models vs. Static Models In general, Dynamic System Models represent
systems that have internal dynamics, while Static Models represent static input/output relationships.
Numeric Models vs. Generalized Models Numeric Models are the basic numeric representation of
linear systems with fixed coefficients. Generalized Models represent systems with tunable or uncertain
components.

Concepts What Are Model Objects? on page 1-3

Dynamic System Models on page 1-7
Numeric Models on page 1-9

Dynamic System Models

Dynamic System Models generally represent systems that have internal dynamics or memory of past
states (such as integrators).
Most commands for analyzing linear systems, such as bode, margin,and ltiview, work on most Dynamic
System Model objects. For Generalized Models, analysis commands use the current value of tunable
parameters and the nominal value of uncertain parameters.
ThefollowingtableliststheDynamicSystemModels.
Model Family

Numeric LTI models Basic numeric representation of linear systems

Identified LTI models
Representations of linear systems with tunable coefficients, whose values can be identified using
measured input/output data.
Identified nonlinear models Representations of nonlinear systems with tunable coefficients, whose
values can be identified using input/output data. Limited support for commands that analyze linear
systems.
Model Types tf
zpk
ss
frd
pid
pidstd
idtf
idss
idfrd
idgrey
idpoly
idproc
idnlarx
idnlhw
idnlgrey
Model Family
Generalized LTI models Representations of systems that includetunableoruncertain coefficients
Dynamic Control Design Blocks Tunable, uncertain, or switch components for constructing models
of control systems
Model Types genss
genfrd
uss
ufrd
ltiblock.gain ltiblock.tf ltiblock.ss ltiblock.pid ltiblock.pid2 ultidyn
udyn
loopswitch

Concepts Numeric Linear Time Invariant (LTI) Models on page 1-9

Identified LTI Models on page 1-9
Identified Nonlinear Models on page 1-10

Numeric Models
Numeric Linear Time Invariant (LTI) Models
Numeric LTI models are the basic numeric representation of linear systems or components of linear
systems. Use numeric LTI models for modeling dynamic components, such as transfer functions or
state-space models, whose coefficients are fixed, numeric values. You can use numeric LTI models for
linear analysis or control design tasks.
The following table summarizes the available types of numeric LTI models.
Model Type Description
tf Transfer function model in polynomial form
zpk Transfer function model in zero-pole-gain (factorized) form
ss State-space model
frd Frequency response data model pid Parallel-form PID controller
pidstd Standard-form PID controller

Identified LTI Models

Identified LTI Models represent linear systems with coefficients that are identified using measured input/
output data. You can specify initial values and constraints for the estimation of the coefficients.
Thefollowing table summarizes the available types of identified LTI models.
Model Type Description
idtf Transfer function model in polynomial form, with identifiable parameters
idss State-space model, with identifiable parameters idpoly Polynomial input-output model, with
identifiable parameters
Model Type Description
idproc Continuous-time process model, with identifiable parameters
idfrd Frequency-response model, with identifiable parameters
idgrey Linear ODE (grey-box) model, with identifiable parameters

Identified Nonlinear Models

Identified Nonlinear Models represent nonlinear systems with coefficients that are identified using
measured input/output data. You can specify initial values and constraints for the estimation of the
coefficients.
The following table summarizes the available types of identified nonlinear models.
Model Type idnlarx
idnlgrey
idnlhw
Description
Nonlinear ARX model, with identifiable parameters

Nonlinear ODE (grey-box) model, with identifiable parameters Hammerstein-Wiener model, with
identifiable parameters

About Identified Linear Models

In this section...
What are IDLTI Models? on page 1-11
Measured and Noise Component Parameterizations on page 1-12 Linear Model Estimation on page
1-16

What are IDLTI Models?

System Identification Toolbox software uses objects to represent a variety of linear and nonlinear model
structures. These linear model objects are collectively known as Identified Linear Time-Invariant
(IDLTI) models.
IDLTI models contain two distinct dynamic components:
Measured component Describes the relationship between the measured inputs and the measured
output (G)
Noise component Describes the relationship between the disturbances at the output and the
measured output (H)
Models that only have the noise component Hare called time-series or signal models. Typically, you
create such models using time-series data that consist of one or more outputs y(t)with no corresponding
input.
The total output is the sum of the contributions from the measured inputs and the disturbances:
y=Gu+He,where u represents the measured inputs and e the disturbance. e(t) is modeled as zero-mean
Gaussian white noise with variance . The following figure illustrates an IDLTI model.
e

IDLTIH
u

G
y
WhenyousimulateanIDLTImodel,youstudytheeffectofinput u(t) (and possibly initial conditions) on the
output y(t). The noise e(t) is not considered. However, with finite-horizon prediction of the output, both
the measured and the noise components of the model contribute towards computation of the (predicted)
response.

H-1G

y_predicted

y_measured1-H-1
One-step ahead prediction model corresponding to a linear identified model (y = Gu+He)

Measured and Noise Component Parameterizations

The various linear model structures provide different ways of parameterizing the transfer functions Gand
H. When you construct an IDLTI model or estimate a model directly using input-output data, you can
configure the structure of both G and H, as described in the following table: State Represents an
identified space state-space model structure, model governed by the equations: (idss)
xAx Bu Ke
yCx Du e
where the transfer function between the measured
input u and output y is
Gs CsI A B D1 and
the noise transfer function is Hs CsI A K I1 .
Polynomial Represents a polynomial model model such as ARX, ARMAX and BJ. (idpoly) An
ARMAX model, for example,
uses the input-output equation Ay(t) = Bu(t)+Ce(t), so that the measured transfer function
G is Gs A B1 ,whilethe noise transfer function is Hs A C1 .
The ARMAX model is a special configuration of the general polynomial model whose governing
equation is:
Construction: Use idssto create a model, specifying values of state-space matrices A, B, C, D and K as
input arguments (using NaNsto denote unknown entries).
Estimation: Use ssest
or n4sid,specifying
name-value pairs for various configurations, such as,
canonical parameterization of the measured dynamics ('Form'/'canonical'),
denoting absence of
feedthrough by fixing D to zero ('Feedthrough'/false), and absence of noise
dynamics by fixing K to zero ('DisturbanceModel'/'none').

Construction: Use idpolyto create a model using values of active polynomials as input arguments. For
example, to create an Output-Error model which usesG= B/Fas the measured component and has a
trivial noise component (H= 1). enter:
y = idpoly([],B,[],[],F)
Estimation: Use the armax, arx,or bj, specifying the orders
BC
Ay t() Fut() Det()
The autoregressive component, A, is common between
the measured and noise
components. The polynomials B and F constitute the
measured component while the polynomials C and D constitute the noise component.
of the polynomials as input arguments. For example, bj requires you to specify the orders of the B, C,
D,and F polynomials to construct a model with governing equation
BC
yt() Fut() Det()
Transfer function model
(idtf)
Represents an identified transfer function model, which has no dynamic elements to model noise
behavior. This object uses the trivial noise model H(s) = I. The governing equation is
num
yt()
denut et() ()
Process model
(idproc)
Represents a process model, which provides options to represent the noise dynamics as either first- or
second-order ARMA process (that is, H(s)= C(s)/A(s),where C(s) and A(s) are monic polynomials of
equal degree). The measured component, G(s),isrepresented Construction: Use idtfto create a model,
specifying values of the numerator
and denominator coefficients as input arguments. The numerator and denominator vectors constitute the
measured componentG= num(s)/den(s).Thenoise component is fixed toH= 1.
Estimation: Use tfest,
specifying the number of poles and zeros of the measured component G.
For process (and grey-box) models, the noise component is often treated as an on-demand
extensiontoanotherwise measured component-centric representation. For these models, you can add a
noise component by using the
by a transfer function expressed in pole-zero form.

DisturbanceModelestimation option. For example:

model = procest(data,'P1D')
estimates a model whose equation is:
1
sTd () ().
ys K Ts euses
() ()
p1 1

To add a second order noise component to the model, use:

Options = procestOptions(`Disturbance model = procest(data, `P1D', Options)
This model has the equation:
1sTd () 1 cs
ys K Ts eus1 dses()
() ()p1 1 1
where the coefficients c1 and
d1 parameterize the noise
component of the model. If
you are constructing a process
model using the idproc
command, specify the structure
of the measured component
using the Typeinput argument
and the noise component by
using the NoiseTFname-value
pair. For example,
model=idproc('P1','Kp',1,'Tp1',1,'Noi
creates the process model y(s) = 1/(s+1) u(s) + (s + 0.1)/(s + 0.5) e(s)
Sometimes, fixing coefficients or specifying bounds on the parameters are not sufficient. For example,
you may have unrelated parameter dependencies in the model or parameters may be a function of a
different set of parameters that you want to identify exclusively. For example, in a mass-spring-damper
system, the Aand Bparameters both depend on the mass of the system. To achieve such parameterization
of linear models, you can use grey-box modeling where you establish the link between the actual
parameters and model coefficients by writing an ODE file. To learn more, see Grey-Box Model
Estimation.

Linear Model Estimation

You typically use estimation to create models in System Identification Toolbox. You execute one of the
estimation commands, specifying as input arguments the measured data, along with other inputs
necessary to define the structure of a model. To illustrate, the following example uses the state space
estimation command, ssest, to create a state space model. The first input argument dataspecifies the
measured input-output data. The second input argument specifies the order of the model.
sys = ssest(data, 4)

The estimation function treats the noise variable e(t) as prediction error the residual portion of the
output that cannot be attributed to the measured inputs. All estimation algorithms work to minimize a
weighted norm of e(t) over the span of available measurements. The weighting function is defined by
the nature of the noise transfer function H and the focus of estimation, such as simulation or prediction
error minimization.
Black Box (Cold Start) Estimation on page 1-17
Structured Estimations on page 1-17
Estimation Options on page 1-18
Estimation Report on page 1-19
Black Box (Cold Start) Estimation
In a black-box estimation, you only have to specify the order to configure the structure of the model.
sys = estimator(data, orders)
where estimatoris the name of an estimation command to use for the desired model type.
For example, you use tfestto estimate transfer function models, arxfor ARX-structure polynomial
models, and procestfor process models.
The first argument, data, is time- or frequency domain data represented as an iddataor idfrdobject. The
second argument, orders, represents one or more numbers whose definitions depends upon the model
type:
For transfer functions, ordersrefers to the number of poles and zeros.
For state-space models, ordersis a scalar that refers to the number of states.
For process models, ordersis a string denoting the structural elements of a process model, such as, the
number of poles and presence of delay and integrator.
When working with the GUI, you specify the orders in the appropriate edit fields of corresponding
model estimation dialogs.
Structured Estimations
In some situations, you want to configure the structure of the desired model more closely than what is
achieved by simply specifying the orders. In such cases, you construct a template model and configure
its properties. You then pass that template model as an input argument to the estimation commands in
place of orders.
To illustrate, the following example assigns initial guess values to the numerator and the denominator
polynomials of a transfer function model, imposes minimum and maximum bounds on their estimated
values, and then passes the object to the estimator function.
% initial guess for numerator
num = [1 2] den = [1 2 1 1]
% initial guess for the denominator
sys = idtf(num, den);
% set min bound on den coefficients to 0.1 sys.Structure.den.Minimum = [1 0.1 0.1 0.1]; sysEstimated =
tfest(data, sys);

The estimation algorithm uses the provided initial guesses to kick-start the estimation and delivers a
model that respects the specified bounds.
You can use such a model template to also configure auxiliary model properties such as input/output
names and units. If the values of some of the models parameters are initially unknown, you can use
NaNsfor them in the template.
Estimation Options
There are many options associated with a models estimation algorithm that configure the estimation
objective function, initial conditions and numerical search algorithm, among other things. For every
estimation command, estimator, there is a corresponding option command named estimatorOptions. To
specify options for a particular estimator command, such as tfest, use the options command that
corresponds to the estimation command, in this case, tfestOptions. The options command returns an
options set that you then pass as an input argument to the corresponding estimation command.
For example, to estimate an Output-Error structure polynomial model, you use oe.Tospecify
simulationas the focus and lsqnonlinas the search method, you use oeOptions:
load iddata1 z1
Options = oeOptions('Focus','simulation','SearchMethod','lsqnonlin'); sys= oe(z1, [2 2 1], Options);
For information about how to view the list of options used to create an estimated model, see Estimation
Report on page 1-19.
Estimation Report
The model returned by the estimation command contains information about the estimation operation in
the Reportproperty. Information includes:
Information on initial conditions used for estimation
Termination conditions for iterative estimation algorithms under a category called Termination
Final prediction error (FPE), fit % and mean-square error (MSE) under a category called Fit
The nature of estimation data under a category called DataUsed
All estimated quantities - parameter values and their covariance, initial state values for state-space
models and values of input level for process models under a category called Parameters
The option set used for estimation under a category called OptionsUsed
To illustrate, this example displays the contents of the Report.OptionsUsed property of the model
estimated in Estimation Options on page 1-18.
sys.Report
Status: 'Estimated using POLYEST with Focus = "prediction"' Method: 'OE'
InitialCondition: 'zero'
Fit: [1x1 struct]
Parameters: [1x1 struct]
OptionsUsed: [1x1 idoptions.polyest]

RandState: [1x1 struct]

DataUsed: [1x1 struct]
Termination: [1x1 struct]
Exploring one of the categories provides additional data:
sys.Report.Fit shows:
FitPercent: 70.5666 FPE: 1.7558 LossFcn: 1.7094 MSE: 1.6865

Linear Model Structures

In this section...
About System Identification Toolbox Model Objects on page 1-21 Available Linear Models on page
1-22
When to Construct a Model Structure Independently of Estimation on page 1-24
Commands for Constructing Model Structures on page 1-25 Model Properties on page 1-26
See Also on page 1-29

About System Identification Toolbox Model Objects

Objects are instances of model classes. Each class is a blueprint that defines the following information
about your model:
How the object stores data
Which operations you can perform on the object
This toolbox includes nine classes for representing models. For example, idpolyrepresents linear inputoutput polynomial models, and idss represents linear state-space models. For a complete list of available
model objects, see Available Linear Models on page 1-22 and Available Nonlinear Models on page
1-32.
Model properties define how a model object stores information. Model objects store information about a
model, including the mathematical form of a model, names of input and output channels, units, names
and values of estimated parameters, parameter uncertainties, algorithm specifications, and estimation
information. For example, an idpolymodel has an InputName property for storing one or more input
channel names. Different model objects have different properties.
The allowed operations on an object are called methods.InSystem Identification Toolbox software, some
methods have the same name but apply to multiple model objects. For example, bodecreates a bode plot
for all linear model objects. However, other methods are unique to a specific model object. For example,
the estimation method canonis unique to the state-space idssmodel.
Every class has a special method for creating objects of that class, called the constructor. Using a
constructor creates an instance of the corresponding class or instantiates the object. The constructor
name is the same as the class name. For example, idpolyis both the name of the class representing linear
black-box polynomial models and the name of the constructor for instantiating the model object.

Available Linear Models

A linear model is often sufficient to accurately describe the system dynamics and, in most cases, you
should first try to fit linear models. Available linear structures include transfer functions and state-space
models, summarized in the following table.
Model Type Usage Learn More Transfer function (idtf)
Use this structure to represent transfer functions:
Identifying Transfer Function Models on page 3-113
num
y
denue
where num and den are polynomials of arbitrary lengths. You can specify initial guesses for, and
estimate,num, den, and transport delays.
Process model ( idproc) Use this structure to represent process models that are low order transfer
functions expressed in pole-zero form. They include Identifying Process Models on page 3-26
integrator, delay, zero, and up to 3 poles.
State-space model Use this structure (idss) to represent known state-space structures and black-box
structures. You can fix certain parameters to known values and estimate the remaining parameters. You
can also prescribe minimum/maximum bounds on the values of thefreeparameters. If you need to
specify parameter dependencies or parameterize the state-space matrices using your own parameters, use
a grey-box model.
Polynomial models Usetorepresentlinear (idpoly) transfer functions based on the general form inputoutput polynomial form: Identifying
Input-Output
Polynomial Models on page 3-45
BC
Ay =+
FuDe
where A, B, C, D and F are polynomials with coefficients that the toolbox estimates from data.
Identifying
State-Space Models on page 3-79
Typically, you begin modeling using simpler forms of this generalized structure (such as
ARX: AyBu and=+
B
OE: y=+Fue )and,if necessary, increase the model complexity.

Grey-box model Usetorepresent (idgrey) arbitrary

parameterizations of state-space models. For example, you can use this structure to represent your
ordinary differential or difference equation (ODE) and to define parameter dependencies. Estimating
Linear Grey-Box Models on page 5-6

When to Construct a Model Structure Independently of Estimation

You use model constructors to create a model object at the command line by specifying all required
model properties explicitly.
You must construct the model object independently of estimation when you want to:
Simulate or analyze the effect of models parameters on its response, independent of estimation.
Specify an initial guess for specific model parameter values before estimation. Specify bounds on
parameter values, or set up the auxiliary model information in advance, or both. Auxiliary model
information includes specifying input/output names, units, Notes, user data, etc.
In most cases, you can use the estimation commands to both construct and estimate the modelwithout
having to construct the model object independently. For example, the estimation command tfestcreates a
transfer function model using data and some information on the order of the model - the number of poles
and zeros. For information about how to both construct and estimate models with a single command, see
Model Estimation Commands on page 1-40.
In case of grey-box models, you must always construct the model object first and then estimate the
parameters of the ordinary differential or difference equation.

Commands for Constructing Model Structures

The following table summarizes the model constructors available in the System Identification Toolbox
product for representing various types of models.
After model estimation, you can recognize the corresponding model objects in the MATLAB Workspace
browser by their class names. The name of the constructor matches the name of the object it creates.
For information about how to both construct and estimate models with a single command, see Model
Estimation Commands on page 1-40.
Summary of Model Constructors
Model Constructor idfrd
idproc
Resulting Model Class
Nonparametric frequency-response model.
Continuous-time, low-order transfer functions (process models).
Summary of Model Constructors (Continued)
Model Constructor idpoly

Resulting Model Class

Linear input-output polynomial models:
idss
idtf
idgrey
idnlgrey
idnlarx
idnlhw
ARX
ARMAX
Output-Error
Box-Jenkins
Linear state-space models.
Linear transfer function models.
Linear ordinary differential or difference equations (grey-box models). You write a function that
translates user parameters to state-space matrices. Can also be viewed as state-space models with userspecified parameterization.
Nonlinear ordinary differential or difference equation (grey-box models). You write a function or MEXfile to represent the governing equations.
Nonlinear ARX models, which define the predicted output as a nonlinear function of past inputs and
outputs.
Nonlinear Hammerstein-Wiener models, which include a linear dynamic system with nonlinear static
transformations of inputs and outputs.
For more information about when to use these commands, see When to Construct a Model Structure
Independently of Estimation on page 1-24.

Model Properties
Categories of Model Properties on page 1-27
Viewing Model Properties and Estimated Parameters on page 1-28
Categories of Model Properties
The way a model object stores information is defined by the properties of the corresponding model
class.
Each model object has properties for storing information that are relevant only to that specific model
type. The idtf, idgrey, idpoly, idproc,and idssmodel objects are based on the idltisuperclass and inherit
all idlti properties.

Similarly, the nonlinear models idnlarx, idnlhw,and idnlgreyare based on the idnlmodelsuperclass and
inherit all idnlmodelproperties.
In general, all model objects have properties that belong to the following categories:
Names of input and output channels, such as InputNameand OutputName
Sampling interval of the model, such as Ts
Units for time or frequency
Model order and mathematical structure (for example, ODE or nonlinearities)
Properties that store estimation results (Report)
User comments, such as Notesand Userdata
Estimation algorithm information (nonlinear models only)
- Algorithm
Structure includes fields that specify the estimation method. Algorithm includes another structure, called
Advanced, which provides additional flexibility for setting the search algorithm. Different fields apply to
different estimation techniques.
For linear parametric models, Algorithmspecifies the frequency weighing of the estimation using the
Focusproperty.
For information about getting help on object properties, see the model reference pages.
Viewing Model Properties and Estimated Parameters The following table summarizes the commands
for viewing and changing model property values. Property names are not case sensitive. You do not need
to type the entire property name if the first few letters uniquely identify the property.
Task
View all model
properties and their values
Access a specific model
property Use dot notation
Command get
Example
Load sample data, compute an ARX model, and list the model properties:
load iddata8
m_arx=arx(z8,[4 3 2 3 0 0 0]); get(m_arx)
View the A matrix containing the estimated parameters in the previous model:
Access model uncertainty information For properties, such as Report,that
are configured like structures, use dot notation of the form model.PropertyName.FieldName.
FieldNameis the name of any field of the property.

getpvecand
polydata
m_arx.a ViewthemethodusedinARXmodel estimation:
m_arx.Report.Method
View the A polynomial and 1 standard uncertainty of an ARX model:
[a,~,~,~,~,da] = polydata(m_arx)
Task Command
Change setor dot notation model
property
values
Example
Change the input delays for all three input channels to[111]for an ARX model:
set(m_arx,'InputDelay',[1 1 1])
or m_arx.InputDelay = [1 1 1]

See Also
Validate each model directly after estimation to help fine-tune your modeling strategy. When you do not
achieve a satisfactory model, you can try a different model structure and order, or try another
identification algorithm. For more information about validating and troubleshooting models, see
Validating Models After Estimation on page 8-3.

Imposing Constraints on Model Parameter Values

All identified linear (IDLTI) models, except idfrd,containa Structure property. The Structureproperty
contains the adjustable entities (parameters) of the model. Each parameter has attributes such as value,
minimum/maximum bounds, and free/fixed status that allow you to constrain them to desired values or a
range of values during estimation. You use the Structureproperty to impose constraints on the values of
various model parameters.
The Structureproperty contains the essential parameters that define the structure of a given model:
For identified transfer functions, includes the numerator, denominator and delay parameters
For polynomial models, includes the list of active polynomials
For state-space models, includes the list of state-space matrices
For information about other model types, see the model reference pages.
For example, the following example constructs an idtfmodel, specifying values for the numand
denparameters:
num =[12];
den= [122];
sys = idtf(num,den)

You can update the value of the numand denproperties after you create the object as follows:
new_den = [1 1 10]; sys.den = new_den;
To fix the denominator to the value you specified (treat its coefficients as fixed parameters), use the
Structureproperty of the object as follows:
sys.Structure.den.Value = new_den; sys.Structure.den.Free = false(1,3);
For a transfer function model, the num, den,and ioDelaymodel properties are simply pointers to the
Valueattribute of the corresponding parameter in the Structureproperty.
IDTF Model Properties Parameters
num double vector
den double vector
ioDelay scalar num:
Value:
Minimum: Maximum: Free:
Scale:
Info:
double vector double vector double vector logical vector double vector struct
Structure
InputDelay scalar
Ts scalar
. . den:
Value:
Minimum: Maximum: Free:
Scale:
Info:
..
. . ioDelay:
Value:
Minimum: Maximum:
Free:
Scale:
Info:
Similar relationships exist for other model structures. For example, the a property of a state-space model
contains the double value of the state matrix. It is an alias to the Aparameter value stored in
Structure.a.Value.

Available Nonlinear Models

System Identification Toolbox provides several nonlinear black-box model structures, which have
traditionally been useful for representing dynamic systems.
Model Type
Nonlinear ARX model (idnlarxobject)
Linear models
with input/output nonlinearities, or Hammerstein-Wiener model
(idnlhwobject)
Nonlinear grey-box model
(idnlgreyobject)
Usage
Usetorepresent
nonlinear extensions of linear models. This structure allows you to model complex
nonlinear behavior using flexible nonlinear functions,suchas
wavelet andsigmoid networks.
Usetorepresentlinear models with static
nonlinearities.
Learn More
Identifying Nonlinear ARX Models on page 4-8
Identifying
Hammerstein-Wiener Modelson page 4-48
Usetorepresent
nonlinear ODEs with unknown parameters. Estimating Nonlinear Grey-Box Models on page 5-17

Recommended Model Estimation Sequence

System identification is an iterative process, where you identify models with different structures from
data and compare model performance. You start by estimating the parameters of simple model
structures. If the model performance is poor, you gradually increase the complexity of the model
structure. Ultimately, you choose the simplest model that best describes the dynamics of your system.
Another reason to start with simple model structures is that higher-order models are not always more
accurate. Increasing model complexity increases the uncertainties in parameter estimates and typically
requires more data (which is common in the case of nonlinear models).

Note Model structure is not the only factor that determines model accuracy. If your model is poor, you
might need to preprocess your data by removing outliers or filtering noise. For more information, see
Ways to Prepare Data for System Identification on page 2-6.
Estimate impulse-response and frequency-response models first to gain insight into the system dynamics
and assess whether a linear model is sufficient. Then, estimate parametric models in the following order:
Transfer function, ARX polynomial and state-space models provide the simplest structures. Estimation
of ARX and state-space models let you determine the model orders.
1

In the System Identification Tool GUI .ChoosetoestimatetheTransfer function models, ARX

polynomial models and the state-space model using the n4sidmethod.
At thecommandline.Usethe tfest, arx,andthe n4sidcommands, respectively.
For more information, see Identifying Input-Output Polynomial Models on page 3-45 and Identifying
State-Space Models on page 3-79.
ARMAX and BJ polynomial models provide more complex structures and require iterative estimation.
Try several model orders and keep the model ordersaslowaspossible.
2

In the System Identification Tool GUI.SelecttoestimatetheBJand ARMAX polynomial models.

At the command line.Usethe bjor armaxcommands.
For more information, see Identifying Input-Output Polynomial Models on page 3-45.
3Nonlinear ARX or Hammerstein-Wiener models provide nonlinear structures. For more information,
see Nonlinear Model Identification.
For general information about choosing you model strategy, see Overview. For information about
validating models, see Validating Models After Estimation on page 8-3.

Supported Models for Time- and Frequency-Domain Data

In this section...
Supported Models for Time-Domain Data on page 1-35 Supported Models for Frequency-Domain
Data on page 1-36 See Also on page 1-37

Supported Models for Time-Domain Data

Continuous-Time Models
You can directly estimate the following types of continuous-time models:
Transfer function models. See Identifying Transfer Function Models on page 3-113.
Process models. See Identifying Process Models on page 3-26.
State-space models. See Identifying State-Space Models on page 3-79.
You can also use d2cto convert an estimated discrete-time model into a continuous-time model.
Discrete-Time Models
You can estimate all linear and nonlinear models supported by the System Identification Toolbox
product as discrete-time models, except process models, which are defined only in continuous-time..

ODEs (Grey-Box Models)

You can estimate both continuous-time and discrete-time models from time-domain data for linear and
nonlinear differential and difference equations.
Nonlinear Models
You can estimate discrete-time Hammerstein-Wiener and nonlinear ARX models from time-domain
data.
You can also estimate nonlinear grey-box models from time-domain data. See Estimating Nonlinear
Grey-Box Models on page 5-17.

Supported Models for Frequency-Domain Data

There are two types of frequency-domain data:
Frequency response data
Frequency domain input/output signals which are Fourier Transforms of the corresponding time domain
signals.
The data is considered continuous-time if its sample time (Ts)is 0,andis considered discrete-time if the
sample time is nonzero.
Continuous-Time Models
You can estimate the following types of continuous-time models directly:
Transfer function models using continuous- or discrete-time data. See Identifying Transfer Function
Models on page 3-113.
Process models using continuous- or discrete-time data. See Identifying Process Models on page
3-26.
Input-output polynomial models of output-error structure using continuous time data. See Identifying
Input-Output Polynomial Models on page 3-45.
State-space models using continuous- or discrete-time data. From continuous-time frequency-domain
data, you can only estimate continuous-time models.
You can also use d2cto convert an estimated discrete-time model into a continuous-time model.
Discrete-Time Models
You can estimate all linear model types supported by the System Identification Toolbox product as
discrete-time models, except process models, which are defined in continuous-time only. For estimation
of discrete-time models, you mustusediscrete-timedata.
The noise component of a model cannot be estimated using frequency domain data, with the exception
of ARX models. Thus, the K matrix of an identified state-space model, the noise component, is zero. An
identified polynomial model has output-error (OE) or ARX structure; BJ/ARMAX or other polynomial
structure with nontrivial values of C or D polynomials cannot be estimated.
ODEs (Grey-Box Models)
For linear grey-box models, you can estimate both continuous-time and discrete-time models from
frequency-domain data. The noise component of the model, the K matrix, cannot be estimated using
frequency domain data; it remains fixed to 0.
Nonlinear grey-box models are supported only for time-domain data.

Nonlinear Black-Box Models

Nonlinear black box (nonlinear ARX and Hammerstein-Wiener models) cannot be estimated using
frequency domain data.

See Also
Supported Continuous- and Discrete-Time Models on page 1-38

Supported Continuous- and Discrete-Time Models

For linear and nonlinear ODEs (grey-box models), you can specify any ordinary differential or
difference equation to represent your continuous-time or discrete-time model in state-space form,
respectively. In the linear case, both time-domain and frequency-domain data are supported. In the
nonlinear case, only time-domain data is supported.
For black-box models, the following tables summarize supported continuous-time and discrete-time
models.
Supported Continuous-Time Models
Model Type
Identifying Transfer Function Models on page 3-113
Description
Estimate continuous-time transfer function models directly using tfestfrom either time- and frequencydomain data. If you estimated a discrete-time transfer function model from time-domain data, then use
d2cto transform it into a continuous-time model.
Low-order transfer functions (process models)
Linear input-output polynomial models
Estimate low-order process models for up to three free poles from either time- or frequency-domain
data.
To get a linear, continuous-time model of arbitrary structure from time-domain data, you can estimate a
discrete-time model, and then use d2cto transform it into a continuous-time model.
You can estimate only polynomial models of Output Error structure using continuous-time frequency
domain data.. Other structures that include noise models, such as Box-Jenkins (BJ) and ARMAX, are
not supported for frequency-domain data.
Supported Continuous-Time Models (Continued)
Model Type
State-space models
Description
Estimate continuous-time state-space models directly using the estimation commands from either timeand frequency-domain data.
If you estimated a discrete-time state-space model from time-domain data, then use d2cto transform it
into a continuous-time model.

Linear ODEs (grey-box) models

Nonlinear ODEs (grey-box) models
If the MATLAB file returns continuous-time model matrices, then estimate the ordinary differential
equation (ODE) coefficients using either time- or frequency-domain data.
If the MATLAB file returns continuous-time output and state derivative values, estimate arbitrary
differential equations (ODEs) from time-domain data.
Supported Discrete-Time Models
Model Type
Linear, input-output polynomial models
Description
Estimate arbitrary-order, linear parametric models from time- or frequency-domain data.
To get a discrete-time model, your data sampling interval must be set to the (nonzero) value you used to
sample in
your experiment.
Nonlinear Model Identification Estimate from time-domain data only. Linear ODEs (grey-box)
models
Nonlinear ODEs (grey-box) models
If the MATLAB file returns discrete-time model matrices, thenestimate ordinary difference equation
coefficients from time-domain or discrete-time frequency-domain data.
If the MATLAB file returns discrete-time output and state update values, estimate ordinary difference
equations from time-domain data.

Model Estimation Commands

In most cases, a model can be created by using a model estimation command on a dataset. For example,
SSEST(DATA, Nx) creates a continuous-time state-space model of order Nx using the input/output of
frequency response data DATA.
Note For ODEs (grey-box models), you must first construct the model structure and then apply an
estimation command (either greyestor pem)to the resulting model object.
The following table summarizes System Identification Toolbox estimation commands. For detailed
information about using each command, see the corresponding reference page.
Commands for Constructing and Estimating Models
Model Type
Transfer function models Process models
Linear input-output polynomial models
State-space models

Time-series models
Nonlinear ARX models Hammerstein-Wiener models
Estimation Commands
tfest
procest
armax (ARMAX only)
arx(ARX only)
bj(BJ only)
iv4(ARX only)
oe(OE only)
polyest(for all models)
n4sid
ssest
ar
arx(for multiple outputs)
ivar
nlarx(for nonlinear time-series models)
nlarx
nlhw

Modeling Multiple-Output Systems

In this section...
About Modeling Multiple-Output Systems on page 1-41 Modeling Multiple Outputs Directly on
page 1-42
Modeling Multiple Outputs as a Combination of Single-Output Models on page 1-42
Improving Multiple-Output Estimation Results by Weighing Outputs During Estimation on page 1-43

About Modeling Multiple-Output Systems

You can estimate multiple-output model directly using all the measured inputs and outputs, or you can
try building models for subsets of the most important input and output channels. To learn more about
each approach, see:
Modeling Multiple Outputs Directly on page 1-42
Modeling Multiple Outputs as a Combination of Single-Output Models on page 1-42
Modeling multiple-output systems is more challenging because input/output couplings require additional
parameters to obtain a good fit and involve more complex models. In general, a model is better when
more data inputs are included during modeling. Including more outputs typically leads to worse
simulation results because it is harder to reproduce the behavior of several outputs simultaneously.

If you know that some of the outputs have poor accuracy and should be less important during estimation,
you can control how much each output is weighed in the estimation. For more information, see
Improving Multiple-Output Estimation Results by Weighing Outputs During Estimation on page 1-43.

Modeling Multiple Outputs Directly

You can perform estimation with linear and nonlinear models for multiple-outputdata.
Tip Estimating multiple-output state-space models directly generally produces better results than
estimating other types of multiple-output models directly.

Modeling Multiple Outputs as a Combination of Single-Output Models

You may findthat it is harder for a single model to explain the behavior of several outputs. If you get a
poor fit estimating a multiple-output model directly,you can try building models for subsets of the most
important input and outputchannels.
Use this approach when no feedback is present in the dynamic system and there areno couplings
between the outputs. If you are unsure about the presenceof feedback, see How to Analyze Data Using
the advice Command on page 2-94.
To construct partial models, use subreferencing to create partial data sets, such that each data set
contains all inputs and one output. For more information about creating partial data sets, see the
following sections in the SystemIdentification Toolbox Users Guide:
For working in the System Identification Tool GUI, see Creating Data Sets from aSubset of Signal
Channels on page 2-37.
For working at the command line, see the Select Data Channels, I/O Data and Experiments in iddata
Objects on page 2-63.
After validating the single-output models, use vertical concatenation to combine these partial models
into a single multiple-output model. For more information about concatenation, see Increasing Number
of Channels or Data Points of iddata Objects on page 2-67 or Adding Input or Output Channels in
idfrd Objects on page 2-80.
You can try refining the concatenated multiple-output model using the original (multiple-output) data
set.

Improving Multiple-Output Estimation Results by Weighing Outputs During

Estimation
When estimating linear and nonlinear black-box models for multiple-output systems, you can control the
relative importance of output channels during the estimation process. The ability to control how much
each output is weighed during estimation is useful when some of the measured outputs have poor
accuracy or should be treated as less important during estimation. For example, if you have already
modeled one output well, you might want to focus the estimation on modeling the remaining outputs.
Similarly, you mightwanttorefineamodelforasubsetofoutputs.

For linear models, use the OutputWeightestimation to indicate the desired output weighting. If you set
this option to 'noise', an automatic weighting, equal to the inverse of the estimated noise variance, is
used for model estimation. You can also specify a custom weighting matrix, which must be a positive
semi-definite matrix.
Note
The OutputWeightoption is not available for polynomial models, except ARX models, since their
estimation algorithm estimates the parameters oneoutputatatime.
Transfer function (idtf) and process models (idproc)ignore OutputWeight when they contain nonzero or
free transport delays. In the presence of delays, the estimation is carried out one output at a time.
For nonlinear models, you can specify output weights directly in the estimation command using the
Criterionand Weightingestimation options. You must set the Criterionfield to Trace, and set the
Weightingfield to the matrix that contains the output weights. The Tracecriterion minimizes the
weighted sum of the prediction errors using the weights specified by Weighting.
For more information about the OutputWeight estimation option, see arxOptions, ssestOptions,
tfestOptions, procestOptions,etc. Formore information about the Algorithmfields for nonlinear
estimation, see the idnlarxand idnlhwreference pages.
Note For multiple-output idnlarxmodels containing neuralnetor treepartitionnonlinearity estimators,
output weighting is ignored because each output is estimated independently.

Data Import and Processing

Supported Data on page 2-3
Ways to Obtain Identification Data on page 2-5
Ways to Prepare Data for System Identification on page 2-6
Requirements on Data Sampling on page 2-8
Representing Data in MATLAB Workspace on page 2-9
Importing Data into the GUI on page 2-17
Representing Time- and Frequency-Domain Data Using iddata Objects on page 2-55
Representing Frequency-Response Data Using idfrd Objects on page 2-76
Analyzing Data Quality on page 2-85
Selecting Subsets of Data on page 2-96
Handling Missing Data and Outliers on page 2-100
Handling Offsets and Trends in Data on page 2-104
How to Detrend Data Using the GUI on page 2-108
How to Detrend Data at the Command Line on page 2-109
Resampling Data on page 2-111
Resampling Data Using the GUI on page 2-117
Resampling Data at the Command Line on page 2-118

Filtering Data on page 2-120

How to Filter Data Using the GUI on page 2-122
How to Filter Data at the Command Line on page 2-126
Generating Data Using Simulation on page 2-130
Transforming Between Time- and Frequency-Domain Data on page 2-134
Manipulating Complex-Valued Data on page 2-144

Supported Data
System Identification Toolbox software supports estimation of linear models from both time- and
frequency-domain data. For nonlinear models, this toolbox supports only time-domain data. For more
information, see Supported Models for Time- and Frequency-Domain Data on page 1-35.
The data can have single or multiple inputs and outputs, and can be either real or complex.
Your time-domain data should be sampled at discrete and uniformly spaced time instants to obtain an
input sequence
u={u(T),u(2T),...,u(NT)}
and a corresponding output sequence
y={y(T),y(2T),...,y(NT)}
u(t) and y(t) are the values of the input and output signals at time t, respectively.
This toolbox supports modeling both single- or multiple-channel input-output data or time-series data.
Supported Data Time-domain I/O data
Time-series data
Description
One or more input variables u(t) and one or more output variables y(t), sampled as a function of time.
Time-domain data can be either real or complex
Contains one or more outputs y(t) and no measured input. Can be time-domain or frequency-domain
data.
Supported Data Description
Frequency-domain data Fourier transform of the input and output time-domain signals. The data is the
set of input and output signals in frequency domain; the frequency grid need not be uniform.
Frequency-response data Complex frequency-response values for a linear system characterized by its
transfer function G, measurable directly using a spectrum analyzer. Also called frequency function
data.Representedby frdor idfrd objects. The data sample time may be zero or nonzero. The frequency
vector need not be uniformly spaced.
Note If your data is complex valued, see Manipulating Complex-Valued Data on page 2-144 for
information about supported operations for complex data.

Ways to Obtain Identification Data

You can obtain identification data by:
Measuring input and output signals from a physical system.
Your data must capture the important system dynamics, such as dominant time constants. After
measuring the signals, organize the data into variables, as described in Representing Data in MATLAB
Workspace on page 2-9. Then, import it in the System Identification Tool GUI or represent it as a data
object for estimating models at the command line.
Generating an input signal with desired characteristics, such as a random Gaussian or binary signal or a
sinusoid, using idinput. Then, generate an output signal using this input to simulate a model with known
coefficients. For more information, see Generating Data Using Simulation on page 2-130.
Using input/output data thus generated helps you study the impact of input signal characteristics and
noise on estimation.
Logging signals from Simulinkmodels.
This technique is useful when you want to replace complex components in your model with identified
models to speed up simulations or simplify control design tasks. For more information on how to log
signals, see Export Signal Data Using Signal Logging in the Simulink documentation.

Ways to Prepare Data for System Identification

Before you can perform any task in this toolbox, your data must be in the MATLAB workspace. You can
import the data from external data files or manually create data arrays at the command line. For more
information about importing data, see Representing Data in MATLAB Workspace on page 2-9.
The following tasks help to prepare your data for identifying models from data:
Represent data for system identification
You can represent data in the format of this toolbox by doing one of the following:
For working in the GUI, import data into the System Identification Tool GUI.
See Importing Data into the GUI on page 2-17.
For working at the command line, create an iddataor idfrdobject.
For time-domain or frequency-domain data, see Representing Time- and Frequency-Domain Data
Using iddata Objects on page 2-55. For frequency-response data, see Representing FrequencyResponse Data Using idfrd Objects on page 2-76.
To simulate data with and without noise, see Generating Data Using Simulation on page 2-130.
Analyze data quality
You can analyze your data by doing either of the following:
Plotting data to examine both time- and frequency-domain behavior. See Analyzing Data Quality on
page 2-85.
Using the advicecommandtoanalyzethedataforthepresenceofconstant offsets and trends, delay, possible
feedback, and signal excitation levels. SeeHowtoAnalyzeDataUsingtheadviceCommandonpage2-94.
Preprocess data

Review the data characteristics for any of the following features to determine if there is a need for
preprocessing:
Missing or faulty values (also known as outliers). For example, you might see gaps that indicate
missing data, values that do not fit with the rest of the data, or noninformative values.
See Handling Missing Data and Outliers on page 2-100.
Offsets and drifts in signal levels (low-frequency disturbances).
See Handling Offsets and Trends in Data on page 2-104 for information about subtracting means and
linear trends, and Filtering Data on page 2-120 for information about filtering.
High-frequency disturbances above the frequency interval of interest for the system dynamics.
See Resampling Data on page 2-111 for information about decimating and interpolating values, and
Filtering Data on page 2-120 for information about filtering.
Select a subset of your data
You can use data selection as a way to clean the data and exclude parts with noisy or missing
information. You can also use data selection to create independent data sets for estimation and
validation.
To learn more about selecting data, see Selecting Subsets of Data on page 2-96.
Combine data from multiple experiments
You can combine data from several experiments into a single data set. The model you estimate from a
data set containing several experiments describes the average system that represents these experiments.
To learn more about creating multiple-experiment data sets, see Creating Multiexperiment Data Sets in
the GUI on page 2-39 or Creating Multiexperiment Data at the Command Line on page 2-61.

Requirements on Data Sampling

A sampling intervalisthetimebetweensuccessivedatasamples. Itis sometimes also referred to as sampling
time or sample time.
The System Identification Tool GUI only supports uniformly sampled data.
The System Identification Toolbox product provides limited support for nonuniformly sampled data. For
more information about specifying uniform and nonuniform time vectors, see Constructing an iddata
Object for Time-Domain Data on page 2-56.

Representing Data in MATLAB Workspace

In this section...
Time-Domain Data Representation on page 2-9 Time-Series Data Representation on page 2-10
Frequency-Domain Data Representation on page 2-11

Time-Domain Data Representation

Time-domain data consists of one or more input variables u(t) and one or more output variables y(t),
sampled as a function of time. If there is no input variable, see Time-Series Data Representation on
page 2-10.
You must organize time-domain input/output data in the following format:
For single-input/single-output (SISO) data, the sampled data values must be double column vectors.
For multi-input/multi-output (MIMO) data with Nuinputs and Nyoutputs, and Nsnumber of data samples
(measurements):
- The input data must be an Ns-byNumatrix
- The output data must be an Ns-byNymatrix
To use time-domain data for identification, you must know the sampling interval. If you are working
with uniformly sampled data, use the actual sampling interval from your experiment. Each data value is
assigned a time instant, which is calculated from the start time and sampling interval. You can work with
nonuniformly sampled data only at the command line by specifying a vector of time instants using the
SamplingInstantsproperty of iddata, as described in Constructing an iddata Object for Time-Domain
Data on page 2-56.
For continuous-time models, you must also know the input intersample behavior, such as zero-order
hold and first-order-hold.
For more information about importing data into MATLAB, see Data Import and Export.
After you have the variables in the MATLAB workspace, import them into the System Identification
Tool GUI or create a data object for working at the command line. For more information, see Importing
Time-Domain Data into the GUI on page 2-18 and Representing Time- and Frequency-Domain Data
Using iddata Objects on page 2-55.

Time-Series Data Representation

Time-series data is time-domain or frequency-domain data that consist of one or more outputs y(t) with
no corresponding input. For more information on how to obtain identification data, see Ways to Obtain
Identification Data on page 2-5.
You must organize time-series data in the following format:
For single-input/single-output (SISO) data, the output data values must be a column vector.
For data with Nyoutputs, the output is an Ns-byNymatrix, where Nsis the number of output data samples
(measurements).
To use time-series data for identification, you also need the sampling interval. If you are working with
uniformly sampled data, use the actual sampling interval from your experiment. Each data value is
assigned a sample time, which is calculated from the start time and the sampling interval. If you are
working with nonuniformly sampled data at the command line, you can specify a vector of time instants
using the iddata SamplingInstants property, as described in Constructing an iddata Object for TimeDomain Data on page 2-56. Note that model estimation cannot be performed using non-uniformly
sampled data.

For more information about importing data into the MATLAB workspace, see Data Import and
Export.
After you have the variables in the MATLAB workspace, import them into the System Identification
Tool GUI or create a data object for working at the command line. For more information, see Importing
Time-Domain Data into the GUI on page 2-18 and Representing Time- and Frequency-Domain Data
Using iddata Objects on page 2-55.
For information about estimating time-series model parameters, see Time-Series Model Identification.

Frequency-Domain Data Representation

Frequency-domain data consists of either transformed input and output time-domain signals or system
frequency response sampled as a function of the independent variable frequency.
Frequency-Domain Input/Output Signal Representation on page 2-11
Frequency-Response Data Representation on page 2-13
Frequency-Domain Input/Output Signal Representation
What Is Frequency-Domain Input/Output Signal? on page 2-11
How to Represent Frequency-Domain Data in MATLAB on page 2-12
What Is Frequency-Domain Input/Output Signal?. Frequency-domain data is the Fourier transform
of the input and output time-domain signals. For continuous-time signals, the Fourier transform over the
entire time axis is defined as follows:
Yiw() ()yte
Uiw() ()ute f iwtdt

f
f

iwtdt

In the context of numerical computations, continuous equations are replaced by their discretized
equivalents to handle discrete data values. For a discrete-time system with a sampling interval T, the
frequency-domain output Y(eiw) and input U(eiw) is the time-discrete Fourier transform (TDFT):
N Ye
k=1

iwT

= () ()T ykTeiwkT

In this example, k = 1,2,...,N,where Nis the number of samples in the sequence.

Note This form only discretizes the time. The frequency is continuous.

In practice, the Fourier transform cannot be handled for all continuous frequencies and you must specify
a finite number of frequencies. The discrete Fourier transform (DFT) of time-domain data for N equally
spaced frequencies between 0 and the sampling frequency 2/N is:
N Ye
iw T=() ()nniw kT

ykTe

k=1

2
n
wn ==

TnN1

The DFT is useful because it can be calculated very efficiently using the fast Fourier transform (FFT)
method. Fourier transforms of the input and output data are complex numbers.
For more information on how to obtain identification data, see Ways to Obtain Identification Data on
page 2-5.
How to Represent Frequency-Domain Data in MATLAB.You must organize frequency-domain data
in the following format:
Input and output
- For single-input/single-output (SISO) data:
The input data must be a column vector containing the values
ue
ikT

Z()
The output data must be a column vector containing the values
ye

ikT

Z()
k=1, 2, ..., Nf,where Nfis the number of frequencies.
- For multi-input/multi-output data with Nuinputs, Nyoutputs and Nffrequency measurements:
The input data must be an Nf-byNumatrix
The output data must be an Nf-byNymatrix
Frequencies
- Must be a column vector.
For more information about importing data into the MATLAB workspace, see Data Import and
Export.

After you have the variables in the MATLAB workspace, import them into the System Identification
Tool GUI or create a data object for working at the command line. For more information, see Importing
Frequency-Domain Input/Output Signals into the GUI on page 2-23 and Representing Timeand
Frequency-Domain Data Using iddata Objects on page 2-55.
Frequency-Response Data Representation
What Is Frequency-Response Data? on page 2-13
How to Represent Frequency-Response Data in MATLAB on page 2-15
What Is Frequency-Response Data?. Frequency-response data,alsocalled frequency-function data,
consists of complex frequency-response values for a linear system characterized by its transfer function
G. Frequency-response data tells you how the system handles sinusoidal inputs. You can measure
frequency-response data values directly using a spectrum analyzer, for example, which provides a
compact representation of the input-output relationship (compared to storing input and output
independently). The transfer function G is an operator that takes the input u of a linear system to the
output y:
=yGu
For a continuous-time system, the transfer function relates the Laplace transforms of the input U(s) and
output Y(s):
Ys GsUs= () ()
In this case, the frequency function G(iw) is the transfer function evaluated on the imaginary axis s=iw.
For a discrete-time system sampled with a time interval T,thetransfer function relates the Z-transforms
of the input U(z) and output Y(z):
Yz GzU z= () ()
In this case, the frequency function G(eiwT) is the transfer function G(z) evaluated on the unit circle. The
argument of the frequency function G(eiwT) is scaled by the sampling interval T to make the frequency
function periodic
with the sampling frequency 2T.
When the input to the system is a sinusoid of a specific frequency, the output is also a sinusoid with the
same frequency. The amplitude of the output is G times the amplitude of the input. The phase of the
shifted from the input by = arg G. G is evaluated at the frequency of the input sinusoid.
Frequency-response data represents a (nonparametric) model of the relationship between the input and
the outputs as a function of frequency. You might use such a model, which consists of a table or plot of
values, to study the system frequency response. However, this model is not suitable for simulation and
prediction. You should create parametric model from the frequency-response data.
For more information on how to obtain identification data, see Ways to Obtain Identification Data on
page 2-5.
How to Represent Frequency-Response Data in MATLAB.You can represent frequency-response
data in two ways:
Complex-values G(ei) , for given frequencies
Amplitude G and phase shift = arg Gvalues

You can import both the formats directly in the System Identification Tool GUI. At the command line,
you must represent complex data using an frd or idfrdobject. If the data is in amplitude and phase
format, convert it to complex frequency-response vector using h() = A()ej().
You must organize frequency-response data in the following format:
Rrequency-ResponseFor Single-Input
Single-Output (SISO) Data Data
Representation Complex Values
For Multi-Input Multi-Output (MIMO) Data
Frequency function must be a column vector.
Frequency values must be a column vector.
Amplitude and phase shift values
Amplitude and phase must each be a column vector.
Frequency values must be a column vector.
Frequency function must be an Ny-byNu-byNfarray, where
N uis the number of inputs, Ny

is the number of outputs, and Nfis the number of frequency measurements.

Frequency values must be a column vector.

Amplitude and phase must each be an Ny-byNu-byNfarray, where
N uis the number of inputs, Ny

is the number of outputs, and Nfis the number of frequency measurements.

Frequency values must be a column vector.

For more information about importing data into the MATLAB workspace, see Data Import and
Export.
After you have the variables in the MATLAB workspace, import them into the System Identification
Tool GUI or create a data object for working at the command line. For more information about
importing data into the GUI, see Importing Frequency-Response Data into the GUI on page 2-26. To
learn more about creating a data object, see Representing Frequency-Response Data Using idfrd
Objects on page 2-76.

Importing Data into the GUI

In this section...
Types of Data You Can Import into the GUI on page 2-17 Importing Time-Domain Data into the
GUI on page 2-18 Importing Frequency-Domain Data into the GUI on page 2-22 Importing Data
Objects into the GUI on page 2-30 Specifying the Data Sampling Interval on page 2-34 Specifying
Estimation and Validation Data on page 2-35 Preprocessing Data Using Quick Start on page 2-36
Creating Data Sets from a Subset of Signal Channels on page 2-37 Creating Multiexperiment Data
Sets in the GUI on page 2-39 Managing Data in the GUI on page 2-46

Types of Data You Can Import into the GUI

You can import the following types of data from the MATLAB workspace into the System Identification
Tool GUI:
Importing Time-Domain Data into the GUI on page 2-18
Importing Frequency-Domain Input/Output Signals into the GUI on page 2-23
Importing Frequency-Response Data into the GUI on page 2-26
Importing Data Objects into the GUI on page 2-30
To open the GUI, type the following command in the MATLAB Command Window:
ident
In the Import data list, select the type of data to import from the MATLAB workspace, as shown in the
following figure.

For
an example of importing data into the System Identification Tool GUI, see the Getting Started
documentation.

Importing Time-Domain Data into the GUI

Before you can import time-domain data into the System Identification Tool GUI, you must import the
data into the MATLAB workspace, as described in Time-Domain Data Representation on page 2-9.
Note Your time-domain data must be sampled at equal time intervals. The input and output signals must
have the same number of data samples. To import data into the GUI:
1Type the following command in the MATLAB Command Window to open the GUI:
ident

In the System Identification Tool window, select Import data > Time domain data. This action opens
the Import Data dialog box.
2

Specify the following options:

Note For time series, only import the output signal and enter []for the input.
Input Enter the MATLAB variable name (column vector or matrix) or a MATLAB expression that
represents the input data. The expression must evaluate to a column vector or matrix.
Output Enter the MATLAB variable name (column vector or matrix) or a MATLAB expression
that represents the output data. The expression must evaluate to a column vector or matrix.
Data name Enter the name of the data set, which appears in the System Identification Tool window
after the import operation is completed.
Starting timeEnterthestartingvalueofthetimeaxisfortimeplots.
Sampling interval Enter the actual sampling interval in the experiment. For more information
about this setting, see Specifying the Data Sampling Interval on page 2-34.
Tip The System Identification Toolbox product uses the sampling interval during model estimation and
to set the horizontal axis on time plots. If you transform a time-domain signal to a frequency-domain
signal, the Fourier transforms are computed as discrete Fourier transforms (DFTs) using this sampling
interval.
4(Optional) In the Data Information area, click More to expand the dialog
box and enter the following settings:
Input Properties
InterSample This options specifies the behavior of the input signals between samples during data
acquisition. It is used when transforming models from discrete-time to continuous-time and when
resampling the data.
zoh(zero-order hold) indicates that the input was piecewise-constant during data acquisition.
foh(first-order hold) indicates that the output was piecewise-linear during data acquisition.
bl(bandwidth-limited behavior) specifies that the continuous-time input signal has zero power above
the Nyquist frequency (equal to the inverse of the sampling interval).

Note See the d2cand c2dreference pages for more information about transforming between discrete-time
and continuous-time models.
PeriodEnter Infto specify a nonperiodic input. If the underlying time-domain data was periodic over
an integer number of periods, enter the period of the input signal.
Note If your data is periodic, always include a whole number of periods for model estimation.
Channel Names
InputEnterastringtospecifythenameofoneormoreinput channels.
Tip Naming channels helps you to identify data in plots. For multivariable input-output signals, you can
specify the names of individual Input and Output channels, separated by commas.
OutputEnterastringtospecifythenameofoneormoreoutput channels.
Physical Units of Variables
Input Enter a string to specify the input units.
Tip When you have multiple inputs and outputs, enter a comma-separated list of Input and Output
units corresponding to each channel.
Output Enter a string to specify the output units.
Notes Enter comments about the experiment or the data. For example, you might enter the
experiment name, date, and a description of experimental conditions. Models you estimate from this
data inherit your data notes.
5Click
6Click

Import. This action adds a new data icon to the System Identification Tool window.
Close to close the Import Data dialog box.

Importing Frequency-Domain Data into the GUI

Importing Frequency-Domain Input/Output Signals into the GUI on page 2-23
Importing Frequency-Response Data into the GUI on page 2-26
Importing Frequency-Domain Input/Output Signals into the GUI
Frequency-domain data consists of Fourier transforms of time-domain data (a function of frequency).
Before you can import frequency-domain data into the System Identification Tool GUI, you must import
the data into the MATLAB workspace, as described in Frequency-Domain Input/Output Signal
Representation on page 2-11.
Note The input and output signals must have the same number of data samples.
To import data into the GUI:
1Type the following command in the MATLAB Command Window to open the GUI:
ident

In the System Identification Tool window, select Import data > Freq. domain data. This action
opens the Import Data dialog box.
3Specify the following options:
2

Input Enter the MATLAB variable name (column vector or matrix) or a MATLAB expression that
represents the input data. The expression must evaluate to a column vector or matrix.
Output Enter the MATLAB variable name (column vector or matrix) or a MATLAB expression
that represents the output data. The expression must evaluate to a column vector or matrix.
Frequency Enter the MATLAB variable name of a vector or a MATLAB expression that represents
the frequencies. The expression must evaluate to a column vector.
The frequency vector must have the same number of rows as the input and output signals.
Data name Enter the name of the data set, which appears in the System Identification Tool window
after the import operation is completed.
Frequency unitEnter Hzfor Hertz or keep the rad/sdefault value.
Sampling interval Enter the actual sampling interval in the experiment. For continuous-time data,
enter 0. For more information about this setting, see Specifying the Data Sampling Interval on page
2-34.
4(Optional)

In the Data Information area, click More to expand the dialog box and enter the following
optional settings:
Input Properties
InterSample This options specifies the behavior of the input signals between samples during data
acquisition. It is used when transforming models from discrete-time to continuous-time and when
resampling the data.
zoh(zero-order hold) indicates that the input was piecewise-constant during data acquisition.
foh(first-order hold) indicates that the output was piecewise-linear during data acquisition.
bl(bandwidth-limited behavior) specifies that the continuous-time input signal has zero power above
the Nyquist frequency (equal to the inverse of the sampling interval).
Note See the d2cand c2dreference page for more information about transforming between discrete-time
and continuous-time models.
PeriodEnter Infto specify a nonperiodic input. If the underlying time-domain data was periodic over
an integer number of periods, enter the period of the input signal.
Note If your data is periodic, always include a whole number of periods for model estimation.
Channel Names
InputEnterastringtospecifythenameofoneormoreinput channels.

Tip Naming channels helps you to identify data in plots. For multivariable input and output signals, you
can specify the names of individual Input and Output channels, separated by commas.
OutputEnterastringtospecifythenameofoneormoreoutput
Input Enter a string to specify the input units.
Tip When you have multiple inputs and outputs, enter a comma-separated list of Input and Output
units corresponding to each channel.
Output Enter a string to specify the output units.
Notes Enter comments about the experiment or the data. For example, you might enter the
experiment name, date, and a description of experimental conditions. Models you estimate from this
data inherit your data notes.
5Click
6Click

Import. This action adds a new data icon to the System Identification Tool window.
Close to close the Import Data dialog box.
Importing Frequency-Response Data into the GUI
Prerequisite on page 2-26
Importing Complex-Valued Frequency-Response Data on page 2-26
Importing Amplitude and Phase Frequency-Response Data on page 2-28
Prerequisite. Before you can import frequency-response data into the System Identification Tool GUI,
you must import the data into the MATLAB workspace, as described in Frequency-Response Data
Representation on page 2-13.
Importing Complex-Valued Frequency-Response Data. To import frequency-response data consisting
of complex-valued frequency values at specified frequencies:
1Type

the following command in the MATLAB Command Window to open the GUI:
ident
2In the System Identification Tool window, select Import data > Freq. domain data. This action opens
the Import Data dialog box.
3In the Data Format for Signals list, select Freq. Function (Complex).
4Specify the following options:
Freq. Func. EntertheMATLABvariablenameoraMATLAB expression that represents the complex
frequency-response data G(eiw).
Frequency Enter the MATLAB variable name of a vector or a MATLAB expression that represents
the frequencies. The expression must evaluate to a column vector.
Data name Enter the name of the data set, which appears in the System Identification Tool window
after the import operation is completed.
Frequency unitEnter Hzfor Hertz or keep the rad/sdefault value.

 Sampling interval Enter the actual sampling interval in the experiment. For continuous-time data,
enter 0. For more information about this setting, see Specifying the Data Sampling Interval on page
2-34.
5(Optional)

In the Data Information area, click More to expand the dialog box and enter the following
optional settings:
Channel Names
InputEnterastringtospecifythenameofoneormoreinput channels.
Tip Naming channels helps you to identify data in plots. For multivariable input and output signals, you
can specify the names of individual Input and Output channels, separated by commas.
OutputEnterastringtospecifythenameofoneormoreoutput
Input Enter a string to specify the input units.
Tip When you have multiple inputs and outputs, enter a comma-separated list of Input and Output
units corresponding to each channel.
Output Enter a string to specify the output units.
Notes Enter comments about the experiment or the data. For example, you might enter the
experiment name, date, and a description of experimental conditions. Models you estimate from this
data inherit your data notes.
6Click
7Click

Import. This action adds a new data icon to the System Identification Tool window.
Close to close the Import Data dialog box.

Importing Amplitude and Phase Frequency-Response Data. To import frequency-response data

consisting of amplitude and phase values at specified frequencies:
1Type

the following command in theMATLAB Command Window to open the GUI:

ident
2In the System Identification Tool window, select Import data > Freq. domain data. This action opens
the Import Data dialog box.
3In the Data Format for Signals list, select Freq. Function (Amp/Phase).
4Specify the following options:
Amplitude Enter the MATLAB variable name or a MATLAB expression that represents the
amplitude G.
Phase (deg) Enter the MATLAB variable name or a MATLAB expression that represents the phase
= arg G.
Frequency Enter the MATLAB variable name of a vector or a MATLAB expression that represents
the frequencies. The expression must evaluate to a column vector.
Data name Enter the name of the data set, which appears in the System Identification Tool window
after the import operation is completed.

 Frequency unitEnter Hzfor Hertz or keep the rad/sdefault value.

Sampling interval Enter the actual sampling interval in the experiment. For continuous-time data,
enter 0. For more information about this setting, see Specifying the Data Sampling Interval on page
2-34.
5(Optional)

In the Data Information area, click More to expand the dialog box and enter the following
optional settings:
Channel Names
InputEnterastringtospecifythenameofoneormoreinput channels.
Tip Naming channels helps you to identify data in plots. For multivariable input and output signals, you
can specify the names of individual Input and Output channels, separated by commas.
OutputEnterastringtospecifythenameofoneormoreoutput channels.
Physical Units of Variables
Input Enter a string to specify the input units.
Tip When you have multiple inputs and outputs, enter a comma-separated list of Input and Output
units corresponding to each channel.
Output Enter a string to specify the output units.
Notes Enter comments about the experiment or the data. For example, you might enter the
experiment name, date, and a description of experimental conditions. Models you estimate from this
data inherit your data notes.
6Click
7Click

Import. This action adds a new data icon to the System Identification Tool window.
Close to close the Import Data dialog box.

Importing Data Objects into the GUI

You can import the System Identification Toolbox iddataand idfrddata objects into the System
Identification Tool GUI.
Before you can import a data object into the System Identification Tool GUI, you must create the data
object in the MATLAB workspace, as described in Representing Time- and Frequency-Domain Data
Using iddata Objects on page 2-55 or Representing Frequency-Response Data Using idfrd Objects on
page 2-76.
Note You can also import a Control System Toolbox frdobject. Importing an frdobject converts it to an
idfrdobject.
Select Import data > Data object to open the Import Data dialog box.
Import iddata, idfrd,or frddata object in the MATLAB workspace.
To import a data object into the GUI: 1Type the following command in the MATLAB Command
Window to open the GUI:
ident
2In the System Identification Tool window, select Import data > Data object.

This action opens the Import Data dialog box. IDDATA or IDFRD/FRD is already
selected in the Data Format for Signals list.
3Specify the following options:
ObjectEnterthenameoftheMATLABvariablethatrepresentsthe data object in the MATLAB
workspace. Press Enter.
Data name Enter the name of the data set, which appears in the System Identification Tool window
after the import operation is completed.
(Only for time-domain iddataobject) Starting timeEnterthe starting value of the time axis for time
plots.
(Only for frequency domain iddataor idfrdobject) Frequency unit Enter the frequency unit for
response plots.
Sampling interval Enter the actual sampling interval in the experiment. For more information
about this setting, see Specifying the Data Sampling Interval on page 2-34.
Tip The System Identification Toolbox product uses the sampling interval during model estimation and
to set the horizontal axis on time plots. If you transform a time-domain signal to a frequency-domain
signal, the Fourier transforms are computed as discrete Fourier transforms (DFTs) using this sampling
interval.
4(Optional)

In the Data Information area, click More to expand the dialog box and enter the following
optional settings:
(Only for iddataobject) Input Properties
InterSample This options specifies the behavior of the input signals between samples during data
acquisition. It is used when transforming models from discrete-time to continuous-time and when
resampling the data.
zoh(zero-order hold) indicates that the input was piecewise-constant during data acquisition.
foh(first-order hold) indicates that the input was piecewise-linear during data acquisition.
bl(bandwidth-limited behavior) specifies that the continuous-time input signal has zero power above
the Nyquist frequency (equal to the inverse of the sampling interval).
Note See the d2cand c2dreference page for more information about transforming between discrete-time
and continuous-time models.
PeriodEnter Infto specify a nonperiodic input. If the underlying time-domain data was periodic over
an integer number of periods, enter the period of the input signal.

Note If your data is periodic, always include a whole number of periods for model estimation.
Channel Names
InputEnterastringtospecifythenameofoneormoreinput channels.
Tip Naming channels helps you to identify data in plots. For multivariable input and output signals, you
can specify the names of individual Input and Output channels, separated by commas.
OutputEnterastringtospecifythenameofoneormoreoutput channels.
Physical Units of Variables
Input Enter a string to specify the input units.
Tip When you have multiple inputs and outputs, enter a comma-separated list of Input and Output
units corresponding to each channel.
Output Enter a string to specify the output units.
Notes Enter comments about the experiment or the data. For example, you might enter the
experiment name, date, and a description of experimental conditions. Models you estimate from this
data inherit your data notes.
5Click
6Click

Import. This action adds a new data icon to the System Identification Tool window.
Close to close the Import Data dialog box.

Specifying the Data Sampling Interval

When you import data into the GUI, you must specify the data sampling interval.
The sampling interval is the time between successive data samples in your experiment and must be the
numerical time interval at which your data is sampled in any units. For example, enter 0.5if your data
was sampled every 0.5 s,andenter 1if your data was sampled every1s.
You can also use the sampling interval as a flag to specify continuous-time data. When importing
continuous-time frequency domain or frequency-response data, set the Sampling interval to 0.
The sampling interval is used during model estimation. For time-domain data, the sampling interval is
used together with the start time to calculate the sampling time instants. When you transform timedomain signals to frequency-domain signals (see the fftreference page), the Fourier transforms are
computed as discrete Fourier transforms (DFTs) for this sampling interval. In addition, the sampling
instants are used to set the horizontal axis on time plots.

Sampling Interval in the Import Data dialog box

Specifying Estimation and Validation Data

You should use different data sets to estimate and validate your model for best validation results.
In the System Identification Tool GUI, Working Data refers to estimation data. Similarly, Validation
Data refers to the data set you use to validate a model. For example, when you plot the model output,
the input to the model is the input signal from the validation data set. This plot compares model output
to the measured output in the validation data set. Selecting Model resids performs residual analysis
using the validation data.
To specify Working Data, drag and drop the corresponding data icon into the Working Data rectangle,
as shown in the following figure. Similarly, to specify Validation Data, drag and drop the
corresponding data icon into the Validation Data rectangle.

Drag and drop estimate data set.

Drag and drop validate data set.

Preprocessing Data Using Quick Start

As a preprocessing shortcut for time-domain data, select Preprocess > Quick start to simultaneously
perform the following four actions:
Subtract the mean value from each channel.
Note For information about when to subtract mean values from the data, see Handling Offsets and
Trends in Data on page 2-104.
Split data into two parts.
Specifythefirstpartasestimationdataformodels(or Working Data).
Specify the second part as Validation Data.

Creating Data Sets from a Subset of Signal Channels

You can create a new data set in the System Identification Tool GUI by extracting subsets of input and
output channels from an existing data set.
To create a new data set from selected channels:
In the System Identification Tool GUI, drag the icon of the data from which you want to select
channels to the Working Data rectangle.
2Select Preprocess > Select channels to open the Select Channels dialog box.
1

The Inputs list displays the input channels and the Outputs list
displays the output channels in the selected data set.
In the Inputs list, select one or more channels in any of following ways:
Select one channel by clicking its name.
3

Select adjacent channels by pressing the Shift key while clicking the first and last channel names.
Select nonadjacent channels by pressing the Ctrl key while clicking each channel name.
Tip To exclude input channels and create time-series data, clear all selections by holding down the Ctrl
key and clicking each selection. To reset selections, click Revert.
4In the Outputs list, select one or more channels in any of following ways:
Select one channel by clicking its name.
Select adjacent channels by pressing the Shift key while clicking the first and last channel names.
Select nonadjacent channels by pressing the Ctrl key while clicking each channel name.
Tip To reset selections, click Revert.
5In the Data name field, type the name of the new data set. Use a name that is unique in the Data Board.
6Click Insert to add the new data set to the Data Board in the System Identification Tool GUI.
7Click Close.

Creating Multiexperiment Data Sets in the GUI

Why Create Multiexperiment Data? on page 2-39
Limitations on Data Sets on page 2-40
Merging Data Sets on page 2-40
Extracting Specific Experiments from a Multiexperiment Data Set into a New Data Set on page 2-44
Why Create Multiexperiment Data?
You can create a time-domain or frequency-domain data set in the System Identification Tool GUI that
includes several experiments. Identifying models for multiexperiment data results in an average model.

Experiments can mean data that was collected during different sessions, or portions of the data collected
during a single session. In the latter situation, you can create multiexperiment data by splitting a single
data set into multiple segments that exclude corrupt data, and then merge the good data segments.
Limitations on Data Sets
You can only merge data sets that have all of the following characteristics:
Same number of input and output channels.
Different names. The name of each data set becomes the experiment name in the merged data set.
Same input and output channel names.
Same data domain (that is, time-domain data or frequency-domain data only).
Merging Data Sets
You can merge data sets using the System Identification Tool GUI. For example, suppose that you want
to combine the data sets tdata, tdata2, tdata3, tdata4showninthefollowingfigure.

GUI Contains Four Data Sets to Merge

To merge data sets in the GUI:
In the Operations area, select <--Preprocess > Merge experiments from the drop-down menu to
open the Merge Experiments dialog box.
1

In the System Identification Tool window, drag a data set icon to the Merge Experiments dialog box,to
the drop them here to be merged rectangle.
2

The name of the data set is added to the List of sets. Repeat for each data set you want to merge.

tdata and tdata2 to Be Merged

Tip To empty the list, click Revert.
In the Data name field, type the name of the new data set. This name must be unique in the Data
Board.
4Click Insert to add the new data set to the Data Board in the System Identification Tool window.
3

Data Board Now Contains tdatam with Merged Experiments 5Click Close to
close the Merge Experiments dialog box.
Tip To get information about a data set in the System Identification Tool GUI, right-click the data icon
to open the Data/model Info dialog box.
Extracting Specific Experiments from a Multiexperiment Data Set into a New Data Set
When adata set already consists of several experiments, you can extract one ormore of these experiments
into a new data set, using the System Identification Tool GUI.
For example, suppose that tdatamconsists of four experiments.
To create a new data set that includes only the first and third experiments in this data set:
1In the System Identification Tool window, drag and drop the tdatamdata icon to the Working Data
rectangle.

tdatam Is Set to Working Data

In the Operations area, select Preprocess > Select experiments from the drop-down menu to open
the Select Experiment dialog box.
3In the Experiments list, select one or more data sets in either of the following ways:
2

Select one data set by clicking its name.

Select adjacent data sets by pressing the Shift key while clicking the first and last names.
Select nonadjacent data sets by pressing the Ctrl key while clicking each name.

4In the Data name field, type the name of the new data set. This
name must be unique in the Data Board.
5Click Insert to add the new data set to the Data Board in the System Identification Tool GUI.
6Click Close to close the Select Experiment dialog box.

Managing Data in the GUI

Viewing Data Properties on page 2-48
Renaming Data and Changing Display Color on page 2-49
Distinguishing Data Types on page 2-51
Organizing Data Icons on page 2-51
Deleting Data Sets on page 2-52
Exporting Data to the MATLAB Workspace on page 2-53
Viewing Data Properties
You can get information about each data set in the System Identification Tool GUI by right-clicking the
corresponding data icon.
The Data/model Info dialog box opens. This dialog box describes the contents and the properties of the
corresponding data set. It also displays any associated notes and the command-line equivalent of the
operations you used to create this data.
Tip To view or modify properties for several data sets, keep this window open and right-click each data
set in the System Identification Tool GUI. The Data/model Info dialog box updates as you select each
data set.
Data object
description
History of
syntax that
created this
object

To displays the data properties in the MATLAB

Command Window, click Present.
Renaming Data and Changing Display Color
You can rename data and change its display color by double-clicking the data icon in the System
Identification Tool GUI.
The Data/model Info dialog box opens. This dialog box describes both the contents and the properties of
the data. The object description area displays thesyntaxoftheoperationsyouusedtocreatethedataintheGUI.
TheData/modelInfodialogboxalsoletsyourenamethedatabyenteringa new name in the Data name field.
You can also specify a new display color using three RGB values in the Color field. Each value is
between 0to 1and indicates the relative presence of red, green, and blue, respectively. For more
information about specifying default data color, see Customizing the System Identification Tool GUI
on page 12-17.
TipAsanalternativetousingthreeRGBvalues,youcanenterany one of the following letters in single quotes:
'y' 'r' 'b' 'c' 'g' 'm' 'k'
These strings represent yellow, red, blue, cyan, green, magenta, and black, respectively.
Specify name
for data set
Specify color
used to display
data set

Information About the Data

You can enter comments about the origin and state of the data in the Diary And Notes area. For
example, you might want to include the experiment name, date, and the description of experimental
conditions. When you estimate models from this data, these notes are associated with the models.
Clicking Present display portions of this information in the MATLAB Command Window.
Distinguishing Data Types
The background color of a data icon is color-coded, as follows:
White background represents time-domain data.
Blue background represents frequency-domain data.
Yellow background represents frequency-response data.
Time-domain data
Frequency-domain
data
Frequency-response

data

Colors Representing Type of Data

Organizing Data Icons
You can rearrange data icons in the System Identification Tool GUI by dragging and dropping the icons
to empty Data Board rectangles in the GUI.
Note You cannot drag and drop a data icon into the model area on the right.
When you need additional space for organizing data or model icons, select Options > Extra model/data
board in the System Identification Tool GUI. This action opens an extra session window with blank
rectangles for data and models. The new window is an extension of the current session and does not
represent a new session.
Tip When you import or create data sets and there is insufficient space for the icons, an additional
session window opens automatically.
You can drag and drop data between the main System Identification Tool GUI and any extra session
windows.

Type comments in the Notes field to describe the data sets. When you save a session, as described in
Saving, Merging, and Closing Sessions on page 12-6, all additional windows and notes are also saved.

Deleting Data Sets

To delete data sets in the System Identification Tool GUI, drag and drop the corresponding icon into
Trash.Movingitemsto Trash does not permanently delete these items.
Note You cannot delete a data set that is currently designated as Working Data or Validation Data.
You must first specify a different data set in the System Identification Tool GUI to be Working Data or
Validation Data,as described in Specifying Estimation and Validation Data on page 2-35. To restore a
data set from Trash, drag its icon from Trash to the Data or Model Board in the System Identification
Tool window. You can view the Trash contents by double-clicking the Trash icon.
Note You must restore data to the Data Board; you cannot drag data icons to the Model Board.

To permanently delete all items in Trash,select

Options > Empty trash.
Exiting a session empties the Trash automatically.
Exporting Data to the MATLAB Workspace
The data you create in the System Identification Tool GUI is not available in the MATLAB workspace
until you export the data set. Exporting to the MATLAB workspace is necessary when you need to
perform an operation on the data that is only available at the command line.
To export a data set to the MATLAB workspace, drag and drop the corresponding icon to the To
Workspace rectangle.
When you export data to the MATLAB workspace, the resulting variables have the same name as in the
System Identification Tool GUI. For example, the following figure shows how to export the timedomain data object datad.

Exporting Data to the MATLAB Workspace

In this example, the MATLAB workspace contains a variable named data after export.

Representing Time- and Frequency-Domain Data Using iddata Objects

In this section...
iddata Constructor on page 2-55
iddata Properties on page 2-58
Creating Multiexperiment Data at the Command Line on page 2-61
Select Data Channels, I/O Data and Experiments in iddata Objects on page 2-63
Increasing Number of Channels or Data Points of iddata Objects on page 2-67
Managing iddata Objects on page 2-69

iddata Constructor
Requirements for Constructing an iddata Object on page 2-55
Constructing an iddata Object for Time-Domain Data on page 2-56
Constructing an iddata Object for Frequency-Domain Data on page 2-57
Requirements for Constructing an iddata Object
To construct an iddataobject, you must have already imported data into the MATLAB workspace, as
described in Representing Data in MATLAB Workspace on page 2-9.
Constructing an iddata Object for Time-Domain Data Use the following syntax to create a timedomain iddataobject data:
data = iddata(y,u,Ts)
You can also specify additional properties, as follows:
data = iddata(y,u,Ts,'Property1',Value1,...,'PropertyN',ValueN)
For more information about accessing object properties, see Properties.

In this example, Tsisthesamplingtime,orthetimeinterval,between successive data samples. For uniformly

sampled data, Tsis a scalar value equal to the sampling interval of your experiment. The default time unit
is seconds, but you can set it to a new value using the TimeUnitproperty. For more information about
iddatatime properties, see Modifying Time and Frequency Vectors on page 2-69.
For nonuniformly sampled data, specify Tsas [], and set the value of the SamplingInstantsproperty as a
column vector containing individual time values. For example:
data = iddata(y,u,Ts,[],'SamplingInstants',TimeVector)
Where TimeVectorrepresents a vector of time values.
Note You can modify the property SamplingInstantsby setting it to a new vector with the length equal to
the number of data samples.
To represent time-series data, use the following syntax:
ts_data = iddata(y,[],Ts)
where yis the output data, []indicates empty input data, and Tsis the sampling interval.
The following example shows how to create an iddataobject using single-input/single-output (SISO)
data from dryer2.mat.Theinputand output each contain 1000 samples with the sampling interval of
0.08second.
load dryer2 % Load input u2 and output y2. data = iddata(y2,u2,0.08) % Create iddata object.
MATLAB returns the following output:
Time domain data set with 1000 samples. Sampling interval: 0.08
Outputs Unit (if specified) y1
Inputs Unit (if specified) u1
The default channel name 'y1'is assigned to the first and only output channel. When y2contains several
channels, the channels are assigned default names 'y1','y2','y2',...,'yn'. Similarly, the default channel
name 'u1'is assigned to the first and only input channel. For more information about naming channels,
see Naming, Adding, and Removing Data Channels on page 2-73.
Constructing an iddata Object for Frequency-Domain Data Frequency-domain data is the Fourier
transform of the input and output signals at specific frequency values. To represent frequency-domain
data, use the following syntax to create the iddataobject:
data = iddata(y,u,Ts,'Frequency',w)
'Frequency' is an iddataproperty that specifies the frequency values w, where wis the frequency column
vector that defines the frequencies at which the Fourier transform values of yand uare computed. Tsis the
time interval between successive data samples in seconds for the original time-domain data. w, y,and
uhave the same number of rows.
Note You must specify the frequency vector for frequency-domain data.
For more information about iddatatime and frequency properties, see Modifying Time and Frequency
Vectors on page 2-69.
To specify a continuous-time system, set Tsto 0.
You can specify additional properties when you create the iddataobject, as follows:

data = iddata(y,u,Ts,'Property1',Value1,...,'PropertyN',ValueN) For more information about accessing

object properties, see Properties.

iddata Properties
To view the properties of the iddataobject, use the getcommand. For example, type the following
commands at the prompt:
load dryer2
data = iddata(y2,u2,0.08); get(data)
% Load input u2 and output y2 % Create iddata object
% Get property values of data Domain: 'Time'
Name: []
OutputData: [1000x1 double] y: 'Same as OutputData'
OutputName: {'y1'}
OutputUnit: {''}
InputData: [1000x1 double] u: 'Same as InputData' InputName: {'u1'}
InputUnit: {''}
Period: Inf
InterSample: 'zoh'
Ts: 0.0800
Tstart: []
SamplingInstants: [1000x0 double] TimeUnit: ''
ExperimentName: 'Exp1'
Notes: []
UserData: []
For a complete description of all properties, see the iddatareference page.
You can specify properties when you create an iddataobject using the constructor syntax:
data = iddata(y,u,Ts,'Property1',Value1,...,'PropertyN',ValueN)
To change property values for an existing iddataobject, use the setcommand or dot notation. For
example, to change the sampling interval to 0.05,type the following at the prompt:
set(data,'Ts',0.05)
or equivalently:
data.ts = 0.05
Property names are not case sensitive. You do not need to type the entire property name if the first few
letters uniquely identify the property. Tip You can use data.yas an alternative to data.OutputDatato
access the output values, or use data.uas an alternative to data.InputDatato access the input values.
An iddataobject containing frequency-domain data includes frequency-specific properties, such as
Frequencyfor the frequency vector and Unitsfor frequency units (instead of Tstartand
SamplingIntervals).

To view the property list, type the following command sequence at the prompt:
% Load input u2 and output y2
load dryer2;
% Create iddata object
data = iddata(y2,u2,0.08);
% Take the Fourier transform of the data
% transforming it to frequency domain data = fft(data)
% Get property values of data
get(data)
Domain: 'Frequency'
Name: []
OutputData: [501x1 double] y: 'Same as OutputData' OutputName: {'y1'}
OutputUnit: {''}
InputData: [501x1 double] u: 'Same as InputData' InputName: {'u1'}
InputUnit: {''}
Period: Inf
InterSample: 'zoh'
Ts: 0.0800
Units: 'rad/s'
Frequency: [501x1 double] TimeUnit: ''
ExperimentName: 'Exp1'
Notes: []
UserData: []

Creating Multiexperiment Data at the Command Line

Why Create Multiexperiment Data Sets? on page 2-61
Limitations on Data Sets on page 2-62
Entering Multiexperiment Data Directly on page 2-62
Merging Data Sets on page 2-62
Adding Experiments to an Existing iddata Object on page 2-63
Why Create Multiexperiment Data Sets?
You can create iddataobjects that contain several experiments. Identifying models for an iddataobject
with multiple experiments results in an average model.
In the System Identification Toolbox product, experiments can either mean data collected during
different sessions, or portions of the data collected during a single session. In the latter situation, you can
create a multiexperiment iddataobject by splitting the data from a single session into multiple segments
to exclude bad data, and merge the good data portions.
Note The idfrdobject does not support the iddataequivalent of multiexperiment data.
Limitations on Data Sets
You can only merge data sets that have all of the following characteristics:

 Same number of input and output channels.

Same input and output channel names.
Same data domain (that is, time-domain data or frequency-domain data).
Entering Multiexperiment Data Directly
To construct an iddataobject that includes N data sets, you can use this syntax:
data = iddata(y,u,Ts)
where y, u,and Tsare 1-byN cell arrays containing data from the different experiments. Similarly, when
you specify Tstart, Period, InterSample,and SamplingInstantsproperties of the iddataobject, you must
assign their valuesas1-byN cell arrays.
Merging Data Sets
Create a multiexperiment iddataobject by merging iddataobjects, where each contains data from a single
experiment or is a multiexperiment data set. For example, you can use the following syntax to merge
data:
load iddata1 load iddata3 % Loads iddata object z1 % Loads iddata object z3
z = merge(z1,z3) % Merges experiments z1 and z3 into % the iddata object z
These commands create an iddataobject that conatains two experiments, where the experiments are
assigned default names 'Exp1'and 'Exp2', respectively.
Adding Experiments to an Existing iddata Object You can add experiments individually to an
iddataobject as an alternative approach to merging data sets.
Forexample,toaddtheexperimentsinthe iddataobject dat4to data,use the following syntax:
data(:,:,:,'Run4') = dat4
This syntax explicitly assigns the experiment name 'Run4'to the new experiment. The
Experimentproperty of the iddataobject stores experiment names.
For more information about subreferencing experiments in a multiexperiment data set, see
Subreferencing Experiments on page 2-66.

Select Data Channels, I/O Data and Experiments in iddata Objects

Subreferencing Input and Output Data on page 2-63
Subreferencing Data Channels on page 2-65
Subreferencing Experiments on page 2-66
Subreferencing Input and Output Data
Subreferencing data and its properties lets you select data values and assign new data and property
values.
Use the following general syntax to subreference specific data values in iddataobjects:

data(samples,outputchannels,inputchannels,experimentname)
In this syntax, samplesspecify one or more sample indexes, outputchannels and inputchannelsspecify
channel indexes or channel names, and experimentnamespecifies experiment indexes or names.
Forexample,toretrievesamples 5through 30in the iddataobject dataand store them in a new iddataobject
data_sub, use the following syntax:
data_sub = data(5:30)
You can also use logical expressions to subreference data. For example, to retrieve all data values from a
single-experiment data set that fall between sample instants 1.27and 9.3in the iddataobject dataand
assign them to data_sub, use the following syntax:
data_sub = data(data.sa>1.27&data.sa<9.3)
Note You do not need to type the entire property name. In this example, sa in data.sauniquely identifies
the SamplingInstantsproperty.
You can retrieve the input signal from an iddataobject using the following commands:
u = get(data,'InputData')
or
data.InputData
or
data.u % u is the abbreviation for InputData
Similarly, you can retrieve the output data using
data.OutputData
or
data.y % y is the abbreviation for OutputData
Subreferencing Data Channels
Use the following general syntax to subreference specific data channels in iddataobjects:
data(samples,outputchannels,inputchannels,experiment)
In this syntax, samplesspecify one or more sample indexes, outputchannels and inputchannelsspecify
channel indexes or channel names, and experimentnamespecifies experiment indexes or names.
To specify several channel names, you must use a cell array of name strings.
For example, suppose the iddataobject datacontains three output channels (named y1, y2,and y3), and
four input channels (named u1, u2, u3,and u4). To select all data samples in y3, u1,and u4, type the
following command at the prompt:
% Use a cell array to reference % input channels 'u1' and 'u4' data_sub = data(:,'y3',{'u1','u4'})
or equivalently
% Use channel indexes 1 and 4 % to reference the input channels data_sub = data(:,3,[1 4])
Tip Use a colon (:) to specify all samples or all channels, and the empty matrix ([]) to specify no
samples or no channels.

If you want to create a time-series object by extracting only the output data from an iddataobject, type
the following command:
data_ts = data(:,:,[])
You can assign new values to subreferenced variables. For example, the following command assigns the
first 10 values of output channel 1of datato values in samples 101through 110in the output channel 2of
data1.Italso assigns the values in samples 101through 110in the input channel 3of data1 to the first 10
values of input channel 1of data.
data(1:10,1,1) = data1(101:110,2,3)
Subreferencing Experiments
Use the following general syntax to subreference specific experiments in iddataobjects:
data(samples,outputchannels,inputchannels,experimentname)
In this syntax, samplesspecify one or more sample indexes, outputchannels and inputchannelsspecify
channel indexes or channel names, and experimentnamespecifies experiment indexes or names.
When specifying several experiment names, you must use a cell array of name strings. The iddataobject
stores experiments name in the ExperimentName property.
For example, suppose the iddataobject datacontains five experiments with default names, Exp1, Exp2,
Exp3, Exp4,and Exp5. Use the following syntax to subreference the first and fifth experiment in data:
data_sub = data(:,:,:,{'Exp1','Exp5'}) % Using experiment name
or
data_sub = data(:,:,:,[1 5]) % Using experiment index
Tip Use a colon (:) to denote all samples and all channels, and the empty matrix ([]) to specify no
samples and no channels.
Alternatively, you can use the getexpcommand. The following example shows how to subreference the
first and fifth experiment in data:
data_sub = getexp(data,{'Exp1','Exp5'}) % Using experiment name
or
data_sub = getexp(data,[1 5]) % Using experiment index
The following example shows how to retrieve the first 100 samples of output channels 2 and 3 and input
channels 4 to 8 of Experiment 3:
dat(1:100,[2,3],[4:8],3)

Increasing Number of Channels or Data Points of iddata Objects

iddata Properties Storing Input and Output Data on page 2-67
Horizontal Concatenation on page 2-67
Vertical Concatenation on page 2-68

iddata Properties Storing Input and Output Data

The InputData iddataproperty stores column-wise input data, and the OutputDataproperty stores columnwise output data. For more information about accessing iddataproperties, see iddata Properties on
page 2-58.
Horizontal Concatenation
Horizontal concatenation of iddataobjects creates a new iddataobject that appends all
InputDatainformation and all OutputData.Thistypeof concatenation produces a single object with more
input and output channels. For example, the following syntax performs horizontal concatenation on the
iddataobjects data1,data2,...,dataN:
data = [data1,data2,...,dataN]
[data1.InputData,data2.InputData,...,dataN.InputData] data.OutputData =
[data1.OutputData,data2.OutputData,...,dataN.OutputData]
For horizontal concatenation, data1,data2,...,dataNmust have the same number of samples and
experiments , and the sameTsand Tstartvalues.
The channels in the concatenated iddataobject are named according to the following rules:
Combining default channel names. If you concatenate iddataobjects with default channel names,
such as u1and y1, channels in the new iddata object are automatically renamed to avoid name
duplication.
Combining duplicate input channels. If data1,data2,...,dataN have input channels with duplicate
user-defined names, such that dataK contains channel names that are already present in dataJwithJ<
K,the dataKchannels are ignored.
Combining duplicate output channels. If data1,data2,...,dataN have input channels with duplicate
user-defined names, only the output channels with unique names are added during the concatenation.
Vertical Concatenation
Vertical concatenation of iddataobjects creates a new iddataobject that vertically stacks the input and
output data values in the corresponding data channels. The resulting object has the same number of
channels, but each channel contains more data points. For example, the following syntax creates a
dataobject such that its total number of samples is the sum of the samples in data1,data2,...,dataN.
data = [data1;data2;... ;dataN]
[data1.InputData;data2.InputData;...;dataN.InputData] data.OutputData =
[data1.OutputData;data2.OutputData;...;dataN.OutputData]
For vertical concatenation, data1,data2,...,dataNmust have the same number of input channels, output
channels, and experiments.

Managing iddata Objects

Modifying Time and Frequency Vectors on page 2-69
Naming, Adding, and Removing Data Channels on page 2-73
Subreferencing iddata Objects on page 2-74
Concatenating iddata Objects on page 2-75
Modifying Time and Frequency Vectors
The iddataobject stores time-domain data or frequency-domain data and has several properties that
specify the time or frequency values. To modify the time or frequency values, you must change the
corresponding property values.
Note You can modify the property SamplingInstantsby setting it to a new vector with the length equal to
the number of data samples. For more information, see Constructing an iddata Object for Time-Domain
Data on page 2-56.
The following tables summarize time-vector and frequency-vector properties, respectively, and provides
usage examples. In each example, datais an iddataobject.
Note Property names are not case sensitive. You do not need to type the entire property name if the first
few letters uniquely identify the property.
iddata Time-Vector Properties
Property Description
Ts Sampling time interval.
Syntax Example To set the sampling interval to 0.05:
For a single experiment, Tsis a scalar value.
For multiexperiement data with Ne experiments, Tsis a 1-byNecell array, and each cell contains the
sampling interval of the corresponding experiment.
Tstart Starting time of the experiment.
set(data,'ts',0.05)
or
data.ts = 0.05
To change starting time of the first data sample to 24:
For a single
experiment, Tsis a scalar value.
For multiexperiement data with Ne
experiments, Tsis a 1-byNecell array, and each cell contains the sampling interval of the corresponding
experiment.
data.Tstart = 24
Time units are set by the property TimeUnit.

iddata Time-Vector Properties (Continued)

Property Description
SamplingInstants Time values in the time vector, computed from the properties Tstartand Ts.
For a single
experiment,
SamplingInstants is an N-by1vector.
For multiexperiement data with Ne experiments, this property is a 1-byNe cell array, and each cell
contains the sampling instants of the corresponding experiment.
TimeUnit Unit of time. Specify as one of the following: 'nanoseconds', 'microseconds', 'milliseconds',
'seconds', 'minutes', 'hours', 'days', 'weeks', 'months', and 'years'.
Syntax Example
To retrieve the time vector for iddataobject data,use:
get(data,'sa')
To plot the input data as a function of time:
plot(data.sa,data.u)
Note sais the first two letters of the
SamplingInstants property that uniquely identifies this property.
To change the unit ofthetimevectorto milliseconds:
data.ti = 'milliseconds'
iddata Frequency-Vector Properties
Property Description
Frequency Frequency values at which the Fourier transforms of the signals are defined.
Syntax Example
To specify 100frequency values in log space, ranging between 0.1and 100,usethefollowing syntax:
For a single
experiment, Frequency is a scalar value.
For multiexperiement data with Ne
experiments,
Frequency is a 1-byNecell array, and each cell contains the frequencies of the corresponding experiment.
FrequencyUnit Unit of Frequency. Specify as one of the following: be one of the following: 'rad/
TimeUnit', 'cycles/TimeUnit', 'rad/s', 'Hz', 'kHz', 'MHz', 'GHz',and, 'rpm'. Default: `rad/TimeUnit'

For multi-experiement data with Neexperiments, Unitsis a 1-byNecell array, and each cell contains the
frequency unit for each experiment.
data.freq =
logspace(-1,2,100)
Set the frequency unit to Hz:
data.FrequencyUnit = 'Hz'
Note that changing the frequency unit does not scale the frequency vector. For a proper translation of
units, use chgFreqUnit.
Naming, Adding, and Removing Data Channels
What Are Input and Output Channels? on page 2-73
Naming Channels on page 2-73
Adding Channels on page 2-74
Modifying Channel Data on page 2-74
What Are Input and Output Channels?. A multivariate system might contain several input variables
or several output variables, or both. When an input or output signal includes several measured variables,
these variables are called channels.
Naming Channels. The iddataproperties InputNameand OutputNamestore the channel names for the
input and output signals. When you plot the data, you use channel names to select the variable displayed
on the plot. If you have multivariate data, it is helpful to assign a name to each channel that describes the
measured variable. For more information about selecting channels on a plot, see Selecting Measured
and Noise Channels in Plots on page 12-16.
You can use the setcommand to specify the names of individual channels. For example, suppose
datacontains two input channels (voltage and current) and one output channel (temperature). To set these
channel names, use the following syntax:
set(data,'InputName',{'Voltage','Current'}, 'OutputName','Temperature')
Tip You can also specify channel names as follows:
data.una = {'Voltage','Current') data.yna = 'Temperature'
unais equivalent to the property InputName,and ynais equivalent to OutputName.
If you do not specify channel names when you create the iddataobject, the toolbox assigns default
names. By default, the output channels are named 'y1','y2',...,'yn', and the input channels are named
'u1','u2',...,'un'.
Adding Channels.You can add data channels to an iddataobject.
For example, consider an iddataobject named datathat contains an input signal with four channels. To
add a fifth input channel, stored as the vector Input5, use the following syntax:
data.u(:,5) = Input5;

Input5 must have the same number of rows as the other input channels. In this example,
data.u(:,5)references all samples as (indicated by :)ofthe input signal uand sets the values of the fifth
channel. This channel is created when assigning its value to Input5.
You can also combine input channels and output channels of several iddata objects into one iddataobject
using concatenation. For more information, see Increasing Number of Channels or Data Points of
iddata Objects on page 2-67.
Modifying Channel Data. After you create an iddataobject, you can modify or remove specific input
and output channels, if needed. You can accomplish this by subreferencing the input and output matrices
and assigning new values.
For example, suppose the iddataobject datacontains three output channels (named y1, y2,and y3), and
four input channels (named u1, u2, u3,and u4). To replace datasuch that it only contains samples in y3,
u1,and u4,type the following at the prompt:
data = data(:,3,[1 4])
The resulting data object contains one output channel and two input channels.
Subreferencing iddata Objects
See Select Data Channels, I/O Data and Experiments in iddata Objects on page 2-63.
Concatenating iddata Objects
See Increasing Number of Channels or Data Points of iddata Objects on page 2-67.

Representing Frequency-Response Data Using idfrd Objects

In this section...
idfrd Constructor on page 2-76
idfrd Properties on page 2-77
Select I/O Channels and Data in idfrd Objects on page 2-79 Adding Input or Output Channels in
idfrd Objects on page 2-80 Managing idfrd Objects on page 2-83
Operations That Create idfrd Objects on page 2-83

idfrd Constructor
The idfrdrepresents complex frequency-response data. Before you can create an idfrdobject, you must
import your data as described in Frequency-Response Data Representation on page 2-13.
Note The idfrdobject can only encapsulate one frequency-response data set. It does not support the
iddataequivalent of multiexperiment data.
Use the following syntax to create the data object fr_data:
fr_data = idfrd(response,f,Ts)
Suppose that nyis the number of output channels, nuis the number of input channels, and nfis a vector of
frequency values. responseis an ny-bynu-bynf3-D array. fis the frequency vector that contains the

frequencies of the response.Tsis the sampling time, which is used when measuring or computing the
frequency response. If you are working with a continuous-time system, set Tsto 0.
response(ky,ku,kf) ,where ky, ku,and kfreference the kth output, input, and frequency value,
respectively, is interpreted as the complex-valued frequency response from input kuto output kyat
frequency f(kf). Note When you work at the command line, you can only create idfrdobjects from
complex values of G(eiw).ForaSISOsystem, responsecan be a vector.
You can specify object properties when you create the idfrdobject using the constructor syntax:
fr_data = idfrd(response,f,Ts,
'Property1',Value1,...,'PropertyN',ValueN)

idfrd Properties
To view the properties of the idfrdobject, you can use the getcommand. The following example shows
how to create an idfrdobject that contains 100 frequency-response values with a sampling time interval
of 0.08 s and get its properties:
% Create the idfrd data object fr_data = idfrd(response,f,0.08)
% Get property values of data get(fr_data)
responseand fare variables in the MATLAB Workspace browser, representing the frequency-response
data and frequency values, respectively.
MATLAB returns the following object properties and values:
ans =
Name: ''
Frequency: [100x1 double] ResponseData: [1x1x100 double] SpectrumData: []
CovarianceData: []
NoiseCovariance: []
Units: 'rad/s'
Ts: 0.0800
InputDelay: 0
EstimationInfo: [1x1 struct]
InputName: {'u1'}
OutputName: {'y1'}
InputUnit: {''}
OutputUnit: {''}
Notes: []
UserData: []
For a complete description of all idfrdobject properties, see the idfrd reference page.
To change property values for an existing idfrdobject, use the setcommand or dot notation. For example,
to change the name of the idfrdobject, type the following command sequence at the prompt:

% Set the name of the f_data object set(fr_data,'name','DC_Converter')

% Get fr_data properties and values get(fr_data)
Property names are not case sensitive. You do not need to type the entire property name if the first few
letters uniquely identify the property. If you import fr_datainto the System Identification Tool GUI, this
data has the name DC_Converterin the GUI, and not the variable name fr_data.
MATLAB returns the following object properties and values:
ans =
Name: 'DC_Converter' Frequency: [100x1 double] ResponseData: [1x1x100 double] SpectrumData: []
CovarianceData: []
NoiseCovariance: []
Units: 'rad/s'
Ts: 0.0800
InputDelay: 0
EstimationInfo: [1x1 struct]
InputName: {'u1'}
OutputName: {'y1'}
InputUnit: {''}
OutputUnit: {''}
Notes: []
UserData: []

Select I/O Channels and Data in idfrd Objects

You can reference specific data values in the idfrdobject using the following syntax:
fr_data(outputchannels,inputchannels)
Reference specific channels by name or by channel index.
Tip Use a colon (:) to specify all channels, and use the empty matrix ([])to specify no channels.
For example, the following command references frequency-response data from input channel 3to output
channel 2:
fr_data(2,3)
You can also access the data in specific channels using channel names. To list multiple channel names,
use a cell array. For example, to retrieve the power output, and the voltage and speed inputs, use the
following syntax:
fr_data('power',{'voltage','speed'})
To retrieve only the responses corresponding to frequency values between 200 and 300, use the
following command:
fr_data_sub = fselect(fr_data,[200:300])

You can also use logical expressions to subreference data. For example, to retrieve all frequencyresponse values between frequencies 1.27and 9.3in the idfrdobject fr_data,usethefollowingsyntax:
fr_data_sub = fselect(fr_data,fr_data.f>1.27&fr_data.f<9.3)
Tip Use endto reference the last sample number in the data. For example, data(77:end).
Note You do not need to type the entire property name. In this example, fin fr_data.funiquely identifies
the Frequencyproperty of the idfrdobject.

Adding Input or Output Channels in idfrd Objects

About Concatenating idfrd Objects on page 2-81
Horizontal Concatenation of idfrd Objects on page 2-81
Vertical Concatenation of idfrd Objects on page 2-82
Concatenating Noise Spectrum Data of idfrd Objects on page 2-82
About Concatenating idfrd Objects
The horizontal and vertical concatenation of idfrdobjects combine information in the
ResponseDataproperties of these objects. ResponseDatais an ny-bynu-bynfarray that stores the response
of the system, where nyis the number of output channels, nuis the number of input channels, and nf is a
vector of frequency values (see Properties).
Horizontal Concatenation of idfrd Objects
The following syntax creates a new idfrdobject datathat contains the horizontal concatenation of
data1,data2,...,dataN:
data = [data1,data2,...,dataN]
data contains the frequency responses from all of the inputs in data1,data2,...,dataNto the same outputs.
The following diagram is a graphical representation of horizontal concatenation of frequency-response
data. The (j,i,:)vector of the resulting response data represents the frequency response from the ith input
to the jth output at all frequencies.
u1

Data 1y1 Data 2y1 u2 2-by-2-by-nfy2 u3 2-by-1-by-nfy2

Combined

u1

Horizonal Concatenation y1Same inputs u2 of Data 1 and Data 2y2 outputs u3 2-by-3-by-nf

Note Horizontal concatenation of idfrdobjects requires that they have the same outputs and frequency
vectors. If the output channel names are different and their dimensions are the same, the concatenation
operation resets the output names to their default values.
Vertical Concatenation of idfrd Objects
The following syntax creates a new idfrdobject datathat contains the vertical concatenation of
data1,data2,...,dataN:
data = [data1;data2;... ;dataN]
The resulting idfrdobject datacontains the frequency responses from the same inputs in
data1,data2,...,dataNto all the outputs. The following diagram is a graphical representation of vertical

concatenation of frequency-response data. The (j,i,:)vector of the resulting response data represents the
frequency response from the ith input to the jth output at all frequencies.
u1

Data 1y1 u1 Data 1y3 u2 2-by-2-by-nfy2 u2 1-by-2-by-nf

Same
u1 Vertical Concatenationy1
of Data 1 and Data 2
y2
inputs

u2 3-by-2-by-nf y3 Combined outputs

Note Vertical concatenation of idfrdobjects requires that they have the same inputs and frequency
vectors. If the input channel names are different and their dimensions are the same, the concatenation
operation resets the input names to their default values.
Concatenating Noise Spectrum Data of idfrd Objects When the SpectrumDataproperty of individual
idfrdobjectsisnotempty, horizontal and vertical concatenation handle SpectrumData,asfollows.
In case of horizontal concatenation, there is no meaningful way to combine the SpectrumDataof
individual idfrdobjects and the resulting SpectrumData property is empty. An empty property results
because each idfrdobject has its own set of noise channels, where the number of noise channels equals
the number of outputs. When the resulting idfrdobject contains the same output channels as each of the
individual idfrdobjects, it cannot accommodate the noise data from all the idfrdobjects.
In case of vertical concatenation, the toolbox concatenates individual noise models diagonally. The
following shows that data.SpectrumDatais a block diagonal matrix of the power spectra and cross
spectra of the output noise in the system:
?data1.s 0?
data s
=
??
??
?

0dataN s?
? ? sin data.sis the abbreviation for the SpectrumDataproperty name.

Managing idfrd Objects

Subreferencing idfrd Objects on page 2-83
Concatenating idfrd Objects on page 2-83
Subreferencing idfrd Objects
See Select I/O Channels and Data in idfrd Objects on page 2-79.

Concatenating idfrd Objects

See Adding Input or Output Channels in idfrd Objects on page 2-80.

Operations That Create idfrd Objects

The following operations create idfrdobjects:
Constructing idfrdobjects.
Estimating nonparametric models using etfe, spa,and spafdr.Formore information, see Identifying
Frequency-Response Models on page 3-8.
Converting the Control System Toolbox frdobject. For more information, see Using Identified Models
for Control Design Applications on page 10-2.
Converting any linear dynamic system using the idfrdcommand.
For example:
sys_idpoly = idpoly([1 2 1],[0 2],'Ts',1);
G = idfrd(sys_idpoly,linspace(0,pi,128))

Analyzing Data Quality

In this section...
Is Your Data Ready for Modeling? on page 2-85
Plotting Data in the GUI Versus at the Command Line on page 2-86 How to Plot Data in the GUI on
page 2-86
How to Plot Data at the Command Line on page 2-92 How to Analyze Data Using the advice
Command on page 2-94

Is Your Data Ready for Modeling?

Before you start estimating models from data, you should check your data for the presence of any
undesirable characteristics. For example, you might plot the data to identify drifts and outliers. You plot
analysis might lead you to preprocess your data before model estimation.
The following data plots are available in the toolbox:
Time plot Shows data values as a function of time.
Tip You can infer time delays from time plots, which are required inputs to most parametric models. A
time delayisthetimeintervalbetweenthe change in input and the corresponding change in output.
Spectral plot Shows a periodogram that is computed by taking the absolute squares of the Fourier
transforms of the data, dividing by the number of data points, and multiplying by the sampling interval.
Frequency-response plot For frequency-response data, shows the amplitude and phase of the
frequency-response function on a Bode plot. For time- and frequency-domain data, shows the empirical
transfer function estimate (see etfe).

See Also
How to Analyze Data Using the advice Command on page 2-94 Ways to Prepare Data for System
Identification on page 2-6

Plotting Data in the GUI Versus at the Command Line

The plots you create using the System Identification Tool GUI provide options that are specific to the
System Identification Toolbox product, such as selecting specific channel pairs in a multivariate signals
or converting frequency units between Hertz and radians per second. For more information,
seeHowtoPlotDataintheGUIonpage2-86.
Theplotsyoucreateusingtheplotcommands,suchas plot,and bodeare displayed in the standard MATLAB
Figure window, which provides options for formatting, saving, printing, and exporting plots to a variety
of file formats. To learn about plotting at the command line, see How to Plot Data at the Command
Line on page 2-92. For more information about working with Figure window, see Figure Windows.

HowtoPlotDatainthe GUI
How to Plot Data in the GUI on page 2-86
Manipulating a Time Plot on page 2-88
Manipulating Data Spectra Plot on page 2-89
Manipulating a Frequency Function Plot on page 2-91
How to Plot Data in the GUI
After importing data into the System Identification Tool GUI, as described in Importing Data into the
GUI on page 2-17, you can plot the data.
To create one or more plots, select the corresponding check box in the Data Views area of the System
Identification Tool GUI.
An active data icon has a thick line in the icon, while an inactive data set has a thin line. Only active data
sets appear on the selected plots. To toggle including and excluding data on a plot, click the
corresponding icon in the System Identification Tool GUI. Clicking the data icon updates any plots that
are currently open.
Thick lines indicate active data sets included in plots.
All three available data plots are
selected.
When you have several data sets, you can view different input-output channel pair by selecting that pair
from the Channel menu. For more information about selecting different input and output pairs, see
Selecting Measured and Noise Channels in Plots on page 12-16.

In this example, dataand

dataffareactiveandappearonthethreeselected plots.
To close a plot, clear the corresponding check box in the System Identification Tool GUI.
Tip To get information about working with a specific plot, select a help topic from the Help menu in the
plot window.
Manipulating a Time Plot
The Time plot only shows time-domain data. In this example, data1is displayed on the time plot
because, of the three data sets, it is the only one that contains time-domain input and output.

Time Plot of data1

Thefollowingtablesummarizesoptionsthatarespecifictotimeplots,which you can select from the plot

window menus. For general information about working with System Identification Toolbox plots, see
Working with Plots on page 12-13.
Time Plot Options
Action
Toggle input display between piece-wise continuous (zero-order hold) and linear interpolation (firstorder hold) between samples.
Command
Select Style > Staircase input for zero-order hold or Style > Regular input for first-order hold.
Note This option only affects the display and not the intersample behavior specified when importing the
data.
Manipulating Data Spectra Plot The Data spectra plot shows a periodogram or a spectral estimate of
data1 and data3fd.
The periodogram is computed by taking the absolute squares of the Fourier transforms of the data,
dividing by the number of data points, and multiplying by the sampling interval. The spectral estimate
for time-domain data is a smoothed spectrum calculated using spa. For frequency-domain data, the Data
spectra plot shows the square of the absolute value of the actual data, normalized by the sampling
interval.
The top axes show the input and the bottom axes show the output. The vertical axis of each plot is
labeled with the corresponding channel name.

Periodograms of data1 and data3fd

Data Spectra Plot Options

Action
Toggle display between periodogram and spectral estimate.
Change frequency units.
Toggle frequency scale between linear and logarithmic.
Toggle amplitude scale between linear and logarithmic.
Command
Select Options > Periodogram or Options > Spectral analysis. Select Style > Frequency (rad/s) or
Style > Frequency (Hz).
Select Style > Linear frequency scale or Style > Log frequency scale.
Select Style > Linear amplitude scale or Style > Log amplitude scale.
Manipulating a Frequency Function Plot
For time-domain data, the Frequency function plot shows the empirical transfer function estimate
(etfe). For frequency-domain data, the plot shows the ratio of output to input data.
Thefrequency-responseplotshowstheamplitudeandphaseplotsof the corresponding frequency response.
For more information about frequency-response data, see Frequency-Response Data Representation
on page 2-13.

Frequency Functions of data1 and data3fd

Frequency Function Plot Options
Action
Change frequency units.
Toggle frequency scale between linear and logarithmic.
Toggle amplitude scale between linear and logarithmic.
Command
Select Style > Frequency (rad/s) or Style > Frequency (Hz).

Select Style > Linear frequency scale or Style > Log frequency scale.
Select Style > Linear amplitude scale or Style > Log amplitude scale.

How to Plot Data at the Command Line

The following table summarizes the commands available for plotting time-domain, frequency-domain,
and frequency-response data.
Commands for Plotting Data
Command Description
bode , For frequency-response data bodeplot only. Shows the magnitude and phase of the frequency
response on a logarithmic
frequency scale of a Bode plot.
Example
To plot idfrddata:
bode(idfrd_data)
or:
plot The type of plot corresponds to the type of data. For example, plotting time-domain data generates a
time plot, and plotting frequency-response data generates a
frequency-response plot.
When plotting time- or frequency-domain inputs
bodeplot(idfrd_data) To plot iddataor idfrd data:
plot(data)
Commands for Plotting Data (Continued)
Command Description Example
and outputs, the top axes show the output and the bottom axes show the input.
All plot commands display the data in the standard MATLAB Figure window. For more information
about working with the Figure window, see the Figure Windows.
To plot portions of the data, you can subreference specific samples (see Select Data Channels, I/O Data
and Experiments in iddata Objects on page 2-63 and Select I/O Channels and Data in idfrd Objects
on page 2-79. For example:
plot(data(1:300)) For time-domain data, to plot only the input data as a function of time, use the
following syntax:
plot(data(:,[],:)
When data.intersample = 'zoh', the input is piece-wise constant between sampling points on the plot. For
more information about properties, see the iddatareference page.
You can generate plots of the input data in the time domain using:
plot(data.SamplingInstants,data.u)

To plot frequency-domain data, you can use the following syntax:

semilogx(data.Frequency,abs(data.u))
When you specify to plot a multivariable iddataobject, each input-output combination is displayed one at
a time in the same MATLAB Figure window. You must press Enter to update the Figure window and
view the next channel combination. To cancel the plotting operation, press Ctrl+C. Tip To plot specific
input and output channels, use plot(data(:,ky,ku)), where kyand kuare specific output and input channel
indexes or names. For more information about subreferencing channels, see Subreferencing Data
Channels on page 2-65.
To plot several iddatasets d1,...,dN,use plot(d1,...,dN). Input-output channels with the same experiment
name, input name, and output name are always plotted in the same plot.

How to Analyze Data Using the advice Command

You can use the advicecommand to analyze time- or frequency- domain data before estimating a model.
The resulting report informs you about the possible need to preprocess the data and identifies potential
restrictions on the model accuracy. You should use these recommendations in combination with plotting
the data and validating the models estimated from this data.
Note advicedoes not support frequency-response data.
Before applying the advicecommand to your data, you must have represented your data as an
iddataobject. For more information, see Representing Timeand Frequency-Domain Data Using iddata
Objects on page 2-55.
If you are using the System Identification Tool GUI, you must export your data to the MATLAB
workspace before you can use the advicecommand on this data. For more information about exporting
data, see Exporting Models from the GUI to the MATLAB Workspace on page 12-12.
Use the following syntax to get advice about an iddataobject data:
advice(data)
For more information about the advicesyntax, see the advicereference page.
Advice provide guidance for these kinds of questions:
Doesitmakesensetoremoveconstantoffsetsandlineartrendsfromthe data?
What are the excitation levels of the signals and how does this affects the model orders?
Is there an indication of output feedback in the data? When feedback is present in the system, only
prediction-error methods work well for estimating closed-loop data.
Is there an indication of nonlinearity in the process that generated the data?
See Also advice
delayest
detrend
feedback pexcit

Selecting Subsets of Data

In this section...
Why Select Subsets of Data? on page 2-96
Extract Subsets of Data Using the GUI on page 2-97 Extract Subsets of Data at the Command Line
on page 2-99

Why Select Subsets of Data?

You can use data selection to create independent data sets for estimation and validation.
You can also use data selection as a way to clean the data and exclude parts with noisy or missing
information. For example, when your data contains missing values, outliers, level changes, and
disturbances, you can select one or more portions of the data that are suitable for identification and
exclude the rest.
If you only have one data set and you want to estimate linear models, you should split the data into two
portions to create two independent data sets for estimation and validation, respectively. Splitting the data
is selecting parts of the data set and saving each part independently.
You can merge several data segments into a single multiexperiment data set and identify an average
model. For more information, see Importing Data into the GUI on page 2-17 or Representing Timeand Frequency-Domain Data Using iddata Objects on page 2-55.
Note Subsets of the data set must contain enough samples to adequately represent the system, and the
inputs must provide suitable excitation to the system.
Selecting potions of frequency-domain data is equivalent to filtering the data. For more information
about filtering, see Filtering Data on page 2-120.

Extract Subsets of Data Using the GUI

Ways to Select Data in the GUI on page 2-97
Selecting a Range for Time-Domain Data on page 2-97
Selecting a Range of Frequency-Domain Data on page 2-99
Ways to Select Data in the GUI
You can use System Identification Tool GUI to select ranges of data on a time-domain or frequencydomain plot. Selecting data in the frequency domain is equivalent to passband-filtering the data.
After you select portions of the data, you can specify to use one data segment for estimating models and
use the other data segment for validating models. For more information, see Specifying Estimation and
Validation Data on page 2-35.
Note Selecting <--Preprocess > Quick start performs the following actions simultaneously:

 Remove the mean value from each channel.

Split the data into two parts.
Specifythefirstpartasestimationdata(or Working Data).
Specify the second part as Validation Data.
Selecting a Range for Time-Domain Data
You can select a range of data values on a time plot and save it as a new data set in the System
Identification Tool GUI.
Note Selecting data does not extract experiments from a data set containing multiple experiments. For
more information about multiexperiment data, see Creating Multiexperiment Data Sets in the GUI on
page 2-39. To extract a subset of time-domain data and save it as a new data set:
1Import

time-domain data into the System Identification Tool GUI, as described in Importing Data into
the GUI on page 2-17.
2Drag the data set you want to subset to the Working Data area.
If your data contains multiple I/O channels, in the Channel menu, select
thechannelpairyouwanttoview. Theupperplotcorrespondstotheinput signal, and the lower plot
corresponds to the output signal.
3

Although you view only one I/O channel pair at a time, your data selection is applied to all channels in
this data set.
4Select the data of interest in either of the following ways:
Graphically Draw a rectangle on either the input-signal or the output-signal plot with the mouse to
select the desired time interval. Your selection appears on both plots regardless of the plot on which you
draw the rectangle. The Time span and Samples fields are updated to match the selected region.
By specifying the Time span Edit the beginning and the end times in seconds. The Samples field is
updated to match the selected region. For example:
28.5 56.8
By specifying the Samples range Edit the beginning and the end indices of the sample range. The
Time span field is updated to match the selected region. For example:
342 654
Note To clear your selection, click Revert.
5In the Data name field, enter the name of the data set containing the selected data.
6 Click Insert. This action saves the selection
7To select another range, repeat steps 4 to 6.

as a new data set and adds it to the Data Board.

Selecting a Range of Frequency-Domain Data

Selecting a range of values in frequency domain is equivalent to filtering the data. For more information
about data filtering, see Filtering Frequency-Domain or Frequency-Response Data in the GUI on page
2-123.

Extract Subsets of Data at the Command Line

Selecting ranges of data values is equivalent to subreferencing the data.
For more information about subreferencing time-domain and frequency-domain data, see Select Data
Channels, I/O Data and Experiments in iddata Objects on page 2-63.
For more information about subreferencing frequency-response data, see Select I/O Channels and Data
in idfrd Objects on page 2-79.

Handling Missing Data and Outliers

In this section...
Handling Missing Data on page 2-100
Handling Outliers on page 2-101
Extract and Model Specific Data Segments on page 2-102 See Also on page 2-103

Handling Missing Data

Data acquisition failures sometimes result in missing measurements both in the input and the output
signals. When you import data that contains missing values using the MATLAB Import Wizard, these
values are automatically set to NaN. NaNserves as a flag for nonexistent or undefined data. When you
plot data on a time-plot that contains missing values, gaps appear on the plot where missing data exists.
You can use misdatato estimate missing values. This command linearly interpolates missing values to
estimate the first model. Then, it uses this model to estimate the missing data as parameters by
minimizing the output prediction errors obtained from the reconstructed data. You can specify the model
structure you want to use in the misdataargument or estimate a default-order model using the
n4sidmethod. For more information, see the misdatareference page.
Note You can only use misdataon time-domain data stored in an iddata object. For more information
about creating iddataobjects, see Representing Time- and Frequency-Domain Data Using iddata
Objects on page 2-55.
For example, suppose yand uare output and input signals that contain NaNs. This data is sampled at
0.2s. The following syntax creates a new iddata object with these input and output signals.
dat = iddata(y,u,0.2) % y and u contain NaNs % representing missing data
Apply the misdatacommand to the new data object. For example:
dat1 = misdata(dat);
plot(dat,dat1) % Check how the missing data % was estimated on a time plot

Handling Outliers
Malfunctions can produce errors in measured values, called outliers.Such outliers might be caused by
signal spikes or by measurement malfunctions. If you do not remove outliers from your data, this can
adversely affect the estimated models.

To identify the presence of outliers, perform one of the following tasks:

Before estimating a model, plot the data on a time plot and identify values that appear out of range.
After estimating a model, plot the residuals and identify unusually large values. For more information
about plotting residuals, see Residual Analysis on page 8-24. Evaluate the original data that is
responsible for large residuals. For example, for the model Modeland validation data Data, you can use
the following commands to plot the residuals:
% Compute the residuals
E = resid(Model,Data) % Plot the residuals
plot(E)
Next, try these techniques for removing or minimizing the effects of outliers:
Extracttheinformativedataportionsintosegmentsandmergethem into one multiexperiment data set (see
Extract and Model Specific Data Segments on page 2-102). For more information about selecting and
extracting data segments, see Selecting Subsets of Data on page 2-96. Tip The inputs in each of the
data segments must be consistently exciting the system. Splitting data into meaningful segments for
steady-state data results in minimum information loss. Avoid making data segments too small.
Manually replace outliers with NaNs and then use the misdatacommand to reconstruct flagged data.
This approach treats outliers as missing data and is described in Handling Missing Data on page
2-100. Use this method when your data contains several inputs and outputs, and when you have
difficulty finding reliable data segments in all variables.
Remove outliers by prefiltering the data for high-frequency content because outliers often result from
abrupt changes. For more information about filtering, see Filtering Data on page 2-120.
Note The estimation algorithm can handle outliers by assigning a smaller weight to outlier data. A
robust error criterion applies an error penalty that is quadratic for small and moderate prediction errors,
and is linear for large prediction errors. Because outliers produce large prediction errors, this approach
gives a smaller weight to the corresponding data points during model estimation. Set the
ErrorThresholdestimation option (see Advanced.ErrorThresholdin, for example, polyestOptions) to a
nonzero value to activate the correction for outliers in the estimation algorithm.

Extract and Model Specific Data Segments

This example shows how to create a multi-experiment, time-domain data set by merging only the
accurate data segments and ignoring the rest.
Assume that the data has poor or no measurements for some sample ranges (for example 341499). You
cannot simply concatenate the good data segments because the transients at the connection points
compromise the model. Instead, you must create a multiexperiment iddataobject, where each experiment
corresponds to a good segment of data, as follows:
% Plot the data in a MATLAB Figure window plot(data)
% Create multiexperiment data set

% by merging data segments

datam = merge(data(1:340),... data(500:897),... data(1001:1200),... data(1550:2000));
% Model the multiexperiment data set % using "experiments" 1, 2, and 4 m =
n4sid(getexp(datam,[1,2,4]))
% Validate the model by comparing its output to % the output data of experiment 3
compare(getexp(datam,3),m)

See Also
To learn more about the theory of handling missing data and outliers, see the chapter on preprocessing
data in System Identification: Theory for the User, Second Edition, by Lennart Ljung, Prentice Hall
PTR, 1999.

Handling Offsets and Trends in Data

In this section...
When to Detrend Data on page 2-104
Alternatives for Detrending Data in GUI or at the Command-Line on page 2-105
Next Steps After Detrending on page 2-107

When to Detrend Data

Detrending is removing means, offsets, or linear trends from regularly sampled time-domain inputoutput data signals. This data processing operation helps you estimate more accurate linear models
because linear models cannot capture arbitrary differences between the input and output signal levels.
The linear models you estimate from detrended data describe the relationship between the change in
input signals and the change in output signals.
For steady-state data, you should remove mean values and linear trends from both input and output
signals.
For transient data, you should remove physical-equilibrium offsets measured prior to the excitation input
signal.
Remove one linear trend or several piecewise linear trends when the levels drift during the experiment.
Signal drift is considered a low-frequency disturbance and can result in unstable models.
You should not detrend data before model estimation when you want:
Linear models that capture offsets essential for describing important system dynamics. For example,
when a model contains integration behavior, you could estimate a low-order transfer function (process
model) from nondetrended data. For more information, see Identifying Process Models on page 3-26.
Nonlinear black-box models, such as nonlinear ARX or
Hammerstein-Wiener models. For more information, see Nonlinear Model Identification.

Tip When signals vary around a large signal level, you can improve computational accuracy of
nonlinear models by detrending the signal means.
Nonlinear ODE parameters (nonlinear grey-box models). For more information, see Estimating
Nonlinear Grey-Box Models on page 5-17.
To simulate or predict the linear model response at the system operating conditions, you can restore the
removed trend to the simulated or predicted model output using the retrendcommand.
For more information about handling drifts in the data, see the chapter on preprocessing data in System
Identification: Theory for the User,Second Edition, by Lennart Ljung, Prentice Hall PTR, 1999.
Examples
How to Detrend Data Using the GUI on page 2-108 How to Detrend Data at the Command Line on
page 2-109

Alternatives for Detrending Data in GUI or at the Command-Line

You can detrend data using the System Identification Tool GUI and at the command line using the
detrendcommand.
Both the GUI and the command line let you subtract the mean values and one linear trend from steadystate time-domain signals.
However, the detrendcommand provides the following additional functionality (not available in the
GUI):
Subtracting piecewise linear trends at specified breakpoints. A breakpoint is a time value that defines
the discontinuities between successive linear trends.
Subtracting arbitrary offsets and linear trends from transient data signals.
Saving trend information to a variable so that you can apply it to multiple data sets.
To learn how to detrend data, see:
How to Detrend Data Using the GUI on page 2-108
How to Detrend Data at the Command Line on page 2-109
As an alternative to detrending data beforehand, you can specify the offsets levels as estimation options
and use them directly with the estimation command.
For example, suppose your data has an input offset, u0, and an output offset, y0. There are two ways to
perform a linear model estimation (say, a transfer function model estimation) using this data:
Using detrend:
T=getTrend(data)
T.InputOffset = u0;
T.OutputOffset = y0;
datad = detrend(data, T);

model = tfest(datad, np);

Specify offsets as estimation options:
opt = tfestOptions('InputOffset',u0, 'OutputOffset', y0);
model = tfest(data, np, opt)
The advantage of this approach is that there is a record of offset levels in the model in
model.Report.OptionsUsed. The limitation of this approach is that it cannot handle linear trends, which
can only be removed from the data by using detrend.

Next Steps After Detrending

After detrending your data, you might do the following:
Perform other data preprocessing operations. See Ways to Prepare Data for System Identification on
page 2-6.
Estimate a linear model. See Linear Model Identification.

How to Detrend Data Using the GUI

Before you can perform this task, you must have regularly-sampled, steady-state time-domain data
imported into the System Identification Tool GUI. See Importing Time-Domain Data into the GUI on
page 2-18). For transient data, see How to Detrend Data at the Command Line on page 2-109.
Tip You can use the shortcut Preprocess > Quick start to perform several operations: remove the mean
value from each signal, split data into two halves, specify the first half as model estimation data (or
Working Data), and specify the second half as model Validation Data.
1In

the System Identification Tool, drag the data set you want to detrend to the Working Data rectangle.

Detrend the data.

To remove linear trends, select Preprocess > Remove trends.
2

To remove mean values from each input and output data signal, select Preprocess > Remove means.
More About Handling Offsets and Trends in Data on page 2-104

How to Detrend Data at the Command Line

In this section...
Detrending Steady-State Data on page 2-109 Detrending Transient Data on page 2-109 See Also
on page 2-110

Detrending Steady-State Data

Before you can perform this task, you must have time-domain data as an iddataobject. See
Representing Time- and Frequency-Domain Data Using iddata Objects on page 2-55.

Note If you plan to estimate models from this data, your data must be regularly sampled.
Use the detrendcommandtoremovethesignalmeansorlineartrends:
[data_d,T]=detrend(data,Type)
where datais the data to be detrended. The second input argument Type=0removes signal means or
Type=1removes linear trends. data_dis the detrended data. Tis a TrendInfoobject that stores the values of
the subtracted offsets and slopes of the removed trends.
More About Handling Offsets and Trends in Data on page 2-104

Detrending Transient Data

Before you can perform this task, you must have
Time-domain data as an iddataobject. See Representing Time- and Frequency-Domain Data Using
iddata Objects on page 2-55. Note If you plan to estimate models from this data, your data must be
regularly sampled.
Values of the offsets you want to remove from the input and output data. If you do not know these
values, visually inspect a time plot of your data. For more information, see How to Plot Data at the
Command Line on page 2-92.
1Create

a default object for storing input-output offsets that you want to remove from the data.
T = getTrend(data)
where Tis a TrendInfoobject.
2Assign offset values to T.
T.InputOffset=I_value; T.OutputOffset=O_value;
where I_valueis the input offset value, and O_valueis the input offset value.
3Remove the specified offsets from data.
data_d = detrend(data,T)
where the second input argument Tstores the offset values as its properties.
More About Handling Offsets and Trends in Data on page 2-104

See Also
detrend
TrendInfo

Resampling Data
In this section...
WhatIsResampling?onpage2-111
Resampling Data Without Aliasing Effects on page 2-112 See Also on page 2-116

What Is Resampling?
Resampling data signals in the System Identification Toolbox product applies an antialiasing (lowpass)
FIR filter to the data and changes the sampling rate of the signal by decimation or interpolation.
If your data is sampled faster than needed during the experiment, you can decimate it without
information loss. If your data is sampled more slowly than needed, there is a possibility that you miss
important information about the dynamics at higher frequencies. Although you can resample the data at
a higher rate, the resampled values occurring between measured samples do not represent new measured
information about your system. Instead of resampling, repeat the experiment using a higher sampling
rate.
Tip You should decimate your data when it contains high-frequency noise outside the frequency range
of the system dynamics.
Resampling takes into account how the data behaves between samples, which you specify when you
import the data into the System Identification Tool GUI (zero-order or first-order hold). For more
information about the data properties you specify before importing the data, see Importing Data into the
GUI on page 2-17.
You can resample data using the System Identification Tool GUI or the resamplecommand. You can only
resample time-domain data at uniform time intervals.
Examples
Resampling Data Using the GUI on page 2-117 Resampling Data at the Command Line on page
2-118

Resampling Data Without Aliasing Effects

Typically, you decimate a signal to remove the high-frequency contributions that result from noise from
the total energy. Ideally, you want to remove the energy contribution due to noise and preserve the
energy density of the signal.
The command resampleperforms the decimation without aliasing effects. This command includes a
factor of T to normalize the spectrum and preserve the energy density after decimation. For more
information about spectrum normalization, see Spectrum Normalization on page 3-14.
If you use manual decimation instead of resampleby picking every fourth sample from the signal, for
examplethe energy contributions from higher frequencies are folded back into the lower
frequencies("aliasing"). Because the total signal energy is preserved by this operation and this energy
must now be squeezed into a smaller frequency range, the amplitude of the spectrum at each frequency
increases. Thus, the energy density of the decimated signal is not constant.
The following example illustrates how resampleavoids folding effects:
% Construct fourth-order MA-process
m0 = idpoly(1,[ ],[1 1 1 1]); % a time series (no input) model
% Generate error signal e = idinput(2000,'rgs');

% Simulate the output using error signal

sim_opt = simOptions('AddNoise',true,'NoiseData',e); y = sim(m0,zeros(2000,0),sim_opt)
y = iddata(y,[],1)
% Estimate signal spectrum g1 = spa(y);
% Estimate spectrum of modified signal including % every fourth sample of the original signal. % This
command automatically sets Ts to 4. g2 = spa(y(1:4:2000));
% Plot frequency response to view folding effects h = spectrumplot(g1,g2,g1.Frequency);
opt = getoptions(h);
opt.FreqScale='linear';
opt.FreqUnits='Hz';
setoptions(h,opt)
% Estimate spectrum after prefiltering that does not % introduce folding effects
g3 = spa(resample(y,1,4));
figure
spectrumplot(g1,g3,g1.Frequency,opt)

Folding
Effects with Manual Decimation
Use resampleto decimate the signal before estimating the spectrum and plot the frequency response, as
follows:
g3 = spa(resample(y,1,4));
figure
spectrumplot(g1,g3,g1.Frequency,opt)
The following figure shows that the estimated spectrum of the resampled signal has the same amplitude
as the original spectrum. Thus, there is no indication of folding effects when you use resampleto
eliminate aliasing.

No Folding
Effects When Using resample
Examples
Resampling Data Using the GUI on page 2-117
Resampling Data at the Command Line on page 2-118

See Also
For a detailed discussion about handling disturbances, see the chapter on preprocessing data in System
Identification: Theory for the User,Second Edition, by Lennart Ljung, Prentice Hall PTR, 1999.

Resampling Data Using the GUI

Use the System Identification Tool GUI to resample time-domain data. To specify additional options,
such as the prefilter order, see Resampling Data at the Command Line on page 2-118.
The System Identification Tool GUI uses idresampto interpolate or decimate the data. For more
information about this command, type help idresampat the prompt.
To create a new data set by resampling the input and output signals:
1Import time-domain data into the System Identification Tool GUI, as described in Importing Data into

the GUI on page 2-17.

2Drag the data set you want to resample to the Working Data area.
3In the Resampling factor field, enter the factor by which to multiply the current sampling interval:
For decimation (fewer samples), enter a factor greater than 1to increase the
samplingintervalbythisfactor.
For interpolation (more samples), enter a factor less than 1to decrease the sampling interval by this
factor.
Default = 1.
4In the Data name field, type the name of the new data set. Choose a name that is unique in the Data
Board.
5Click Insert to add the new data set to the Data Board in the System Identification Toolbox window.
6Click Close to close the Resample dialog box.
More About
Resampling Data on page 2-111

Resampling Data at the Command Line

Use resampleto decimate and interpolate time-domain iddataobjects. You can specify the order of the
antialiasing filter as an argument.
Note resampleuses the Signal Processing Toolbox command, when this toolbox is installed on your
computer. If this toolbox is not installed, use idresampinstead. idresamponly lets you specify the filter
order, whereas resamplealso lets you specify filter coefficients and the design parameters of the Kaiser
window.
To create a new iddataobject datarby resampling data,usethefollowing syntax:
datar = resample(data,P,Q,filter_order)
In this case, Pand Qare integers that specify the new sampling interval: the new sampling interval is Q/
Ptimes the original one. You can also specify the order of the resampling filter as a fourth argument
filter_order,whichis an integer (default is 10). For detailed information about resample,seethe
corresponding reference page.
For example, resample(data,1,Q)results in decimation with the sampling interval modified by a factor Q.
The next example shows how you can increase the sampling rate by a factor of 1.5 and compare the
signals:
plot(u)
ur = resample(u,3,2); plot(u,ur)
When the Signal Processing Toolbox product is not installed, using resample calls idresampinstead.
idresampuses the following syntax:
datar = idresamp(data,R,filter_order)
In this case, R=Q/P, which means that data is interpolated by a factor Pand then decimated by a factor Q.
To learn more about idresamp,type help idresamp.

The data.InterSampleproperty of the iddataobject is taken into account during resampling (for example,
first-order hold or zero-order hold). For more information, see iddata Properties on page 2-58.
More About Resampling Data on page 2-111

Filtering Data
In this section...
Supported Filters on page 2-120
Choosing to Prefilter Your Data on page 2-120 See Also on page 2-121

Supported Filters
You can filter the input and output signals through a linear filter before estimating a model in the System
Identification Tool GUI or at the command line. How you want to handle the noise in the system
determines whether it is appropriate to prefilter the data.
The filter available in the System Identification Tool GUI is a fifth-order (passband) Butterworth filter.
If you need to specify a custom filter, use the idfiltcommand.
Examples
How to Filter Data Using the GUI on page 2-122 How to Filter Data at the Command Line on page
2-126

Choosing to Prefilter Your Data

Prefiltering data can help remove high-frequency noise or low-frequency disturbances (drift). The latter
application is an alternative to subtracting linear trends from the data, as described in Handling Offsets
and Trends in Data on page 2-104.
In addition to minimizing noise, prefiltering lets you focus your model on specific frequency bands. The
frequency range of interest often corresponds to a passband over the breakpoints on a Bode plot. For
example, if you are modeling a plant for control-design applications, you might prefilter the data to
specifically enhance frequencies around the desired closed-loop bandwidth. Prefiltering the input and
output data through the same filter does not change the input-output relationship for a linear system.
However, prefiltering does change the noise characteristics and affects the estimated model of the
system.
To get a reliable noise model, avoid prefiltering the data. Instead, set the Focusproperty of the estimation
algorithm to Simulation.
Note When you prefilter during model estimation, the filtered data is used to only model the input-tooutput dynamics. However, the disturbance model is calculated from the unfiltered data.
Examples
How to Filter Data Using the GUI on page 2-122 How to Filter Data at the Command Line on page
2-126

See Also
To learn how to filter data during linear model estimation instead, you can set the Focusproperty of the
estimation algorithm to Filterand specify the filter characteristics.
For more information about prefiltering data, see the chapter on preprocessing data in System
Identification: Theory for the User, Second Edition, by Lennart Ljung, Prentice Hall PTR, 1999.
For practical examples of prefiltering data, see the section on posttreatment of data in Modeling of
Dynamic Systems, by Lennart Ljung and Torkel Glad, Prentice Hall PTR, 1994.

HowtoFilterDataUsingtheGUI
In this section...
Filtering Time-Domain Data in the GUI on page 2-122
Filtering Frequency-Domain or Frequency-Response Data in the GUI on page 2-123

Filtering Time-Domain Data in the GUI

The System Identification Tool GUI lets you filter time-domain data using a fifth-order Butterworth
filter by enhancing or selecting specific passbands.
To create a filtered data set:
1Import time-domain data into the System Identification Tool GUI, as described in Importing Data into
the GUI on page 2-17.
2Drag the data set you want to filter to the Working Data area.
3Select <--Preprocess > Filter. By default, this selection shows a periodogram of the input and output
spectra (see the etfereference page).
Note Todisplay smoothed spectral estimates instead of the periodogram, select Options > Spectral
analysis. This spectral estimate is computed using spaand your previous settings in the Spectral Model
dialog box. To change these settings, select <--Estimate > Spectral model in the System Identification
Tool GUI, and specify new model settings.
If your data contains multiple input/output channels, in the Channel menu, select the channel pair you
want to view. Although you view only one channel pair at a time, the filter applies to all input/output
channels in this data set.
4

5Select

the data of interest using one of the following ways:

Graphically Draw a rectangle with the mouse on either the input-signal or the output-signal plot to
select the desired frequency interval. Your selection is displayed on both plots regardless of the plot on
which you draw the rectangle. The Range field is updated to match the selected region. If you need to
clear your selection, right-click the plot.
Specify the Range Edit the beginning and the end frequency values. For example:
8.5 20.0(rad/s).

Tip To change the frequency units from rad/sto Hz,select Style > Frequency (Hz). To change the
frequency units from Hzto rad/s, select Style > Frequency (rad/s).
In the Range is list, select one of the following:
Pass band Allows data in the selected frequency range.
Stop band Excludes data in the selected frequency range.
6

7Click Filter to preview the filtered results. If you are satisfied, go to step 8. Otherwise,
8In the Data name field, enter the name of the data set containing the selected data.
9Click Insert to save the selection as a new data set and add it to the Data Board.
10To select another range, repeat steps 5 to 9.

return to step 5.

More About Filtering Data on page 2-120

Filtering Frequency-Domain or Frequency-Response Data in the GUI

For frequency-domain and frequency-response data, filtering is equivalent to selecting specific data
ranges.
To select a range of data in frequency-domain or frequency-response data:
1Import data into the System Identification Tool GUI, as described in Importing Data into the GUI on
page 2-17.
2Drag the data set you want you want to filter to the Working Data area.
3Select <--Preprocess > Select range. This selection displays one of the following plots:
Frequency-domain data Plot shows the absolute of the squares of the input and output spectra.
Frequency-response data Top axes show the frequency response magnitude equivalent to the ratio
of the output to the input, and the bottom axes show the ratio of the input signal to itself, which has the
value of 1at all frequencies.
If your data contains multiple input/output channels, in the Channel menu, select the channel pair you
want to view. Although you view only one channel pair at a time, the filter applies to all input/output
channels in this data set.
4

5Select

the data of interest using one of the following ways:

Graphically Draw a rectangle with the mouse on either the input-signal or the output-signal plot to
select the desired frequency interval. Your selection is displayed on both plots regardless of the plot on
which you draw the rectangle. The Range field is updated to match the selected region.
If you need to clear your selection, right-click the plot.
Specify the Range Edit the beginning and the end frequency values.
For example:
8.5 20.0(rad/s).
Tip If you need to change the frequency units from rad/sto Hz,select Style > Frequency (Hz). To
change the frequency units from Hzto rad/s, select Style > Frequency (rad/s).

In the Range is list, select one of the following:

Pass band Allows data in the selected frequency range.
Stop band Excludes data in the selected frequency range.
6

7In the Data name field, enter the name of the data set containing the selected data.
8Click Insert. This action saves the selection as a new data set and adds it to the Data
9To select another range, repeat steps 5 to 8.

Board.

More About Filtering Data on page 2-120

HowtoFilterDataattheCommandLine
In this section...
Simple Passband Filter on page 2-126 Defining a Custom Filter on page 2-127 Causal and
Noncausal Filters on page 2-128

Simple Passband Filter

Use idfiltto apply passband and other custom filters to a time-domain or a frequency-domain
iddataobject.
In general, you can specify any custom filter. Use this syntax to filter an iddataobject datausing the filter
called filter:
fdata = idfilt(data,filter)
In the simplest case, you can specify a passband filter for time-domain data using the following syntax:
fdata = idfilt(data,[wl wh])
In this case, w1and whrepresent the low and high frequencies of the passband, respectively.
You can specify several passbands, as follows:
filter=[[w1l,w1h];[ w2l,w2h]; ....;[wnl,wnh]]
The filter is an n-by2matrix, where each row defines a passband in radians per second.
To define a stopband between ws1and ws2,use
filter = [0 ws1; ws2 Nyqf]
where, Nyqfis the Nyquist frequency. For time-domain data, the passband filtering is cascaded
Butterworth filters of specified order. The default filter order is 5. The Butterworth filteristhesameas
butterin the Signal Processing Toolbox product. For frequency-domain data, select the indicated portions
of the data to perform passband filtering.
More About Filtering Data on page 2-120

Defining a Custom Filter

Use idfiltto apply passband and other custom filters to a time-domain or a frequency-domain
iddataobject.
In general, you can specify any custom filter. Use this syntax to filter an iddataobject datausing the filter
called filter:
fdata = idfilt(data,filter)

You can define a general single-input/single-output (SISO) system for filtering time-domain or
frequency-domain data. For frequency-domain only, you can specify the (nonparametric) frequency
response of the filter.
You use this syntax to filter an iddataobject datausing a custom filter specified by filter:
fdata = idfilt(data,filter)
filtercanbealsoanyofthefollowing:
filter = idm
filter = {num,den} filter = {A,B,C,D}
idmis a SISO identified linear model or LTI object. For more information about LTI objects, see the
Control System Toolbox documentation.
{num,den}defines the filter as a transfer function as a cell array of numerator and denominator filter
coefficients.
{A,B,C,D}is a cell array of SISO state-space matrices.
Specifically for frequency-domain data, you specify the frequency response of the filter:
filter = Wf
Here, Wfis a vector of real or complex values that define the filter frequency response, where the inputs
and outputs of dataat frequency data.Frequency(kf)are multiplied by Wf(kf). Wfis a column vector with
the length equal to the number of frequencies in data.
When datacontains several experiments, Wfis a cell array with the length equal to the number of
experiments in data.
More About Filtering Data on page 2-120

Causal and Noncausal Filters

For time-domain data, the filtering is causal by default. Causal filters typically introduce a phase shift in
the results. To use a noncausal zero-phase filter (corresponding to filtfiltin the Signal Processing
Toolbox product), specify a third argument in idfilt:
fdata = idfilt(data,filter,'noncausal')
For frequency-domain data, the signals are multiplied by the frequency response of the filter. With the
filters defined as passband filters, this calculation gives ideal, zero-phase filtering (brick wall filters).
Frequencies that have been assigned zero weight by the filter (outside the passband or via frequency
response) are removed.
When you apply idfiltto an idfrddata object, the data is first converted to a frequency-domain
iddataobject (see Transforming Between Frequency-Domain and Frequency-Response Data on page
2-141). The result is an iddataobject.
More About Filtering Data on page 2-120

Generating Data Using Simulation

In this section...
Commands for Generating Data Using Simulation on page 2-130 Create Periodic Input Data on
page 2-131
Generate Output Data Using Simulation on page 2-132 Simulating Data Using Other MathWorks
Products on page 2-133

Commands for Generating Data Using Simulation

You can generate input data and then use it with a model to create output data.
Simulating output data requires that you have a model with known coefficients. For more information
about commands for constructing models, see Commands for Constructing Model Structures on page
1-25.
To generate input data, use idinputto construct a signal with the desired characteristics, such as a random
Gaussian or binary signal or a sinusoid. idinputreturns a matrix of input values.
The following table lists the commands you can use to simulate output data. For more information about
these commands, see the corresponding reference pages.
Commands for Generating Data
Command Description Example
idinput Constructs a signal
with the desired
characteristics, such as a random Gaussian or binary signal or a
u = iddata([],...
idinput(400,'rbs',[0 0.3]));
Commands for Generating Data (Continued)
Command Description Example
sinusoid, and returns a matrix of input values.
sim Simulates response data based on existing linear or nonlinear parametric model in the MATLAB
workspace.
To simulate the model output yfor a given input, use the following command:
y = sim(m,data)
mis the model object name, and datais input data matrix or iddataobject.

Create Periodic Input Data

This example shows how to create a periodic random Gaussian input signal using idinput.
1Create a periodic input for one input and consisting of five periods, where each period is 300 samples.
per_u = idinput([300 1 5]); 2Create an iddataobject using the periodic input and leaving the output

empty.
u = iddata([],per_u,'Period',.300); 3View the data characteristics in time- and frequency-domain.
% Plot data in time-domain. plot(u);
% Plot the spectrum.
spectrum(spa(u));
4(Optional)

Simulate model output using the data.

% Construct a polynomial model. m0 =idpoly([1 -1.5 0.7],[0 1 0.5]); % Simulate model output with
Gaussian noise. sim_opt = simOptions('AddNoise',true); sim(m0,u,sim_opt);

Generate Output Data Using Simulation

This example shows how to generate output data by simulating a model using an input signal created
using idinput.
You use the generated data to estimate a model of the same order as the model used to generate the data.
Then, you check how closely both models match to understand the effects of input data characteristics
and noise on the estimation.
1Create

an ARMAX model with known coefficients.

A = [1 -1.2 0.7];
B = {[0 1 0.5 0.1],[0 1.5 -0.5],[0 -0.1 0.5 -0.1]}; C=[1 000 0];
Ts = 1;
m = idpoly(A,B,C,'Ts',1);
The leading zeros in the Bmatrix indicate the input delay (nk), which is 1 for each input channel.
2Construct a pseudorandom binary input data.
u = idinput([200,3],'prbs');
3Simulate model output with noise using the input data.
sim(m0,u,simOptions('AddNoise',true))
4Represent the simulation data as an iddataobject.
iodata = iddata(y,u,m.Ts);
5(Optional) Estimate a model of the same order as musing iodata.
na =2;nb= [3 23];nc= 4;nk= [1 11]; me = armax(iodata,[na,nb,nc,nk]);
Use bode(m,me)and compare(iodata,me)to check how closely meand m match.

Simulating Data Using Other MathWorks Products

You can also simulate data using the Simulink and Signal Processing Toolbox software. Data simulated
outside the System Identification Toolbox product must be in the MATLAB workspace as double
matrices. For more information about simulating models using the Simulink software, see Simulating
Identified Model Output in Simulink on page 11-5.

Transforming Between Time- and Frequency-Domain Data

In this section...
Transforming Data Domain in the GUI on page 2-134
Transforming Data Domain at the Command Line on page 2-139

Transforming Data Domain in the GUI

Transforming Time-Domain Data on page 2-134
Transforming Frequency-Domain Data on page 2-136
Transforming Frequency-Response Data on page 2-137
See Also on page 2-139
Transforming Time-Domain Data
In the System Identification Tool GUI, time-domain data has an icon with a white background. You can
transform time-domain data to frequency-domain or frequency-response data. The frequency values of
the resulting frequency
vector range from 0to the Nyquist frequency fS = Ts ,where Tsis the sampling interval.
Transforming from time-domain to frequency-response data is equivalent to estimating a model from the
data using the spafdrmethod.
1In the System Identification Tool GUI, drag the icon of the data you want to transform to the Working
Data rectangle.
In the Operations area, select <--Preprocess > Transform data in the dropdownmenutoopentheTransformDatadialogbox.
3In the Transform to list, select one of the following:
2

Frequency FunctionCreateanew idfrdobject using the spafdr method. Gotostep4.

Frequency Domain DataCreateanew iddataobject using the fft method. Gotostep6.

4In the Frequency Spacing list, select the spacing of the frequencies at which the frequency function is
estimated:
linear Uniform spacing of frequency values between the endpoints.
logarithmic Base-10 logarithmic spacing of frequency values between the endpoints.
5In the Number of Frequencies field, enter the number of frequency values.
6In the Name of new data field, type the name of the new data set. This name must be unique in the

Data Board.
7Click Transform to add the new data set to the Data Board in the System Identification Tool GUI.
Transforming Frequency-Domain Data
In the System Identification Tool GUI, frequency-domain data has an icon with a green background.
You can transform frequency-domain data to time-domain or frequency-response (frequency-function)
data.
Transforming from time-domain or frequency-domain data to frequency-response data is equivalent to
estimating a nonparametric model of the data using the spafdrmethod.
1In

the System Identification Tool GUI, drag the icon of the data you want to transform to the Working
Data rectangle.
2Select <--Preprocess > Transform data.
3In the Transform to list, select one of the following:
Frequency FunctionCreateanew idfrdobject using the spafdr method. Gotostep4.
Time Domain DataCreateanew iddataobject using the ifft (inverse fast Fourier transform) method.
Go to step 6.
4In the Frequency Spacing list, select the spacing of the frequencies at which the frequency function is
estimated:
linear Uniform spacing of frequency values between the endpoints.
logarithmic Base-10 logarithmic spacing of frequency values between the endpoints.
5In the Number of Frequencies field, enter the number of frequency values.
6In the Name of new data field, type the name of the new data set. This name must be unique in the
Data Board.
7Click Transform to add the new data set to the Data Board in the System Identification Tool GUI.
Transforming Frequency-Response Data
In the System Identification Tool GUI, frequency-response data has an icon with a yellow background.
You can transform frequency-response data to frequency-domain data (iddataobject) or to frequencyresponse data with a different frequency resolution.
When you select to transform single-input/single-output (SISO) frequency-response data to frequencydomain data, the toolbox creates outputs that equal the frequency responses, and inputs equal to 1.
Therefore, the ratio between the Fourier transform of the output and the Fourier transform of the input is
equal to the system frequency response.
For the multiple-input case, the toolbox transforms the frequency-response data to frequency-domain
data as if each input contributes independently to the entire output of the system and then combines
information. For example, if a system has three inputs, u1, u2,and u3and two frequency samples, the
input matrix is set to:
? 10 0? ?10 0?
?

01 0? ? ? ?01 0? ? ? ?00 1? ?00 1? ? ?

??

In general, for nuinputs and nssamples (the number of frequencies), the input matrix has nucolumns and
(ns? nu)rows.
Note To create a separate experiment for the response from each input, see Transforming Between
Frequency-Domain and Frequency-Response Data on page 2-141.
When you transform frequency-response data by changing its frequency resolution, you can modify the
number of frequency values by changing between linear or logarithmic spacing. You might specify
variable frequency spacing to increase the number of data points near the system resonance frequencies,
and also make the frequency vector coarser in the region outside the system dynamics. Typically, highfrequency noise dominates away from frequencies where interesting system dynamics occur. The
System Identification Tool GUI lets you specify logarithmic frequency spacing, which results in a
variable frequency resolution.
Note The spafdrcommand lets you lets you specify any variable frequency resolution.
1In the System Identification Tool GUI, drag the icon of the data you want to transform to the Working
Data rectangle.
2Select <--Preprocess > Transform data.
In the Transform to list, select one of the following:
Frequency Domain DataCreateanew iddataobject. Gotostep6.
3

Frequency FunctionCreateanew idfrdobject with different resolution (number and spacing of

frequencies) using the spafdrmethod. Go to step 4.
4In

the Frequency Spacing list, select the spacing of the frequencies at which the frequency function is
estimated:
linear Uniform spacing of frequency values between the endpoints.
logarithmic Base-10 logarithmic spacing of frequency values between the endpoints.
5In the Number of Frequencies field, enter the number of frequency values.
6In the Name of new data field, type the name of the new data set. This name must be unique in the
Data Board.
7Click Transform to add the new data set to the Data Board in the System Identification Tool GUI.
See Also
For a description of time-domain, frequency-domain, and frequency-response data, see Representing
Data in MATLAB Workspace on page 2-9.
To learn how to transform data at the command line instead of the GUI, see Transforming Data Domain
at the Command Line on page 2-139.

Transforming Data Domain at the Command Line

Supported Data Transformations on page 2-139
Transforming Between Time and Frequency Domain on page 2-140

Transforming Between Frequency-Domain and Frequency-Response Data on page 2-141

See Also on page 2-142
Supported Data Transformations
The following table shows the different ways you can transform data from one data domain to another. If
the transformation is supported for a given row and column combination in the table, the method used
by the software is listed in the cell at their intersection.
Original Data Format
To Time Domain (iddata object)
Time Domain (iddataobject)
Frequency Domain
(iddataobject) N/A.
Yes, using ifft.
To Frequency To Frequency Domain Function
(iddata object) (idfrd object)
Yes, using fft.Yes,using etfe, spa,or spafdr. N/A. Yes, using etfe, spa,or spafdr.
Frequency Function
(idfrdobject)
No. Yes. Calculation
creates
frequency-domain iddataobject
Yes. Calculates a frequency
function with different
Original Data To Time Domain Format (iddata object) To Frequency Domain
(iddata object) To Frequency Function
(idfrd object)
that has the
same ratio
between output and input as the original idfrd objects response data.
resolution
(number and spacing of
frequencies) using spafdr.
Transforming Between Time and Frequency Domain The iddataobject stores time-domain or
frequency-domain data. The following table summarizes the commands for transforming data between
time and frequency domains.

Command Description Syntax Example

fft Transforms time-domain data to the frequency domain.
You can specify N,the number of frequency values.
ifft Transforms
frequency-domain data to the time domain. Frequencies are linear and equally spaced.
To transform time-domain iddataobject t_datato frequency-domain iddata object f_datawith N frequency
points, use:
f_data =
fft(t_data,N)
To transform
frequency-domainiddata object f_datato
time-domain iddata object t_data,use:
t_data =
ifft(f_data)
Transforming Between Frequency-Domain and
Frequency-Response Data
You can transform frequency-response data to frequency-domain data (iddataobject). The idfrdobject
represents complex frequency-response of the system at different frequencies. For a description of this
type of data, see Frequency-Response Data Representation on page 2-13.
When you select to transform single-input/single-output (SISO) frequency-response data to frequencydomain data, the toolbox creates outputs that equal the frequency responses, and inputs equal to 1.
Therefore, the ratio between the Fourier transform of the output and the Fourier transform of the input is
equal to the system frequency response.
For information about changing the frequency resolution of frequency-response data to a new constant
or variable (frequency-dependent) resolution, see the spafdrreference page. You might use this feature to
increase the number of data points near the system resonance frequencies and make the frequency vector
coarser in the region outside the system dynamics. Typically, high-frequency noise dominates away from
frequencies where interesting system dynamics occur.
Note You cannot transform an idfrdobject to a time-domain iddataobject.
To transform an idfrdobject with the name idfrdobjto a frequency-domain iddataobject, use the
following syntax:
dataf = iddata(idfrdobj)
The resulting frequency-domain iddataobject contains values at the same frequencies as the original
idfrdobject.
For the multiple-input case, the toolbox represents frequency-response data as if each input contributes
independently to the entire output of the system and then combines information. For example, if a

system has three inputs, u1, u2,and u3and two frequency samples, the input matrix is set to: ?10 0? ?10
0?
?

01 0? ? ? ?01 0? ? ? ?00 1? ?00 1? ? ?

??

In general, for nuinputs and nssamples, the input matrix has nucolumns and (ns? nu)rows.
If you have nyoutputs, the transformation operation produces an output matrix has nycolumns and (ns?
nu) rows using the values in the complex frequency response G(iw) matrix (ny-bynu-byns). In this
example, y1is determined by unfolding G(1,1,:), G(1,2,:),and G(1,3,:)into three column vectors and
vertically concatenating these vectors into a single column. Similarly, y2is determined by unfolding
G(2,1,:), G(2,2,:),and G(2,3,:) into three column vectors and vertically concatenating these vectors.
If you are working with multiple inputs, you also have the option of storing the contribution by each
input as an independent experiment in a multiexperiment data set. To transform an idfrdobject with the
name idfrdobjto a multiexperiment data set datf, where each experiment corresponds to each of the
inputs in idfrdobj
datf = iddata(idfrdobj,'me')
In this example, the additional argument 'me'specifies that multiple experiments are created.
By default, transformation from frequency-response to frequency-domain data strips away frequencies
where the response is infor NaN. To preserve the entire frequency vector, use datf =
iddata(idfrdobj,'inf').Formore information, type help idfrd/iddata.
See Also
Transforming from time-domain or frequency-domain data to frequency-response data is equivalent to
creating a frequency-response model from the data. For more information, see Identifying FrequencyResponse Models on page 3-8.

Manipulating Complex-Valued Data

In this section...
Supported Operations for Complex Data on page 2-144 Processing Complex iddata Signals at the
Command Line on page 2-144

Supported Operations for Complex Data

System Identification Toolbox estimation algorithms support complex data. For example, the following
estimation commands estimate complex models from complex data: ar, armax, arx, bj, ivar, iv4, oe, pem,
spa, tfest, ssest, and n4sid.
Model transformation routines, such as freqrespand zpkdata,workfor complex-valued models. However,
they do not provide pole-zero confidence regions. For complex models, the parameter variancecovariance information refers to the complex-valued parameters and the accuracy of the real and
imaginary is not computed separately.

The display commands compareand plotalso work with complex-valued data and models. To plot the
real and imaginary parts of the data separately, use plot(real(data))and plot(imag(data)),respectively.

Processing Complex iddata Signals at the Command Line

If the iddataobject datacontains complex values, you can use the following commands to process the
complex data and create a new iddataobject.
Command Description
abs(data) Absolute value of complex signals in iddataobject.
angle(data) Phase angle (in radians) of each complex signals in iddataobject.
Command
complex(data)
imag(data)
isreal(data)
real(data)
Description
For time-domain data, this command makes the iddata object complexeven when the imaginary parts
are zero. For frequency-domain data that only stores the values for nonnegative frequencies, such that
realdata(data)=1, it adds signal values for negative frequencies using complex conjugation.
Selects the imaginary parts of each signal in iddata object.
1 when data(time-domain or frequency-domain) contains only real input and output signals, and returns
0when data(time-domain or frequency-domain) contains complex signals.
Real part of complex signals in iddataobject. realdata(data) Returns a value of 1when datais a realvalued, time-domain signal, and returns 0otherwise.
For example, suppose that you create a frequency-domain iddataobject Datf by applying fftto a realvalued time-domain signal to take the Fourier transform of the signal. The following is true for Datf:
isreal(Datf) = 0 realdata(Datf) = 1

Linear Model Identification

Black-Box Modeling on page 3-3
Identifying Frequency-Response Models on page 3-8
Identifying Impulse-Response Models on page 3-17
Identifying Process Models on page 3-26
Identifying Input-Output Polynomial Models on page 3-45
Identifying State-Space Models on page 3-79
Identifying Transfer Function Models on page 3-113
Refining Linear Parametric Models on page 3-130

Refine ARMAX Model with Initial Parameter Guesses at Command Line on page 3-133
Refine Initial ARMAX Model at Command Line on page 3-134
Extracting Numerical Model Data on page 3-136
Transforming Between Discrete-Time and Continuous-Time Representations on page 3-139
Continuous-Discrete Conversion Methods on page 3-143
Effect of Input Intersample Behavior on Continuous-Time Models on page 3-153
Transforming Between Linear Model Representations on page 3-156
Subreferencing Models on page 3-159
Concatenating Models on page 3-164
Merging Models on page 3-168
Building and Estimating Process Models Using System Identification Toolbox on page 3-169
Determining Model Order and Delay on page 3-195
Model Structure Selection: Determining Model Order and Input Delay on page 3-196
Frequency Domain Identification: Estimating Models Using Frequency Domain Data on page 3-211
Building Structured and User-Defined Models Using System Identification Toolbox on page 3-235

Black-Box Modeling
In this section...
Selecting Black-Box Model Structure and Order on page 3-3 When to Use Nonlinear Model
Structures? on page 3-5 Black-Box Estimation Example on page 3-5

Selecting Black-Box Model Structure and Order

Black-box modeling is useful when your primary interest is in fitting the data regardless of a particular
mathematical structure of the model. The toolbox provides several linear and nonlinear black-box model
structures, which have traditionally been useful for representing dynamic systems. These model
structures vary in complexity depending on the flexibility you need to account for the dynamics and
noise in your system. You can choose one of these structures and compute its parameters to fit the
measured response data.
Black-box modeling is usually a trial-and-error process, where you estimate the parameters of various
structures and compare the results. Typically, you start with the simple linear model structure and
progress to more complex structures. You might also choose a model structure because you are more
familiar with this structure or because you have specific application needs.
The simplest linear black-box structures require the fewest options to configure:
Transfer function, with a given number of poles and zeros.
Linear ARX model, which is the simplest input-output polynomial model.

State-space model, which you can estimate by specifying the number of model states
Estimation of some of these structures also uses noniterative estimation algorithms, which further
reduces complexity.
You can configure a model structure using the model order. The definition of model order varies
depending on the type of model you select. For example, if you choose a transfer function
representation, the model order is related to the number of poles and zeros. For state-space
representation, the model order corresponds to the number of states. In some cases, such as for linear
ARX and state-space model structures, you can estimate the model order from the data.
If the simple model structures do not produce good models, you can select more complex model
structures by:
Specifying a higher model order for the same linear model structure. Higher model order increases the
model flexibility for capturing complex phenomena. However, unnecessarily high orders can make the
model less reliable.
Explicitly modeling the noise:
y(t)=Gu(t)+He(t)
where H models the additive disturbance by treating the disturbance as the output of a linear system
driven by a white noise source e(t).
Using a model structure that explicitly models the additive disturbance can help to improve the accuracy
of the measured component G.Furthermore, such a model structure is useful when your main interest is
using the model for predicting future response values.
Using a different linear model structure.
See Linear Model Structureson page 1-21 in the Users Guide.
Using a nonlinear model structure.
Nonlinear models have more flexibility in capturing complex phenomena than linear models of similar
orders. See Available Nonlinear Models on page 1-32 in Users Guide.
Ultimately, you choose the simplest model structure that provides the best fit to your measured data. For
more information, see Estimating Linear Models Using Quick Start.
Regardless of the structure you choose for estimation, you can simplify the model for your application
needs. For example, you can separate out the measured dynamics (G) from the noise dynamics (H) to
obtain a simpler model that represents just the relationship between y and u. You can also linearize a
nonlinear model about an operating point.

When to Use Nonlinear Model Structures?

A linear model isoften sufficient to accurately describe the system dynamics and, in most cases, you
should first try to fit linear models. If the linear model output does notadequately reproduce the
measured output, you might need to use a nonlinear model.

You can assessthe need to use a nonlinear model structure by plotting the response of the system to an
input. If you notice that the responses differ depending onthe input level or input sign, try using a
nonlinear model. For example, if the output response to an input step up is faster than the response to a
step down, you might need a nonlinear model.
Before building a nonlinear model of a system that you know is nonlinear, try transforming the input and
output variables such that the relationship between the transformed variables is linear. For example,
consider a system that has current and voltage as inputs to an immersion heater, and the temperature of
the heated liquid as an output. The output depends on the inputs via the power of the heater, which is
equal to the product of current and voltage. Instead of building a nonlinear model for this two-input and
one-output system, you can create a new input variable by taking the product of current and voltage and
then build a linear model that describes the relationship between power and temperature.
If you cannot determine variable transformations that yield a linear relationship between input and
output variables, you can use nonlinear structures such as Nonlinear ARX or Hammerstein-Wiener
models. For a list of supported nonlinear model structures and when to use them, see Available
Nonlinear Models on page 1-32 in Users Guide.

Black-Box Estimation Example

You can use the GUI or commands to estimate linear and nonlinear models of various structures. In most
cases, you choose a model structure and estimate the model parameters using a single command.
Consider the mass-spring-damper system, described in About Dynamic Systems and Models. If you
do not know the equation of motion of this system, you can use a black-box modeling approach to build
a model. For example, you can estimate transfer functions or state-space models by specifying the orders
of these model structures.
A transfer function is a ratio of polynomials:

Gs bbsbs2 ()= ++ +
1fs f s2++ + ()
For the mass-spring damper system, this transfer function is: Gs()1

ms 2 cs
which is a system with no zeros and 2 poles.
In discrete-time, the transfer function of the mass-spring-damper system can be:
bz
1

Gz() = 1 fz 1
++()
where the model orders correspond to the number of coefficients of the numerator and the denominator
(nb=1and nf= 2) and the input-output delay equals the lowest order exponent of z1in the numerator
(nk=1).
In continuous-time, you can build a linear transfer function model using the tfestcommand:
m = tfest(data, 2, 0)
where datais your measured input-output data, represented as an iddata object and the model order is the
set of number of poles (2) and the number of zeros (0).
Similarly, you can build a discrete-time model Output Error structure using the following command:
m = oe(data, [1 2 1])
The model order is [ nb nf nk]=[12 1]. Usually, you do not know the model orders in advance. You
should try several model order values until you find the orders that produce an acceptable model.
Alternatively, you can choose a state-space structure to represent the mass-spring-damper system and
estimate the model parameters using the ssestor the n4sidcommand:
m = ssest(data, 2)
where order= 2represents the number of states in the model.
In black-box modeling, you do not need the systems equation of motiononly a guess of the model
orders.
For more information about building models, see Steps for Using the System Identification Tool GUI
on page 12-2 and Model Estimation Commands on page 1-40 in the Users Guide.

Identifying Frequency-Response Models

In this section...
What Is a Frequency-Response Model? on page 3-8
Data Supported by Frequency-Response Models on page 3-9 How to Estimate Frequency-Response
Models in the GUI on page 3-9
How to Estimate Frequency-Response Models at the Command Line on page 3-11
Selecting the Method for Computing Spectral Models on page 3-11 Controlling Frequency
Resolution of Spectral Models on page 3-12 Spectrum Normalization on page 3-14

What Is a Frequency-Response Model?

You can estimate frequency-response models and visualize the responses on a Bode plot, which shows
the amplitude change and the phase shift as a function of the sinusoid frequency.

The frequency-response function describes the steady-state response of a system to sinusoidal inputs.
For a linear system, a sinusoidal input of a specific frequency results in an output that is also a sinusoid
with the same frequency, but with a different amplitude and phase. The frequency-response function
describes the amplitude change and phase shift as a function of frequency.
For a discrete-time system sampled with a time interval T,the frequency-response model G(z) relates the
Z-transforms of the input U(z) and output Y(z):
Yz GzU z= () ()
In other words, the frequency-response function, G(eiwT),istheLaplace
transformoftheimpulseresponsethatisevaluatedontheimaginaryaxis. The frequency-response function is
the transfer function G(z) evaluated on the unit circle.
The estimation result is an idfrdmodel, which stores the estimated frequency response and its
covariance.

Data Supported by Frequency-Response Models

You can estimate spectral analysis models from data with the following characteristics:
Complex or real data.
Time- or frequency-domain iddataor idfrddata object. To learn more about estimating time-series
models, see Time-Series Model Identification.
Single- or multiple-output data.

How to Estimate Frequency-Response Models in the GUI

You must have already imported your data into the GUI and performed any necessary preprocessing
operations. For more information, see Data Preparation.
To estimate frequency-response models in the System Identification Tool GUI:
1In the System Identification Tool GUI, select Estimate > Spectral models to open the Spectral Model
dialog box.
In the Method list, select the spectral analysis method you want to use. For information about each
method, see Selecting the Method for Computing Spectral Models on page 3-11.
2

3Specify

the frequencies at which to compute the spectral model in one of the following ways:

In the Frequencies field, enter either a vector of values, a MATLAB expression that evaluates to a
vector, or a variable name of a vector in the MATLAB workspace. For example, logspace(-1,2,500).
Use the combination of Frequency Spacing and Frequencies to construct the frequency vector of
values:
In the Frequency Spacing list, select Linearor Logarithmic frequency spacing.

Note For etfe,onlythe Linearoption is available.

In the Frequencies field, enter the number of frequency points.
For time-domain data, the frequency ranges from 0 to the Nyquist frequency. For frequency-domain
data, the frequency ranges from the smallest to the largest frequency in the data set.
In the Frequency Resolution field, enter the frequency resolution, as described in Controlling
Frequency Resolution of Spectral Models on page 3-12. To use the default value, enter defaultor,
equivalently, the empty matrix [].
4

5In

the Model Name field, enter the name of the correlation analysis model. The model name should be
unique in the Model Board.
6Click Estimate to add this model to the Model Board in the System Identification Tool GUI.
7In the Spectral Model dialog box, click Close.
To view the frequency-response plot, select the Frequency resp check box in the System
Identification Tool GUI. For more information about working with this plot, see Frequency Response
Plots on page 8-42.
8

To view the estimated disturbance spectrum, select the Noise spectrum check box in the System
Identification Tool GUI. For more information about working with this plot, see Noise Spectrum Plots
on page 8-51.
9

10Validate

the model after estimating it. For more information, see Model Validation.

To export the model to the MATLAB workspace, drag it to the To Workspace rectangle in the System
Identification Tool GUI. You can retrieve the responses from the resulting idfrdmodel object using the
bodeor nyquist command.

How to Estimate Frequency-Response Models at the Command Line

You can use the etfe, spa,and spafdrcommands to estimate spectral models. The following table provides
a brief description of each command and usage examples.
The resultingmodels are stored as idfrdmodel objects. For detailed information about the commands and
their arguments, see the corresponding reference page.
Commands for Frequency Response
Command Description
etfe Estimatesan empirical transfer function using Fourier analysis.
spa Estimates a frequency responsewithafixed frequency resolution using spectral analysis.
spafdr Estimates a frequency response with a variable frequency resolution usingspectral analysis.
Usage
To estimate a model m, use the following syntax:

m=etfe(data) To estimate a model m, use the following syntax:

m=spa(data)
To estimate a model m, use the following syntax:
m=spafdr(data,R,w)
where Ris the resolution vector and wis the frequency vector.
Validate the model after estimating it. For more information, see Model Validation.

Selecting the Method for Computing Spectral Models

This section describes how to select the method for computing spectral models in the estimation
procedures How to Estimate Frequency-Response Models
intheGUIonpage3-9andHowtoEstimateFrequency-ResponseModelsat the Command Line on page
3-11.
You can choose from the following three spectral-analysis methods:
etfe(Empirical Transfer Function Estimate)
For input-output data. This method computes the ratio of the Fourier transform of the output to the
Fourier transform of the input. For time-series data. This method computes a periodogram as the
normalized absolute squares of the Fourier transform of the time series.
ETFE works well for highly resonant systems or narrowband systems. The drawback of this method is
that it requires linearly spaced frequency values, does not estimate the disturbance spectrum, and does
not provide confidence intervals. ETFE also works well for periodic inputs and computes exact
estimates at multiples of the fundamental frequency of the input and their ratio.
spa(SPectral Analysis)
This method is the Blackman-Tukey spectral analysis method, where windowed versions of the
covariance functions are Fourier transformed.
spafdr(SPectral Analysis with Frequency Dependent Resolution)
This method is a variant of the Blackman-Tukey spectral analysis method with frequency-dependent
resolution. First, the algorithm computes Fourier transforms of the inputs and outputs. Next, the products
of the transformed inputs and outputs with the conjugate input transform are smoothed over local
frequency regions. The widths of the local frequency regions can vary as a function of frequency. The
ratio of these averages computes the frequency-response estimate.

Controlling Frequency Resolution of Spectral Models

What Is Frequency Resolution? on page 3-13
Frequency Resolution for etfe and spa on page 3-13
Frequency Resolution for spafdr on page 3-13
etfe Frequency Resolution for Periodic Input on page 3-14
This section supports the estimation procedures How to Estimate Frequency-Response Models in the
GUI on page 3-9 and How to Estimate Frequency-Response Models at the Command Line on page
3-11.

What Is Frequency Resolution?

Frequency resolution is the size of the smallest frequency for which details in the frequency response
and the spectrum can be resolved by the estimate. A resolution of 0.1 rad/s means that the frequency
response variations at frequency intervals at or below 0.1 rad/s are not resolved.
Note Finer resolution results in greater uncertainty in the model estimate.
Specifying the frequency resolution for etfeand spais different than for spafdr.
Frequency Resolution for etfe and spa
For etfeand spa, the frequency resolution is approximately equal to the following value:
2? radians ? M
?
sampling interval?
??
M is a scalar integer that sets the size of the lag window. The value of M controls the trade-off between
bias and variance in the spectral estimate.
Thedefaultvalueof M for spais good for systems without sharp resonances. For etfe, the default value of
M gives the maximum resolution.
Alargevalueof M gives good resolution, but results in more uncertain estimates. If a true frequency
function has sharp peak, you should specify higher Mvalues.
Frequency Resolution for spafdr
In case of etfeand spa, the frequency response is defined over a uniform frequency range, 0-Fs/2 radians
per second, where Fsis the sampling frequencyequal to twice the Nyquist frequency. In contrast,
spafdrlets you increase the resolution in a specific frequency range, such as near a resonance frequency.
Conversely, you can make the frequency grid coarser in the region where the noise dominatesat higher
frequencies, for example. Such customizing of the frequency grid assists in the estimation process by
achieving high fidelity in the frequency range of interest.
For spafdr, the frequency resolution around the frequency k is the value R(k). You can enter R(k) in any
one of the following ways:
Scalar value of the constant frequency resolution value in radians per second.
Note The scalar R is inversely related to the M value used for etfeand spa.
Vector of frequency values the same size as the frequency vector.
Expression using MATLAB workspace variables and evaluates to a resolution vector that is the same
size as the frequency vector.
Thedefaultvalueoftheresolutionfor spafdris twice the difference between neighboring frequencies in the
frequency vector.
etfe Frequency Resolution for Periodic Input
If the input data is marked as periodic and contains an integer number of periods (data.Periodis an
integer), etfecomputes the frequency response at
frequencies
T

2k k

where k= 12, ,...,Period.()

For periodic data, the frequency resolution is ignored.

Spectrum Normalization
The spectrum of a signal is the square of the Fourier transform of the signal. The spectral estimate using
the commands spa, spafdr,and etfeis normalized by the sampling interval T:
M
iwT

()= TRkTe WkM( )yy

kM
=

where WM(k) is the lag window, and M is the width of the lag window. The output covariance Ry(kT) is
given by the following discrete representation:
RkT
y

()
1

=ylT kT ylT( )( ) N

l=1

Because there is no scaling in a discrete Fourier transform of a vector, the purpose of T is to relate the
discrete transform of a vector to the physically meaningful transform of the measured signal. This
normalization sets the
units of y() as power per radians per unit time, and makes the frequency

units radians per unit time.

The scaling factor of T is necessary to preserve the energy density of the spectrum after interpolation or
decimation.
By Parsevals theorem, the average energy of the signal must equal the average energy in the estimated
spectrum, as follows:
Ey
2

()
=
1

/
T

()
d
2 /T
SEyt2()

1 /T
2Sdy()/T
Tocomparetheleftsideoftheequation(S1)totherightside(S2), enter the following commands in the
MATLAB Command Window:
load iddata1
% Create time-series iddata object
y = z1(:,1,[]);
% Define sample interval from the data T = y.Ts;
% Estimate frequency response
sp = spa(y);
% Remove spurious dimensions
phiy = squeeze(sp.spec);
% Compute average energy from the estimated % energy spectrum, where S1 is scaled by T S1 =
sum(phiy)/length(phiy)/T
% Compute average energy of the signal S2 = sum(y.y.^2)/size(y,1)
In this code, phiycontains y() between = 0 and = T with the

frequency step given as follows:

? ? ?T ?length(phiy)?
??
MATLAB computes the following values for S1and S2:
S1 =
19.2076 S2 =
19.4646
Thus, the average energy of the signal approximately equals the average energy in the estimated
spectrum.

Identifying Impulse-Response Models

In this section...
What Is Time-Domain Correlation Analysis? on page 3-17 Data Supported by Correlation Analysis
on page 3-17 How to Estimate Impulse-Response Models Using the GUI on page 3-18
How to Estimate Impulse-Response Models at the Command Line on page 3-19
How to Compute Response Values on page 3-21
How to Identify Delay Using Transient-Response Plots on page 3-21 Correlation Analysis
Algorithm on page 3-23

What Is Time-Domain Correlation Analysis?

Time-domain correlation analysis refers to non-parametric estimation of the impulse response of
dynamic systems as a finite impulse response (FIR) model from the data. Correlation analysis assumes a
linear system and does not require a specific model structure.
Impulse response is the output signal that results when the input is an impulse and has the following
definition for a discrete model:
ut() =>00 ut() ==10
Theresponsetoaninput u(t) is equal to the convolution of the impulse response, as follows:
t

yt() =()?h t z u z dz
0

Data Supported by Correlation Analysis

You can estimate impulse-response models from data with the following characteristics:
Real or complex data.
Single- or multiple-output data.
Time- or frequency-domain data with nonzero sampling time.
Time-domain data must be regularly sampled. You cannot use time-series data for correlation analysis.

How to Estimate Impulse-Response Models Using the GUI

Before you can perform this task, you must have:
Imported data into the System Identification Tool GUI. See Importing Time-Domain Data into the
GUI on page 2-18. For supported data formats, see Data Supported by Correlation Analysis on page
3-17.
Performed any required data preprocessing operations. To improve the accuracy of your model, you
should detrend your data. See Ways to Prepare Data for System Identification on page 2-6.
To estimate in the System Identification Tool GUI using time-domain correlation analysis:
1In the System Identification Tool GUI, select Estimate > Correlation models to open the Correlation
Model dialog box.
In the Time span (s) field, specify a scalar value as the time interval over which the impulse or step
response is calculated. For a scalar time span T, the resulting response is plotted from -T/4 to T.
2

Tip You can also enter a 2-D vector in the format [min_value max_value].
3In the Order of whitening filter field, specify the filter order.

The prewhitening filter is determined by modeling the input as an autoregressive process of order N. The
algorithm applies a filter of the form A(q)u(t)=u_F(t). That is, the input u(t) is subjected to an FIR filter
A to produce the filtered signal u_F(t). Prewhitening the input by applying a whitening filter before
estimation might improve the quality of the estimated impulse response g.
The order of the prewhitening filter, N, is the order of the A filter. N equals the number of lags. The
default value of N is 10, which you can also specify as [].
4In

the Model Name field, enter the name of the correlation analysis model. The name of the model
should be unique in the Model Board.
5Click Estimate to add this model to the Model Board in the System Identification Tool GUI.
6In the Correlation Model dialog box, click Close.
Next Steps
Export the model to the MATLAB workspace for further analysis by dragging it to the To Workspace
rectangle in the System Identification Tool GUI.
View the transient response plot by selecting the Transient resp check box in the System
Identification Tool GUI. For more information about working with this plot and selecting to view
impulse- versus step-response, see Impulse and Step Response Plots on page 8-33.

How to Estimate Impulse-Response Models at the Command Line

Before you can perform this task, you must have:
Input/output or frequency-response data. See Representing Time- and Frequency-Domain Data Using
iddata Objects on page 2-55. For supported data formats, see Data Supported by Correlation Analysis
on page 3-17.
Performed any required data preprocessing operations. If you use time-domain data, you can detrend it
before estimation. See Ways to Prepare Data for System Identification on page 2-6.
Use impulseestto compute impulse response models. impulseestestimates a high-order, noncausal FIR
model using correlation analysis. The resulting models are stored as idtfmodel objects and contain
impulse-response coefficients in the model numerator.
To estimate the model mand plot the impulse or step response, use the following syntax:
m=impulseest(data,N); impulse(m,Time);
step(m,Time);
where datais a single- or multiple-output iddataor idfrdobject. Nis a scalar value specifying the order of
the FIR system corresponding to the time range 0:Ts:(N-1)*Ts, where Ts is the data sampling time.
You can also specify estimation options, such as regularizing kernel, pre-whitening filter order and data
offsets, using impulseestOptionsand pass them as an input to impulseest. For example:

opt = impulseestOptions('RegulKernel','TC')); m = impulseest(data,N,opt);

To view the confidence region for the estimated response, use impulseplot and stepplotto create the plot.
Then use showConfidence.
For example:
h = stepplot(m,Time);
showConfidence(h,3) % 3 std confidence region
Note crais an alternative method for computing impulse response from time-domain data only.
Next Steps
Perform model analysis. See Validating Models After Estimation on page 8-3.

How to Compute Response Values

You can use impulseand stepcommands with output arguments to get the numerical impulse- and stepresponse vectors as a function of time, respectively.
To get the numerical response values:
1Compute the FIR model by using impulseest, as described in How to Estimate Impulse-Response
Models at the Command Line on page 3-19.
2Applythefollowingsyntaxontheresultingmodel:
% To compute impulse-response data [y,t,~,ysd] = impulse(model) % To compute step-response data
[y,t,~,ysd] = step(model)
where yistheresponsedata, tis the time vector, and ysdis the standard deviations of the response.

How to Identify Delay Using Transient-Response Plots

You can use transient-response plots to estimate the input delay, or dead time, of linear systems. Input
delay represents the time it takes for the output to respond to the input.
In the System Identification Tool GUI. To view the transient response plot, select the Transient resp
check box in the System Identification Tool GUI. For example, the following step response plot shows a
time delay of about 0.25 s before the system responds to the input.

Step Response Plot

At the command line. You can use the impulseplotcommand to plot the impulse response. The time
delay is equal to the first positive peak in the transient response magnitude that is greater than the
confidence region for positive time values.
For example, the following commands create an impulse-response plot with a 1-standard-deviation
confidence region:
% Load sample data
load dry2
ze = dry2(1:500);
opt = impulseestOptions('RegulKernel','TC');
sys = impulseest(ze,40,opt);
h = impulseplot(sys);
showConfidence(h,1);
Instead of using showConfidence, you can plot the confidence interval interactively, by right-clicking on
the plot and selecting Characteristics > Confidence Region.
The resulting figure shows that the first positive peak of the response magnitude, which is greater than
the confidence region for positive time values, occurs at 0.24 s.

Correlation Analysis Algorithm

Correlation analysis refers to methods that estimate the impulse response of a linear model, without
specific assumptions about model orders.
The impulse response, g, is the systems output when the input is an impulse signal. The output response
to a general input, u(t), is obtained as the convolution with the impulse response. In continuous time:

yt() t gu tWWWd
f

In discrete-time:
f

yt g k u t k
k1

The values of g(k)arethe discrete time impulse response coefficients.

You can estimate the values from observed input-output data in several different ways.
impulseestestimates the first n coefficients using the least-squares method to obtain a finite impulse
response (FIR) model of order n.
Several important options are associated with the estimate:

 Prewhitening The input can be pre-whitened by applying an input-whitening filter of order PWto
the data. This minimizes the effect of the neglected tail (k> n) of the impulse response.
1Afilteroforder

PWis applied such that it whitens the input signal u:

1/A = A(u)e,where Ais a polynomial and eis white noise. 2 The inputs and outputs are filtered using the
filter:
uf = Au , yf = Ay
3 The filtered signals ufand yfare used for estimation.
You can specify prewhitening using the PWname-value pair argument of impulseestOptions.
Regularization The least-squares estimate can be regularized. This means that a prior estimate of
the decay and mutual correlation among g(k)is formed and used to merge with the information about
gfrom the observed data. This gives an estimate with less variance, at the price of some bias. You can
choose one of the several kernels to encode the prior estimate.
This option is essential because, often, the model order ncan be quite large. In cases where there is no
regularization, ncan be automatically decreased to secure a reasonable variance.
You can specify the regularizing kernel using the RegulKernelName-Value pair argument of
impulseestOptions.
Autoregressive Parameters The basic underlying FIR model can be complemented by
NAautoregressive parameters, making it an ARX model. n NA

yt ka yt k
k

11

This gives both better results for small nandallowsunbiasedestimates when data are generated in closed
loop. impulseestuses NA = 5 for t>0 and NA = 0 (no autoregressive component) for t<0.
Noncausal effects Response for negative lags. It may happen that the data has been generated
partly by output feedback:
f

ut() h k y t k r t
k0

where h(k) is the impulse response of the regulator and r is a setpoint or disturbance term. The existence
and character of such feedback h can be estimated in the same way as g, simply by trading places
between y and u in the estimation call. Using impulseestwith an indication of negative
delays,mi impulseest(,,), nk 0, returns a model miwith
an impulse response

h(nk),h(nk- ),...,1012( ),...,g nb>@

alignedsothatitcorrespondstolags nk nk 1012,..,, ,,...,nb >@.Thisis
achieved because the input delay (InputDelay) of model miis nk.
For a multi-input multi-output system, the impulse response g(k)isan ny-bynu matrix, where ny is the
number of outputs and nu is the number of inputs. The ij element of the matrix g(k) describes the
behavior of the ith output after an impulse in thejth input.

Identifying Process Models

In this section...
What Is a Process Model? on page 3-26
Data Supported by Process Models on page 3-27
How to Estimate Process Models Using the GUI on page 3-27 How to Estimate Process Models at
the Command Line on page 3-32 Process Model Structure Specification on page 3-40
Estimating Multiple-Input, Multi-Output Process Models on page 3-41 Disturbance Model Structure
for Process Models on page 3-42 Assigning Estimation Weightings on page 3-43
Specifying Initial Conditions for Iterative Estimation Algorithms on page 3-43

What Is a Process Model?

The structure of a process model is a simple continuous-time transfer function that describes linear
system dynamics in terms of one or more of the following elements:
Static gain Kp.
Oneormoretimeconstants Tpk. For complex poles, the time constant is called Tequal to the inverse
of the natural frequencyand the damping coefficient is (zeta).
Process zero Tz.
Possibletimedelay Tdbefore the system output responds to the input (dead time).
Possible enforced integration.
Process models are popular for describing system dynamics in many industries and apply to various
production environments. The advantages of these models are that they are simple, support transport
delay estimation, and the model coefficients have an easy interpretation as poles and zeros.
You can create different model structures by varying the number of poles, adding an integrator, or
adding or removing a time delay or a zero. You can specify a first-, second-, or third-order model, and
the poles can be real or complex (underdamped modes).
For example, the following model structure is a first-order continuous-time process model, where K is
the static gain, Tp1is a time constant, and Tdis the input-to-output delay:

Gs
()
1 sTp1 K pesTd

Data Supported by Process Models

You can estimate low-order (up to third order), continuous-time transfer functions using regularly
sampled time- or frequency-domain iddataor idfrd data objects. The frequency-domain data may have a
zero sample time.
You must import your data into the MATLAB workspace, as described in Data Preparation.

How to Estimate Process Models Using the GUI

Before you can perform this task, you must have
Imported data into the System Identification Tool GUI. See Importing Time-Domain Data into the
GUI on page 2-18. For supported data formats, see Data Supported by Process Models on page 3-27.
Performed any required data preprocessing operations. If you need to model nonzero offsets, such as
when model contains integration behavior, do not detrend your data. In other cases, to improve the
accuracy of your model, you should detrend your data. See Ways to Prepare Data for System
Identification on page 2-6.

1In

the System Identification Tool GUI, select Estimate > Process models to open the Process Models
dialog box.

If your model contains multiple inputs, select the input channel in the Input list. This list only appears
when you have multiple inputs. For more information, see Estimating Multiple-Input, Multi-Output
Process Models on page 3-41.
2

3In

the Model Transfer Function area, specify the model structure using the following options:
Under Poles, select the number of poles, and then select All realor Underdamped.
Note You need at least two poles to allow underdamped modes (complex-conjugate pair).
Select the Zero check box to include a zero, which is a numerator term other than a constant, or clear
the check box to exclude the zero.
Select the Delay check box to include a delay, or clear the check box to exclude the delay.
Select the Integrator check box to include an integrator (self-regulating
process),orclearthecheckboxtoexcludetheintegrator.
The Parameterareashowsasmanyactiveparametersasyouincludedin the model structure.
Note By default, the model Name is set to the acronym that reflects the model structure, as described in
Process Model Structure Specification on page 3-40.

In the Initial Guess area, select Auto-selectedto calculate the initial parameter values for the
estimation. The Initial Guess column in the Parameter table displays Auto. If you do not have a good
guess for the parameter values, Autoworks better than entering an ad hoc value.
4

(Optional) If you approximately know a parameter value, enter this value in the Initial Guess column
of the Parameter table. The estimation algorithm uses this value as a starting point. If you know a
parameter value exactly, enter this value in the Initial Guess column, and also select the corresponding
Known check box in the table to fix its value.
5

If youknow the range of possible values for a parameter, enter these values into the corresponding
Bounds field to help the estimation algorithm.
For example, the following figure shows that the delay value Tdis fixed at 2sandisnotestimated.

In the Disturbance Model list, select one of the available options. For more information about each
option, see Disturbance Model Structure for Process Models on page 3-42.
6

In the Focus list, select how to weigh the relative importance of the fit at different frequencies. For
more information about each option, see Assigning Estimation Weightings on page 3-43.
7

In the Initial state list, specify how you want the algorithm to treat initial states. For more information
about the available options, see Specifying Initial Conditions for Iterative Estimation Algorithms on
page 3-43.
8

Tip If you get a bad fit, you might try setting a specific method for handling initial states, rather than
choosing it automatically.
In the Covariance list, select Estimateif you want the algorithm to compute parameter uncertainties.
Effects of such uncertainties are displayed on plots as model confidence regions.
To omit estimating uncertainty, select None. Skipping uncertainty computation might reduce
computation time for complex models and large data sets.
9

10In

the Model Name field, edit the name of the model or keep the default. The name of the model
should be unique in the Model Board.
To view the estimation progress, select the Display Progress check box. This opens a progress viewer
window in which the estimation progress is reported.
11

12Click

Estimate to add this model to the Model Board in the System Identification Tool GUI.

To stop the search and save the results after the current iteration has been completed, click Stop
Iterations. To continue iterations from the current model, click the Continue button to assign current
parameter values as initial guesses for the next search.
13

Next Steps
Validate the model by selecting the appropriate check box in the Model Views area of the System
Identification Tool GUI. For more information about validating models, see Validating Models After
Estimation on page 8-3.
Refine the model by clicking the Value > Initial Guess button to assign current parameter values as
initial guesses for the next search, edit the Name field, and click Estimate.
Export the model to the MATLAB workspace for further analysis by dragging it to the To Workspace
rectangle in the System Identification Tool GUI.

How to Estimate Process Models at the Command Line

Prerequisites on page 3-33
UsingprocesttoEstimateProcessModelsonpage3-33
Example Estimating Process Models with Free Parameters at the Command Line on page 3-34
Example Estimating Process Models with Fixed Parameters at the Command Line on page 3-36
Prerequisites
Before you can perform this task, you must have

 Input-output data as an iddataobject or frequency response data as frd or idfrdobjects. See

Representing Time- and Frequency-Domain Data Using iddata Objects on page 2-55. For supported
data formats, see Data Supported by Process Models on page 3-27.
Performed any required data preprocessing operations. When working with time domain data, if you
need to model nonzero offsets, such as when model contains integration behavior, do not detrend your
data. In other cases, to improve the accuracy of your model, you should detrend your data. See Ways to
Prepare Data for System Identification on page 2-6.
Using procest to Estimate Process Models
You can estimate process models using the iterative estimation method procestthat minimizes the
prediction errors to obtain maximum likelihood estimates. The resulting models are stored as
idprocmodel objects.
You can use the following general syntax to both configure and estimate process models:
m = procest(data,mod_struc,opt)
datais the estimation data and mod_strucis one of the following:
A string that represents the process model structure, as described in Process Model Structure
Specification on page 3-40.
Atemplate idprocmodel. optis an option set for configuring the estimation of the process model, such
as handling of initial conditions, input offset and numerical search method.
Tip You do not need to construct the model object using idproc before estimation unless you want to
specify initial parameter guesses, minimum/maximum bounds, or fixed parameter values, as described
in Example Estimating Process Models with Fixed Parameters at the Command Line on page 3-36.
For more information about validating a process model, see Validating Models After Estimation on
page 8-3.
You can use procestto refine parameter estimates of an existing process model, as described in Refining
Linear Parametric Models on page 3-130.
For detailed information, see procestand idproc.
Example Estimating Process Models with Free Parameters at the Command Line
This example demonstrates how to estimate the parameters of a first-order process model:
Gs
()
1 sTp1
K

pesTd

This process has two inputs and the response from each input is estimated by a first-order process
model. All parameters are free to vary.
Usethefollowingcommandstoestimateamodel mfrom sample data:
% Load sample data
load co2data
% Sampling interval is 0.5 min (known) Ts = 0.5;

% Split data set into estimation data ze % and validation data zv

ze = iddata(Output_exp1,Input_exp1,Ts,...
'TimeUnit','min'); zv = iddata(Output_exp2,Input_exp2,Ts,...
'TimeUnit','min');
Estimate model with one pole, a delay, and a first order disturbance component for this model. The data
contains known offsets. Specify them using InputOffsetand OutputOffsetoptions.
opt = procestOptions;
opt.InputOffset = [170;50]; opt.OutputOffset = -45;
opt.Display = 'on';
opt.DisturbanceModel = 'arma1'; m=procest(ze,'p1d',opt)
MATLAB computes the following output:
m=
Process model with 2 inputs: y = G11(s)u1 + G12(s)u2 From input "u1" to output "y1":
Kp
G11(s) = ---------- * exp(-Td*s)
1+Tp1*s
Kp = 2.6542
Tp1 = 0.15451 Td = 2.3185
From input "u2" to output "y1": Kp
G12(s) = ---------- * exp(-Td*s) 1+Tp1*s
Kp = 9.9754 Tp1 = 2.0653 Td = 4.9195
An additive ARMA disturbance model exists for output "y1": y = G u + (C/D)e
C(s) = s + 2.677 D(s) = s + 0.6237
Parameterization: 'P1D' 'P1D'
Number of free coefficients: 8
Use "getpvec", "getcov" for parameters and their uncertainties.
Status:
Estimated using PROCEST on time domain data "ze". Fit to estimation data: 91.07% (prediction focus)
FPE: 2.431, MSE: 2.413
Use dot notation to get the value of any model parameter. For example, to get the value of dc gain
parameter Kp,type:
m.Kp
Example Estimating Process Models with Fixed Parameters at the Command Line
When you know the values of certain parameters in the model and want to estimate only the values you
do not know, you must specify the fixed parameters after creating the idprocmodel object.

Use the following commands to prepare the data and construct a process model with one pole and a
delay:
% Load sample data
load co2data
% Sampling interval is 0.5 min (known) Ts = 0.5;
% Split data set into estimation data ze % and validation data zv
ze = iddata(Output_exp1,Input_exp1,Ts,...
'TimeUnit','min'); zv = iddata(Output_exp2,Input_exp2,Ts,...
'TimeUnit','min'); mod = idproc({'p1d','p1d'},'TimeUnit','min')
The model parameters Kp, Tp1,and Tdare assigned NaNvalues, which means that the parameters have
not yet been estimated from the data.
Use the Structuremodel property to specify the initial guesses for unknown parameters, minimum/
maximum parameter bounds and fix known parameters.
Set the value of Kpfor the second transfer function to 10and specify it as a fixed parameter. Initialize the
delay values for the two transfer functions to 2 and 5 minutes, respectively. Specify them as free
estimation parameters.
mod.Structure(2).Kp.Value = 10; mod.Structure(2).Kp.Free = false;
mod.Structure(1).Td.Value = 2; mod.Structure(2).Td.Value = 5;
To estimate Tp1and Tdonly, use the following command:
mod_proc = procest(ze, mod, opt) MATLAB computes the following result:
mod_proc =
Process model with 2 inputs: y = G11(s)u1 + G12(s)u2 From input "u1" to output "y1":
Kp
G11(s) = ---------- * exp(-Td*s)
1+Tp1*s
Kp = 2.7448
Tp1 = 0.40544 Td = 1.9745
From input "u2" to output "y1": Kp
G12(s) = ---------- * exp(-Td*s) 1+Tp1*s
Kp = 10
Tp1 = 2.0734
Td = 4.92
An additive ARMA disturbance model exists for output "y1": y = G u + (C/D)e
C(s) = s + 2.702 D(s) = s + 0.6309
Parameterization: 'P1D' 'P1D'

Number of free coefficients: 7

Use "getpvec", "getcov" for parameters and their uncertainties.
Status:
Estimated using PROCEST on time domain data "ze". Fit to estimation data: 91.06% (prediction focus)
FPE: 2.435, MSE: 2.419
In this case, the value of Kpis fixed at 12,but Tp1and Tdare estimated.

Process Model Structure Specification

This topic describes how to specify the model structure in the estimation procedures How to Estimate
Process Models Using the GUI on page 3-27
andHowtoEstimateProcessModelsattheCommandLineonpage3-32.
In the System Identification Tool GUI. Specify the model structure by selecting the number of real or
complex poles, and whether to include a zero, delay, and integrator. The resulting transfer function is
displayed in the Process Models dialog box.
At the command line. Specify the model structure using an acronym that includes the following letters
and numbers:
(Required) Pfor a process model
(Required) 0, 1, 2or 3for the number of poles
(Optional) Dto include a time-delay term e sTd
(Optional) Zto include a process zero (numerator term)
(Optional) Uto indicate possible complex-valued (underdamped) poles
(Optional) Ito indicate enforced integration
Typically, you specify the model-structure acronym as a string argument in the estimation command
procest:
to estimate the following structure: procest(data,'P1D')
Gs
()
1 sTp1
K

pesTd

procest(data,'P2ZU')to estimate the following structure:

Gs KsT1+() = pz
12
++
22

sT s T
ww

procest(data,'P0ID')to estimate the following structure:

Gs
()
=
K

pesTd s
procest(data,'P3Z')to estimate the following structure:
=
Gs KsT
1+

pz()

+() +()11 1+()

For more information about estimating models, see How to Estimate Process Models at the Command
Line on page 3-32.

Estimating Multiple-Input, Multi-Output Process Models

If your model contains multiple inputs, multiple outputs, or both, you can specify whether to estimate
the same transfer function for all input-output pairs, or a different transfer function for each. The
information in this section supports the estimation procedures How to Estimate Process Models Using
the GUI on page 3-27 and How to Estimate Process Models at the Command Line on page 3-32.
In the System Identification Tool GUI. To fit a data set with multiple inputs, or multiple outputs, or
both, in the Process Models dialog box, configure the process model settings for one input-output pair at
a time. Use the input and output selection lists to switch to a different input/output pair.
If you want the same transfer function to apply to all input/output pairs, select the All same check box.
To apply a different structure to each channel, leave this check box clear, and create a different transfer
function for each input.
At the command line. Specify the model structure as a cell array of acronym strings in the estimation
command procest. For example, use this command to specify the first-order transfer function for the first
input, and a second-order model with a zero and an integrator for the second input:
m = idproc({'P1','P2ZI'}) m = procest(data,m)
To apply the same structure to all inputs, define a single structure in idproc.

Disturbance Model Structure for Process Models

This section describes how to specify a noise model in the estimation procedures How to Estimate
Process Models Using the GUI on page 3-27
andHowtoEstimateProcessModelsattheCommandLineonpage3-32.
In addition to the transfer function G, a linear system can include an additive noise term He,asfollows:
yGu He
=+
where e is white noise.

You can estimate only the dynamic model G, or estimate both the dynamic model and the disturbance
model H. For process models, H is a rational transfer function C/D,wherethe C and D polynomials for a
first- or second-order ARMA model.
In the GUI. To specify whether to include or exclude a noise model in the Process Models dialog box,
select one of the following options from the Disturbance Model list:
None The algorithm does not estimate a noise model (C=D=1). This option also sets Focus to
Simulation.
Order 1 Estimates a noise model as a continuous-time, first-order ARMA model.
Order 2 Estimates a noise model as a continuous-time, second-order ARMA model.
At the command line. Specify the disturbance model using the procestOptionsoption set. For example,
use this command to estimate a first-order transfer function and a first-order noise model:
opt = procestOptions;
opt.DisturbanceModel = 'arma1'; model = procest(data, 'P1D', opt);
For a complete list of values for the DisturbanceModelmodel property, see the procestOptionsreference
page.

Assigning Estimation Weightings

You can specify how the estimation algorithm weighs the fit at various frequencies. This information
supports the estimation procedures How to Estimate Process Models Using the GUI on page 3-27 and
How to Estimate Process ModelsattheCommandLineonpage3-32.
In the System Identification Tool GUI. Set Focus to one of the following options:
PredictionUses the inverse of the noise model H to weigh the relative importance of how closely to
fit the data in various frequency ranges. Corresponds to minimizing one-step-ahead prediction, which
typically favors the fit over a short time interval. Optimized for output prediction applications.
Simulation Uses the input spectrum to weigh the relative importance of the fit in aspecific frequency
range. Does not use the noise model to weigh the relative importance of how closely to fit the data in
various frequency ranges. Optimized for output simulation applications.
Stability Behaves the same way as the Predictionoption, but also forces the model to be stable. For
more information about model stability, see Unstable Models on page 8-94.
FilterSpecify a custom filter to open the Estimation Focus dialog box, where you can enter a filter,
as described in Simple Passband Filter on page 2-126 or Defining a Custom Filter on page 2-127.
This prefiltering applies only for estimating the dynamics from input to output. The disturbance model is
determined from the estimation data.
At the command line. Specify the focus using the procestOptionsoption set. For example, use this
command to optimize the fit for simulation and estimate a disturbance model:

opt = procestOptions('DisturbanceModel','arma2', 'Focus','simulation'); model = procest(data, 'P1D',

opt);

Specifying Initial Conditions for Iterative Estimation Algorithms

You can optionally specify how the iterative algorithm treats initial conditions for estimation of model
parameters. This information supports the estimation procedures How to Estimate Process Models
Using the GUI on page 3-27 andHowtoEstimateProcessModelsattheCommandLineonpage3-32.
In the System Identification Tool GUI. Set Initial condition to one of the following options:
Zero Sets all initial states to zero.
Estimate Treats the initial states as an unknown vector of parameters and estimates these states from
the data.
Backcast Estimates initial states using a backward filtering method (least-squares fit).
U-level est Estimates both the initial conditions and input offset levels. For multiple inputs, the input
level for each input is estimated individually. Useifyouincludedanintegratorinthetransferfunction.
Auto Automatically chooses one of the preceding options based on the estimation data. If the initial
conditions have negligible effect on the prediction errors, they are taken to be zero to optimize algorithm
performance.
At the command line. Specify the initial conditions using the InitialConditionmodel estimation option,
configured using the procestOptionscommand. For example, use this command to estimate a first-order
transfer function and set the initial states to zero:
opt = procestOptions('InitialCondition','zero'); model = procest(data, 'P1D', opt)

Identifying Input-Output Polynomial Models

In this section...
What Are Polynomial Models? on page 3-45
Data Supported by Polynomial Models on page 3-52
Preliminary Step Estimating Model Orders and Input Delays on page 3-53
How to Estimate Polynomial Models in the GUI on page 3-61 How to Estimate Polynomial Models
at the Command Line on page 3-64
Polynomial Sizes and Orders of Multi-Output Polynomial Models on page 3-68
Assigning Estimation Weightings on page 3-72
Specifying Initial States for Iterative Estimation Algorithms on page 3-73 Polynomial Model
Estimation Algorithms on page 3-73 Estimate Models Using armax on page 3-74

What Are Polynomial Models?

Polynomial Model Structure on page 3-46
Understanding the Time-Shift Operator q on page 3-47
Different Configurations of Polynomial Models on page 3-47
Continuous-Time Representation of Polynomial Models on page 3-51
Multi-Output Polynomial Models on page 3-51
Polynomial Model Structure
A polynomial model uses a generalized notion of transfer functions to express the relationship between
the input, u(t), the output y(t), and the noise e(t) using the equation:
nu

iBq()

Cq()

Aq yt() ()= Fqut nk() +

Dq

et() ()i=1 iii

The variables A, B, C, D,and F are polynomials expressed in the time-shift operator q^-1. uiis the ith
input, nu is the total number of inputs, and nkiis the ith input delay that characterizes the transport delay.
The variance of the white noise e(t) is assumed to be . For more information about the time-shift
operator, see Understanding the Time-Shift Operator q on page 3-47.
In practice, not all the polynomials are simultaneously active. Often, simpler forms, such as ARX,
ARMAX, Output-Error, and Box-Jenkins are employed. You also have the option of introducing an
integrator in the noise source so that the general model takes the form:
Aq yt() ()
nu Bq () Cq() 1
i
i1

et()Fqut nk Dq()1 q1

iii

For more information, see Different Configurations of Polynomial Models on page 3-47.
You can estimate polynomial models using time or frequency domain data.
For estimation, you must specify the model order as a set of integers that represent the number of
coefficients for each polynomial you include in your selected structurena for A, nb for B, nc for C, nd
for D,and nf for F.Youmust also specify the number of samples nk corresponding to the input
delaydead timegiven by the number of samples before the output responds to the input.
The number of coefficients in denominator polynomials is equal to the number of poles, and the number
of coefficients in the numerator polynomials is equal to the number of zeros plus 1. When the dynamics
from u(t) to y(t) contain a delay of nk samples, then the first nk coefficients of B are zero. For more
information about the family of transfer-function models, see the corresponding section in System
Identification: Theory for the User,Second Edition, by Lennart Ljung, Prentice Hall PTR, 1999.

Understanding the Time-Shift Operator q

The general polynomial equation is written in terms of the time-shift operator q1. To understand this
time-shift operator, consider the following discrete-time difference equation:
yt a yt T a yt++ =2T)
12

but T b ut() ( )
+
12

where y(t) is the output, u(t) is the input, and T is the sampling interval. q-1 is a time-shift operator that
compactly represents such difference equations using1qut ut T):
=
1yt a q yt a q yt2
()++ =

b ()1 1qut bq ut
+

or
Aq yt =))() ( ) ()
In this case, Aq()1 a q1 +2and Bq b q1 2.=+a q
1 =+b q
Note This q description is completely equivalent to the Z-transform form: q corresponds to z.
Different Configurations of Polynomial Models These model structures are subsets of the following
general polynomial equation:
nuiBq()

Cq()

Aq yt() ()= Fqut nk() +

Dq

et() ()i=1 iii

The model structures differ by how many of these polynomials are included in the structure. Thus,
different model structures provide varying levels of flexibility for modeling the dynamics and noise
characteristics.
The following table summarizes common linear polynomial model structures supported by the System
Identification Toolbox product. If you have a specific structure in mind for your application, you can
decide whether the dynamics and the noise have common or different poles. A(q) corresponds to poles
that are common for the dynamic model and the noise model. Using common poles for dynamics and
noise is useful when the disturbances enter the system at the input. Fidetermines the poles unique to the
system dynamics, and D determines the poles unique to the disturbances.
Model Equation Description Structure
ARX

nu

The noise model is1 and theA

Aq yt() () B qu t nk() +et() noise is coupled to the dynamics=

ii i

model. ARX does not let youi=1

model noise and dynamics independently. Estimate an ARX model to obtain a simple model at good
signal-to-noise
ratios.
ARIX Extends the ARX structure by including an integrator in the Ay Bu1 e noise source, e(t). This is
useful1 q1 in cases where the disturbance is
not stationary. ARMAX Extends the ARX structure by nu providing more flexibility for

Aq yt() () =() +C qet() () modeling noise using the C

ii i

parameters (a moving average ofi=1

white noise). Use ARMAX when the dominating disturbances enter at the input. Such disturbances are
called load disturbances.
ARIMAX Extends theARMAX structure Ay BuC 1 e by including an integrator in the 1
q

noise source, e(t). This is useful in cases where the disturbance is not stationary.

Box-Jenkins Provides completely independent (BJ)nu Bq() Cq() parameterization for the

yt() =Fqut nk() + Dq et() () dynamics and the noise using

rational polynomial functions.=1

Use BJ models when the noise does not enter at the input, but is primary a measurement disturbance,
This structure provides additional flexibility for modeling noise.
i iii

Output-Error Use when you want to (OE)nu Bq() parameterize dynamics, but yt()=Fqut nk et()
donotwanttoestimateanoise

model.i=1iii
Note In this case, the noise models is H = 1in the general equation and the white noise source e(t) affects
only the output.
The polynomial models can contain one or more outputs and zero or more inputs.

The System Identification Tool GUI supports direct estimation of ARX, ARMAX, OE and BJ models.
You can add a noise integrator to the ARX, ARMAX and BJ forms. However, you can use polyestto
estimate all five polynomial or any subset of polynomials in the general equation. For more information
about working with pem, see Using polyest to Estimate Polynomial Models on page 3-66.
Continuous-Time Representation of Polynomial Models In continuous time, the general frequencydomain equation is written in terms of the Laplace transform variable s, which corresponds to a
differentiation operation:
AsY s
() ()
=+
FsUs Cs
()
() ()
In the continuous-time case, the underlying time-domain model is a differential equation and the model
order integers represent the number of estimated numerator and denominator coefficients. For example,
na=3 and nb=2 correspond to the following model:
4
13
22

As s asasa
=+ + +

Bs b s b12
=+
You can only estimate continuous-time polynomial models directly using continuous-time frequencydomain data. In this case, you must set the Tsdata property to 0 to indicate that you have continuoustime frequency-domain data, and use the oecommand to estimate an Output-Error polynomial model.
Continuous-time models of other structures such as ARMAX or BJ cannot be estimated. You can obtain
those forms only by direct construction (using idpoly), conversion from other model types, or by
converting a discrete-time model into continuous-time (d2c). Note that the OE form represents a transfer
function expressed as a ratio of numerator (B) and denominator (F) polynomials. For such forms
consider using the transfer function models, represented by idtfmodels. You can estimate transfer
function models using both time and frequency domain data. In addition to the numerator and
denominator polynomials, you can also estimate transport delays. See idtfand tfestfor more information.
Multi-Output Polynomial Models
You can create multi-output polynomial models by using the idpolycommand or estimate them using ar,
arx, bj, oe, armax,and polyest.IntheGUI,you can estimate such models by choosing a multi-output data
set and setting the orders appropriately in the Polynomial Models dialog box. For more details on the
orders of multi-output models, see Polynomial Sizes and Orders of Multi-Output Polynomial Models
on page 3-68.

Data Supported by Polynomial Models

Types of Supported Data on page 3-52
Designating Data for Estimating Continuous-Time Models on page 3-52
DesignatingData for Estimating Discrete-Time Models on page 3-53
Types of Supported Data
You can estimate linear, black-box polynomial models from data with the following characteristics:
Time- or frequency-domain data (iddataor idfrddata objects).
Note For frequency-domain data, you can only estimate ARX and OE models.
To estimate polynomial models for time-series data, see Time-Series Model Identification.
Real data or complex data in any domain.
Single-output and multiple-output.
You must import your data into the MATLAB workspace, as described in Data Preparation.
Designating Data for Estimating Continuous-Time Models To geta linear, continuous-time model of
arbitrary structure for time-domain data, you can estimate a discrete-time model, and then use d2cto
transform it toa continuous-time model.
For continuous-time frequency-domain data, you can estimate directly only Output-Error (OE)
continuous-time models. Other structures include noise models, which is not supported for frequencydomain data.
Tip To denote continuous-time frequency-domain data, set the data sampling interval to 0. You can set
the sampling interval when you import data into the GUI or set the Tsproperty of the data object at the
command line.
Designating Data for Estimating Discrete-Time Models You can estimate arbitrary-order, linear statespace models for both time- or frequency-domain data.
Set the data property Tsto:
0, for frequency response data that is measured directly from an experiment.
Equal to the Tsof the original data, for frequency response data obtained by transforming time-domain
iddata(using spaand etfe).
Tip You can set the sampling interval when you import data into the GUI or set the Tsproperty of the
data object at the command line.

Preliminary Step Estimating Model Orders and Input Delays

Why Estimate Model Orders and Delays? on page 3-53
Estimating Orders and Delays in the GUI on page 3-54
Estimating Model Orders at the Command Line on page 3-57
Estimating Delays at the Command Line on page 3-59
Selecting Model Orders from the Best ARX Structure on page 3-59
Why Estimate Model Orders and Delays?
To estimate polynomial models, you must provide input delays and model orders. If you already have

insight into the physics of your system, you can specify the number of poles and zeros.
In most cases, you do not know the model orders in advance. To get initial model orders and delays for
your system, you can estimate several ARX models with a range of orders and delays and compare the
performance of these models. You choose the model orders that correspond to the best model
performance and use these orders as an initial guess for further modeling.
Because this estimation procedure uses the ARX model structure, which includes the A and B
polynomials, you only get estimates for the na, nb,and nk parameters. However, you can use these
results as initial guesses for the corresponding polynomial orders and input delays in other model
structures, such as ARMAX, OE, and BJ.
If the estimated nk is too small, the leading nb coefficients are much smaller than their standard
deviations. Conversely, if the estimated nk is too large, there is a significant correlation between the
residuals and the input for lags that correspond to the missing B terms. For information about residual
analysis plots, see Residual Analysis on page 8-24.
Estimating Orders and Delays in the GUI
The following procedure assumes that you have already imported your data into the GUI and performed
any necessary preprocessing operations. For more information, see Represent Data.
To estimate model orders and input delays in the System Identification Tool GUI:
In the System Identification Tool GUI, select Estimate > Polynomial and State Space Models to
open the Polynomials and State Space Models dialog box.
1

The ARX model is already selected by default in the Structure list.

Note For time-series models, select the AR model structure. 2Edit the Orders field to specify a range of
poles, zeros, and delays. For example, enter the following values for na, nb,and nk:
[1:10 1:10 1:10]

Tip As a shortcut for entering 1:10for each required model order, click Order Selection.

Click Estimate to open the ARX Model Structure Selection window, which displays the model
performance for each combination of model parameters. The following figure shows an example plot.
3

Select a rectangle that represents the optimum parameter combination and click Insert to estimates a
model with these parameters. For information about using this plot, see Selecting Model Orders from
the Best ARX Structure on page 3-59.
4

This action adds a new model to the Model Board in the System Identification Tool GUI. The default
name of the parametric model contains the model type and the number of poles, zeros, and delays. For
example, arx692is an ARX model with na=6, nb=9, and a delay of two samples.
5Click

Close to close the ARX Model Structure Selection window.

Note You cannot estimate model orders when using multi-output data. for estimating other model
structures, as described in How to Estimate Polynomial Models in the GUI on page 3-61.
Estimating Model Orders at the Command Line
You can estimate model orders using the struc, arxstruc,and selstruc commands in combination.
If you are working with a multiple-output system, you must use the struc, arxstruc,and
selstruccommands one output at a time. You must subreference the correct output channel in your
estimation and validation data sets.
For each estimation, you use two independent data setsan estimation data set and a validation data set.
These independent data set can be from different experiments, or data subsets from a single experiment.
For more information about subreferencing data, see Select Data Channels, I/O Data and Experiments
in iddata Objects on page 2-63 and Select I/O Channels and Data in idfrd Objects on page 2-79.
For an example of estimating model orders for a multiple-input system, see Estimating Delays in the
Multiple-Input System in System Identification Toolbox Getting Started Guide.

struc. The struccommand creates a matrix of possible model-order combinations for a specified range of
na, nb,and nkvalues.
For example, the following command defines the range of model orders and delays na=2:5, nb=1:5,and
nk=1:5:
NN = struc(2:5,1:5,1:5))
arxstruc. The arxstruccommand takes the output from struc,estimates an ARX model for each model
order, and compares the model output to the measured output. arxstrucreturns the loss for each model,
which is the normalized sum of squared prediction errors.
For example, the following command uses the range of specified orders NN to compute the loss function
for single-input/single-output estimation data data_eand validation data data_v:
V = arxstruc(data_e,data_v,NN)
Each row in NNcorresponds to one set of orders:
[na nb nk]
selstruc. The selstruccommand takes the output from arxstrucand opens the ARX Model Structure
Selection window to guide your choice of the model order with the best performance.
For example, to open the ARX Model Structure Selection window and interactively choose the optimum
parameter combination, use the following command:
selstruc(V)
For more information about working with the ARX Model Structure Selection window, see Selecting
Model Orders from the Best ARX Structure on page 3-59.
To find the structure that minimizes Akaikes Information Criterion, use the following command:
nn = selstruc(V,'AIC')
where nncontains the corresponding na, nb,and nkorders.
Similarly, to find the structure that minimizes the Rissanens Minimum Description Length (MDL), use
the following command:
nn = selstruc(V,'MDL')
To select the structure with the smallest loss function, use the following command:
nn = selstruc(V,0)
for estimating other model structures, as described in Using polyest to Estimate Polynomial Models
on page 3-66.
Estimating Delays at the Command Line
The delayestcommand estimates the time delay in a dynamic system by estimating a low-order, discretetime ARX model and treating the delay as an unknown parameter.
By default, delayestassumes that na=nb=2and that there is a good signal-to-noise ratio, and uses this
information to estimate nk.
To estimate the delay for a data set data, type the following at the prompt:
delayest(data)

If your data has a single input, MATLAB computes a scalar value for the input delayequal to the
number of data samples. If your data has multiple inputs, MATLAB returns a vector, where each value is
the delay for the corresponding input signal.
To compute the actual delay time, you must multiply the input delay by the sampling interval of the data.
You can also use the ARX Model Structure Selection window to estimate input delays and model order
together, as described in Estimating Model Orders at the Command Line on page 3-57.
SelectingModel Ordersfrom theBestARX Structure You generate the ARX Model Structure Selection
window for your data to select the best-fit model.
For a procedure on generating this plot in the System Identification Tool GUI, see Estimating Orders
and Delays in the GUI on page 3-54. To open this plot at the command line, see Estimating Model
Orders at the Command Line on page 3-57.
The following figure shows a sample plot in the ARX Model Structure Selection window.
You use this plot to select the best-fit model.

The horizontal axis is the total number of parameters na+ nb.

The vertical axis, called Unexplained output variance (in %),is the portion of the output not
explained by the modelthe ARX model prediction error for the number of parameters shown on the
horizontal axis.
The prediction error is the sum of the squares of the differences between the validation data output and
the model one-step-ahead predicted output.
nkis the delay.

Three rectangles are highlighted on the plot in green, blue, and red. Each color indicates a type of bestfit criterion, as follows:
RedBestfitminimizesthesumofthesquaresofthedifferencebetween the validation data output and the
model output. This rectangle indicates the overall best fit.
Green Best fit minimizes Rissanen MDL criterion.
Blue Best fit minimizes Akaike AIC criterion.
In the ARX Model Structure Selection window, click any bar to view the orders that give the best fit.
The area on the right is dynamically updated to show the orders and delays that give the best fit.
For more information about the AIC criterion, see Akaikes Criteria for Model Validation on page
8-86.

How to Estimate Polynomial Models in the GUI

Prerequisites
Before you can perform this task, you must have:
Imported data into the System Identification Tool GUI. See Importing Time-Domain Data into the
GUI on page 2-18. For supported data formats, see Data Supported by Polynomial Models on page
3-52.
Performed any required data preprocessing operations. To improve the accuracy of your model, you
should detrend your data. Removing offsets and trends is especially important for Output-Error (OE)
models and has less impact on the accuracy of models that include a flexible noise model structure, such
as ARMAX and Box-Jenkins. See Ways to Prepare Data for System Identification on page 2-6.
Select a model structure, model orders, and delays. For a list of available structures, see What Are
Polynomial Models? on page 3-45 For more information about how to estimate model orders and
delays, see Estimating Orders and Delays in the GUI on page 3-54. For multiple-output models, you
must specify order matrices in the MATLAB workspace, as described in Polynomial Sizes and Orders
of Multi-Output Polynomial Models on page 3-68.
In the System Identification Tool GUI, select Estimate > Polynomial to open the Polynomial and
State Space Models dialog box.
2In the Structure list, select the polynomial model structure you want to estimate from the following
options:
1

ARX:[na nb nk]
ARMAX:[na nb nc nk]
OE:[nb nf nk]
BJ:[nb nc nd nf nk]
This action updates the options in the Polynomial and State Space Models dialog box to correspond with
this model structure. For information about each model structure, see What Are Polynomial Models?
on page 3-45.

Note For time-series data, only AR and ARMA models are available. For more information about
estimating time-series models, see Time-Series Model Identification.
3In

the Orders field, specify the model orders and delays, as follows:

For single-output polynomial models. Enter the model orders and delays according to the sequence
displayed in the Structure field. For multiple-input models, specify nband nkas row vectors with as
many elements as there are inputs. If you are estimating BJ and OE models, you must also specify nfas a
vector.
For example, for a three-input system, nbcan be [1 2 4], where each element corresponds to an input.
For multiple-output models. Enter the model orders, as described in Polynomial Sizes and Orders of
Multi-Output Polynomial Models on page 3-68.
Tip To enter model orders and delays using the Order Editor dialog box, click Order Editor.
(ARX models only) Select the estimation Method as ARX or IV (instrumental variable method). For
information about the algorithms, see Polynomial Model Estimation Algorithms on page 3-73.
4

5(ARX,

ARMAX, and BJ models only) Check the Add noise integration check box to add an integrator
to the noise source, e.
Specify the delay using the Input delay edit box. The value must be a vector of length equal to the
number of input channels in the data. For discrete-time estimations (any estimation using data with
nonzero sample-time), the delay must be expressed in the number of lags. These delays are separate
from the in-model delays specified by the nkorder in the Orders edit box.
6

7In

the Name field, edit the name of the model or keep the default.

In the Focus list, select how to weigh the relative importance of the fit at different frequencies. For
more information about each option, see Assigning Estimation Weightings on page 3-72.
8

In the Initial state list, specify how you want the algorithm to treat initial conditions. For more
information about the available options, see Specifying Initial Conditions for Iterative Estimation
Algorithms on page 3-43.
9

Tip If you get an inaccurate fit, try setting a specific method for handling initial states rather than
choosing it automatically.
In the Covariance list, select Estimateif you want the algorithm to compute parameter uncertainties.
Effects of such uncertainties are displayed on plots as model confidence regions.
10

To omit estimating uncertainty, select None. Skipping uncertainty computation for large, multiple-output
models might reduce computation time.

(ARMAX, OE, and BJ models only) To view the estimation progress in the MATLAB Command
Window, select the Display progress check box. This launches a progress viewer window in which
estimation progress is reported.
11

12 Click Estimate to add this model to the Model Board in the System Identification Tool GUI.
13(Prediction-error method only) To stop the search and save the results after the current iteration

has
been completed, click Stop Iterations.To continue iterations from the current model, click the Continue
iter button to assign current parameter values as initial guesses for the next search.
Next Steps
Validate the model by selecting the appropriate check box in the Model Views area of the System
Identification Tool GUI. For more information about validating models, see Validating Models After
Estimation on page 8-3.
Export the model to the MATLAB workspace for further analysis by dragging it to the To Workspace
rectangle in the System Identification Tool GUI.
Tip For ARX and OE models, you can use the exported model for initializing a nonlinear estimation at
the command line. This initialization may improve the fit of the model. See Using Linear Model for
Nonlinear ARX Estimation on page 4-28, and Using Linear Model for Hammerstein-Wiener
Estimation on page 4-63.

How to Estimate Polynomial Models at the Command Line

Usingarxandiv4toEstimateARXModelsonpage3-65
Using polyest to Estimate Polynomial Models on page 3-66
Prerequisites
Before you can perform this task, you must have
Input-output data as an iddataobject or frequency response data as an frdor idfrdobject. See
Representing Time- and Frequency-Domain Data Using iddata Objects on page 2-55. For supported
data formats, see Data Supported by Polynomial Models on page 3-52.
Performed any required data preprocessing operations. To improve the accuracy of results when using
time domain data, you can detrend the data or specify the input/output offset levels as estimation
options. Removing offsets and trends is especially important for Output-Error (OE) models and has less
impact on the accuracy of models that include a flexible noise model structure, such as ARMAX and
Box-Jenkins. See Ways to Prepare Data for System Identification on page 2-6.
Select a model structure, model orders, and delays. For a list of available structures, see What Are
Polynomial Models? on page 3-45 For more information about how to estimate model orders and
delays, see Estimating Model Orders at the Command Line on page 3-57 and Estimating Delays at
the Command Line on page 3-59. For multiple-output models, you must specify order matrices in the
MATLAB workspace, as described in Polynomial Sizes and Orders of Multi-Output Polynomial
Models on page 3-68.

Using arx and iv4 to Estimate ARX Models

You can estimate single-output and multiple-output ARX models using the arxand iv4commands. For
information about the algorithms, see Polynomial Model Estimation Algorithms on page 3-73.
You can use the following general syntax to both configure and estimate ARX models:
% Using ARX method
m = arx(data,[na nb nk],opt) % Using IV method
m = iv4(data,[na nb nk],opt)
datais the estimation data and [na nb nk]specifies the model orders, as discussed in What Are
Polynomial Models? on page 3-45.
The third input argument optcontains the options for configuring the estimation of the ARX model, such
as handling of initial conditions and input offsets. You can create and configure the option set optusing
the arxOptions and iv4Optionscommands. The three input arguments can also be followed by name and
value pairs to specify optional model structure attributes such as InputDelay, ioDelay,and
IntegrateNoise.
To get discrete-time models, use the time-domain data (iddataobject).
Note Continuous-time polynomials of ARX structure are not supported.
For more information about validating you model, see Validating Models After Estimation on page
8-3.
You can use pemor polyestto refine parameter estimates of an existing polynomial model, as described
in Refining Linear Parametric Models on page 3-130.
For detailed information about these commands, see the corresponding reference page.
Tip You can use the estimated ARX model for initializing a nonlinear estimation at the command line,
which improves the fit of the model. See Using Linear Model for Nonlinear ARX Estimation on page
4-28.
Using polyest to Estimate Polynomial Models
You can estimate any polynomial model using the iterative prediction-error estimation method polyest.
For Gaussian disturbances of unknown variance, this method gives the maximum likelihood estimate.
The resulting models are stored as idpolymodel objects.
Use the following general syntax to both configure and estimate polynomial models:
m = polyest(data, [na nb nc nd nf nk], opt,Name,Value)
where datais the estimation data. na, nb, nc, nd, nfare integers that specify the model orders, and
nkspecifies the input delays for each input.For more information about model orders, see What Are
Polynomial Models? on page 3-45.
Tip You do not need to construct the model object using idpolybefore estimation.
If you want to estimate the coefficients of all five polynomials, A, B, C, D,and F, you must specify an
integer order for each polynomial. However, if you want to specify an ARMAX model for example,

which includes only the A, B, and C polynomials, you must set ndand nfto zero matrices of the
appropriate size. For some simpler configurations, there are dedicated estimation commands such as arx,
armax, bj,and oe, which deliver the required model by using just the required orders. For example,
oe(data, [nb nf nk],opt) estimates an output-error structure polynomial model.
Note To get faster estimation of ARX models, use arxor iv4instead of polyest.
In addition to the polynomial models listed in What Are Polynomial Models? on page 3-45, you can
use polyestto model the ARARX structurecalled the generalized least-squares modelby setting
nc=nf=0. You can also model the ARARMAX structurecalled the extended matrix modelby setting
nf=0.
The third input argument, opt, contains the options for configuring the estimation of the polynomial
model, such as handling of initial conditions, input offsets and search algorithm. You can create and
configure the option set optusing the polyestOptionscommand. The three input arguments can also be
followed by name and value pairs to specify optional model structure attributes such as InputDelay,
ioDelay,and IntegrateNoise.
For ARMAX, Box-Jenkins, and Output-Error modelswhich can only be estimated using the iterative
prediction-error methoduse the armax, bj, and oeestimation commands, respectively. These
commands are versions of polyestwith simplified syntax for these specific model structures, as follows:
m = armax(Data,[na nb nc nk]) m = oe(Data,[nb nf nk])
m = bj(Data,[nb nc nd nf nk])
Similar to polyest, you can specify as input arguments the option set configured using commands
armaxOptions, oeOptions,and bjOptionsfor the estimators armax, oe,and bjrespectively. You can also
use name and value pairs to configure additional model structure attributes.
Tip If your data is sampled fast, it might help to apply a lowpass filter to the data before estimating the
model, or specify a frequency range for the Focus property during estimation. For example, to model
only data in the frequency range 0-10 rad/s, use the Focusproperty, as follows:
opt = oeOptions('Focus',[0 10]) m = oe(Data, [nb nf nk], opt)
For more information about validating your model, see Validating Models After Estimation on page
8-3.
You can use pemor polyestto refine parameter estimates of an existing polynomial model (of any
configuration), as described in Refining Linear Parametric Models on page 3-130.
For more information, see polyest, pemand idpoly.

Polynomial Sizes and Orders of Multi-Output Polynomial Models

For a model with Ny (Ny > 1)outputsand Nu inputs, the polynomials A, B, C, D,and F are specified as
cell arrays of row vectors. Each entry in the cell array contains the coefficients of a particular
polynomial that relates input, output, and noise values. Orders are matrices of integers used as input
arguments to the estimation commands.

A Ny-byNyarray of row vectors A{i,j}contains coefficients of relation between output yiand output yj
B Ny-byNuarray of row vectors B{i,j}contain coefficients of relations between output yiand input uj
na :
Ny-byNy
matrix
such
that each
entry
contains
the
degree
of the
corresponding A
polynomial.
nk:
Ny-byNumatrix
such
that each
entry
contains
the
number
of
leading
fixed
zeros
of the
corresponding B
polynomial
(input
delay).
nb:
Ny-byNumatrix
such
nb(i,j)
C,D Ny-by-1 array of row vectors C{i}and D{i}contain coefficients of relations between output yiand
noise ei
F Ny-byNuarray of row vectors F{i,j}contains coefficients of relations between output yiand input uj
=
length(B{i,j})nk(i,j).

nc and
ndare
Ny-by-1
matrices
such
that each
entry
contains
the
degree
of the
corresponding
C and D
polynomial,
respectively.
nf:
Ny-byNumatrix
such
that each
entry
contains
the
degree
of the
corresponding
F
polynomial.
For more information, see idpoly.
For example, consider the ARMAX set of equations for a 2 output, 1 input model:
y (t) + 0.5 y (t-1) + 0.9 y (t-1) + 0.1 y (t-2) = u(t) + 5 u(t-1) + 2 u(t-2) + e (t) + 0.01 e 11 2 2 11
y (t) + 0.05 y (t-1) + 0.3 y (t-2)
2222 2 2= 10 u(t-2) + e (t) + 0.1 e (t-1) + 0.02 e (t-2)2
y1andy2represent the two outputs and u represents the input variable. e1 and e2represent the white noise
disturbances on the outputs, y1and y2, respectively. To represent these equations as an ARMAX form
polynomial using idpoly,configurethe A, B,and Cpolynomialsasfollows:
A = cell(2,2);
A{1,1} = [1 0.5]; A{1,2} = [0 0.9 0.1]; A{2,1} = [0]; A{2,2} = [1 0.05 0.3];
B = cell(2,1);
B{1,1} = [1 5 2]; B{2,1} = [0 0 10];
C = cell(2,1);
C{1} = [1 0.01]; C{2} = [1 0.1 0.02];

model = idpoly(A,B,C)
modelis a discrete-time ARMAX model with unspecified sample-time.
When estimating such models, you need to specify the orders of these polynomials as input arguments.
In the System Identification Tool GUI. You can enter the matrices directly in the Orders field.
At the command line. Define variables that store the model order matrices and specify these variables
in the model-estimation command.
Tip To simplify entering large matrices orders in the System Identification Tool GUI, define the variable
NN=[NA NB NK]at the command line. You can specify this variable in the Orders field.

Assigning Estimation Weightings

You can specify how the estimation algorithm weighs the fit at various frequencies. This information
supports the estimation procedures How to Estimate Polynomial Models in the GUI on page 3-61 and
Using polyest to Estimate Polynomial Models on page 3-66.
In the System Identification Tool GUI. Set Focus to one of the following options:
PredictionUses the inverse of the noise model H to weigh the relative importance of how closely to
fit the data in various frequency ranges. Corresponds to minimizing one-step-ahead prediction, which
typically favors the fit over a short time interval. Optimized for output prediction applications.
Simulation Uses the input spectrum to weigh the relative importance of the fit in aspecific frequency
range. Does not use the noise model to weigh the relative importance of how closely to fit the data in
various frequency ranges. Optimized for output simulation applications.
Stability Estimates the best stable model. For more information about model stability, see Unstable
Models on page 8-94.
FilterSpecify a custom filter to open the Estimation Focus dialog box, where you can enter a filter,
as described in Simple Passband Filter on page 2-126 or Defining a Custom Filter on page 2-127.
This prefiltering applies only for estimating the dynamics from input to output. The disturbance model is
determined from the unfiltered estimation data.
At the command line. Specify the focus as an estimation option (created using polyestOptions,
oeOptionsetc.) usingthesameoptionsasintheGUI. For example, use this command to estimate an ARX
model and emphasize the frequency content related to the input spectrum only:
opt = arxOptions('Focus', 'simulation'); m = arx(data,[2 2 3],opt)
This Focussetting might produce more accurate simulation results, provided theorders picked are
optimal for the given data..

Specifying Initial States for Iterative Estimation Algorithms

When you use the pemor polyestto estimate ARMAX, Box-Jenkins (BJ), Output-Error (OE), you must
specify how the algorithm treats initial conditions.

This information supports the estimation procedures How to Estimate Polynomial Models in the GUI
on page 3-61 and Using polyest to Estimate Polynomial Models on page 3-66.
In the System Identification Tool GUI. For ARMAX, OE, and BJ models, set Initial state to one of
the following options:
Auto Automatically chooses Zero, Estimate,or Backcastbased on the estimation data. If initial states
have negligible effect on the prediction errors, the initial states are set to zero to optimize algorithm
performance.
Zero Sets all initial states to zero.
Estimate Treats the initial states as an unknown vector of parameters and estimates these states from
the data.
Backcast Estimates initial states using a smoothing filter.
At the command line. Specify the initial conditions as an estimation option. Use polyestOptionsto
configure options for the polyestcommand, armaxOptionsfor the armaxcommand etc. Set the
InitialConditionoption to the desired value in the option set. For example, use this command to estimate
an ARMAX model and set the initial states to zero:
opt = armaxOptions('InitialCondition','zero') m = armax(data,[2 2 2 3],opt)
For a complete list of values for the InitialConditionestimation option, see the armaxOptionsreference
page.

Polynomial Model Estimation Algorithms

For linear ARX and AR models, you can choose between the ARX and IV algorithms. ARX implements
the least-squares estimation method that uses QR-factorization for overdetermined linear equations. IV is
the instrument variable method. For more information about IV, see the section on variance-optimal
instruments in System Identification: Theory for the User, Second Edition, by Lennart Ljung, Prentice
Hall PTR, 1999.
The ARX and IV algorithms treat noise differently. ARX assumes white noise. However, the
instrumental variable algorithm, IV, is not sensitive to noise color. Thus, use IV when the noise in your
system is not completely white and it is incorrect to assume white noise. If the models you obtained
using ARX are inaccurate, try using IV.
Note AR models apply to time-series data, which has no input. For more information, see Time-Series
Model Identification. For more information about working with AR and ARX models, see Identifying
Input-Output Polynomial Models on page 3-45.

Estimate Models Using armax

Thisexampleshowshowtoestimatealinear,polynomialmodelwithan ARMAX structure for a three-input
and single-output (MISO) system using the iterative estimation method armax. For a summary of all
available estimation commands in the toolbox, see Model Estimation Commands on page 1-40.

1Load

a sample data set z8with three inputs and one output, measured at 1-second intervals and
containing 500 data samples:
load iddata8
2Use armaxto both construct the idpolymodel object, and estimate the parameters:
nu

Aq yt() () B qu t nk() + C qet() ()=

ii i
i=1

Typically you try different model orders and compare results, ultimately choosing the simplest model
that best describes the system dynamics. The following command specifies the estimation data set,
z8,andtheordersof the A, B,and C polynomials as na, nb,and nc, respectively. nkof [0 0 0] specifies that
there is no input delay for all three input channels.
opt = armaxOptions;
opt.Focus = 'simulation';
opt.SearchOption.MaxIter = 50;
opt.SearchOption.Tolerance = 1e-5; na = 4;
nb =[32 3];
nc = 4;
nk =[00 0];
m_armax = armax(z8, [na nb nc nk], opt)
Focus , Tolerance,and MaxIterare estimation options that configure the estimation objective function and
the attributes of the search algorithm. The Focusoption specifies whether the model is optimized for
simulation or prediction applications. The Toleranceand MaxItersearch options specify when to stop
estimation. For more information about these properties, see the armaxOptionsreference page.
armax is a version of polyestwith simplified syntax for the ARMAX model structure. The armaxmethod
both constructs the idpolymodel object and estimates its parameters.
3To view information about the resulting model object, type the following at
the prompt:
m_armax
MATLAB returns the following information about this model object:
m_armax =
Discrete-time ARMAX model: A(z)y(t) = B(z)u(t) + C(z)e(t) A(z) = 1 - 1.284 z^-1 + 0.3048 z^-2 + 0.2648 z^-3 - 0.05708
z^-4
B1(z) = -0.07547 + 1.087 z^-1 + 0.7166 z^-2
B2(z) = 1.019 + 0.1142 z^-1
B3(z) = -0.06739 + 0.06828 z^-1 + 0.5509 z^-2
C(z) = 1 - 0.06096 z^-1 - 0.1296 z^-2 + 0.02489 z^-3 - 0.04699 z^-4
Sample time: 1 seconds
Parameterization:
Polynomial orders: na=4 nb=[3 2 3] nc=4 nk=[0 0 0]

Number of free coefficients: 16

Use "polydata", "getpvec", "getcov" for parameters and their uncertainties.
Status:
Estimated using ARMAX on time domain data "z8". Fit to estimation data: 80.86% (simulation focus) FPE: 1.056, MSE:
0.9888

m_armaxis an idpolymodel object. The coefficients represent estimated parameters of this polynomial
model.
Tip You can use present(m_armax)to show additional information about the model, including parameter
uncertainties.
4To view all property values for this model, type the following command:
get(m_armax)
MATLAB returns the following information:
a: [1 -1.2836 0.3048 0.2648 -0.0571] b: {1x3 cell}
c: [1 -0.0610 -0.1296 0.0249 -0.0470] d: 1
f: {[1] [1] [1]}
Variable: 'z^-1'
ioDelay: [0 0 0]
IntegrateNoise: 0
Structure: [1x1 pmodel.polynomial]
NoiseVariance: 0.9899
Report: [1x1 idresults.polyest]
InputDelay: [3x1 double]
OutputDelay: 0
Ts: 1
TimeUnit: 'seconds'
InputName: {3x1 cell}
InputUnit: {3x1 cell}
InputGroup: [1x1 struct]
OutputName: {'y1'}
OutputUnit: {''}
OutputGroup: [1x1 struct]
Name: ''
Notes: {}
UserData: []
The Reportmodel property contains detailed information on the estimation results. To view the
properties and values inside Report,usedotnotation. For example:
5

m_armax.Report
This action displays the contents of estimation report such as model quality measures (Fit), search
termination criterion (Termination), and a record of estimation data
(DataUsed)andoptions(OptionsUsed).

Status: 'Estimated using POLYEST with Focus = "simulation"' Method: 'ARMAX' InitialCondition: 'zero' Fit: [1x1 struct]
Parameters: [1x1 struct] OptionsUsed: [1x1 idoptions.polyest]
RandState: [1x1 struct] DataUsed: [1x1 struct] Termination: [1x1 struct]

Identifying State-Space Models

In this section...
What Are State-Space Models? on page 3-79
Data Supported by State-Space Models on page 3-83
Supported State-Space Parameterizations on page 3-83 Estimate State-Space Model With Order
Selection on page 3-83 How to Estimate State-Space Models in the GUI on page 3-89 How to
Estimate State-Space Models at the Command Line on page 3-92 How to Estimate FreeParameterization State-Space Models on page 3-98
How to Estimate State-Space Models with Canonical Parameterization on page 3-99
How to Estimate State-Space Models with Structured Parameterization on page 3-100
How to Estimate the State-Space Equivalent of ARMAX and OE Models on page 3-108
Assigning Estimation Weightings on page 3-110
Specifying Initial States for Iterative Estimation Algorithms on page 3-111 State-Space Model
Estimation Algorithms on page 3-112

What Are State-Space Models?

Definition of State-Space Models on page 3-80
Continuous-Time Representation on page 3-80
Discrete-Time Representation on page 3-81
Relationship Between Continuous-Time and Discrete-Time State Matrices on page 3-81
State-Space Representation of Transfer Functions on page 3-82
Definition of State-Space Models
State-space models are models that use state variables to describe a system by a set of first-order
differential or difference equations, rather than by one or more nth-order differential or difference
equations. State variables x(t) can be reconstructed from the measured input-output data, but are not
themselves measured during an experiment.
The state-space model structure is a good choice for quick estimation because it requires only one user
input, the model order, n.
The model order is an integer equal to the dimension of x(t) and relates to, but is not necessarily equal
to, the number of delayed inputs and outputs used in the corresponding linear difference equation.
Continuous-Time Representation
In continuous-time, the state-space description has the following form:
xt () =+ +Fxt Gut Kwt() yt() =+ +Hx t Dut wt() xx=00

It is often easier to define a parameterized state-space model in continuous time because physical laws
are most often described in terms of differential equations. In this case, the matrices F, G, H,and D
contain elements with physical significancefor example, material constants. x0 specifies the initial
states. You can estimate continuous-time state-space model using both timeand frequency-domain data.
Note K = 0 gives the state-space representation of an Output-Error model. For more information about
Output-Error models, see What Are Polynomial Models? on page 3-45.
Discrete-Time Representation
Discrete-time state-space models provide the same type of linear difference relationship between the
inputs and the outputs as the linear ARX model, but are rearranged such that there is only one delay in
the expressions. The discrete-time state-space model structure is often written in the innovations form
that describes noise:
x kT( ) () () ()
+= + +
ykT() () () ()
=+ +
x() = x0
where T is the sampling interval, u(kT) is the input at time instant kT,and y(kT) is the output at time
instant kT.
Note K=0 gives the state-space representation of an Output-Error model. For more information about
Output-Error models, see What Are Polynomial Models? on page 3-45.
Relationship Between Continuous-Time and Discrete-Time State Matrices
The relationships between the discrete state-space matrices A, B, C, D,and K and the continuous-time
state-space matrices F, G, H, D,and K are given for piece-wise-constant input, as follows:
AeFT
=
T

=BeGdF
0

CH
=
These relationships assume that the input is piece-wise-constant over time intervals kT t< +k () .
The exact relationship between K and K is complicated. However, for short sampling intervals T, the
following approximation works well: T

= KeKdF
0

State-Space Representation of Transfer Functions For linear models, the general symbolic model

description is given by:

yGu He
=+
G is a transfer function that takes the input u to the output y. H is a transfer function that describes the
properties of the additive output noise model.
The discrete-time state-space representation is given by the following equation:
x kT( ) () () ()
+= + +
ykT() () () ()
=+ +
x() = x0
where T is the sampling interval, u(kT) is the input at time instant kT,and y(kT) is the output at time
instant kT.
The relationships between the transfer functions and the discrete-time state-space matrices are given by
the following equations:
Gq CqInx 1 A B D= + Hq CqI
nx

= +
1 A K I

ny

where Inxis the nx-bynx identity matrix, Inyis the nx-bynx identity matrix, and ny is the dimension of y
and e. The state-space representation in the continuous-time case is similar.

Data Supported by State-Space Models

You can estimatelinear state-space models from data with the following characteristics:
Real data or complex data in any domain
Single-output and multiple-output
Time- or frequency-domain data
To estimate state-space models for time-series data, see Time-Series Model Identification.
You must first import your data into the MATLAB workspace, as described in Data Preparation.

Supported State-Space Parameterizations

The SystemIdentification Toolbox product supports the following parameterizations that indicate which
parameters are estimated and which remain fixed at specific values:
Free parameterization results in the estimation of all system matrix elements A, B, C, D,and K.
Canonical forms of A, B, C, D,and K matrices.
Canonical parameterization represents a state-space system in a reduced-parameter form where many
entries of the A, B and C matrices are fixed tozerosandones. Thefreeparametersappearinonlyafewofthe
rows and columns in the system matrices.

 Structured parameterization lets you specify the fixed values of specific parameters and exclude
these parameters from estimation. You choose which entries of the system matrices to estimate and
which to treat as fixed.
Completely arbitrary mapping of user-chosen parameters to state-space matrices. For more
information, see Estimating Linear Grey-Box Models on page 5-6.

Estimate State-Space Model With Order Selection

Estimate Model With Selected Order in the GUI on page 3-84
Estimate Model With Selected Order at the Command Line on page 3-87
Using the Model Order Selection Window on page 3-88
To estimate a state-space model, you must provide a value of its order, which represents the number of
states. When you do not know the order, you can search and select an order using the following
procedures.
Estimate Model With Selected Order in the GUI
You must have already imported your data into the GUI, as described in Importing Data into the GUI
on page 2-17.
To estimate model orders for a specific input delay:
1In the System Identification Tool GUI, select Estimate > State Space Models to open the Polynomial
and State Space Models dialog box.
2In the Structure list, select State Space: n.
3Edit the Orders field to specify a range of orders for a specific delay. For example, enter the following
values for n:
1:10

Tip As a shortcut for entering 1:10,click Order Selection.

4Verify

that the Method is set to N4SID.

Click Estimate to open the Model Order Selection window, which displays the relative measure of
how much each state contributes to the input-output behavior of the model (log of singular values of the
covariance matrix). The following figure shows an example plot.
5

Select the rectangle that represents the cutoff for the states on the left that provide a significant
contribution to the input-output behavior, and click Insert to estimate a model with this order. Red
indicates the recommended choice. In the previous figure, states 1 and 2 provide the most significant
contribution. The contributions to the right of state 2 drop significantly. For information about using the
Model Order Selection window, see Using the Model Order Selection Window on page 3-88.
6

This action adds a new model to the Model Board in the System Identification Tool GUI. The default
name of the parametric model combines the string n4sand the selected model order.
7Click

Close to close the Model Order Selection window.

You can use this model as an initial guess for estimating other state-space models, as described in How
to Estimate State-Space Models in the GUI on page 3-89.
Note You can specify additional attributes of the model structure when searching for best orders, such as
input delays, the presence of feedthrough, the canonical form and the model sample time.
Estimate Model With Selected Order at the Command Line You can estimate the state-space model
with a selected order using the n4sid and ssestcommands.
Use following syntax to specify the range of model orders to try for a specific input delay.
m = n4sid(data,n1:n2); where datais the estimation data set, n1and n2specify the range of orders, and
nkspecifies the input delay.
This command opens the Model Order Selection window. For information about using this plot, see
Using the Model Order Selection Window on page 3-88.
Alternatively, you can use the ssestcommand to open the Model Order Selection window, as follows:
m = ssest(data, nn)

where nn = [n1,n2,...,nN]specifies the vector or range of orders you want to try. n4sidestimates a model
whose sample time matches that of data, hence a discrete-time model for time domain data.
ssestestimates a continuous-time model by default. You can change the default setting by including the
Tsname and value pair input arguments in the estimation command. For example, to estimate a discretetime model of optimal order using ssest,assuming Data.Ts>0,type:
model = ssest(data,nn,'Ts',data.Ts); To omit opening the Model Order Selection window and
automatically select the best order, use the following syntax:
m = n4sid(data,'best')
Using the Model Order Selection Window
You can generate the Model Order Selection window for your data to select the number of states that
provide the highest relative contribution to the input-output behavior of the model (log of singular
values of the covariance matrix).
For a procedure on generating this plot in the System Identification Tool GUI, see Estimate Model
With Selected Order in the GUI on page 3-84. To open this plot at the command line, see Estimate
Model With Selected Order at the Command Line on page 3-87.
The following figure shows a sample Model Order Selection window.

The horizontal axis corresponds to the model order n. The vertical axis, called Log of Singular values,
shows the singular values of a covariance matrix constructed from the observed data.
You use this plot to decide which states provide a significant relative contribution to the input-output
behavior, and which states provide the smallest contribution. Based on this plot, select the rectangle that
represents the cutoff for the states on the left that provide a significant contribution to the input-output
behavior. The recommended choice is shown in red.

For example, in the previous figure, states 1 and 2 provide the most significant contribution. However,
the contributions of the states to the right of state 2 drop significantly. This sharp decrease in the log of
the singular values after n=2 indicates that using two states is sufficient to get an accurate model.

How to Estimate State-Space Models in the GUI

Prerequisites on page 3-89
Estimating State-Space Models in the GUI on page 3-89
Next Steps on page 3-91
Prerequisites
Before you can perform this task, you must have
Imported data into the System Identification Tool GUI. See Importing Time-Domain Data into the
GUI on page 2-18. For supported data formats, see Data Supported by State-Space Models on page
3-83.
Performed any required data preprocessing operations. To improve the accuracy of your model, you
should detrend your data. See Ways to Prepare Data for System Identification on page 2-6.
Select a model order. For more information about how to estimate models with selected orders, see
Estimate State-Space Model With Order Selection on page 3-83.
Estimating State-Space Models in the GUI
To estimate a state-space model with free parameterization in the System Identification Tool GUI:
In the System Identification Tool GUI, select Estimate > State Space Models to open the Polynomial
and State Space Models dialog box.
2In the Structure list, select State Space: n.
1

This action updates the options in the Polynomial and State Space Models dialog box to correspond with
this model structure. For information about each model structure, see What Are State-Space Models?
on page 3-79.
3In

the Orders field, specify the model order, for example, as 4.

Tip To enter model order using the Order Editor dialog box, click Order Editor. Using this editor, you
can also configure options that influence the estimation such as N4Weightand N4Horizon.
Select the estimation Method as N4SID or PEM. For more information about these methods, StateSpace Model Estimation Algorithms on page 3-112.
4

Choosetoestimateeitheracontinuous-timeordiscrete-timemodel. Youcannotestimateadiscretetimemodeliftheworkingdatais continuous-time frequency domain data.

5

Specify whether the model has feedthrough or not. In the Feedthrough edit box, enter a logical vector
of length equal to number of inputs. For example, for a model with 2 inputs, you may enter [true
6

false],which means that the first input has direct feedthrough to the models outputs while the second
input does not.
7Select

the type of state-space parameterization using the Form list.

Specify any input delays using the Input delay edit box. Specify the value as a vector of length equal
to number of inputs. For continuous-time models, the delay must be specified in seconds. For discretetime models, specify the value in terms on number of lags (multiples of sample data time).
8

9In

the Name field, edit the name of the model or keep the default. The name of the model should be
unique in the Model Board.
In the Focus list, select how to weigh the relative importance of the fit at different frequencies. For
more information about each option, see Assigning Estimation Weightings on page 3-110.
10

In the Initial state list, specify how you want the algorithm to treat initial states. For more
information about the available options, see Specifying Initial States for Iterative Estimation
Algorithms on page 3-111.
11

Tip If you get an inaccurate fit, try setting a specific method for handling initial states rather than
choosing it automatically.
In the Covariance list, select Estimateif you want the algorithm to compute parameter uncertainties.
Effects of such uncertainties are displayed on plots as model confidence regions.
12

To omit estimating uncertainty, select None. Skipping uncertainty computation reduces computation
time for complex models and large data sets.
(PEM only) To view the estimation progress, select the Display progress check box. This launches a
progress viewer window in which estimation progress is reported.
13

14Click

Estimate to add this model to the System Identification Tool GUI.

(PEM only) To stop the search and save the results after the current iteration has been completed,
click Stop Iterations. To continue iterations from the current model, click the Continue iter button to
assign current parametervaluesasinitialguessesforthenextsearch.
15

Next Steps
Validate the model by selecting the appropriate check box in the Model Views area of the System
Identification Tool GUI. For more information about validating models, see Validating Models After
Estimation on page 8-3.
Export the model to the MATLAB workspace for further analysis by dragging it to the To Workspace
rectangle in the System Identification Tool GUI.

How to Estimate State-Space Models at the Command Line

Prerequisites on page 3-92
Black Box vs.Structured State-Space Model Estimation on page 3-92
Estimating State-Space Models Using ssest and n4sid on page 3-93
Choosing theStructure of A, B, C Matrices on page 3-94
Choosing Between Continuous-Time and Discrete-Time Representations on page 3-95
Choosing toEstimateD,K,andX0Matricesonpage3-95
Prerequisites
Before youcan perform this task, you must have
Input-output data as an iddataobject or frequency response data as an frdor idfrdobject. See
Representing Time- and Frequency-Domain Data Using iddata Objects on page 2-55. For supported
data formats, see Data Supported by State-Space Models on page 3-83.
Performed any required data preprocessing operations. To improve the accuracy of results when using
time domain data, you can detrend the data or specify the input/output offset levels as estimation
options. See Ways to Prepare Data for System Identification on page 2-6.
Selecta model order. For more information about how to estimate model orders, see Estimate StateSpace Model With Order Selection on page 3-83.
Black Box vs. Structured State-Space Model Estimation You can estimate state-space models in two
ways, depending upon your prior knowledge of the nature of the system and your requirements.
Black Box Estimation on page 3-93
Structured Estimation on page 3-93
Black Box Estimation. You specify the model order, and, optionally, additional model structure
attributes that configure the overall structure of the state-space matrices. In this way of estimating
models, you call ssest or n4sidwith data and model order as primary input arguments, and use namevalue pairs to specify any additional attributes, such as model sample time, presence of feedthrough,
absence of noise component, etc. You do not work directly with the coefficients of the A, B, C, D, K,and
X0 matrices.
Structured Estimation. You create and configure an idssmodel that contains the initial values for all the
system matrices. You use the Structure property of the idssmodel to specify all the parameter constraints.
For example, you can designate certain coefficients of system matrices as fixed and impose minimum/
maximum bounds on the values of the others. For quick configuration of the parameterization and
whether to estimate feedthrough and disturbance dynamics, use ssform.
After configuring the idssmodel with desired constraints, you specify this model as an input argument to
the ssestcommand. You cannot use n4sidfor structured estimation.
Note

 The structured estimation approach is also referred to as grey-box modeling. However, in this toolbox,
the grey box modeling terminology is used only when referring to idgreyand idnlgreymodels.
Using the structured estimation approach, you cannot specify relationships among state-space
coefficients. Each coefficient is essentially considered to be independent of others. For imposing
dependencies, or to use more complex forms of parameterization, use the idgreymodel and the
associated greyestestimator.
Estimating State-Space Models Using ssest and n4sid You can estimate continuous-time and discretetime state-space model using the iterative estimation command ssestthat minimizes the prediction errors
to obtain maximum-likelihood values. You can also use the noniterative subspace estimator n4sid.
Use the following general syntax to both configure and estimate state-space models:
m = ssest(data,n,opt,Name,Value)
where datais the estimation data, nis the model order. optcontains the options for configuring the
estimation of the state-space models. These options include the handling of the initial conditions, input
offset, estimation focus and search algorithm options. opt can be followed by name and value pair input
arguments that specify optional model structure attributes such as the presence of feedthrough, the
canonical form of the model, and input delay. For more information, see ssest.
As an alternative to ssest,youcanuse n4sid:
m = n4sid(data,n,opt,Name,Value)
Unless the sample time is specified as a name and value pair input argument, ssestestimates a
continuous-time model, while n4sidestimates a model whosesampletimematchesthatofdata.
Note ssestuses n4sidto initialize the state-space matrices, and takes longer than n4sidto estimate a model
but typically provides better fit to data.
For more information about estimating model order, see Estimate Model With Selected Order at the
Command Line on page 3-87.
For information about validating your model, see Validating Models After Estimation on page 8-3
Choosing the Structure of A, B, C Matrices
By default, all entries of the A, B,and C state-space matrices are treated as free parameters. Using the
Formname and value pair input argument of ssest, you can choose various canonical forms that employ
fewer parameters, such as the companion and modal forms.
For more information about estimating a specific state-space parameterization, see the following topics:
How to Estimate Free-Parameterization State-Space Models on page 3-98
How to Estimate State-Space Models with Canonical Parameterization on page 3-99
How to Estimate State-Space Models with Structured Parameterization on page 3-100
Choosing Between Continuous-Time and Discrete-Time Representations
For estimation of state-space models, you have the option of switching the model sample time between
zero and that of the estimation data. You can do this using the Tsname and value pair input argument.

 By default, ssestestimates a continuous-time model. If you are using data set with nonzero sample
time, data, which includes all time domain data, you can also estimate a discrete-time model by using:
model = ssest(data,nx,'Ts',data.Ts);
If you are using continuous-time frequency-domain data, you cannot estimate a discrete-time model.
By default, n4sidestimates a model whose sample time matches that of the data. Thus, for time-domain
data, n4siddelivers a discrete-time model. You can estimate a continuous-time model by using:
model = n4sid(data,nx,'Ts',0);
Choosing to Estimate D, K, and X0 Matrices
For state-space models with any parameterization, you can specify whether to estimate the D, K and X0
matrices, which represent the input-to-output feedthrough, noise model and the initial states,
respectively.
For state-space models with structured parameterization, you can also specify to estimate the D matrix.
However, for free and canonical forms, the structure of the D matrix is set based on your choice of
'Feedthrough'name and value pair input argument.
DMatrix.By default, the Dmatrixisnotestimatedanditsvalueisfixedto zero, except for static models.
Black box estimation: Use the Feedthroughname and value pair input argument to denote the
presence or absence of feedthrough from individual inputs. For example, in case of a two input model
such that there is feedthrough from only the second output, use model =
n4sid(data,n,'Feedthrough',[false true]);.
Structured estimation: Configure the values of the
init_sys.Structure.d,where init_sysis an idssmodel that represents the desired model structure. To force
no feedthrough for the i-th input, set:
init_sys.Structure.d.Value(:,i) = 0; init_sys.Structure.d.Free = true; init_sys.Structure.d.Free(:,i) = false;
The first line specifies the value of the i-th column of D as zero. The next line specifies all the elements
of D as free, estimable parameters. The last line specifies that the i-th column of the D matrix is fixed for
estimation.
Alternatively, use ssformwith 'Feedthrough'name-value pair..
KMatrix.K represents the noise matrix of the model, such that the noise component of the model is:.
xAx Ke nyCxe
For frequency-domain data, no noise model is estimated and K is set to 0. For time-domain data, K is
estimated by default in the black box estimation setup.
Black box estimation: Use the DisturbanceModelname and value pair input argument to indicate if
the disturbance component is fixed to zero (specify Value = `none') or estimated as a free parameter
(specify Value = `estimate'). For example, use model = n4sid(data,n,'DisturbanceModel','none').
Structured estimation: Configure the value of the
init_sys.Structure.kparameter, where init_sysis an idss

model that represents the desired model structure. You can fix some K matrix coefficients to known
values and prescribe minimum/maximum bounds for free coefficients. For example, to estimate only the
first column of the K matrix for a two output model:
kpar = init_sys.Structure.k;
kpar.Free(:,1) = true;
kpar.Free(:,2) = false;
kpar.Value(:,2) = 0; % second column value is fixed to zero init_sys.Structure.k = kpar;
Alternatively, use ssform.
When not sure how to easily fix or free all coefficients of K, initially you can omit estimating the noise
parameters in K to focus on achieving a reasonable model for the system dynamics. After estimating the
dynamic model, you can use ssestto refine the model while configuring the K parameters to be free. For
example:
init_sys = ssest(data, n,'DisturbanceModel','none'); init_sys.Structure.k.Free = true;
sys = ssest(data, init_sys);
where init_sysis the dynamic model without noise.
To set K to zero in an existing model, you can set its Valueto 0and Free flag to false:
m.Structure.k.Value = 0; m.Structure.k.Free = false;
X0 Matrices. The initial state vector X0 is obtained as the by-product of model estimation. The
n4sidand ssestcommands return the value of X0 as their second output arguments. You can choose how
to handle initial conditions during model estimation by using the InitialStateestimation option. Use
n4sidOptions(for n4sid)and ssestOptions(for ssest) to create the estimation option set. For example, in
order to hold the initial states to zero during estimation using n4sid:
opt = n4sidOptions;
opt.InitialState = 'zero'; [m,X0] = n4sid(data,n,opt);
The returned X0variable is a zero vector of length n.
When you estimate models using multiexperiment data, the X0matrix contains as many columns as data
experiments.
For a complete list of values for the InitialStatesoption, see Specifying Initial States for Iterative
Estimation Algorithms on page 3-111.

How to Estimate Free-Parameterization State-Space Models

The default parameterization of the state-space matrices A, B, C, D,and K is free; that is, any elements in
the matrices are adjustable by the estimation routines. Because the parameterization of A, B,and C is
free, a basis for the state-space realization is automatically selected to give well-conditioned
calculations.
To estimate the disturbance model K, you must use time domain data.

Suppose that you have no knowledge about the internal structure of the discrete-time state-space model.
To quickly get started, use the following syntax:
m = ssest(data)
where datais your estimation data. This command estimates a continuous-time state-space model for an
automatically selected order between 1 and 10.
To find a model of a specific order n, use the following syntax:
m = ssest(data,n)
The iterative algorithm ssestis initialized by the subspace method n4sid. You can use n4siddirectly, as an
alternative to ssest:
m = n4sid(data,n)

How to Estimate State-Space Models with Canonical Parameterization

What Is Canonical Parameterization? on page 3-99
Estimating Canonical State-Space Models on page 3-99
What Is Canonical Parameterization?
Canonical parameterization represents a state-space system in a reduced parameter form where the
many entries of A, B and C matrices are fixed to zeros and ones. The free parameters appear in only a
few of the rows and columnsin system matrices A, B, C,and D, and the remaining matrix elements are
fixed to zeros and ones.
The software supports the following canonical forms:
Companionform: The characteristic polynomial appears in the rightmost column ofthe A matrix.
Modal decomposition form: The state matrix A is block diagonal, with each block corresponding to a
cluster of nearby modes.
Observability canonical form: The free parameters appear only in select rows of the A matrix.
For more information about the distribution of free parameters in the observability canonical form, see
the appendix on identifiability of black-box multivariable model structures in System Identification:
Theory for the User, Second Edition, by Lennart Ljung, Prentice Hall PTR, 1999 (equation 4A.16).
Note The modal form has a certain symmetry in its block diagonal elements. If you update the
parameters of a model of this form (as a structured estimation using ssest), the symmetry is not
preserved, even though the updated model is still block-diagonal.
Estimating Canonical State-Space Models
You can estimate state-space models with chosen parameterization at the command line.
To specify a canonical form for A, B, C,and D,usethe 'Form'name and value pair input argument in the
estimator syntax, as follows:
m = ssest(data, n, 'Form', 'canonical')
If you have time-domain data, the preceding command estimates a continuous-time model. If you want a
discrete-time model, specify the data sample time using the 'Ts'name and value pair input argument:

md = ssest(data, n, 'Form', 'canonical', 'Ts', data.Ts)

If you have continuous-time frequency-domain data, you can only estimate a continuous-time model.

How to Estimate State-Space Models with Structured Parameterization

What Is Structured Parameterization? on page 3-100
Specifying the State-Space Structure on page 3-101
Are Grey-Box Models Similar to State-Space Models with Structured Parameterization? on page
3-103
Example Estimating Structured Discrete-Time State-Space Models on page 3-103
Example Estimating Structured Continuous-Time State-Space Models on page 3-105
What Is Structured Parameterization?
Structured parameterization lets you exclude specific parameters from estimation by setting these
parameters to specific values. This approach is useful when you can derive state-space matrices from
physical principles and provide initial parameter values based on physical insight. You can use this
approach to discover what happens if you fix specific parameter values or if you free certain parameters.
In the case of structured parameterization, there are two stages to the estimation procedure:
1Using the idsscommand to specify the structure of the state-space matrices and the initial values of the
free parameters
2Using the ssestestimation command to estimate the free model parameters
This approach differs from estimating models with free and canonical parameterizations, where it is not
necessary to specify initial parameter values before the estimation. For free parameterization, there is no
structure to specify because it is assumed to be unknown. For canonical parameterization, the structure
is fixed to a specific form.
Note To estimate structured state-space models in the System Identification Tool GUI, define the
corresponding model structures at the command line and import them into the System Identification
Tool GUI.
Specifying the State-Space Structure
To specify the state-space model structure, first define the A, B, C, D,and K matrices in the MATLAB
workspace.
To define a discrete-time state-space structure, use the following syntax:
m = idss(A,B,C,D,K,'Ts',T)
where A, B, C, D,and Kspecify the initial values for the free parameters. Tis the sampling interval.
Similarly, to define a continuous-time state-space structure, use the following syntax:
m = idss(A,B,C,D,K,'Ts',0)
In the continuous-time case, you must set the sampling interval property Ts to zero.
After you create the nominal model structure, you must specify which parameters to estimate and which
to set to specific values. To do this, use the Structureproperty of the model. Structurecontains parameters

for the five state-space matrices, A, B, C, D,and K. For each parameter, you can set the following
attributes:
Value Parameter values.
Minimum Minimum value that the parameter can assume during estimation.
Maximum Maximum value that the parameter can assume during estimation.
Free Boolean specifying whether the parameter is a free estimation variable. If you want to fix the
value of a parameter during estimation, set the corresponding Free = false. For example, if A is a 3-by-3
matrix of the model sys, sys.Structure.a.Free = eye(3)fixes all of the off-diagonal entries in A, to the
values specified in sys.Structure.a.Value.Inthis case, only the diagonal entries in A are estimable.
Scale Scale of the parameters value. Scaleis not used in estimation.
Info Structure array for storing parameter units and labels. The structure has Labeland Unitfields.
Use these fields for your convenience, to store strings that describe parameter units and labels.
For example, suppose that you constructed a nominal state-space model m with the following Amatrix:
A = [2 0; 0 3]
Suppose you want to fix A(1,2)=A(2,1)=0. Todothis,use:
m.Structure.a.Value(1,2) = 0; m.Structure.a.Value(2,1) = 0; m.Structure.a.Free(1,2) = false;
m.Structure.a.Free(2,1) = false;
Alternatively, to quickly configure the parameterization and whether to estimate feedthrough and
disturbance dynamics, use ssform.
The estimation algorithm only estimates the parameters in Afor which m.Structure.a.Freeis true.
Finally, use ssestto estimate the model, as described in How to Estimate State-Space Models at the
Command Line on page 3-92.
Use physical insight, whenever possible, to initialize the parameters for the iterative search algorithm.
Because it is possible that the numerical minimization gets stuck in a local minimum, try several
different initialization values for the parameters. For random initialization, use the initcommand. When
the model structure contains parameters with different orders of magnitude, try to scale the variables so
that the parameters are all roughly thesamemagnitude.
The iterative search computes gradients of the prediction errors with respect to the parameters using
numerical differentiation. The step size is specified by the nuderstcommand. The default step size is
equal to 104times the absolute value of a parameter or equal to 107, whichever is larger. To specify a
different step size, edit the nuderstMATLAB file.
Are Grey-Box Models Similar to State-Space Models with Structured Parameterization?
Structured parameterization state-space models are similar to grey-box modeling. However, the statespace models are simpler to estimate than grey-box models. The structured estimation approach is also
referred to as grey-box modeling. However, in this toolbox, the grey box modeling terminology is
used only when referring to idgreyand idnlgreymodels. Using the structured estimation approach, you
cannot specify relationships among state-space coefficients. Each coefficient is essentially considered to
be independent of others. For imposing dependencies, or to use more complex forms of
parameterization, use the idgreymodel and the associated greyest estimator.

To learn more about grey-box models, see Grey-Box Model Estimation.

Example Estimating Structured Discrete-Time State-Space Models
In this example, you estimate the unknown parameters (,,, , )
1 234 5

in the following discrete-time model:

xt
+= ?
1 1? 2? 4? ?01?xt()+ ??3?ut() + ??5?et() ? ? ? ? ? ?

yt() = []xt et() ()

+
x
()
=?
0? ?0?
??
Suppose that the nominal values of the unknown parameters (,,, , )
1 234 5

are -1, 2, 3, 4,and 5,respectively.

The discrete-time state-space model structure is defined by the following equation:
x kT( ) () () ()
+= + +
ykT() () () ()
=+ +
x() = x0
To construct and estimate the parameters of this discrete-time state-space model:
1Construct the parameter matrices and initialize the parameter values using the nominal parameter
values:
A = [1,-1;0,1]; B = [2;3];
C = [1,0];
D=0;
K = [4;5];
2Construct

the state-space model object:

m = idss(A,B,C,D,K);
3Specify the parameter values in the structure matrices that you do not want to estimate:
S = m.Structure;
S.a.Free(1,1) = false; S.a.Free(2,:) = false; S.c.Free = false; m.Structure = S;
D is initialized, by default, as a fixed value and K and B are initialized as free values. Suppose you want
to fix the initial states to known zero values. To enforce this, configure the InitialStateestimation option:
opt = ssestOptions;
opt.InitialState = 'zero';

4Estimate

the model structure:

m = ssest(data, m, opt)
where datais name of the iddataobject containing time-domain or frequency-domain data. The iterative
search starts with the nominal values in the A, B, C, D,and Kmatrices.
Example Estimating Structured Continuous-Time State-Space Models
In this example, you estimate the unknown parameters (,, )inthe 12 3 following continuous-time
model:
xt
()
=?
01? 0? ?0 ?xt()+ ?? ?ut() ? ?? ?
yt()10?= ??01?x t et)
+
??
x()3?= ?? 0?
??
This equation represents an electrical motor, whereyt x t= () is the
11

angular position of the motor shaft, andyt x t= () is the angular velocity.

22

The parameter

is the inverse time constant of the motor, and

2

static gain from the input to the angular velocity.

is the

The motor is at rest at t=0, but its angular position 3is unknown. Suppose that the approximate nominal
values of the unknown parameters are 1 =1
and 2 = 025 . The variance of the errors in the position measurement is 0.01, and the variance in the
angular velocity measurements is 0.1.For more information about this example, see the section on statespace models in System Identification: Theory for the User, Second Edition, by Lennart Ljung, Prentice
Hall PTR, 1999.
The continuous-time state-space model structure is defined by the following equation:
xt() Fxt Gut Kwt()
=+ +
yt() Hx t Dut wt()
=+ +
xx=00

To construct and estimate the parameters of this continuous-time state-space model:

1Construct the parameter matrices and initialize the parameter values using the nominal parameter
values:
Note The following matrices correspond to continuous-time representation. However, to be consistent
with the idssobject property name, this example uses A, B,and Cinstead of F, G,and H.
A = [0 1;0 -1]; B = [0;0.25]; C = eye(2);
D = [0;0];
K = zeros(2,2); x0 = [0;0];
2Construct

the continuous-time state-space model object:

m = idss(A,B,C,D,K,'Ts',0);
3Specify the parameter values in the structure matrices that you do not want to estimate:
S = m.Structure;
S.a.Free(1,:) = false; S.a.Free(2,1) = false; S.b.Free(1) = false; S.c.Free = false;
S.d.Free = false;
S.k.Free = false;
m.Structure = S;
m.NoiseVariance = [0.01 0; 0 0.1];
The initial state is partially unknown. Use the InitialStateoption of the ssestOptionsoption set to
configure the estimation behavior of X0:
opt = ssestOptions;
opt.InitialState = idpar(x0); opt.InitialState.Free(2) = false;
4Estimate

the model structure:

m = ssest(data, m, opt)
where datais name of the iddataobject containing time-domain or frequency-domain data. For this
example, you may use the data set dcmotordatathat is available in the folder matlabroot/toolbox/ident/
iddemos/data/. The iterative search for a minimum is initialized by the parameters in the nominal model
m.The continuous-time model is sampled using the same sampling interval as the data during estimation.
5To

simulate this system using the sampling interval T = 0.1for input u and the noise realization e, use
the following commands:
e = randn(300,2);
u = idinput(300);
simdat = iddata([],u,'Ts',0.1);
simopt = simOptions('AddNoise', true, 'NoiseData', e) y = sim(m,simdat,simopt)
The continuous system is sampled using Ts=0.1for simulation purposes. The noise sequence is scaled
according to the matrix m.NoiseVariance..
If you discover that the motor was not initially at rest, you can estimate x2(0) by setting the second

element of the InitialStateparametertobefree:

opt.InitialState.Free(2) = true; m_new = ssest(data, m, opt)

How to Estimate the State-Space Equivalent of ARMAX and OE Models

This example shows how to estimate ARMAX and OE form models using the state-space estimation
approach.
You can estimate the equivalent of ARMAX and output-error (OE) multiple-output models using statespace model structures. For the ARMAX case, specify to estimate theK matrix for the state-space model.
For the OE case, set K = 0. Convert the resulting model into idpolymodels to see them in the commonly
defined ARMAX or Output-Error forms.
Load measured data.
load iddata1 z1
Estimate state-space models.
mss_noK = n4sid(z1, 2,'DisturbanceModel','none'); mss = n4sid(z1,2);
mss_noK is a second order state-space model with no disturbance model used during estimation. mssis
also a second order state-space model, but with an estimated noise component. Both models use the
measured data set z1for estimation.
Convert the state-space models to polynomial models.
mOE = idpoly(mss_noK); mARMAX = idpoly(mss); Converting to polynomial models results in the
parameter covariance information for mOEand mARMAXto be lost.
You can use one of the following to recompute the covariance:
Zero-iteration update using the same estimation data.
translatecovas a Gauss approximation formula based translation of covariance of mss_noKand mssinto
covariance of mOEand mARMAX.
Reestimate mOEand mARMAXfor the parameters of the polynomial model using a zero iteration
update.
opt = polyestOptions;
opt.SearchOption.MaxIter = 0;
mOE = polyest(z1,mOE);
mARMAX = polyest(z1,mARMAX);
The options object, opt, specifies a zero iteration update for mOEand mARMAX. Consequently, the
model parameters remain unchanged and only their covariance information is updated.
Alternatively, you can use translatecovto convert the estimated models into polynomial form:
fcn = @(x)idpoly(x);
mOE = translatecov(fcn, mss_noK) mARMAX = translatecov(fcn, mss)
Because polyestand translatecovuse different computation algorithms, the covariance data obtained by
running a zero-iteration update may not match that obtained using translatecov.

View the uncertainties of the model parameters.

present(mOE)
present(mARMAX)
Tip You can use a state-space model with K = 0 (Output-Error (OE) form) for initializing a
Hammerstein-Wiener estimation at the command line. This initialization may improve the fit of the
model. See Using Linear Model for Hammerstein-Wiener Estimation on page 4-63.
For more information about ARMAX and OE models, see Identifying Input-Output Polynomial
Models on page 3-45.

Assigning Estimation Weightings

You can specify how the estimation algorithm weighs the fit at various frequencies. This information
supports the estimation procedures How to Estimate State-Space Models in the GUI on page 3-89 and
How to Estimate State-Space Models at the Command Line on page 3-92.
In the System Identification Tool GUI. Set Focus to one of the following options:
Prediction Uses the inverse of the noise model H to weigh the relative importance of how closely to
fit the data in various frequency ranges. Corresponds to minimizing one-step-ahead prediction, which
typically favors the fit over a short time interval. Optimized for output prediction applications.
Simulation Uses the input spectrum to weigh the relative importance of the fit in a specific
frequency range. Does not use the noise model to weigh
therelativeimportanceofhowcloselytofitthedatainvariousfrequency ranges. Optimized for output
simulation applications.
Stability Estimates the best stable model. For more information about model stability, see Unstable
Models on page 8-94.
Filter Specify a custom filter to open the Estimation Focus dialog box, where you can enter a filter,
as described in Simple Passband Filter on page 2-126 or Defining a Custom Filter on page 2-127.
This prefiltering applies only for estimating the dynamics from input to output. The disturbance model is
determined from the estimation data.
At the command line. Specify the focus as an estimation option using the same options as in the GUI.
For example, use this command to emphasize the fit between the 5 and 8 rad/s:
opt = ssestOptions;
opt.Focus = [5 8];
model = ssest(data,4,opt);
For more information on the 'Focus'option, see the reference page for ssestOptions.

Specifying Initial States for Iterative Estimation Algorithms

If you estimate state-space models using the iterative estimation algorithm ssest, you can specify how
the algorithm treats initial states. This information supports the estimation procedures How to Estimate
State-Space Models in the GUI on page 3-89 and How to Estimate State-Space Models at the
Command Line on page 3-92.
In the System Identification Tool GUI. Set Initial state to one of the following options:
Auto Automatically chooses Zero, Estimate,or Backcastbased on the estimation data. If initial states
have negligible effect on the prediction errors, the initial states are set to zero to optimize algorithm
performance.
Zero Sets all initial states to zero.
Estimate Treats the initial states as an unknown vector of parameters and estimates these states from
the data.
Backcast Estimates initial states using a backward filtering method (least-squares fit).
At the command line. Specify the method for handling initial states using the InitialStateestimation
option. For example, to estimate a fourth-order state-space model and set the initial states to be
estimated from the data:
opt = ssestOptions('InitialState','estimate'); m = ssest(data, 4, opt)
For a complete list of values for the InitialStatemodel property, see the ssestOptionsreference page.

State-Space Model Estimation Algorithms

For linear state-space models, you can use the subspace method, called N4SID. You can use the
subspace method N4SID to get an initial model (see the n4sidreference page), and then try to refine the
initial estimate using the iterative prediction-error method SSEST (see the ssestreference page).
N4SID is faster than SSEST, but is typically less accurate and robust, and requires additional arguments
that might be difficult to specify.
You can use the iterative prediction-error minimization (PEM)algorithmfor all linear and nonlinear
model types.

Identifying Transfer Function Models

In this section...
What are Transfer Function Models? on page 3-113
Data Supported by Transfer Function Models on page 3-115
How to Estimate Transfer Function Models in the System Identification Tool on page 3-115
How to Estimate Transfer Function Models at the Command Line on page 3-121
Transfer Function Structure Specification on page 3-122 Estimate Transfer Function Models by
Specifying Number of Poles on page 3-123
How to Estimate Transfer Function Models with Transport Delay to Fit Given Frequency Response

Data on page 3-123

How to Estimate Transfer Function Models With Prior Knowledge of Model Structure and Constraints
on page 3-124
How to Estimate Transfer Function Models with Unknown Transport Delays on page 3-126
Specifying Initial Conditions for Iterative Estimation Algorithms on page 3-127
Estimating Transfer Functions with Delays on page 3-128

What are Transfer Function Models?

Definition of Transfer Function Models
Transfer function models describe the relationship between the inputs and outputs of a system using a
ratio of polynomials. The model order is equal to the order of the denominator polynomial. The roots of
the denominator polynomial are referred to as the model poles. The roots of the numerator polynomial
are referred to as the model zeros.
The parameters of a transfer function model are its poles, zeros and transport delays.
Continuous-Time Representation
In continuous-time, a transfer function model has the form:
num s()
Ys()
den s Us Es()
Where, Y(s), U(s)and E(s) represent the Laplace transforms of the output, input and noise, respectively.
num(s)and den(s) represent the numerator and denominator polynomials that define the relationship
between the input and the output.
Discrete-Time Representation
In discrete-time, a transfer function model has the form:
1
yt() ()
den q() () ()
num q
()
1 01

bq
12

}
den
(()
2

112

qaqaq }1 2
The roots of num(q^-1)and den(q^-1) are expressed in terms of the lag variable q^-1.
Delays
In continuous-time, input and transport delays are of the form:
num s() s W () ()Ys()
den seUs Es
Where represents the delay.
In discrete-time:

num
yt()
denutW ) () where num and den are polynomials in the lag operator q^(-1).

Data Supported by Transfer Function Models

You can estimate transfer function models from data with the following characteristics:
Real data or complex data
Single-output and multiple-output
Time- or frequency-domain data
Note that you cannot use time-series data for transfer function model identification.
You must first import your data into the MATLAB workspace, as described in Data Preparation.

How to Estimate Transfer Function Models in the System Identification Tool

Import data into the System Identification Tool. See Importing Time-Domain Data into the GUI on
page 2-18. For supported data formats, see Data Supported by Transfer Function Models on page
3-115.
1

Perform any required data preprocessing operations. If input and/or output signals contain nonzero
offsets, consider detrending your data. See Ways to Prepare Data for System Identification on page
2-6.
2

3In

the System Identification Tool, select Estimate > Transfer Function Models

The Transfer Functions dialog box opens.

4Specify

the number of

poles and zeros of the transfer function as nonnegative integers.

For systems that are multiple input, multiple output, or both:
To use the same number of poles or zeros for all the input/output pairs, specify a scalar.
To use a different number of poles and zeros for the input/output pairs, specify an ny-bynu matrix. ny is
the number of outputs and nu is the number of inputs.
Alternatively, click
.

The Model Orders dialog box opens where you specify the number of poles and zeros for each input/
output pair. Use the Output list to select an output.

5Select

Continuous-time or Discretetime to specify whether the model is a continuous- or discrete-time transfer function.
For discrete-time models, the number of poles and zeros refers to the roots of the numerator and
denominator polynomials expressed in terms of the lag variable q^-1.

6(For

discrete-time models only) Specify whether to estimate the model feedthrough.

A discrete-time model with 2 poles and 3 zeros takes the following form:
1 12 3

Hz bbz bz bz
11 2
12

az
When the model has direct feedthrough, b0is a free parameter whose value is estimated along with the
rest of the model parameters b1, b2, b3, a1, a2. When the model has no feedthrough, b0is fixed to zero.
For SISO models, select the Feedthrough check box.

For models that are multi input, multi output or both, click Feedthrough.

The Model Orders dialog box opens, where you specify to estimate the feedthrough for each input/
output pair separately. Use the Output list to select an output.

Expand the I/O Delay section to specify nominal values and constraints for transport delays for
different input/output pairs.
7

Use the Output list to select an output. Select the Fixed check box to specify a transport delay as a fixed
value. Specify its nominal value in the Delay field.

8Expand

the Estimation Options section to specify estimation options.

In the Minimum and Maximum fields, specify the frequency range over which the transfer function
model must fit the data.
Select Display progress to view the progress of the optimization.
Select Estimate covariance to estimate the covariance of the transfer function parameters.
Specify how to treat the initial conditions in the Initial condition list. For more information, see
Specifying Initial Conditions for Iterative Estimation Algorithms on page 3-127.
Specify the algorithm used to initialize the values of the numerator and denominator coefficients in the
Initialization method list.
'iv' Instrument Variable approach.
'svf' State Variable Filters approach.
'gpmf' Generalized Poisson Moment Functions approach.
'n4sid' Subspace state-space estimation approach.
`all' Combination of all of the above approaches. The software tries all the above methods and
selects the method that yields the smallest value of prediction error norm.
9Click

Iterations Options to specify the options for controlling the iterations. The Options for Iterative
Minimization dialog box opens.

In this dialog box, you can specify the following

iteration options:
Search Method Method used by the iterative search algorithm. Search method is autoby default.
The descent direction is calculated using gn(Gauss-Newton), gna(Adaptive Gauss-Newton), lm
(Levenberg-Marquardt) and grad(Gradient Search) successively at each iteration until a sufficient
reduction in error is achieved.
Output weighting Weighting applied to the loss function to be minimized. Use this option for
multi-output estimations only. Specify as 'noise'or a positive semidefinite matrix of size equal the
number of outputs.
Maximum number of iterations Maximum number of iterations to use during search.
Termination tolerance Tolerance value when the iterations should terminate.
Error threshold for outlier penalty Robustification of the quadratic criterion of fit.
10Click Estimate to estimate the model. A new model gets added to the System Identification Tool.
Validate the model by selecting the appropriate check box in the Model Views area of the System
Identification Tool. For more information about validating models, see Validating Models After
Estimation on page 8-3.
11

12Export

the model to the MATLAB workspace for further analysis. Drag the model to the To
Workspace rectangle in the System Identification Tool.

How to Estimate Transfer Function Models at the Command Line

Before you estimate a transfer function model, you must have:

 Input/Output data. See Representing Time- and Frequency-Domain Data Using iddata Objects on
page 2-55. For supported data formats, see Data Supported by Transfer Function Models on page
3-115.
Performed any required data preprocessing operations. You can detrend your data before estimation.
For more information, see Ways to Prepare Data for System Identification on page 2-6.
Alternatively, you can specify the input/output offset for the data using an estimation option set. Use
tfestOptionsto create the estimation option set. Use the InputOffsetand OutputOffsetname and value
pairs to specify the input/output offset.
Estimate continuous-time and discrete-time transfer function models using tfest. The output of tfestis an
idtfobject, which represents the identified transfer function.
The general workflow in estimating a transfer function model is:
1Createadataobject(iddataor idfrd) that captures the experimental data.
2(Optional) Specify estimation options using tfestOptions.
3(Optional) Create a transfer function model that specifies the expected model structure and any
constraints on the estimation parameters.
4Use tfestto identify the transfer function model, based on the data. 5Validate the model. See Model
Validation.

Transfer Function Structure Specification

You can use a priori knowledge of the expected transfer function model
structuretoinitializetheestimation. The Structureproperty of an idtfmodel contains parameters that allow
you to specify the values and constraints for the numerator, denominator and transport delays.
For example, specify a third-order transfer function model that contains an integrator and has a transport
delay of at most 1.5 seconds:
init_sys = idtf([nan nan],[1 2 1 0]); init_sys.Structure.ioDelay.Maximum = 1.5;
init_sys.Structure.den.Free(end) = false;
int_sys is an idtfmodel with three poles and one zero. The denominator coefficient for the s^0term is
zero and implies that one of the poles is an integrator.
init_sys.Structure.ioDelay.Maximum = 1.5 constrains the transport delay to a maximum of 1.5 seconds.
The last element of the denominator coefficients (associated with the s^0term) is not a free estimation
variable. This constraint forces one of the estimated poles to be ats= 0. For more information regarding
configuring the initial parameterization of an estimated transfer function, see Structurein idtf.

Estimate Transfer Function Models by Specifying Number of Poles

This example shows how to identify a transfer function containing a specified number of poles for given
data.
Load time domain system response data and use it to estimate a transfer function for the system.

load iddata1 z1; np = 2;

sys = tfest(z1,np);
z1is an iddataobject that contains time-domain, input-output data.
npspecifies the number of poles in the estimated transfer function.
sysis an idtfmodel containing the estimated transfer function.
To see the numerator and denominator coefficients of the resulting estimated model sys,enter:
sys.num sys.den
To view the uncertainty in the estimates of the numerator and denominator and other information, use
tfdata.

How to Estimate Transfer Function Models with Transport Delay to Fit Given
Frequency Response Data
This example shows how to identify a transfer function to fit a given frequency response data (FRD)
containing additional phase roll off induced by input delay.
Obtain frequency response data. For this example, use bodeto obtain the magnitude and phase response
data for the following system:
Hs e
5s s02
3221

sss
Use 100 frequency points, ranging from 0.1 rad/s to 10 rad/s, to obtain the frequency response data. Use
frdto create a frequency response data object.
freq = logspace(-1,1,100);
[mag, phase] = bode(tf([1 .2],[1 2 1 1],'InputDelay',.5),freq); data = frd(mag.*exp(1j*phase*pi/
180),freq);
datais an iddataobject that contains frequency response data for the described system.
Estimate a transfer function using data. Specify an unknown transport delay for the identified transfer
function.
np = 3;
nz = 1;
iodelay = NaN;
sys = tfest(data,np,nz,iodelay)
npand nzspecify the number of poles and zeros in the identified transfer function, respectively.
iodelayspecifies an unknown transport delay for the identified transfer function.
sysis an idtfmodel containing the identified transfer function.

How to Estimate Transfer Function Models With Prior

KnowledgeofModelStructureand Constraints
This example shows how to estimate a transfer function for given data when the structure of the
expected model is known. This example also shows how to apply constraints to the numerator and
denominator coefficients.
Load time domain data.
load iddata1 z1;
z1.y = cumsum(z1.y);
cumsumintegrates the output data of z1. The estimated transfer function should therefore contain an
integrator.
Create a transfer function model with the expected structure.
init_sys = idtf([100 1500],[1 10 10 0]);
int_sysis an idtfmodel with three poles and one zero. The denominator coefficient for the s^0term is
zero. Therefore, int_syscontains an integrator.
Specify constraints on the numerator and denominator coefficients of the transfer function model. To do
so, configure fields in the Structureproperty:
init_sys.Structure.num.Minimum = eps; init_sys.Structure.den.Minimum = eps;
init_sys.Structure.den.Free(end) = false;
The constraints specify that the numerator and denominator coefficients are nonnegative. Additionally,
the last element of the denominator coefficients (associated with the s^0term) is not an estimable
parameter. This constraint forces one of the estimated poles to be ats= 0.
Create an estimation option set that specifies using the LevenbergMarquardt search method.
opt = tfestOptions('SearchMethod', 'lm');
Estimate a transfer function for z1using init_sysand the estimation option set.
sys = tfest(z1,init_sys,opt);
tfest uses the coefficients of init_systo initialize the estimation of sys. Additionally, the estimation is
constrained by the constraints you specify in the Structureproperty of init_sys.Theresulting idtfmodel
syscontains the parameter values that result from the estimation.

How to Estimate Transfer Function Models with Unknown Transport Delays

This example shows how to estimate a transfer function for given data with unknown transport delays.
This example also shows how to apply an upper bound on the unknown transport delays.
Create a transfer function model with the expected numerator and denominator structure and delay
constraints.
For this example, the experiment data consists of two inputs and one output. Both transport delays are
unknown and have an identical upper bound. Additionally, the transfer functions from both inputs to the
output are identical in structure.

init_sys = idtf(NaN(1,2),[1, NaN(1,3)],'ioDelay',NaN); init_sys.Structure(1).ioDelay.Free = true;

init_sys.Structure(1).ioDelay.Maximum = 7;
init_sys is an idtfmodel describing the structure of the transfer function from one input to the output. The
transfer function consists of one zero, three poles and a transport delay. The use of NaNindicates
unknown coefficients.
init_sys.Structure(1).ioDelay.Free = trueindicates that the transport delay is not fixed.
init_sys.Structure(1).ioDelay.Maximum = 7sets the upper bound for the transport delay to 7 seconds.
init_sys = [init_sys, init_sys];
init_sysnow contains the transfer function from both inputs to the output.
Load time domain system response data and detrend .
load co2data;
Ts = 0.5;
data = iddata(Output_exp1,Input_exp1,Ts); T = getTrend(data);
T.InputOffset = [170, 50];
T.OutputOffset = mean(data.y(1:75)); data = detrend(data, T);
datais an iddataobject and has a sample rate of 0.5 seconds.
Identify a transfer function for the measured data using the specified delay constraints.
sys = tfest(data,init_sys); sysis an idtfmodel containing the identified transfer function.

Specifying Initial Conditions for Iterative Estimation Algorithms

If you estimate transfer function models using tfest, you can specify how the algorithm treats initial
conditions.
In the System Identification Tool, set Initial condition to one of the following options:
auto Automatically chooses Zero, Estimate,or Backcastbased on the estimation data. If initial
conditions have negligible effect on the prediction errors, the initial conditions are set to zero to optimize
algorithm performance.
Zero Sets all initial conditions to zero.
EstimateTreatstheinitialconditionsasanestimationparameters.
Backcast Estimates initial conditions using a backward filtering method (least-squares fit).
At the command line. Specify the initial conditions by using an estimation option set. Use
tfestOptionsto create the estimation option set. For example, create an options set that sets the initial
conditions to zero:
opt = tfestOptions('InitialCondition','zero); For more information, see tfestOptions.

Estimating Transfer Functions with Delays

The tfestcommand supports estimation of IO delays. In the simplest case, if you specify NaNas the value
for the ioDelayinput argument, tfestestimates the corresponding delay value.
load iddata1 z1
sys = tfest(z1, 2, 2, NaN); % 2 poles, 2 zeros, unknown transport delay
If you want to assign an initial guess to the value of delay or prescribe bounds for its value, you must
first create a template idtfmodel and configure ioDelayusing the models Structureproperty:
sys0 = idtf([nan nan nan],[1 nan nan]);
sys0.Structure.ioDelay.Value = 0.1; % initial guess
sys0.Structure.ioDelay.Maximum = 1; % maximum allowable value for delay
sys0.Structure.ioDelay.Free = true; % treat delay as estimatable quantity sys = tfest(data, sys0)
If estimation data is in the time-domain, the delays are not estimated iteratively. If a finite initial value is
specified, that value is retained as is with no iterative updates. The same is true of discrete-time
frequency domain data. Thus in the example above, if datahas a nonzero sample time, the estimated
value of delay in the returned model sysis 0.1 (same as the initial guess specified for sys0). The delays
are updated iteratively only for continuos-time frequency domain data. If, on the other hand, a finite
initial value for delay is not specified (e.g., sys0.Structure.ioDelay.Value =NaN), then a value for delay
is determined using the delayestfunction, regardless of the nature of the data.
Determination of delay as a quantity independent of the models poles and zeros is a difficult task.
Estimation of delays becomes especially difficult for multi-input or multi-output data. It is strongly
recommended that you perform some investigation to determine delays before estimation. You can use
functions such as delayest, arxstruc, selstrucand impulse response analysis to determine delays. Often,
physical knowledge of the system or dedicated transient tests (how long does it take for a step change in
input to show up in a measured output?) will reveal the value of transport delays. Use the results of such
analysis to assign initial guesses as well as minimum and maximum bounds on the estimated values of
delays.
For an example of specifying constraints on delays, see How to Estimate Transfer Function Models
with Unknown Transport Delays on page 3-126

Refining Linear Parametric Models

In this section...
When to Refine Models on page 3-130
What You Specify to Refine a Model on page 3-130
How to Refine Linear Parametric Models in the GUI on page 3-131
How to Refine Linear Parametric Models at the Command Line on page 3-132

When to Refine Models

There are two situations where you can refine estimates of linear parametric models.

In the first situation, you have already estimated a parametric model and wish to update the values of its
free parameters to improve the fit to the estimation data. This is useful if your previous estimation
terminated because of search algorithm constraints such as maximum number of iterations or function
evaluations allowed reached. However, if your model captures the essential dynamics, it is usually not
necessary to continue improving the fitespecially when the improvement is a fraction of a percent.
In the second situation, you might have constructed a model using one of the model constructors
described in Commands for Constructing Model Structures on page 1-25. In this case, you built initial
parameter guesses into the model structure and wish to refine these parameter values.

What You Specify to Refine a Model

When you refine a model, you must provide two inputs:
Parametric model
Data You can either use the same data set for refining the model as the one you originally used to
estimate the model, or you can use a different data set.

How to Refine Linear Parametric Models in the GUI

The following procedure assumes that the model you want to refine is already in the System
Identification Tool GUI. You might have estimated this model in the current session or imported the
model from the MATLAB workspace. For information about importing models into the GUI, see
Importing Models into the GUI on page 12-8.
To refine your model:
1In the System Identification Tool GUI, verify that you have the correct data set in the Working Data
area for refining your model.
If you are using a different data set than the one you used to estimate the model, drag the correct data set
into the Working Data area. For more information about specifying estimation data, see Specifying
Estimation and Validation Data on page 2-35.
Select Estimate > Polynomial Models or Estimate > State Space Models to open the Polynomial
and State Space Models dialog box, if this dialog box is not already open.
2

3In the Linear Parametric Models dialog box, select By Initial Modelfrom
4Enter the model name into the Initial model field, and press Enter.

the Structure list.

The model name must be in the Model Board of the System Identification Tool GUI or a variable in the
MATLAB workspace. This model need not be a state-space or polynomial model; it could also be a
process model (idproc) or a transfer function model (idtf).
Tip As a shortcut for specifying a model in the Model Board, you can drag the model icon from the
System Identification Tool GUI into the Initial model field.

When you enter the model name, algorithm settings in the Polynomial and State Space Models dialog
box override the initial model settings. 5Modify the iteration options, displayed in the Polynomial and
State Space Models and Transfer Functions dialog box, if necessary.
6Click Estimate to refine the model.
7Validate the new model.
Tip To continue refining the model using additional iterations, click Continue iter. This action
continues parameter estimation using the most recent model.

How to Refine Linear Parametric Models at the Command Line

If you are working at the command line, you can use pemto refine parametric model estimates. You can
also use the various model-structure specific estimators ssestfor idssmodels, polyestfor idpolymodels,
tfestfor idtfmodels, and greyestfor idgreymodels.
The general syntax for refining initial models is as follows:
m = pem(data,init_model)
pemuses the properties of the initial model.
You can also specify the estimation options configuring the objective function and search algorithm
settings. For more information, see the reference page of the estimating function.

Refine ARMAX Model with Initial Parameter Guesses at Command Line

This example shows how to refine models for which you have initial parameter guesses. Estimate an
ARMAX model for the data by initializing the A, B, and C polynomials.
In this case, you must first create a model object and set the initial parameter values in the model
properties. Next, you provide this initial model as input to armax, polyest,or pem, which refine the initial
parameter guesses using the data.
load iddata8
% Define model parameters.
% Leading zeros in B indicate input delay (nk),
% which is 1 for each input channel.
A = [1 -1.2 0.7];
B{1} = [0 1 0.5 0.1]; % first input
B{2} = [0 1.5 -0.5]; % second input
B{3} = [0 -0.1 0.5 -0.1]; % third input
C=[1 000 0];
Ts = 1;
% Create model object.
init_model = idpoly(A,B,C,'Ts',1);
% Use polyest to update the parameters of the initial model. model = polyest(z8,init_model)
% Compare the two models. compare(z8,init_model,model)

For more information about estimating polynomial models, see Identifying Input-Output Polynomial
Models on page 3-45.

Refine Initial ARMAX Model at Command Line

This example shows to estimate an initial model and refine it using pem.
Load measured data.
load iddata8;
Split the data into an initial estimation data set and a refinement data set.
init_data = z8(1:100);
refine_data = z8(101:end);
init_datais an iddataobject containing the first 100 samples from z8and refine_datais an iddataobject
representing the remaining data in z8.
Estimate an ARMAX model.
na=4;
nb=[3 2 3];
nc=2;
nk=[0 0 0];
sys = armax(init_data,[na nb nc nk]);
armaxuses the default algorithm properties to estimate sys.
Refine the estimated model by specifying the estimation algorithm options. Specify stricter tolerance
and increase the maximum iterations.
opt = armaxOptions;
opt.SearchOption.Tolerance = 1e-5;
opt.SearchOption.MaxIter = 50;
refine_sys = pem(refine_data,sys,opt)
Compare the fit of the initial and refined models.
compare(refine_data,sys,refine_sys)

refine_sysprovides a closer fit to the data than sys. You can similarly use polyestor armaxto refine the
estimated model.

See Also armaxpempolyest

Extracting Numerical Model Data
You can extract the following numerical data from linear model objects:
Coefficients and uncertainty
For example, extract state-space matrices (A, B, C, Dand K) for state-space models, and polynomials (a,
b, c, dand f) for polynomial models.
If you estimated model uncertainty data, this information is stored in the model in the form of the
parameter covariance matrix. You can fetch the covariance matrix (in its raw or factored form) using the
getcovcommand. The covariance matrix represents uncertainties in parameter estimates
andisusedtocompute:
- Confidence bounds on model output plots, Bode plots, residual plots, and pole-zero plots
- Standard deviation in individual parameter values. For example, one standard deviation in the
estimated value of the Apolynomial in an ARX model, returned by the polydatacommand and displayed
by the presentcommand.

Dynamic and noise models

For linear models, the general symbolic model description is given by:
yGu He
=+
G is an operator that takes the measured inputs u to the outputs and captures the system dynamics, also
called the measured model. H is an operator that describes the properties of the additive output
disturbance and takes the hypothetical (unmeasured) noise source inputs e to the outputs, also called the
noise model. When you estimate a noise model, the toolbox includes one noise
channeleforeachoutputinyoursystem.
The following table summarizes the commands for extracting model coefficients and uncertainty.
Commands for Extracting Model Coefficients and Uncertainty Data
Command freqresp
polydata
idssdata
tfdata
zpkdata
Description
Extracts
frequency-response data (H) and corresponding covariance (CovH)from any linear identified model.
Extracts polynomials (such as A)from
any linear identified model. The polynomial uncertainties (such as dA) are returned only for
idpolymodels.
Extracts state-space matrices (such as
A) from any linear
identified model. The matrix uncertainties (such as dA)are
returned only for idss models.
Extracts numerator and denominator
polynomials (num, den) and their uncertainties (dnum, dden)fromany linear identified model.
Extracts zeros, poles, and gains (Z, P, K)and their covariances (covZ, covP, covK)fromany linear
identified model.
Syntax
[H,w,CovH] = freqresp(m)
[A,B,C,D,F,dA,dB,dC,dD,dF] = ... polydata(m)
[A,B,C,D,K,X0,...
dA,dB,dC,dD,dK,dX0] = ... idssdata(Model) [Num,Den,Ts,dNum,dDen] = ... tfdata(Model)
[Z,P,K,Ts,covZ,covP,covK] = ... zpkdata(m)
You can also extract numerical model data by using dot notation to access model properties. For
example, m.Adisplays the A polynomial coefficients from model m. Alternatively, you can use the
getcommand, as follows: get(m,'A').

Tip To view a list of model properties, type get(model).

You can operate on extracted model data as you would on any other MATLAB vectors, matrices and cell
arrays. You can also pass these numerical values to Control System Toolbox commands, for example, or
Simulink blocks.

Transforming Between Discrete-Time and Continuous-Time Representations

In this section...
Why Transform Between Continuous and Discrete Time? on page 3-139 Using the c2d, d2c, and d2d
Commands on page 3-139
Specifying Intersample Behavior on page 3-141
Effects on the Noise Model on page 3-141

Why Transform Between Continuous and Discrete Time?

Transforming between continuous-time and discrete-time representations is useful, for example, if you
have estimated a discrete-time linear model and require a continuous-time model instead for your
application.
You can use c2dand d2cto transform any linear identified model between continuous-time and discretetime representations. d2dis useful is you want to change the sampling interval of a discrete-time model.
All of these operations change the sampling interval, which is called resampling the model.
These commands do not transform the estimated model uncertainty. If you want to translate the
estimated parameter covariance during the conversion, use translatecov.
Note c2dand d2dcorrectly approximate the transformation of the noise model only when the sampling
interval Tis small compared to the bandwidth of the noise.

Using the c2d, d2c, and d2d Commands

The following table summarizes the commands for transforming between continuous-time and discretetime model representations.
Command Description
c2d Converts
continuous-time models to discrete-time models.You cannot use c2dfor idprocmodels and for
idgreymodels whose FcnTypeis not 'cd'. Convert these models into idpoly, idtf,or idssmodels before
calling c2d.
d2c Converts parametric discrete-time models to continuous-time models.You cannot use d2cfor
idgreymodels whose FcnTypeis not 'cd'. Convert these models into idpoly, idtf,or idssmodels before
calling d2c.
d2d Resample a linear discrete-time
model and produce an equivalent

discrete-time model with a new sampling interval.

You can use the resampled model to simulate or predict output with a specified time interval.
Usage Example
To transform a continuous-time model mod_cto a discrete-time form, use the following
command:
mod_d = c2d(mod_c,T)
where Tis the sampling interval of the discrete-time model.
To transform a discrete-time model mod_dto a
continuous-time form, use the following command:
mod_c = d2c(mod_d)
To resample a discrete-time model mod_d1to a discrete-time form with a new sampling interval
Ts,usethefollowing command:
mod_d2 = d2d(mod_d1,Ts)
The following commands compare estimated model mand its continuous-time counterpart mcon a Bode
plot:
% Estimate discrete-time ARMAX model % from the data
m = armax(data,[2 3 1 2]);
% Convert to continuous-time form mc = d2c(m);
% Plot bode plot for both models bode(m,mc)

Specifying Intersample Behavior

A sampled signal is characterized only by its values at the sampling instants. However, when you apply
a continuous-time input to a continuous-time system, the output values at the sampling instants depend
on the inputs at the sampling instants and on the inputs between these points. Thus, the InterSampledata
property describes how the algorithms should handle the input between samples. For example, you can
specify the behavior between the samples to be piece-wise constant (zero-order hold, zoh)orlinearly
interpolated between the samples (first order hold, foh). The transformation formulas for c2dand d2care
affected by the intersample behavior of the input.
By default, c2dand d2cuse the intersample behavior you assigned to the estimation data. To override this
setting during transformation, add an extra argument in the syntax. For example:
% Set first-order hold intersample behavior mod_d = c2d(mod_c,T,'foh')

Effects on the Noise Model

c2d, d2c,and d2dchange the sampling interval of both the dynamic model and the noise model.
Resampling a model affects the variance of its noise model.

A parametric noise model is a time-series model with the following mathematical description:
yt() = H q et( ) () Ee2 =
The noise spectrum is computed by the following discrete-time equation:

v() () 2
where is the variance of the white noise e(t),and Trepresents the spectral density of e(t). Resampling
the noise model preserves the spectral density T . The spectral density T is invariant up to the Nyquist
frequency. For more information about spectrum normalization, see Spectrum Normalization on page
3-14.
d2d resampling of the noise model affects simulations with noise using sim. If you resample a model to
a faster sampling rate, simulating this model results in higher noise level. This higher noise level results
from the underlying continuous-time model being subject to continuous-time white noise disturbances,
which have infinite, instantaneous variance. In this case, the underlying continuous-time
modelistheuniquerepresentationfor discrete-timemodels.
Tomaintainthesamelevelofnoiseafterinterpolating
the noise signal, scale the noise spectrum byTNewT ,where T
Old

new

is the new
sampling interval and Toldis the original sampling interval. before applying sim.

Continuous-Discrete Conversion Methods

In this section...
Choosing a Conversion Method on page 3-143 Zero-Order Hold on page 3-144
First-Order Hold on page 3-146
Impulse-Invariant Mapping on page 3-147 Tustin Approximation on page 3-148 Zero-Pole
Matching Equivalents on page 3-151

Choosing a Conversion Method

The c2dcommand discretizes continuous-time models. Conversely, d2c converts discrete-time models to
continuous time. Both commands support several discretization and interpolation methods, as shown in
the following table.
Discretization Method
Zero-Order Hold on page 3-144
First-Order Hold on page 3-146
Impulse-InvariantMappingon page3-147(c2donly)
Use when:
You want an exact discretization in the time domain for staircase inputs.

You want an exact discretization in the time domain for piecewise linear inputs.
You want an exact discretization in thetimedomainforimpulsetrain inputs.
Discretization Method
Tustin Approximation on page 3-148
Zero-Pole Matching Equivalents on page 3-151
Use when:
You want good matching in the frequency domain between the continuous- and discrete-time models.
Your model has important dynamics at some particular frequency.
You have a SISO model, and you want good matching in the frequency domain between the continuousand discrete-time models.

Zero-Order Hold
The Zero-Order Hold (ZOH) method provides an exact match between the continuous- and discrete-time
systems in the time domain for staircase inputs.
The following block diagram illustrates the zero-order-hold discretization Hd(z) of a continuous-time
linear model H(s)

The ZOH block generates the

continuous-time input signal u(t)byholding each sample value u(k)constantoveronesampleperiod:

ut uk [],1()T
() = ss

The signal u(t) is the input to the continuous system H(s). The output y[k] results from sampling
y(t)every Tsseconds.
Conversely, given a discrete system Hd(z), d2cproduces a continuous system H(s). The ZOH
discretization of H(s) coincides with Hd(z).
The ZOH discrete-to-continuous conversion has the following limitations:
d2ccannot convert LTI models with poles at z=0.
For discrete-time LTI models having negative real poles, ZOH d2c conversion produces a continuous
system with higher order. The model order increases because a negative real pole in the z domain maps
to apureimaginaryvalueinthe s domain. Such mapping results in a continuous-time model with complex
data. To avoid this, the software instead introduces a conjugate pair of complex poles in the s domain.

ZOH Method for Systems with Time Delays

You can use the ZOH method to discretize SISO or MIMO continuous-time models with time delays.
The ZOH method yields an exact discretization for systems with input delays, output delays, or transfer
delays.
For systems with internal delays (delays in feedback loops), the ZOH method results in approximate
discretizations. The following figure illustrates a system with an internal delay.

H(s)
e-ts
For such systems, c2dperforms the following actions to compute an approximate ZOH discretization:
kT
1 Decomposes the delay as with 0< Ts.s
2Absorbs the fractional delay into H(s).
Discretizes H(s)to H(z).
k
4 Represents the integer portion of the delay kTsas an internal discrete-time delay z . The final
discretized model appears in the following figure:
3

H(z)
e-srH(s)
z-k
First-Order Hold
The First-Order Hold (FOH) method provides an exact match between the continuous- and discrete-time
systems in the time domain for piecewise linear inputs.
FOH differs from ZOH by the underlying hold mechanism. To turn the input samplesu[k] into a
continuous input u(t), FOH uses linear interpolation between samples:

()
=

[]
+

ut uk tkTs()11()T
T

s ss

This method is generally more accurate than ZOH for systems driven by smooth inputs.
This FOH method differs from standard causal FOH and is more appropriately called triangle
approximation (see [2], p. 228). The method is also known as ramp-invariant approximation.

FOH Method for Systems with Time Delays

You can use the FOH method to discretize SISO or MIMO continuous-time models with time delays.
The FOH method handles time delays in the same way as the ZOH method. See ZOH Method for
Systems with Time Delays on page 3-145.

Impulse-Invariant Mapping
The impulse-invariant mapping produces a discrete-time model with the same impulse response as the
continuous time system. For example, compare the impulse response of a first-order continuous system
with the impulse-invariant discretization:
G = tf(1,[1,1]);
Gd1 = c2d(G,0.01,'impulse'); impulse(G,Gd1)

The impulse response plot

shows that the impulse responses of the continuous anddiscretizedsystemsmatch.
Impulse-Invariant Mapping for Systems with Time Delays You can use impulse-invariant mapping
to discretize SISO or MIMO continuous-time models with time delay, except that the method does not
support ssmodels with internal delays. For supported models, impulse-invariant mapping yields an exact
discretization of the time delay.

Tustin Approximation
The Tustin or bilinear approximation yields the best frequency-domain match between the continuoustime and discretized systems. This method relates the s-domain and z-domain transfer functions using
the approximation:

zesTs|12s / .12s /
In c2dconversions, the discretization Hd(z) of a continuous transfer function H(s)is:
d

()
=
Hz Hs

()
,
s
=
21 Ts z+1
Similarly, the d2cconversion relies on the inverse correspondence

Hs H z =d(), z =+12s / () 12s/

When you convert a state-space model using the Tustin method, the states are not preserved. The state
transformation depends upon the state-space matrices and whether the system has time delays. For
example, for an explicit (E = I) continuous-time model with no time delays, the state vector w[k]ofthe
discretized model is related to the continuous-time state vector x(t)by:
wkT

>@

T Bu kT T

AT xkT s ss s . 22 2

Tsis the sampling time of the discrete-time model. A and B are state-space matrices of the continuoustime model.
Tustin Approximation with Frequency Prewarping If your system has important dynamics at a
particular frequency that you want the transformation to preserve, you can use the Tustin method with
frequency prewarping. This method ensures a match between the continuousand discrete-time responses
at the prewarp frequency.
The Tustin approximation with frequency prewarping uses the following transformation of variables:
d

()
=
Hz Hs

()
,
s
=
Z z1

() z+1
This change of variable ensures the matching of the continuous- and discrete-time frequency responses
at the prewarp frequency ,becauseof the following correspondence:
Hj H e
()

()

Tustin Approximation for Systems with Time Delays You can use the Tustin approximation to
discretize SISO or MIMO continuous-time models with time delays.
By default, the Tustin method rounds any time delay to the nearest multiple of the sampling time.
Therefore, for any time delay tau,theintegerportion of the delay, k*Ts,mapstoadelayof ksampling periods
in the discretized model. This approach ignores the residual fractional delay, tau - k*Ts.
You can to approximate the fractional portion of the delay by a discrete all-pass filter (Thiran filter) of
specified order. To do so, use the FractDelayApproxOrderoption of c2dOptions.
To understand how the Tustin method handles systems with time delays, consider the following SISO
state-space model G(s). The model has input delay i,outputdelay o, and internal delay .

G(s)
e
tis etos H(s)
ets
The following figure shows the general result of discretizing G(s)usingthe Tustin method.

Gd(z)
z-mi Fi(z)Hd(z) Fo(z) z-mo
z-m F(z)
By default, c2dconverts the time delays to pure integer time delays. The c2dcommand computes the
integer delays by rounding each time delay to the nearest multiple of the sample time Ts. Thus, in the
default case, mi= round( i/Ts), mo= round( o/Ts), and m = round( /Ts).. Also in this case, Fi(z)= Fo(z)=
F(z)=1.
If you set FractDelayApproxOrderto a non-zero value, c2dapproximates the fractional portion of the
time delays by Thiran filters Fi(z), Fo(z), and F(z).
The Thiran filters add additional states to the model. The maximum number of additional states for each
delay is FractDelayApproxOrder. For example, for the input delay i, the order of the Thiran filter Fi(z)is:
order(Fi(z)) = max(ceil( i/Ts), FractDelayApproxOrder).
If ceil( i/Ts)< FractDelayApproxOrder,theThiranfilter Fi(z)approximates the entire input delay i.Ifceil(
i/Ts)> FractDelayApproxOrder,theThiran filter only approximates a portion of the input delay. In that
case, c2d represents the remainder of the input delay as a chain of unit delays zm,i where
mi=ceil( i/Ts) FractDelayApproxOrder.
c2duses Thiran filters and FractDelayApproxOrderin a similar way to approximate the output delay oand
the internal delay .
When you discretize tfand zpkmodels using the Tustin method, c2dfirst aggregates all input, output, and
transfer delays into a single transfer delay TOTfor each channel. c2dthen approximates TOTas a Thiran
filter and a
chain of unit delays in the same way as described for each of the time delays in ssmodels.
For more information about Thiran filters, see the thiranreference page and [4].

Zero-Pole Matching Equivalents

The method of conversion by computing zero-pole matching equivalents applies only to SISO systems.
The continuous and discretized systems have matching DC gains. Their poles and zeros are related by
the transformation:
i

ze
sT is

=
where:
ziis the ith pole or zero of the discrete-time system.
siis the ith pole or zero of the continuous-time system.
Tsis the sampling time.

See [2] for more information.

Zero-Pole Matching for Systems with Time Delays You can use zero-pole matching to discretize
SISO continuous-time models with time delay, except that the method does not support ssmodels with
internal delays. The zero-pole matching method handles time delays in the same way as the Tustin
approximation. See Tustin Approximation for Systems with Time Delays on page 3-149.

References
[1] strm, K.J. and B. Wittenmark, Computer-Controlled Systems: Theory and Design,PrenticeHall,1990,pp. 48-52.
[2] Franklin, G.F., Powell, D.J., and Workman, M.L., Digital Control of Dynamic Systems (3rd Edition),
Prentice Hall, 1997.
[3] Smith, J.O. III, Impulse Invariant Method,
Physical Audio Signal Processing, August 2007.
http://www.dsprelated.com/dspbooks/pasp/Impulse_Invariant_Method.html.
[4] T. Laakso, V. Valimaki, Splitting the Unit Delay, IEEE Signal Processing Magazine,Vol. 13,No.
1,p.30-60,1996.

See Also
c2d | d2c | c2dOptions | d2cOptions | d2d | thiran

Effect of Input Intersample Behavior on Continuous-Time Models

The intersample behavior of the input signals influences the estimation, simulation and prediction of
continuous-time models. A sampled signal is characterized only by its values at the sampling instants.
However, when you apply a continuous-time input to a continuous-time system, the output values at the
sampling instants depend on the inputs at the sampling instants and on the inputs between these points.
The iddataand idfrdobjects have an InterSampleproperty which stores how the input behaves between
the sampling instants. You can specify the behavior between the samples to be piecewise constant (zeroorder hold), linearly interpolated between the samples (first-order hold) or band-limited. A band-limited
intersample behavior of the input signal means:
A filtered input signal (an input of finite bandwidth) was used to excite the system dynamics
The input was measured using a sampling device (A/D converter with antialiasing) that reported it to
be band-limited even though the true input entering the system was piecewise constant or linear. In this
case, the sampling devices can be assumed to be a part of the system being modeled.
When input signal is band-limited, the estimation is performed as follows:
Time-domain data is converted into frequency domain data using fftand
thesampletimeofthedataissettozero.

 Discrete-time frequency domain data (IDDATAwith domain = frequency or IDFRDwith sample time
Ts?0) is treated as continuous-time data by setting the sample time Ts to zero.
The resulting continuous-time frequency domain data is used for model estimation. For more
information, see Pintelon, R. and J. Schoukens, System Identification. A Frequency Domain
Approach,section10.2, pp-352-356,Wiley-IEEE Press, New York, 2001.
Similarly, the intersample behavior of the input data affects the results of simulation and prediction of
continuous-time models. simand predict commands use the InterSampleproperty to choose the right
algorithm for computing model response. Say more or give some pointers about which algorithms are
used?
The following example simulates a system using first-order hold (foh) intersample behavior for input
signal.
sys=idtf([-1 -2],[1 2 1 0.5]);
rng('default')
u = idinput([100 1 5],'sine',[],[],[5 10 1]); Ts = 2;
y = lsim(sys, u, (0:Ts:999)', 'foh');
Create an iddataobject for the simulated input-output data:
data = iddata(y,u,Ts);
The default intersample behavior is zero-order hold (zoh):
data.InterSample
ans =
zoh
Estimate a transfer function using this data:
np = 3; % number of poles
nz = 1; % number of zeros
opt = tfestOptions('InitMethod','all','Display','on'); opt.SearchOption.MaxIter = 100;
modelZOH = tfest(data,np,nz,opt);
The model gives about 80% fit to data. The sample time of the data is large enough that intersample
inaccuracy (using zohrather than foh)leadsto significant modeling errors.
Re-estimate the model using fohintersample behavior:
data.InterSample = 'foh';
modelFOH = tfest(data, np, nz,opt); modelFOHis able to retrieve the original system correctly: Compare
the model outputs with data.
compare(data, modelZOH, modelFOH)

modelZOHis compared to data whose intersample behavior is foh.There,its fit decreases to around 70%.

Transforming Between Linear Model Representations

You can transform linear models between state-space and polynomial forms. You can also transform
between frequency-response, state-space, and polynomial forms.
If you used the System Identification Tool GUI to estimate models, you must export the models to the
MATLAB workspace before converting models.
For detailed information about each command in the following table, see the corresponding reference
page.
Commands for Transforming Model Representations
Command Model Type to Convert
idfrd Converts any linear model to an idfrdmodel.
If you have the Control System Toolbox product, this command converts any numeric LTI model too.
Usage Example
To get frequency response of mat default frequencies, use the following command:
m_f = idfrd(m)
To get frequency response at specific frequencies, use the following command:
idpoly Converts any linear identified model, except idfrd,to ARMAX representation if the original
model has a nontrivial noise component, or OE if the noise model is trivial (H=1).
If you have the Control System Toolbox product,

m_f = idfrd(m,f)
To get frequency response for a submodel from input 2to output 3,usethe following command:
m_f = idfrd(m(2,3))
To get an ARMAX model from state-space model m_ss,usethefollowingcommand:
m_p = idpoly(m_ss)
Commands for Transforming Model Representations (Continued)
Command Model Type to Convert Usage Example
this command converts any numeric LTI model, except frd.
idss Converts any linear identified model, except idfrd,to state-space representation. If you have the
Control System Toolbox product, this command converts any numeric LTI model, except frd.
idtf Converts any linear identified model, except idfrd, to transfer function representation. The noise
component of the original model is lost since an idtf objecthasnoelementsto model noise dynamics. If
you have the Control System Toolbox product, this command converts any numeric LTI model, except
frd.
To get a state-space model from an ARX model m_arx,usethefollowingcommand:
m_ss = idss(m_arx)
To get a transfer function from a state-space model m_ss,usethefollowing command:
m_tf = idtf(m_ss)
Note Most transformations among identified models (among idss, idtf, idpoly) causes the parameter
covariance information to be lost, with few exceptions:
Conversion of an idtfmodel to an idpolymodel.
Conversion of an idgreymodel to an idssmodel.
If you want to translate the estimated parameter covariance during conversion, use translatecov.

Subreferencing Models
In this section...
What Is Subreferencing? on page 3-159
Limitation on Supported Models on page 3-159
Subreferencing Specific Measured Channels on page 3-160 Separation of Measured and Noise
Components of Models on page 3-161 Treating Noise Channels as Measured Inputs on page 3-162

What Is Subreferencing?
You can use subreferencing to create models with subsets of inputs and outputs from existing
multivariable models. Subreferencing is also useful when you want to generate model plots for only

certain channels, such as when you are exploring multiple-output models for input channels that have
minimal effect on the output.
The toolbox supports subreferencing operations for idtf, idpoly, idproc, idss,and idfrdmodel objects.
Subreferencing is not supported for idgreymodels. If you want to analyze the sub-model, convert it into
an idssmodel first, and then subreference the I/Os of the idssmodel. If you want a grey-box
representation of a subset of I/Os, create a new idgreymodel that uses an ODE function returning the
desired I/O dynamics.
In addition to subreferencing the model for specific combinations of measured inputs and output, you
can subreference dynamic and noise models individually.

Limitation on Supported Models

Subreferencing nonlinear models is not supported.

Subreferencing Specific Measured Channels

Use the following general syntax to subreference specific input and output channels in models:
model(outputs,inputs)
In this syntax, outputsand inputsspecify channel indexes or channel names.
To select all output or all input channels, use a colon ( :). To select no channels, specify an empty matrix
([]). If you need to reference several channel names, use a cell array of strings.
For example, to create a new model m2from mfrom inputs 1('power')and 4('speed') to output number
3('position'), use either of the following equivalent commands:
m2 = m('position',{'power','speed'})
or
m2 = m(3,[1 4])
For a single-output model, you can use the following syntax to subreference specific input channels
without ambiguity:
m3 = m(inputs)
Similarly, for a single-input model, you can use the following syntax to subreference specific output
channels:
m4 = m(outputs)

Separation of Measured and Noise Components of Models

For linear models, the general symbolic model description is given by:
yGu He
=+
G is an operator that takes the measured inputs u to the outputs and captures the system dynamics.

H is an operator that describes the properties of the additive output disturbance and takes the
hypothetical (unmeasured) noise source inputs to the outputs. H represents the noise model. When you
specify to estimate a noise model, the resulting model include one noise channel e at the input for each
output in your system.
Thus, linear, parametric models represent input-output relationships for two kinds of input channels:
measured inputs and (unmeasured) noise inputs. For example, consider the ARX model given by one of
the following equations:

Aq yt Bqu t nk et() + ()
=
or Bq() 1
yt() =

Aqut() + Aq et() ()

In this case, the dynamic model is the relationship between the measured
input u and output y, G Bq Aq () (). The noise model is the contribution of the input noise e to the output
y,givenby H = 1 ().
Suppose that the model mcontains both a dynamic model G and a noise model H. To create a new model
that only has G and no noise contribution, simply set its NoiseVarianceproperty value to zero value.
To create a new model by subreferencing H due to unmeasured inputs, use the following syntax:
m_H = m(:,[])
This operation creates a time-series model from mby ignoring the measured input.
The covariance matrix of e is given by the model property NoiseVariance, which is the matrix :
= LLT
The covariance matrix of e is related to v,asfollows:
= eLv
where v is white noise with an identity covariance matrix representing independent noise sources with
unit variances.

Treating Noise Channels as Measured Inputs

To study noise contributions in more detail, it might be useful to convert the noise channels to measured
channels using noisecnv:
m_GH = noisecnv(m)
This operation creates a model m_GHthat represents both measured inputs u and noise inputs e, treating
both sources as measured signals. m_GHis a model from uand e to y, describing the transfer functions G
and H.
Converting noise channels to measured inputs loses information about the variance of the innovations e.
For example, step response due to the noise channels does not take into consideration the magnitude of

the noise contributions. To include this variance information, normalize e such that v becomes white
noise with an identity covariance matrix, where
=eLv
To normalize e, use the following command:
m_GH = noisecnv(m,'Norm')
This command creates a model where u and v are treated as measured signals, as follows:
yt() Gut HLv G HL[]?u?=+ = ?v?

??
For example, the scaling by L causes the step responses from v to y to reflect
thesizeofthedisturbanceinfluence.
The converted noise sources are named in a way that relates the noise channel to the corresponding
output. Unnormalized noise sources e are assigned names such as 'e@y1', 'e@y2',..., 'e@yn',where
'e@yn'refers to the noise input associated with the output yn. Similarly, normalized noise sources v, are
named 'v@y1', 'v@y2', ..., 'v@yn'.
If you want to create a model that has only the noise channels of an identified model as its measured
inputs, use the noise2meascommand. It results in a model with y(t)= He or y(t)= HLv,where e or v is
treated as a measured input.
Note When you plot models in the GUI that include noise sources, you can select to view the response
of the noise model corresponding to specific outputs. For more information, see Selecting Measured
and Noise Channels in Plots on page 12-16.

Concatenating Models
In this section...
About Concatenating Models on page 3-164
Limitation on Supported Models on page 3-165
Horizontal Concatenation of Model Objects on page 3-165 Vertical Concatenation of Model Objects
on page 3-165 Concatenating Noise Spectrum Data of idfrd Objects on page 3-166 See Also on
page 3-167

About Concatenating Models

You can perform horizontal and vertical concatenation of linear model objects to grow the number of
inputs or outputs in the model.
When you concatenate identified models, such as idtf, idpoly, idproc,and idssmodel objects, the resulting
model combines the parameters of the individual models. However, the estimated parameter covariance
is lost. If you want to translate the covariance information during concatenation, use translatecov.
Concatenation is not supported for idgreymodels; convert them to idss models first if you want to
perform concatenation.

You can also concatenate nonparametric models, which contain the estimated impulse-response
(idtfobject) and frequency-response (idfrdobject)ofa system.
In case of idfrdmodels, concatenation combines information in the ResponseDataproperties of the
individual model objects. ResponseDatais an ny-bynu-bynfarray that stores the response of the system,
where ny is the number of output channels, nuis the number of input channels, and nfis the number of
frequency values. The (j,i,:)vector of the resulting response data represents the frequency response from
the ith input to the jth output at all frequencies.

Limitation on Supported Models

Concatenation is supported for linear models only.

Horizontal Concatenation of Model Objects

Horizontal concatenation of model objects requires that they have the same outputs. If the output
channel names are different and their dimensions are the same, the concatenation operation resets the
output names to their default values.
The following syntax creates a new model object mthat contains the horizontal concatenation of
m1,m2,...,mN:
m = [m1,m2,...,mN]
m takes allof the inputs of m1,m2,...,mNto the same outputs as in the original models. The following
diagram is a graphical representation of horizontal concatenation of the models.
u1 y1

Model 2y1 u2 Model 1y2 u3y2

Combinedu1 Horizonal Concatenationy1 Same Inputsu2 of Model 1 and Model 2y2 Outputsu3

Vertical Concatenation of Model Objects

Vertical concatenation combines output channels of specified models. Vertical concatenation of model
objects requires that they have the same inputs. If the input channel names are different and their
dimensions are the same, the concatenation operation resets the input channel names to their default
('')values.
The following syntax creates a new model object mthat contains the vertical concatenation of
m1,m2,...,mN:
m = [m1;m2;... ;mN]
m takes the same inputs in the original models to all of the output of m1,m2,...,mN. The following
diagram is a graphical representation of vertical concatenation of frequency-response data.
u1 y1 u1

y3

Model 2u2 Model 1y2 u2

Same
u1
Vertical Concatenationy1 of Model 1 and Model 2 y2
Inputsu2 y3 Combined Outputs

Concatenating Noise Spectrum Data of idfrd Objects

When idfrdmodels are obtained as a result of estimation (such as using spa), the SpectrumDataproperty
is not empty and contains the power spectra and cross spectra of the output noise in the system. For each
output channel, this toolbox estimates one noise channel to explain the difference between the output of
the model and the measured output.
When the SpectrumDataproperty of individual idfrdobjectsisnotempty, horizontal and vertical
concatenation handle SpectrumData,asfollows.
In case of horizontal concatenation, there is no meaningful way to combine the SpectrumDataof
individual idfrdobjects, and the resulting SpectrumData property is empty. An empty property results
because each idfrdobject has its own set of noise channels, where the number of noise channels equals
the number of outputs. When the resulting idfrdobject contains the same output channels as each of the
individual idfrdobjects, it cannot accommodate the noise data from all the idfrdobjects.
In case of vertical concatenation, this toolbox concatenates individual noise models diagonally. The
following shows that m.SpectrumDatais a block diagonal matrix of the power spectra and cross spectra
of the output noise in the system:
?ms 0?
ms
=
??

??

0mN s?
? ? sin m.sis the abbreviation for the SpectrumDataproperty name.

See Also
If you have the Control System Toolbox product, see Combining Model Objects on page 10-5 about
additional functionality for combining models.

Merging Models
You can merge models of the same structure to obtain a single model with parameters that are
statistically weighed means of the parameters of the individual models. When computing the merged
model, the covariance matrices of the individual models determine the weights of the parameters.
You can perform the merge operation for the idtf, idgrey, idpoly, idproc, and idssmodel objects.
Note Each merge operation merges the same type of model object.

Merging models is an alternative to merging data sets into a single multiexperiment data set, and then
estimating a model for the merged data. Whereas merging data sets assumes that the signal-to-noise
ratios are about the same in the two experiments, merging models allows greater variations in model
uncertainty, which might result from greater disturbances in an experiment.
When the experimental conditions are about the same, merge the data instead of models. This approach
is more efficient and typically involves better-conditioned calculations. For more information about
merging data sets into a multiexperiment data set, see Creating Multiexperiment Data at the Command
Line on page 2-61.
For more information about merging models, see the mergereference page.

Building and Estimating Process Models Using System Identification Toolbox

This example shows how to build simple process models using System Identification Toolbox.
Techniques for creating these models and estimating their parameters using experimental data is
described. This example requires Simulink.
Introduction

This example illustrates how to build simple process models often used in process industry. Simple, loworder continuous-time transfer functions are usually employed to describe process behavior. Such
models are described by IDPROC objects which represent the transfer function in a pole-zero-gain form.
Process models are of the basic type Static Gain + Time Constant + Time Delay. They may be
represented as:
or as an integrating process:

where the user can determine the number of real poles (0, 1, 2 or 3), as well as the presence of a zero in
the numerator, the presence of an integrator term (1/s) and the presence of a time delay (Td). In addition,
an underdamped (complex) pair of poles may replace the real poles.
Representation of Process Models using IDPROC Objects

IDPROC objects define process models by using the letters P(for process model), D(for time delay),
Z(for a zero) and I(for integrator). An integer will denote the number of poles. The models are generated
by calling idprocwith astring identifier using these letters.
This should be clear from the following examples.
idproc('P1') % transfer function with only one pole (no zeros or delay) idproc('P2DIZ') % model with 2
poles, delay integrator and delay idproc('P0ID') % model with no poles, but an integrator and a delay
ans =
Process model with transfer function: Kp
G(s) = ---------

1+Tp1*s
Kp = NaN Tp1 = NaN
Parameterization:
'P1'
Number of free coefficients: 2
Use "getpvec", "getcov" for parameters and their uncertainties.
Status:
Created by direct construction or transformation. Not estimated.
ans =
Process model with transfer function: 1+Tz*s
G(s) = Kp * ------------------- * exp(-Td*s) s(1+Tp1*s)(1+Tp2*s)
Kp = NaN Tp1 = NaN Tp2 = NaN
Td = NaN Tz = NaN
Parameterization:
'P2DIZ'
Number of free coefficients: 5
Use "getpvec", "getcov" for parameters and their uncertainties. Status:
Created by direct construction or transformation. Not estimated.
ans =
Process model with transfer function: Kp
G(s) = --- * exp(-Td*s)
s
Kp = NaN Td = NaN
Parameterization:
'P0DI'
Number of free coefficients: 2
Use "getpvec", "getcov" for parameters and their uncertainties.
Status:
Created by direct construction or transformation. Not estimated.
Creating an IDPROC Object (using a Simulink Model as Example)

Consider the system described by the following SIMULINK model:

open_system('iddempr1')

set_param('iddempr1/Random Number','seed','0')

The red part is the system, the blue part is the controller and the reference
signalisasweptsinusoid(achirpsignal). Thedatasamplingtimeissetto0.5 seconds. As observed, the system
is a continuous-time transfer function, and can hence be described using model objects in System
Identification Toolbox, such as idss, idpolyor idproc.
Let us describe the system using idpolyand idprocobjects. Using idpoly object, the system may be
described as:
m0 = idpoly(1,0.1,1,1,[1 0.5],'Ts',0,'InputDelay',1.57,'NoiseVariance',0.01
The IDPOLY form used above is useful for describing transfer functions of arbitrary orders. Since the
system we are considering here is quite simple (one pole and no zeros), and is continuous-time, we may
use the simpler IDPROC object to capture its dynamics:
m0p = idproc('p1d','Kp',0.2,'Tp1',2,'Td',1.57) % one pole+delay, with initi % for gain, pole and delay s
m0p =
Process model with transfer function: Kp
G(s) = ---------- * exp(-Td*s) 1+Tp1*s
Kp = 0.2 Tp1 = 2
Td = 1.57
Parameterization:
'P1D'
Number of free coefficients: 3
Use "getpvec", "getcov" for parameters and their uncertainties.
Status:
Created by direct construction or transformation. Not estimated.
Estimating Parameters of IDPROC Models

Once a system is described by a model object, such as IDPROC, it may be used for estimation of its
parameters using measurement data. As an example, we consider the problem of estimation of

parameters of the Simulink models system (red portion) using simulation data. We begin by acquiring
data for estimation:
sim('iddempr1')
dat1e = iddata(y,u,0.5); % The IDDATA object for storing measurement data
Letuslookatthedata:
plot(dat1e)

We can identify a process model using procestcommand, by providing the same structure information
specified to create IDPROC models. For example, the 1-pole+delay model may be estimated by calling
procestas follows:
m1 = procest(dat1e,'p1d'); % estimation of idproc model using data 'dat1e'.
% Check the result of estimation: m1
m1 =
Process model with transfer function: Kp
G(s) = ---------- * exp(-Td*s) 1+Tp1*s
Kp = 0.20045 Tp1 = 2.0431
Td = 1.499
Parameterization:
'P1D'
Number of free coefficients: 3
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using PROCEST on time domain data "dat1e". Fit to estimation data: 87.34% (prediction
focus) FPE: 0.01068, MSE: 0.01063
To get information about uncertainties, use
present(m1)
m1 =
Process model with transfer function: Kp
G(s) = ---------- * exp(-Td*s) 1+Tp1*s
Kp = 0.20045 +/- 0.00077275 Tp1 = 2.0431 +/- 0.061216
Td = 1.499 +/- 0.040854 Parameterization:
'P1D'
Number of free coefficients: 3
Use "getpvec", "getcov" for parameters and their uncertainties.
Status:
Termination condition: Near (local) minimum, (norm(g) < tol). Number of iterations: 4, Number of
function evaluations: 9
Estimated using PROCEST on time domain data "dat1e". Fit to estimation data: 87.34% (prediction
focus) FPE: 0.01068, MSE: 0.01063
More information in model's "Report" property.
The model parameters, K, Tp1and Tdare now shown with one standard deviation uncertainty range.
Computing Time and Frequency Response of IDPROC Models

The model m1estimated above is an IDPROC model object to which all of the toolboxs model
commands can be applied:
step(m1,m0) %step response of models m1 (estimated) and m0 (actual) legend('m1 (estimated
parameters)','m0 (known parameters)','location','nort
Bode response with confidence region corresponding to 5 s.d. may be computed by doing:
h = bodeplot(m1,m0); showConfidence(h,3)
Similarly, the measurement data may be compared to the models outputs using compareas follows:
compare(dat1e,m0,m1)

Other operations such as sim, impulse, c2darealsoavailable,justasthey are for other model objects.
bdclose('iddempr1')
Accommodating the Effect of Intersample Behavior in Estimation

It may be important (at least for slow sampling) to consider the intersample behavior of the input data.
To illustrate this, let us study the same system as before, but without the sample-and-hold circuit:
open_system('iddempr5')
Simulate this system with the same sampling interval:
sim('iddempr5')
dat1f = iddata(y,u,0.5); % The IDDATA object for the simulated data
We estimate an IDPROC model using data1fwhilealsoimposinganupper bound on the allowable value
delay. We will use lm as search method and also choose to view the estimation progress.
m2_init = idproc('P1D');
m2_init.Structure.Td.Maximum = 2;
opt = procestOptions('SearchMethod','lm','Display','on'); m2 = procest(dat1f,m2_init,opt);
m2
m2 =
Process model with transfer function: Kp
G(s) = ---------- * exp(-Td*s) 1+Tp1*s

Kp = 0.20038 Tp1 = 2.01

Td = 1.31
Parameterization: 'P1D'
Number of free coefficients: 3
Use "getpvec", "getcov" for parameters and their uncertainties.
Status:
Estimated using PROCEST on time domain data "dat1f". Fit to estimation data: 87.26% (prediction
focus) FPE: 0.01067, MSE: 0.01062
This model has a slightly less precise estimate of the delay than the previous one, m1:
[m0p.Td, m1.Td, m2.Td]
step(m0,m1,m2)
legend('m0 (actual)','m1 (estimated with ZOH)','m2 (estimated without ZOH)'
ans =
1.5700 1.4990 1.3100

However, by
telling the estimation process that the intersample behavior is first-order-hold (an approximation to the
true continuous) input, we do better:
dat1f.InterSample = 'foh';
m3 = procest(dat1f,m2_init,opt);

Compare the four models m0 (true) m1 (obtained from zoh input) m2 (obtained for continuous input,
with zoh assumption) and m3 (obtained for the same input, but with foh assumption)
[m0p.Td, m1.Td, m2.Td, m3.Td]
compare(dat1e,m0,m1,m2,m3)
step(m0,m1,m2,m3)
legend('m0','m1','m2','m3')
bdclose('iddempr5')
ans =
1.5700 1.4990 1.3100 1.5570

Modeling a System
Operating in Closed Loop

Let us now consider a more complex process, with integration, that is operated in closed loop:
open_system('iddempr2')

Thetruesystemcanberepresentedby:
m0 = idproc('P2ZDI','Kp',1,'Tp1',1,'Tp2',5,'Tz',3,'Td',2.2);
The process is controlled by a PD regulator with limited input amplitude and a zero order hold device.
The sampling interval is 1 second.

set_param('iddempr2/Random Number','seed','0')
sim('iddempr2')
dat2 = iddata(y,u,1); % IDDATA object for estimation
Two different simulations are made, the first for estimation and the second one for validation purposes.
set_param('iddempr2/Random Number','seed','13')
sim('iddempr2')
dat2v = iddata(y,u,1); % IDDATA object for validation purpose
Letuslookatthedata(estimationandvalidation).
plot(dat2,dat2v)
legend('dat2 (estimation)','dat2v (validation)')

Let us now
perform estimation using dat2.
m2_init = idproc('P2ZDI');
m2_init.Structure.Td.Maximum = 5;
m2_init.Structure.Tp1.Maximum = 2;
opt = procestOptions('SearchMethod','lm','Display','full'); opt.SearchOption.MaxIter = 50;
m2 = procest(dat2, m2_init, opt)
m2 =
Process model with transfer function:
1+Tz*s

G(s) = Kp * ------------------- * exp(-Td*s)

s(1+Tp1*s)(1+Tp2*s)
Kp = -0.89555 Tp1=2
Tp2 = 1388.4
Td = 3.08
Tz = -1519.1
Parameterization:
'P2DIZ'
Number of free coefficients: 5
Use "getpvec", "getcov" for parameters and their uncertainties.
Status:
Estimated using PROCEST on time domain data "dat2". Fit to estimation data: 90.21% (prediction
focus) FPE: 0.1651, MSE: 0.1457
compare(dat2v,m2,m0) % Gives very good agreement with data

bode(m2,m0)
legend({'m2 (est)','m0 (actual)'},'location','west')

impulse(m2,m0)
legend({'m2 (est)','m0 (actual)'})

Compare also
with the parameters of the true system:
present(m2)

[getpvec(m0), getpvec(m2)]
m2 =
Process model with transfer function:
1+Tz*s
G(s) = Kp * ------------------- * exp(-Td*s)
s(1+Tp1*s)(1+Tp2*s)
Kp = -0.89555 +/- 22.268
Tp1 = 2 +/- 0.25413
Tp2 = 1388.4 +/- 16546
Td = 3.08 +/- 0.20786
Tz = -1519.1 +/- 19687
Parameterization:
'P2DIZ'
Number of free coefficients: 5
Use "getpvec", "getcov" for parameters and their uncertainties.
Status:
Termination condition: Maximum number of iterations reached. Number of iterations: 50, Number of
function evaluations: 344
Estimated using PROCEST on time domain data "dat2". Fit to estimation data: 90.21% (prediction
focus) FPE: 0.1651, MSE: 0.1457
More information in model's "Report" property.
ans =
1.0e+03 *
0.0010 -0.0009
0.0010 0.0020
0.0050 1.3884
0.0022 0.0031
0.0030 -1.5191
A word of caution. Identification of several real time constants may sometimes be an ill-conditioned
problem, especially if the data are collected in closed loop.
To illustrate this, let us estimate a model based on the validation data:
m2v = procest(dat2v, m2_init, opt) [getpvec(m0), getpvec(m2), getpvec(m2v)]
m2v =
Process model with transfer function: 1+Tz*s
G(s) = Kp * ------------------- * exp(-Td*s) s(1+Tp1*s)(1+Tp2*s)
Kp = -2.8056 Tp1=2
Tp2 = 4586.1
Td = 2.252 Tz = -1630

Parameterization:
'P2DIZ'
Number of free coefficients: 5
Use "getpvec", "getcov" for parameters and their uncertainties.
Status:
Estimated using PROCEST on time domain data "dat2v". Fit to estimation data: 89.61% (prediction
focus) FPE: 0.1897, MSE: 0.1675
ans =
1.0e+03 *
0.0010 -0.0009 -0.0028
0.0010 0.0020 0.0020
0.0050 1.3884 4.5861
0.0022 0.0031 0.0023
0.0030 -1.5191 -1.6300
This model has much worse parameter values. On the other hand, it performs nearly identically to the
true system m0 when tested on the other data set dat2:
compare(dat2,m0,m2,m2v)

Fixing Known Parameters During Estimation

Suppose we know from other sources that one time constant is 1:

m2v.Structure.Tp1.Value = 1;
m2v.Structure.Tp1.Free = false;
We can fix this value, while estimating the other parameters:
m2v = procest(dat2v,m2v)
%
m2v =
Process model with transfer function:
1+Tz*s
G(s) = Kp * ------------------- * exp(-Td*s)
s(1+Tp1*s)(1+Tp2*s)
Kp = -8.4621 Tp1=1
Tp2 = 13964
Td = 4.201 Tz = -1630.2
Parameterization:
'P2DIZ'
Number of free coefficients: 4
Use "getpvec", "getcov" for parameters and their uncertainties.
Status:
Estimated using PROCEST on time domain data "dat2v". Fit to estimation data: 90.74% (prediction
focus) FPE: 0.1372, MSE: 0.1332
As observed, fixing Tp1to its known value dramatically improves the estimates of the remaining
parameters in model m2v.
This also indicates that simple approximation should do well on the data:
m1x_init = idproc('P2D'); % simpler structure (no zero, no integrator) m1x_init.Structure.Td.Maximum
= 2;
m1x = procest(dat2v, m1x_init)
compare(dat2,m0,m2,m2v,m1x)
m1x =
Process model with transfer function: Kp
G(s) = ----------------- * exp(-Td*s) (1+Tp1*s)(1+Tp2*s)
Kp = 14815 Tp1 = 15070 Tp2 = 3.1505
Td = 1.833
Parameterization: 'P2D'
Number of free coefficients: 4
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using PROCEST on time domain data "dat2v". Fit to estimation data: 91.23% (prediction
focus) FPE: 0.1209, MSE: 0.1193

Thus, the simpler model is able to estimate system output pretty well. However, m1xdoes not contain
any integration, so the open loop long time range behavior will be quite different:
step(m0,m2,m2v,m1x)
legend('m0','m2','m2v','m1x')
bdclose('iddempr2')

Additional
Information

For more information on identification of dynamic systems with System Identification Toolbox visit the
System Identification Toolbox product information page.

Determining Model Order and Delay

Estimation requires you to specify the model order and delay. Many times, these values are not known.
You can determine the model order and delay in one of the following ways:
Guess their values by visually inspecting the data or based on the prior
knowledge of the system.
Estimate delay as a part of idprocor idtfmodel estimation. These models treat delay as an estimable
parameter and you can determine their values by the estimation commands procestand tfest,respectively.
However automatic estimation of delays can cause errors. Therefore, it is recommended that you analyze
the data for delays in advance.
To estimate delays, you can also use one of the following tools:
- Estimate delay using delayest. The choice of the order of the underlying ARX model and the lower/
upper bound on the value of the delay to be estimated influence the value returned by delayest.
- Compute impulse response using impulseest. Plot the impulse response with a confidence interval of
sufficient standard deviations (usually 3). The delay is indicated by the number of response samples that

are inside the statistically zero region (marked by the confidence bound) before the response goes
outside that region.
- Select the model order in n4sidby specifying the model order as a vector.
- Choose the model order of an ARX model using arxstrucor ivstrucand selstruc. These command select
the number of poles, zeros and delay. See Model Structure Selection: Determining Model Order and
Input Delay on page 3-196 for an example of using these tools.

Model Structure Selection: Determining Model Order and Input Delay

This example shows some methods for choosing and configuring the model structure. Estimation of a
model using measurement data requires selection of a model structure (such as state-space or transfer
function) and its order (e.g., numberof poles and zeros) in advance. This choice is influenced by prior
knowledge about the system being modeled, but can also be motivated by an analysis of data itself. This
example describes some options for determining model ordersand input delay.
Introduction

Choosing a model structure is usually the first step towards its estimation. There are various possibilities
for structure - state-space, transfer functions and polynomial forms such as ARX, ARMAX, OE, BJ etc.
If you do not have detailed prior knowledge of your system, such as its noise characteristics and
indication of feedback, the choice of a reasonable structure may not be obvious. Also for a given choice
of structure, the order of the model needs to be specified before the corresponding parameters are
estimated. System Identification Toolbox offers some tools to assist in the task of model order
selection.
The choice of a model order is also influenced by the amount of delay. A good idea of the input delay
simplifies the task of figuring out the orders of other model coefficients. Discussed below are some
options for input delay determination and model structure and order selection.
Choosing and Preparing Example Data for Analysis

This example uses the hair dryer data, also used by iddemo1 ("Estimating Simple Models from Real
Laboratory Process Data"). The process consists of air being fanned through a tube. The air is heated at
the inlet of the tube, and the input is the voltage applied to the heater. The output is the temperature at
the outlet of the tube.
Let us begin by loading the measurement data and doing some basic preprocessing:
loaddry2
Form a data set for estimation of the first half, and a reference set for validation purposes of the second
half:
ze = dry2(1:500); zr = dry2(501:1000);
Detrend each of the sets:
ze = detrend(ze); zr = detrend(zr);
Letuslookataportionoftheestimationdata:
plot(ze(200:350))

Estimating Input
Delay

There are various options available for determining the time delay from input to output. These are:
Using the DELAYESTutility.
Using a non-parametric estimate of the impulse response, using IMPULSEEST.
Using the state-space model estimator N4SIDwith a number of different orders and finding the delay of
the best one.
Using delayest:
Let us discuss the above options in detail. Function delayestreturns an estimate of the delay for a given
choice of orders of numerator and denominator polynomials. This function evaluates an ARX structure:
y(t) + a1*y(t-1) + ... + ana*y(t-na) = b1*u(t-nk) + ...+bnb*u(t-nb-nk+1)
with various delays and chooses the delay value that seems to return the best fit. In this process, chosen
values of naand nbare used.
delay = delayest(ze) % na = nb = 2 is used, by default
delay =
3
A value of 3 is returned by default. But this value may change a bit if the assumed orders of numerator
and denominator polynomials (2 here) is changed. For example:
delay = delayest(ze,5,4)
delay =
2

returns a value of 2. To gain insight into how delayestworks, let us evaluate the loss function for various
choices of delays explicitly. We select a second order model (na=nb=2), which is the default for
delayest, and try out every time delay between 1 and 10. The loss function for the different models are
computed using the validation data set:
V = arxstruc(ze,zr,struc(2,2,1:10));
We now select that delay that gives the best fit for the validation data:
[nn,Vm] = selstruc(V,0); % nn is given as [na nb nk]
The chosen structure was:
nn
nn =
223
which show the best model has a delay of nn(3) = 3.
We can also check how the fit depends on the delay. This information is returned in the second output
Vm. The logarithms of a quadratic loss function are given as the first row, while the indexes na, nband
nkare given as a column below the corresponding loss function.
Vm
Vm =
Columns 1 through 7
-0.1283 -1.3142 -1.8787 -0.2339 0.0084 0.0900 0.1957 2.0000 2.0000 2.0000 2.0000 2.0000 2.0000
2.0000 2.0000 2.0000 2.0000 2.0000 2.0000 2.0000 2.0000 1.0000 2.0000 3.0000 4.0000 5.0000 6.0000
7.0000
Columns 8 through 10
0.2082 0.1728 0.1631
2.0000 2.0000 2.0000
2.0000 2.0000 2.0000
8.0000 9.0000 10.0000
The choice of 3 delays is thus rather clear, since the corresponding loss is minimum.
Using impulse
To gain a better insight into the dynamics, let us compute the impulse response of the system. We will
use the function impulseestto compute a non-parametric impulse response model. We plot this response
with a confidence interval represented by 3 standard deviations.
FIRModel = impulseest(ze); clf
h = impulseplot(FIRModel); showConfidence(h,3)
The filled light-blue region shows the confidence interval for the insignificant response in this
estimation. There is a clear indication that the impulse response "takes off" (leaves the uncertainty
region) after 3 samples. This points to a delay of three intervals.
Using n4sid based state-space evaluation

We may also estimate a family of parametric models to find the delay corresponding to the "best" model.
In case of state-space models, a range of orders may be evaluated simultaneously and user gets
prompted to choose the best order. Execute the following command to invoke n4sidin an interactive
mode:
m = n4sid(ze,1:15); % All orders between 1 and 15. The plot indicates an order of 4 as the best value.
For this choice, let us computetheimpulseresponseofthemodel m:
m = n4sid(ze,4);
showConfidence(impulseplot(m),3)
As with non-parametric impulse response, there is a clear indication that the delay from input to output
is of three samples.
Choosing a Reasonable Model Structure

In lack of any prior knowledge, it is advisable to try out various available choices and use the one that
seems to work the best. State-space models may be a good starting point since only the number of states
needs to be specified in order to estimate a model. Also, a range of orders may be evaluated quickly,
using n4sid, for determining the best order, as described in the next section. For polynomial models, a
similar advantage is realized using the arx estimator. Output-error (OE) models may also be good choice
for a starting polynomial model because of their simplicity.
Determining Model Order

Once you have decided upon a model structure to use, the next task is to determine the order(s). In
general, the aim should be to not use a model order higher than necessary. This can be determined by
analyzing the improvement in %fit as a function of model order. When doing this, it is advisable to use a
separate, independent dataset for validation. Choosing an independent validation data set (zrin our
example) would improve the detection of over-fitting.
In addition to a progressive analysis of multiple model orders, explicit determination of optimum orders
can be performed for some model structures. Functions arxstrucand selstrucmay be used for choosing
the best order for ARX models. For our example, let us check the fit for all 100 combinations of up to 10
b-parameters and up to 10 a-parameters, all with a delay value of 3:
V = arxstruc(ze,zr,struc(1:10,1:10,3));
The best fit for the validation data set is obtained for:
nn = selstruc(V,0)
nn =
1043
Let us check how much the fit is improved for the higher order models. For this, we use the function
selstrucwith only one input. In this case, a plot showing the fit as a function of the number of parameters
used is generated. Theuserisalsopromptedtoenterthenumberofparameters. The routine then selects a
structure with these many parameters that gives the best fit. Note that several different model structures
use the same number of parameters. Execute the following command to choose a model order
interactively:

nns = selstruc(V) %invoke selstruc in an interactive mode The best fit is thus obtained for nn = [4 4 3],
while we see that the improved fit compared to nn = [2 2 3] is rather marginal.
We may also approach this problem from the direction of reducing a higher order model. If the order is
higher than necessary, then the extra parameters are basically used to "model" the measurement noise.
These "extra" poles are estimated with a lower level of accuracy (large confidence interval). If their are
cancelled by a zero located nearby, then it is an indication that this pole-zero pair may not be required to
capture the essential dynamics of the system.
For our example, let us compute a 4th order model: Let us check the pole-zero configuration for this
model. We can also include confidence regions for the poles and zeros corresponding to 3 standard
deviations, in order to determine how accurately they are estimated and also how close the poles and
zeros are to each other.
h = iopzplot(th4); showConfidence(h,3)

The confidence intervals for the two complex-conjugate poles and zeros overlap, indicating they are
likely to cancel each other. Hence, a second order model might be adequate. Based on this evidence, let
us compute a 2nd order ARX model:
We can test how well this model (th2) is capable of reproducing the validation data set. To compare the
simulated output from the two models with the actual output (plotting the mid 200 data points) we use
the compareutility:

compare(zr(150:350),th2,th4)

The plot indicates that there was no significant loss of accuracy in reducing the order from 4 to 2. We
can also check the residuals ("leftovers") of this model, i.e., what is left unexplained by the model.

e = resid(ze,th2);
plot(e(:,1,[])), title('The residuals')

We see that the residuals are quite small compared to the signal level of the output, that they are
reasonably well (although not perfectly) uncorrelated with the input and among themselves. We can thus
be (provisionally) satisfied with the model th2.
Let us now check if we can determine the model order for a state-space structure. As before, we know
the delay is 3 samples. We can try all orders from 1 to 15 with a fixed delay of 3 in n4sid. Execute the
following command to try various orders and choose one interactively.
ms = n4sid(ze,[1:15],'InputDelay',2); %n4sid estimation with variable orders

The default order, indicated in the figure above, is 3, that is in good agreement with our earlier findings.
Finally, we compare how the state-space model ms and the ARX model th2compare in reproducing the
measured output of the validation data:
ms = n4sid(ze,3,'InputDelay',2);
compare(zr,ms,th2)

The comparison plot indicates that the two models are practically identical.
Conclusions

This example described some options for choosing a reasonable model order. Determining delay in
advance can simplify the task of choosing orders. With ARX and state-space structures, we have some
special tools (arxand n4sid estimators) for automatically evaluating a whole set of model orders, and
choosing the best one among them. The information revealed by this exercise (using utilities such as
arxstruc, selstruc, n4sidand delayest)couldbe used as a starting point when estimating models of other
structures, such as BJ and ARMAX.
Additional Information

Formore information on identification of dynamic systems with System Identification Toolbox visit the
System Identification Toolbox product information page.

Frequency Domain Identification: Estimating Models Using Frequency Domain

Data
This example shows how to estimate models using frequency domain data. The estimation and
validation of models using frequency domain data work the same way as they do with time domain data.
This provides a great amount of flexibility in estimation and analysis of models using time and
frequency domain as well as spectral (FRF) data. You may simultaneously estimate models using data in

both domains, compare and combine these models. A model estimated using time domain data may be
validated using spectral data or vice-versa.
Frequency domain data can not be used for estimation or validation of nonlinear models.
Introduction

Frequencydomainexperimentaldataarecommoninmanyapplications. It could be that the data was

collected as frequency response data (frequency functions: FRF) from the process using a frequency
analyzer. It could also be that it is more practical to work with the inputs and outputs Fourier
transforms (FFT of time-domain data), for example to handle periodic or band-limited data. (A bandlimitedcontinuous time signal has no frequency components above the Nyquist frequency). In System
Identification Toolbox, frequency domain I/O data are represented the same way as time-domain data,
i.e., using iddataobjects. The Domain property of the object must be set to Frequency. Frequency
response data are represented as complex vectors or as magnitude/phase vectors as a function of
frequency. IDFRD objects in the toolbox are used to encapsulate FRFs, where a user specifies the
complex response data and a frequency vector. Such IDDATA or IDFRD objects (and also FRD objects
of Control System Toolbox) may be used seamlessly with any estimation routine (such as procest,
tfestetc).
Inspecting Frequency Domain Data

Let us begin by loading some frequency domain data:

load demofr
This MAT-file contains frequency response data at frequencies W,withthe amplitude response AMPand
the phase response PHA. Let us first have a look at the data:
subplot(211), loglog(W,AMP),title('Amplitude Response') subplot(212), semilogx(W,PHA),title('Phase
Response')

This
experimental data will now be stored as an IDFRD object. First transform amplitude and phase to a
complex valued response:
zfr = AMP.*exp(1i*PHA*pi/180);
Ts = 0.1;
gfr = idfrd(zfr,W,Ts);
Ts is the sampling interval of the underlying data. If the data corresponds to continuous time, for
example since the input has been band-limited, use Ts =0.
Note: If you have the Control System Toolbox, you could use an FRD object instead of the IDFRD
object. IDFRD has options for more information, like disturbance spectra and uncertainty measures
which are not available in FRD objects.
The IDFRD object gfrnow contains the data, and it can be plotted and analyzed in different ways. To
view the data, we may use plotor bode:
clf
bode(gfr), legend('gfr')

Estimating Models
Using Frequency Response (FRF) Data

To estimate models, you can now use gfras a data set with all the commands of the toolbox in a
transparent fashion. The only restriction is that noise models cannot be built. This means that for
polynomial models only OE (output-error models) apply, and for state-space models, you have to fixK=
0.
m1 = oe(gfr,[2 2 1]) % Discrete-time Output error (transfer function) model ms = ssest(gfr) %
Continuous-time state-space model with default choice of mproc = procest(gfr,'P2UDZ') % 2nd-order,
continuous-time model with underd compare(gfr,m1,ms,mproc)
m1 =
Discrete-time OE model: y(t) = [B(z)/F(z)]u(t) + e(t) B(z) = 0.9982 z^-1 + 0.4974 z^-2
F(z) = 1 - 1.499 z^-1 + 0.6998 z^-2
Sample time: 0.1 seconds
Parameterization:
Polynomial orders: nb=2 nf=2 nk=1
Number of free coefficients: 4
Use "polydata", "getpvec", "getcov" for parameters and their uncertainti
Status:
Estimated using OE on frequency response data "gfr". Fit to estimation data: 88.04% (prediction focus)
FPE: 0.2501, MSE: 0.2494

ms =
Continuous-time identified state-space model: dx/dt = A x(t) + B u(t) + K e(t) y(t) = C x(t) + D u(t) +
e(t)
A= x1 x2
x1 -0.9005 6.636
x2 -3.308 -2.669
B= u1
x1 -32.9
x2 -28.33
C= x1 x2
y1 -0.5073 0.499
D= u1
y1 0
K= y1
x1 0
x2 0
Parameterization:
FREE form (all coefficients in A, B, C free).
Feedthrough: none
Disturbance component: none
Number of free coefficients: 8
Use "idssdata", "getpvec", "getcov" for parameters and their uncertainti
Status:
Estimated using SSEST on frequency response data "gfr". Fit to estimation data: 73.16% (prediction
focus) FPE: 0.2511, MSE: 1.255
mproc =
Process model with transfer function:
1+Tz*s
G(s) = Kp * ---------------------- * exp(-Td*s) 1+2*Zeta*Tw*s+(Tw*s)^2
Kp = 7.4612 Tw = 0.20246
Zeta = 0.36238 Td = 0
Tz = 0.013617
Parameterization:
'P2DUZ'
Number of free coefficients: 5
Use "getpvec", "getcov" for parameters and their uncertainties. Status:

Estimated using PROCEST on frequency response data "gfr". Fit to estimation data: 73.2% (prediction
focus) FPE: 0.2504, MSE: 1.251

As shown above a variety of linear model types may be estimated in both continuous and discrete time
domains, using spectral data. These models may be validated using, time-domain data. The time-domain
I/O data set ztime, for example, is collected from the same system, and can be used for validation of m1,
msand mproc:

compare(ztime,m1,ms,mproc) %validation in a different domain

We may also look at the residuals to affirm the quality of the model using the validation data ztime. As
observed, the residuals are almost white:
resid(mproc,ztime) % Residuals plot

Condensing Data
Using SPAFDR

An important reason to work with frequency response data is that it is easy to condense the information
with little loss. The command SPAFDR allows you to compute smoothed response data over limited
frequencies, for example with logarithmic spacing. Here is an example where the gfrdata is condensed to
100 logarithmically spaced frequency values. With a similar technique, also
theoriginaltimedomaindatacanbecondensed:
sgfr = spafdr(gfr) % spectral estimation with frequency-dependent resolutio sz = spafdr(ztime); %
spectral estimation using time-domain data clf
bode(gfr,sgfr,sz)
axis([pi/100 10*pi, -272 105])
legend('gfr (raw data)','sgfr','sz','location','southwest')
sgfr =
IDFRD model.
Contains Frequency Response Data for 1 output(s) and 1 input(s), and the sp Response data and
disturbance spectra are available at 200 frequency points
Sample time: 0.1 seconds
Output channels: 'y1'
Input channels: 'u1'
Status:
Estimated using SPAFDR on frequency response data "gfr".

The Bode plots show that the information in the smoothed data has been taken well care of. Now, these
data records with 100 points can very well be used for model estimation. For example:
msm = oe(sgfr,[2 2 1]);
compare(ztime,msm,m1) % msm has the same accuracy as M1 (based on 1000 poin

Estimation Using Frequency-Domain I/O Data

It may be that the measurements are available as Fourier transforms of inputs and output. Such
frequency domain data from the system are given as the signals Y and U. In loglog plots they look like
Wfd = (0:500)'*10*pi/500;
subplot(211),loglog(Wfd,abs(Y)),title('The amplitude of the output')
subplot(212),loglog(Wfd,abs(U)),title('The amplitude of the input')

The frequency
response data is essentially the ratio between Yand U.To collect the frequency domain data as an
IDDATA object, do as follows:
ZFD = iddata(Y,U,'ts',0.1,'Domain','Frequency','Freq',Wfd)
ZFD =
Frequency domain data set with responses at 501 frequencies, ranging from 0 to 31.416 rad/seconds
Sample time: 0.1 seconds
Outputs Unit (if specified)
y1
Inputs Unit (if specified)
u1
Now, again the frequency domain data set ZFDcan be used as data in all estimation routines, just as time
domain data and frequency response data:
mf = ssest(ZFD)
compare(ztime,mf,m1)
mf =
Continuous-time identified state-space model: dx/dt = A x(t) + B u(t) + K e(t) y(t) = C x(t) + D u(t) +
e(t)
A= x1 x2
x1 -1.501 6.791
x2 -3.115 -2.059

B= u1
x1 -28.11
x2 -33.39
C= x1 x2
y1 -0.5844 0.4129
D= u1
y1 0
K= y1
x1 0
x2 0
Parameterization:
FREE form (all coefficients in A, B, C free). Feedthrough: none
Disturbance component: none
Number of free coefficients: 8
Use "idssdata", "getpvec", "getcov" for parameters and their uncertainti
Status:
Estimated using SSEST on frequency domain data "ZFD". Fit to estimation data: 97.21% (prediction
focus) FPE: 0.04263, MSE: 0.04184

Transformations Between Data Representations (Time - Frequency)

Time and frequency domain input-output data sets can be transformed to either domain by using FFT
and IFFT. These commands are adapted to IDDATA objects:
dataf = fft(ztime)
datat = ifft(dataf)
dataf =
Frequency domain data set with responses at 501 frequencies, ranging from 0 to 31.416 rad/seconds
Sample time: 0.1 seconds
Outputs Unit (if specified) y1
Inputs Unit (if specified) u1
datat =
Time domain data set with 1000 samples. Sample time: 0.1 seconds
Outputs Unit (if specified) y1
Inputs Unit (if specified) u1
Time and frequency domain input-output data can be transformed to frequency response data by
SPAFDR, SPA and ETFE:
g1 = spafdr(ztime) g2 = spafdr(ZFD); bode(g1,g2)

g1 =
IDFRD model.
Contains Frequency Response Data for 1 output(s) and 1 input(s), and the sp Response data and
disturbance spectra are available at 100 frequency points
Sample time: 0.1 seconds Output channels: 'y1'
Input channels: 'u1'
Status:
Estimated using SPAFDR on time domain data "ztime".

Frequency
response data can also be transformed to more smoothed data (less resolution and less data) by SPAFDR
and SPA;
g3 = spafdr(gfr);
Frequency response data can be transformed to frequency domain input-output signals by the command
IDDATA:
gfd = iddata(g3)
gfd =
Frequency domain data set with responses at 200 frequencies, ranging from -31.416 to 31.416 rad/
seconds
Sample time: 0.1 seconds
Outputs Unit (if specified) y1
Inputs Unit (if specified) u1
Using Continuous-time Frequency-domain Data to Estimate Continuous-time Models

Time domain data can naturally only be stored and dealt with as discrete-time, sampled data. Frequency
domain data have the advantage that continuous time data can be represented correctly. Suppose that the
underlying continuous time signals have no frequency information above the Nyquist frequency, e.g.
because they are sampled fast, or the input has no frequency component above the Nyquist frequency.
Then the Discrete Fourier transforms (DFT) of the data also are the Fourier transforms of the continuous
time signals, at the chosen frequencies. They can therefore be used to directly fit continuous time
models. In fact, this is the correct way of using band-limited data for model fit.
This will be illustrated by the following example.
Consider the continuous time system:
m0 = idpoly(1,1,1,1,[1 1 1],'ts',0)
m0 =
Continuous-time OE model: y(t) = [B(s)/F(s)]u(t) + e(t) B(s) = 1
F(s) =s^2 +s +1
Parameterization:
Polynomial orders: nb=1 nf=2 nk=0
Number of free coefficients: 3
Use "polydata", "getpvec", "getcov" for parameters and their uncertainti
Status:
Created by direct construction or transformation. Not estimated.
Choose an input with low frequency contents that is fast sampled:
rng(235,'twister');
u = idinput(500,'sine',[0 0.2]); u = iddata([],u,0.1,'intersamp','bl');
0.1 is the sampling interval, and 'bl'indicates that the input is band-limited, i.e. in continuous time it
consists of sinusoids with frequencies below half the sampling frequency. Correct simulation of such a
system should be done in the frequency domain:
uf = fft(u);
uf.ts = 0; % Denoting that the data is continuous time
yf = sim(m0,uf);
%
% Add some noise to the data:
yf.y = yf.y + 0.05*(randn(size(yf.y))+1i*randn(size(yf.y))); dataf = [yf uf] % This is now a continuous
time frequency domain data set.
dataf =
Frequency domain data set with responses at 251 frequencies, ranging from 0 to 31.416 rad/seconds
Sample time: 0 seconds
Outputs Unit (if specified) y1
Inputs Unit (if specified) u1
Look at the data:

plot(dataf)
axis([0 10 0 2.5])

Using dataffor
estimation will by default give continuous time models: State-space:
m4 = ssest(dataf,2); %Second order continuous-time model
For a polynomial model with nb = 2numerator coefficient and nf = 2 estimated denominator coefficients
use:
nb = 2;
nf = 2;
m5 = oe(dataf,[nb nf])
m5 =
Continuous-time OE model: y(t) = [B(s)/F(s)]u(t) + e(t) B(s) = -0.01761 s + 1
F(s) = s^2 + 0.9873 s + 0.9902
Parameterization:
Polynomial orders: nb=2 nf=2 nk=0
Number of free coefficients: 4
Use "polydata", "getpvec", "getcov" for parameters and their uncertainti
Status:
Estimated using OE on frequency domain data "dataf". Fit to estimation data: 70.15% (prediction focus)
FPE: 0.00482, MSE: 0.004723
Compare step responses with uncertainty of the true system m0and the models m4and m5:

clf
h = stepplot(m0,m4,m5);
showConfidence(h,1)
% *Blue: m0, Green: m4, Red: m5.*
% The confidence intervals are shown with patches.

Although it was not necessary in this case, it is generally advised to focus the fit to a limited frequency
band (low pass filter the data) when estimating using continuous time data. The system has a bandwidth
of about 3 rad/s, and was excited by sinusoids up to 6.2 rad/s. A reasonable frequency range
tofocusthefittoisthen[07]rad/s:
m6 = ssest(dataf,2,ssestOptions('Focus',[0 7])) % state space model
m6 =
Continuous-time identified state-space model:
dx/dt = A x(t) + B u(t) + K e(t)
y(t) = C x(t) + D u(t) + e(t)
A=
x1 x2
x1 -0.2714 -1.809
x2 0.4407 -0.7282

B= u1
x1 0.5873
x2 -0.3025
C= x1 x2
y1 0.7565 1.526
D= u1
y1 0
K= y1
x1 0
x2 0
Parameterization:
FREE form (all coefficients in A, B, C free).
Feedthrough: none
Disturbance component: none
Number of free coefficients: 8
Use "idssdata", "getpvec", "getcov" for parameters and their uncertainti
Status:
Estimated using SSEST on frequency domain data "dataf". Fit to estimation data: 87.08% (filter focus)
FPE: 0.004213, MSE: 0.003716
m7 = oe(dataf,[1 2],oeOptions('Focus',[0 7])) % polynomial model of Output
m7 =
Continuous-time OE model: y(t) = [B(s)/F(s)]u(t) + e(t) B(s) = 0.9866
F(s) = s^2 + 0.9791 s + 0.9761
Parameterization:
Polynomial orders: nb=1 nf=2 nk=0
Number of free coefficients: 3
Use "polydata", "getpvec", "getcov" for parameters and their uncertainti
Status:
Estimated using OE on frequency domain data "dataf". Fit to estimation data: 87.04% (filter focus) FPE:
0.004008, MSE: 0.003742
opt = procestOptions('SearchMethod','lsqnonlin','Focus',[0 7]); % Requires m8 =
procest(dataf,'P2UZ',opt) % process model with underdamped poles
m8 =
Process model with transfer function: 1+Tz*s
G(s) = Kp * ---------------------1+2*Zeta*Tw*s+(Tw*s)^2
Kp = 1.0124
Tw = 1.0019

Zeta = 0.5021
Tz = -0.017474
Parameterization:
'P2UZ'
Number of free coefficients: 4
Use "getpvec", "getcov" for parameters and their uncertainties.
Status:
Estimated using PROCEST on frequency domain data "dataf". Fit to estimation data: 87.08% (filter
focus)
FPE: 0.003947, MSE: 0.003716
opt = tfestOptions('SearchMethod','lsqnonlin','Focus',[0 7]); % Requires Op m9 = tfest(dataf,2,opt) %
transfer function with 2 poles
m9 =
From input "u1" to output "y1":
-0.01647 s + 1.003
---------------------s^2 + 0.9948 s + 0.9922
Continuous-time identified transfer function.
Parameterization:
Number of poles: 2 Number of zeros: 1
Number of free coefficients: 4
Use "tfdata", "getpvec", "getcov" for parameters and their uncertainties
Status:
Estimated using TFEST on frequency domain data "dataf". Fit to estimation data: 87.08% (filter focus)
FPE: 0.004067, MSE: 0.003717
h = stepplot(m0,m6,m7,m8,m9); showConfidence(h,1)
legend('show')

Conclusions

We saw how time, frequency and spectral data can seamlessly be used to estimate a variety of linear
models in both continuous and discrete time domains. The models may be validated and compared in
domains different from the ones they were estimated in. The data formats (time, frequency and
spectrum) are interconvertible, using methods such as fft, ifft, spafdr and spa. Furthermore, direct,
continuous-time estimation is achievable by using tfest, ssestand procestestimation routines. The
seamless use of data in any domain for estimation and analysis is an important feature of System
Identification Toolbox.
Additional Information

For more information on identification of dynamic systems with System Identification Toolbox visit the
System Identification Toolbox product information page.

Building Structured and User-Defined Models Using System Identification

Toolbox
This example shows how to estimate parameters in user-defined model structures. Such structures are
specified by IDGREY (linear state-space) or IDNLGREY (nonlinear state-space) models. We shall
investigate how to assign structure, fix parameters and create dependencies among them.
Experiment Data

We shall investigate data produced by a (simulated) dc-motor. We first load the data:
load dcmdata
who

Your variables are:

text u y
The matrix ycontains the two outputs: y1is the angular position of the motor shaft and y2is the angular
velocity. There are 400 data samples and the sampling interval is 0.1 seconds. The input is contained in
the vector u.Itis theinputvoltagetothemotor.
z = iddata(y,u,0.1); % The IDDATA object z.InputName = 'Voltage';
z.OutputName = {'Angle';'AngVel'};

Figure:
Measurement Data: Voltage to Angle

Figure:
Measurement Data: Voltage to Angle
Model Structure Selection

d/dt x= Ax +B u+ K e
y = Cx +D u +e
We shall build a model of the dc-motor. The dynamics of the motor is well known. If we choose x1 as
the angular position and x2 as the angular velocity it is easy to set up a state-space model of the
following character neglecting disturbances: (see Example 4.1 in Ljung(1999):
|0 1 | | 0 |
d/dt x = | | x + | | u
| 0 -th1 | | th2 |
|1 0 |
y= | |x |0 1|
The parameter th1is here the inverse time-constant of the motor and th2 is such that th2/th1is the static
gain from input to the angular velocity. (See Ljung(1987) for how th1and th2relate to the physical
parameters of the motor). We shall estimate these two parameters from the observed data. The model
structure (parameterized state space) described above can be represented in MATLAB using IDSS and
IDGREY models. These models let you perform estimation of parameters using experimental data.
Specification of a Nominal (Initial) Model

If we guess that th1=1 and th2 = 0.28 we obtain the nominal or initial model

A = [0 1; 0 -1]; %initial guess for A(2,2) is -1 B = [0; 0.28]; %initial guess for B(2) is 0.28 C = eye(2);
D = zeros(2,1);
and we package this into an IDSS model object:
ms = idss(A,B,C,D);
The model is characterized by its matrices, their values, which elements are free (to be estimated) and
upper and lower limits of those:
ms.Structure.a
ans =
Name: 'a'
Value: [2x2 double] Minimum: [2x2 double] Maximum: [2x2 double]
Free: [2x2 logical] Scale: [2x2 double]
Info: [2x2 struct] 1x1 param.Continuous
ms.Structure.a.Value ms.Structure.a.Free
ans =
01 0-1
ans =
11 11
Specification of Free (Independent) Parameters Using IDSS Models

So we should now mark that it is only A(2,2) and B(2,1) that are free parameters to be estimated.
ms.Structure.a.Free = [0 0; 0 1];
ms.Structure.b.Free = [0; 1];
ms.Structure.c.Free = 0; % scalar expansion used ms.Structure.d.Free = 0;
set(ms,'Ts',0); % This defines the model to be continuous
The Initial Model

ms % Initial model
ms =
Continuous-time identified state-space model: dx/dt = A x(t) + B u(t) + K e(t) y(t) = C x(t) + D u(t) +
e(t)
A= x1 x2
x101
x2 0 -1
B= u1
x1 0
x2 0.28
C= x1 x2
y110
y201

D= u1
y1 0
y2 0
K= y1 y2
x100
x200
Parameterization:
STRUCTURED form (some fixed coefficients in A, B, C).
Feedthrough: none
Disturbance component: none
Number of free coefficients: 2
Use "idssdata", "getpvec", "getcov" for parameters and their uncertainti
Status:
Created by direct construction or transformation. Not estimated.
Estimation of Free Parameters of the IDSS Model

The prediction error (maximum likelihood) estimate of the parameters is now computed by:
dcmodel = ssest(z,ms,ssestOptions('Display','on')); dcmodel
dcmodel =
Continuous-time identified state-space model: dx/dt = A x(t) + B u(t) + K e(t) y(t) = C x(t) + D u(t) +
e(t)
A= x1 x2
x101
x2 0 -4.013
B= Voltage x1 0 x2 1.002
C= x1 x2
Angle 1 0
AngVel 0 1
D= Voltage Angle 0 AngVel 0
K= Angle AngVel x100 x200
Parameterization:
STRUCTURED form (some fixed coefficients in A, B, C).
Feedthrough: none
Disturbance component: none
Number of free coefficients: 2
Use "idssdata", "getpvec", "getcov" for parameters and their uncertainti Status:
Estimated using SSEST on time domain data "z". Fit to estimation data: [98.35;84.42]% (prediction
focus) FPE: 0.001071, MSE: 0.05967

The estimated values of the parameters are quite close to those used when
thedataweresimulated(-4and1). Toevaluatethemodelsqualitywecan simulate the model with the actual
input by and compare it with the actual output.
compare(z,dcmodel);

We can now, for example plot zeros and poles and their uncertainty regions. We will draw the regions
corresponding to 3 standard deviations, since the model is quite accurate. Note that the pole at the origin
is absolutely certain, since it is part of the model structure; the integrator from angular velocity to
position.

clf
showConfidence(iopzplot(dcmodel),3)

Now, we may make various modifications. The 1,2-element of the A-matrix (fixed to 1) tells us that x2is
the derivative of x1. Suppose that the sensors are not calibrated, so that there may be an unknown
proportionality constant. To include the estimation of such a constant we just "let loose" A(1,2)and reestimate:
dcmodel2 = dcmodel;
dcmodel2.Structure.a.Free(1,2) = 1;
dcmodel2 = pem(z,dcmodel2,ssestOptions('Display','on'));
The resulting model is
dcmodel2
dcmodel2 =
Continuous-time identified state-space model: dx/dt = A x(t) + B u(t) + K e(t) y(t) = C x(t) + D u(t) +
e(t)
A= x1 x2
x1 0 0.9975
x2 0 -4.011
B= Voltage x1 0 x2 1.004

C= x1 x2
Angle 1 0
AngVel 0 1
D= Voltage Angle 0 AngVel 0
K= Angle AngVel x100 x200
Parameterization:
STRUCTURED form (some fixed coefficients in A, B, C).
Feedthrough: none
Disturbance component: none
Number of free coefficients: 3
Use "idssdata", "getpvec", "getcov" for parameters and their uncertainti
Status:
Estimated using SSEST on time domain data "z". Fit to estimation data: [98.35;84.42]% (prediction
focus) FPE: 0.001076, MSE: 0.05966
We find that the estimated A(1,2)iscloseto1. Tocomparethetwomodelwe use the comparecommand:
compare(z,dcmodel,dcmodel2)

Specification of Models with Coupled Parameters Using IDGREY Objects

Suppose that we accurately know the static gain of the dc-motor (from input voltage to angular velocity,
e.g. from a previous step-response experiment. If the static gain is G, and the time constant of the motor
is t, then the state-space model becomes
|0 1| | 0 |
d/dt x= | |x+ | |u
|0 -1/t| | G/t |
|1 0|
y=| |x |0 1|
With Gknown, there is a dependence between the entries in the different matrices. In order to describe
that, the earlier used way with "Free" parameters will not be sufficient. We thus have to write a
MATLAB file which produces the A, B, C, and D, and optionally also the K and X0 matrices as outputs,
for each given parameter vector as input. It also takes auxiliary arguments as inputs, so that the user can
change certain things in the model structure, without having to edit the file. In this case we let the known
static gain Gbe entered as such an argument. The file that has been written has thenamemotor.m.
type motor
function [A,B,C,D,K,X0] = motor(par,ts,aux)
%MOTOR ODE file representing the dynamics of a motor.
%
% [A,B,C,D,K,X0] = MOTOR(Tau,Ts,G)
% returns the State Space matrices of the DC-motor with
% time-constant Tau (Tau = par) and known static gain G. The sample % time is Ts.
%
% This file returns continuous-time representation if input argument Ts % is zero. If Ts>0, a discretetime representation is returned. To make % the IDGREY model that uses this file aware of this
flexibility, set the % value of Structure.FcnType property to 'cd'. This flexibility is useful % for
conversion between continuous and discrete domains required for % estimation and simulation.
%
% See also IDGREY, IDDEMO7.
% L. Ljung
% Copyright 1986-2011 The MathWorks, Inc.
% $Revision: 1.1.10.2 $ $Date: 2011/12/22 18:22:34 $
t = par(1);
G = aux(1); A = [0 1;0 -1/t];
B = [0;G/t];
C = eye(2);
D = [0;0];
K = zeros(2);
X0 = [0;0];
if ts>0 % Sample the model with sample time Ts

s = expm([[A B]*ts; zeros(1,3)]);

A = s(1:2,1:2);
B = s(1:2,3);
end
We now create an IDGREY model object corresponding to this model structure: The assumed time
constant will be
par_guess = 1;
We also give the value 0.25 to the auxiliary variable G(gain) and sampling interval.
aux = 0.25;
dcmm = idgrey('motor',par_guess,'cd',aux,0);
The time constant is now estimated by
dcmm = greyest(z,dcmm,greyestOptions('Display','on'));
We have thus now estimated the time constant of the motor directly. Its value is in good agreement with
the previous estimate.
dcmm
dcmm =
Continuous-time linear grey box model defined by "motor" function: dx/dt = A x(t) + B u(t) + K e(t)
y(t) = C x(t) + D u(t) + e(t)
A= x1 x2
x101
x2 0 -4.006
B= Voltage x1 0 x2 1.001
C= x1 x2
Angle 1 0
AngVel 0 1
D= Voltage Angle 0 AngVel 0
K= Angle AngVel x100 x200
Model parameters: Par1 = 0.2496
Parameterization:
ODE Function: motor
(parameterizes both continuous- and discrete-time equations) Disturbance component: parameterized by
the ODE function Initial state: parameterized by the ODE function
Number of free coefficients: 1
Use "getpvec", "getcov" for parameters and their uncertainties.
Status:
Estimated using GREYEST on time domain data "z". Fit to estimation data: [98.35;84.42]% (prediction
focus) FPE: 0.00107, MSE: 0.05971

With this model we can now proceed to test various aspects as before. The syntax of all the commands is
identical to the previous case. For example, we can compare the idgrey model with the other state-space
model:
compare(z,dcmm,dcmodel)

They are clearly very close.

Estimating Multivariate ARX Models

The state-space part of the toolbox also handles multivariate (several outputs) ARX models. By a
multivariate ARX-model we mean the following:
A(q) y(t) = B(q) u(t) + e(t)
Here A(q) is a ny | ny matrix whose entries are polynomials in the delay operator 1/q. The k-l element is
denoted by:
where:
It is thus a polynomial in 1/qof degree nakl.
Similarly B(q) is a ny | nu matrix, whose kj-element is:

Thereisthusadelayof nkkjfrom input number jto output number k.The

mostcommonwaytocreatethosewouldbetousetheARX-command. The orders are specified as: nn = [na nb
nk]with nabeing a ny-by-nymatrix whose kj-entry is nakj; nband nkare defined similarly.
Lets test some ARX-models on the dc-data. First we could simply build a general second order model:
dcarx1 = arx(z,'na',[2,2;2,2],'nb',[2;2],'nk',[1;1])
dcarx1 =

Discrete-time ARX model:

Model for output "Angle": A(z)y_1(t) = - A_i(z)y_i(t) + B(z)u(t) + e_1(t) A(z) = 1 - 0.5545 z^-1 0.4454 z^-2
A_2(z) = -0.03548 z^-1 - 0.06405 z^-2
B(z) = 0.004243 z^-1 + 0.006589 z^-2
Model for output "AngVel": A(z)y_2(t) = - A_i(z)y_i(t) + B(z)u(t) + e_2(t A(z) = 1 - 0.2005 z^-1 0.2924 z^-2
A_1(z) = 0.01849 z^-1 - 0.01937 z^-2
B(z) = 0.08642 z^-1 + 0.03877 z^-2
Sample time: 0.1 seconds
Parameterization: Polynomial orders: na=[2 2;2 2] nb=[2;2] nk=[1;1]
Number of free coefficients: 12
Use "polydata", "getpvec", "getcov" for parameters and their uncertainti
Status:
Estimated using ARX on time domain data "z".
Fit to estimation data: [97.87;83.44]% (prediction focus) FPE: 0.002197, MSE: 0.06999
The result, dcarx1, is stored as an IDPOLY model, and all previous commands apply. We could for
example explicitly list the ARX-polynomials by:
dcarx1.a
ans =
[1x3 double] [1x3 double] [1x3 double] [1x3 double]
as cell arrays where e.g. the {1,2} element of dcarx1.a is the polynomial A(1,2) described earlier,
relating y2 to y1.
We could also test a structure, where we know that y1is obtained by filtering y2through a first order
filter. (The angle is the integral of the angular velocity). We could then also postulate a first order
dynamics from input to output number 2:
na =[11;0 1];
nb =[0; 1];
nk =[1; 1];
dcarx2 = arx(z,[na nb nk])
dcarx2 =
Discrete-time ARX model:
Model for output "Angle": A(z)y_1(t) = - A_i(z)y_i(t) + B(z)u(t) + e_1(t) A(z) = 1 - 0.9992 z^-1
A_2(z) = -0.09595 z^-1
B(z) = 0
Model for output "AngVel": A(z)y_2(t) = B(z)u(t) + e_2(t) A(z) = 1 - 0.6254 z^-1
B(z) = 0.08973 z^-1
Sample time: 0.1 seconds

Parameterization:
Polynomial orders: na=[1 1;0 1] nb=[0;1] nk=[1;1]
Number of free coefficients: 4
Use "polydata", "getpvec", "getcov" for parameters and their uncertainti
Status:
Estimated using ARX on time domain data "z".
Fit to estimation data: [97.52;81.46]% (prediction focus) FPE: 0.003468, MSE: 0.08862
To compare the different models obtained we use
compare(z,dcmodel,dcmm,dcarx2)

Finally, we could compare the bodeplots obtained from the input to output one for the different models
by using bode:Firstoutput:
dcmm2 = idss(dcmm); % convert to IDSS for subreferencing Second output:
The two first models are more or less in exact agreement. The ARX-models are not so good, due to the
bias caused by the non-white equation error noise. (We had white measurement noise in the
simulations).
Conclusions

Estimation of models with pre-selected structures can be performed using System Identification toolbox.
In state-space form, parameters may be fixed to their known values or constrained to lie within a

prescribed range. If relationship between parameters or other constraints need to be specified, IDGREY
objects may be used. IDGREY models evaluate a user-specified MATLAB file for estimating statespace system parameters. Multi-variate ARX models offer another option for quickly estimating multioutput models with user-specified structure.
Additional Information

For more information on identification of dynamic systems with System Identification Toolbox visit the
System Identification Toolbox product information page.