2018 UKACC 12th International Conference on Control (CONTROL)

Sheffield, UK, 5-7 Sept 2018

Linear System Identification of Longitudinal Vehicle

Dynamics Versus Nonlinear Physical Modelling
Sebastian James Sean R. Anderson
Dept of Automatic Control Dept of Automatic Control
and Systems Engineering and Systems Engineering
University of Sheffield University of Sheffield
Sheffield, UK, S1 3JD Sheffield, UK, S1 3JD
Email: [email protected] Email: [email protected]

Abstract—Mathematical modelling of vehicle dynamics is es- physically derived model and assess to what extent nonlinear
sential for the development of autonomous cars. Many of the modelling is necessary for vehicle control in normal driving.
vehicle models that are used for control design in cars are based To identify the model of longitudinal vehicle dynamics, we
on nonlinear physical models. However, it is not clear, especially
for the case of longitudinal dynamics, whether such nonlinear used experimental data combined with linear system identi-
models are necessary or simpler models can be used. In this fication techniques. The experimental data were collected in
paper, we identify a linear data-driven model of longitudinal a Lancia Delta car which was driven along a 53 km route
vehicle dynamics and compare it to a nonlinear physically derived on public roads in normal conditions. We used a prediction
model. The linear model was identified in continuous-time state- error method (PEM) to directly identify the linear model in
space form using a prediction error method. The identification
data were obtained from a Lancia Delta car, over 53 km of continuous-time state-space form [10], with model initialisa-
normal driving on public roads. The selected linear model was tion using subspace state-space system identification (N4SID)
first order with requested torque, brake and road gradient as [11]. We used a constrained nonlinear optimisation routine to
inputs and car velocity as output. The key results were that 1. estimate parameters for a nonlinear physical model based on
the linear model was accurate, with a variance accounted for the longitudinal component of the well-known bicycle model
(VAF) metric of VAF=96.5%, and 2. the identified linear model
was also superior in accuracy to the nonlinear physical model, of vehicle dynamics [12].
VAF=77.4%. The implication of these results, therefore, is that The results indicated that the linear model was comparable
for longitudinal dynamics, in normal driving conditions, a first in accuracy to the nonlinear physical model and yet much
order linear model is sufficient to describe the vehicle dynamics. simpler and more attractive for control design.
This is advantageous for control design, state estimation and
real-time implementation, e.g. in predictive control. II. M ETHODS
I. I NTRODUCTION A. Experimental data
Mathematical models of car vehicle dynamics are essen- The experimental data used in modelling and identification
tial for the development of future driverless cars and driver were collected in a Lancia Delta car which was driven along a
assistance systems [1]–[4]. Typically, the development of 53 km route on public roads in normal conditions. The route
longitudinal and lateral vehicle control algorithms have been was chosen to incorporate a typical selection of motorways,
based on physically derived models, either fully nonlinear [5] extra-urban and urban roads, roundabouts and intersections.
or based on linearised time-varying models [6], [7]. These non- The route began at Centro Ricerche Fiat in Orbassano, near
linear models lead to relatively complicated control and state Turin (Italy), then went to Pinerolo via Piossasco before
estimation schemes, which are both challenging to implement returning to Orbassano, a distance of 53 km driven in about
in real-time and also challenging to analyse for stability. 42 minutes, and is the same data as described in [3].
Given that firstly, linear feedback control schemes are rela- The following signals were amongst those recorded during
tively tolerant of plant model error (e.g. due to nonlinearities), the journey: longitudinal velocity, accelerator pedal position,
and secondly, nonlinearities in vehicles are relatively weak brake pedal position, selected gear, engine torque and GPS
(although become important at high-g [8]), it would appear co-ordinates. From the GPS co-ordinates, the road elevations
attractive to develop linear models of vehicle dynamics for were obtained (using the Google Maps API), providing the
control design. There are examples of linear models of vehicle approximate road gradient. Signals were sampled at 20 Hz for
dynamics identified from sampled data [9]. However, the the purpose of this modelling study. The elevation and road
relative utility of linear vehicle models have not yet been well gradient signals were smoothed with a third order Butterworth
investigated and compared to nonlinear physically-derived low-pass filter with a cut-off frequency of 0.5 Hz.
models. In this study we used a single circumnavigation of the route,
The novel aim of this paper is to develop a linear model which provided about 1 hour and 20 minutes of data from
of longitudinal vehicle dynamics, compare it to a nonlinear which to parameterise and identify the models. We subdivided

978-1-5386-2864-5/18/$31.00 ©2018 IEEE 146

Authorized licensed use limited to: Purdue University. Downloaded on February 18,2025 at 15:43:31 UTC from IEEE Xplore. Restrictions apply.
Fig. 1. Test route of the Lancia Delta vehicle from the experiment conducted near Turin, Italy over 53 km of public roads in normal driving conditions,
and example data over 100 seconds. The plot of road gradient shows the raw data and the smoothed data obtained from a zero-phase low-pass 3rd order
Butterworth filter.

this into two sections; one of 2000 s duration for model velocity near 0 m/s, care must be taken to ensure that these
identification and another of 1000 s for validation. Fig. 1 frictional terms always act in the opposite direction to v.
shows the control and output for part of the training data as While it is possible to model (for each gear) the relationship
an example. between the position of the accelerator pedal and Fp , the
engine control unit (ECU) provides engine torque signals
B. Physical model of longitudinal vehicle dynamics which can be used to compute the motive force. The ECU
We employ a well known simplification of the physical provides a requested engine torque signal, Ta , which is a
dynamics of a four wheeled vehicle—that the lateral dynamics scaled (0 to 400 Nm) version of the accelerator pedal position.
and the longitudinal dynamics may be decoupled [12]. The When the requested engine torque is scaled by the gearbox
longitudinal model may then be expressed by its force equa- ratio, R, we have a post-gearbox requested torque, Tr = Ta .R.
tion, The ECU also provides the delivered engine torque, Td , which
may be lower than the requested torque if the engine is
M v̇ = Fp (t) + Fb (t) − M g sin(γ(t)) − kD v 2 − M g kR (1)
unable to comply with the request or if the engine torque is
where M is the mass of the vehicle and v is the vehicle’s being controlled by the automatic gearbox during a gear shift.
longitudinal velocity, which we will solve for. Fp (t) is the Additionally, the ECU reports the engine-speed dependent
collective propulsive force due to the engine actuating through torque due to friction in the engine, Tf . The delivered engine-
all of the drive wheels. Fb (t) is the collective braking force. shaft torque, Te , is therefore given by
g is the acceleration due to gravity and γ(t) is the gradient
Te = Td − Tf (2)
of the road. kD is the parameter governing the strength of air
resistance; kR is the static friction parameter. If operating with For control system design, the use of the requested torque,

978-1-5386-2864-5/18/$31.00 ©2018 IEEE 147

Authorized licensed use limited to: Purdue University. Downloaded on February 18,2025 at 15:43:31 UTC from IEEE Xplore. Restrictions apply.
Ta , and its post-gearbox partner, Tr , is most relevant, as the simulation and hence affecting the fit metrics. For this reason,
control system may not have access to the engine-friction after finding a first set of parameter values, we reduced the
signal or the ability to predict when the engine is unable parameter ranges to kD ∈ [0.1, 3], kR ∈ [0.0001, 0.04] and
to deliver the requested torque. However, the most accurate kτ ∈ [5, 15] and re-ran the parameter search using ode23 and
physical model against which to compare alternative models 30 different parameter sets as a start point. ode23 was used
will be given by making use of the engine-shaft torque, when computing all simulations shown below and to generate
Te , along with the gear ratio information. In our physical the reported fit metrics.
modelling, we therefore have Fp given by
C. Linear state-space system identification
Te (t) R[G(t)] ηg Rd ηd
Fp (t) = (3) For system identification, we used a linear continuous-
rw time state-space model to represent the longitudinal vehicle
where Te (t) is the delivered engine-shaft torque, R[G(t)] is dynamics:
the gearbox ratio as a function of the gear, G(t), and Rd is
ẋ(t) = Ax(t) + Bu(t) (7)
the driveline gear ratio. ηg and ηd model the frictional loss
in the gearbox and driveline, respectively and rw is the wheel y(t) = Cx(t) + Du(t) (8)
radius. R[G(t)] is provided by the ECU as a signal, which we where the output y(t) ∈ R is the vehicle velocity at time
combine with Te to give the pre-loss gear-shaft torque, Tg : t; the input u(t) ∈ Rnu is composed of the delivered post-
Tg (t) = Te (t) R[G(t)] (4) gearbox engine torque (Td ) and the brake pressure signal (pb ),
and in some versions of the model also the smoothed road
We collect the terms Rd (unknown), ηd and ηg (unknown; gradient (γ(t)); x(t) ∈ Rnx is the vehicle state vector, where
approximately 0.93) and rw (known; 0.292 m) together into a nx is the model order, which is determined as part of the
single parameter, kτ , so that Fp is, finally, identification procedure. The matrices A, B, C and D are
assumed fully parameterised here and comprise the unknown
Fp (t) = kτ Tg (t) (5)
model parameters.
with the value of kτ to be found by the parameter search. The parameters of the state-space model were estimated
The collective braking force, Fb (t), is governed by the directly in continuous-time using a prediction error method
relationship, (PEM) [10]. Assume all parameters in A, B, C and D are
 collected in the parameter vector θ, so that the estimation
µb M g
 v > 0, kb pb (t) > µb M g problem is defined as
Fb (t) = kb pb (t) v>0 (6)
N

0 otherwise 1 X
e(tk , θ)2

θ̂ = arg min (9)
θ N
k=1
where kb is a known coefficient of braking force in Newtons
per bar of brake pressure (189 N/bar) and pb (t) is the time- where the prediction error is e(tk , θ) = y(tk ) − ŷ(tk , θ) and
varying master brake cylinder pressure. The upper limit on the ŷ(tk , θ) is the simulated output of the state-space model at
available braking force, µb M g, is the force at which the tyres sample-time tk (N is the number of data samples). The PEM
slip along the road surface; the coefficient µb is road-surface is typically solved using numerical search [10].
dependent. Because the data we use here were collected for The numerical search algorithm can be initialised, as here,
normal driving, with no sharp or emergency braking taking using a subspace state-space system identification (N4SID)
place, we set µb to a high, fixed value for all time. algorithm [11]. The N4SID algorithm requires a number of
To estimate the parameters we used the training data and choices (insight into these choices is given in [15]), such
a nonlinear interior-point optimisation method [13], imple- as the forward prediction horizon and the number of past
mented in the MATLAB function fmincon. The parameters inputs and outputs that are used for the prediction - here
were initialised using a multi-start approach, with 100 different these were chosen by estimating multiple models over a range
parameter sets, randomly sampled in the ranges kD ∈ [0, 20], of values and using Akaike’s information criterion (AIC) to
kR ∈ [0, 0.05] and kτ ∈ [1, 100]. We found the choice of ODE select between them. The weighting scheme used in the N4SID
solver to be important in this work. MATLAB’s ode23 solver algorithm was canonical variate analysis (CVA) [16].
[14] operated well for the ‘best’ parameters—those which Parameter estimation was performed in MATLAB using the
would provide a good fit to the data. However, this solver System Identification Toolbox, using the function ssest, which
became computationally inefficient towards the boundaries of conveniently combines the N4SID initialisation of the state-
the chosen parameter ranges. For this reason, to allow the space model with the PEM estimation stage in one function.
parameter search to complete quickly, the ode23s solver [14] Key user-choices for the estimation algorithms are summarised
was employed, which is more effective in computing stiff in Table I. The ssest function, by default, uses a combination
differential equation systems. Unfortunately, ode23s produced of search methods in the PEM stage including Gauss-Newton
occasional, random, quickly corrected deviations from the (GN), adaptive Gauss-Newton (AGN), Levenberg-Marquardt
putative true solution of the system, introducing noise into the (LM) and gradient descent (GD). Also, note for initialisation

978-1-5386-2864-5/18/$31.00 ©2018 IEEE 148

Authorized licensed use limited to: Purdue University. Downloaded on February 18,2025 at 15:43:31 UTC from IEEE Xplore. Restrictions apply.
Fig. 2. Model selection for the linear system identification model, where the model index is as follows: t1o - torque only, first order; tb1o - torque-brake
input, first order; tbg1o - torque-brake-gradient input, first order; t2o - torque only, second order; tb2o - torque-brake input, second order; tbg2o - torque-
brake-gradient input, second order; (a)-(d) Model selection plots, which indicate that the models with torque-brake-gradient inputs should be preferred. Note
that FPE can be misleading because it is based on one-step-ahead prediction errors, unlike the other measures. (e)-(f) Residual analysis plots, which indicate
that the first and second order models have similar auto-correlation in the model residuals and in the gradient input-residual cross-correlation.

that the N4SID identification was performed in discrete-time, TABLE I

then the model was mapped to continuous-time using a zero- S TATE -S PACE S YSTEM I DENTIFICATION A LGORITHM PARAMETERS
order hold. Algorithm Option Method
To select the model order, nx , model orders nx = 1, . . . , 10 Parameter Initialisation N4SID
were initially tested using an analysis of singular values of Initial State Estimated
the input-output covariance matrix [10], which suggested only N4SID Weighting Scheme CVA
models of first or second order should be further investi- N4SID Prediction Horizon Chosen by AIC
gated. First and second order models were then analysed and Loss Function Simulation Error
compared with more in-depth methods, including Model Fit, Loss Function Weighting None
Variance Accounted For (VAF), and residual analysis (see Search Method GN/AGN/LM/GD
section below on model evaluation and validation).
To identify the model, the training data was used (com- D. Model evaluation and validation methods
prising 2000 seconds of normal driving data), which was
consistent with the parameter estimation for the nonlinear Models were evaluated using a normalised fit metric based
physical modelling. A separate validation data set was used on the Euclidean norm of the fit error, where
to validate the model (comprising a further 1000 seconds of
 
||y − ŷ||2
normal driving data). Model Fit = 100 1 − (10)
||y − ȳ||2

978-1-5386-2864-5/18/$31.00 ©2018 IEEE 149

Authorized licensed use limited to: Purdue University. Downloaded on February 18,2025 at 15:43:31 UTC from IEEE Xplore. Restrictions apply.
Fig. 3. Model simulations vs training and validation data. Top two rows: nonlinear physical model (orange) vs observed data (black). Bottom two rows: linear
first order state-space model (blue) vs observed data (black). The residual error is shown in grey in each plot.

where y is the vector of measured output data, ŷ is the vector III. R ESULTS AND D ISCUSSION
of model simulation outputs, and ȳ is the mean of measured
output data. A value of 100% indicates a perfect model fit, The identification of the linear model first required the
a value of 0% indicates a fit equivalent to the mean of the selection of model order. This was done with N4SID for orders
output data, and becomes negative for poor fits. nx = 1, . . . , 10, using the singular values of the input-output
The model fit was also assessed using the variance ac- covariance matrix, which indicated that a first, second or third
counted for (VAF) metric, which is equivalent to the r2 metric, order model should be chosen (results not shown). The first,
defined as   second and third order models were then evaluated with three
var(y − ŷ)
VAF = 100 1 − (11) combinations of input: torque-only, torque-brake and torque-
var(y) brake-gradient. The model evaluation methods indicated that
To evaluate the models in terms of accuracy and complexity, inputs should consist of torque-brake-gradient as these models
Akaike’s final prediction error (FPE) was used, had the best fits to the data [Fig. 2(a)–(d))].
N   Selecting between the first or second order models requires
1 X 1 + np /N
FPE = e1 (tk , θ̂)2 × (12) analysis of the model residuals and consideration of the model
N 1 − np /N
k=1 complexity. Auto- and cross-correlation residues for the first
where np is the number of model parameters and e1 (.) denotes order torque-brake-gradient model are shown in Fig. 2(e).
the one-step-ahead prediction error. These appear slightly larger than those for the second order
Residual analysis was also used in model validation, by model in Fig. 2(f). On this basis we might choose the second
analysing the auto-correlation of the residual error and the order model as the optimum representation of the system,
cross-correlation of the input and residual error [10]. although as there is not a stark difference in the fit quality

978-1-5386-2864-5/18/$31.00 ©2018 IEEE 150

Authorized licensed use limited to: Purdue University. Downloaded on February 18,2025 at 15:43:31 UTC from IEEE Xplore. Restrictions apply.
or the residual analysis, the first order model is preferred for model (VAF=77.4%). Therefore, the conclusion we draw from
its simplicity. this study is that linear models should be investigated as an
The linear first order model with torque-brake-gradient alternative to nonlinear physical models in longitudinal vehicle
inputs (u1, u2, u3) was identified as control.
 
A = −0.0008036 (13) ACKNOWLEDGMENT
 
B = 0.0000016 −0.0000334 −0.0027613 (14) The authors would like to thank the EU for funding support

C = 3499.522

(15) through grant number 731593 (Dreams4Cars), as well as A.
  Saroldi and Centro Ricerche Fiat for their contribution in
D= 0 0 0 (16) collecting the experimental data and Francesco Biral for useful
and the linear second order model with torque-brake-gradient discussion.
inputs (u1, u2, u3) was identified as R EFERENCES
 
−0.2916319 0.0875774 [1] B. Paden, M. p, S. Z. Yong, D. Yershov, and E. Frazzoli, “A Survey
A= (17)
−0.7851958 0.2345142 of Motion Planning and Control Techniques for Self-Driving Urban
  Vehicles,” IEEE Trans. Intell. Veh., vol. 1, no. 1, pp. 33–55, Mar. 2016.
−0.0000108 0.0003421 0.0042201 [2] M. Da Lio, F. Biral, E. Bertolazzi, M. Galvani, P. Bosetti, D. Windridge,
B= (18)
−0.0000325 0.0009946 0.0167971 A. Saroldi, and F. Tango, “Artificial co-drivers as a universal enabling
  technology for future intelligent vehicles and transportation systems,”
C = 3313.131 −1278.718 (19) IEEE Trans. Intell. Transp. Syst., vol. 16, no. 1, pp. 244–263, 2015.
  [3] A. Bisoffi, F. Biral, M. D. Lio, and L. Zaccarian, “Longitudinal Jerk Es-
D= 0 0 0 (20) timation of Driver Intentions for Advanced Driver Assistance Systems,”
IEEE/ASME Trans. Mechatronics, vol. 22, no. 4, pp. 1531–1541, Aug.
The nonlinear physical model parameters were estimated as 2017.
kD = 0.215, kR = 0.0214 and kτ = 12.41. [4] R. Hult, F. E. Sancar, M. Jalalmaab, A. Vijayan, A. Severinson,
Simulation comparison of the nonlinear physical model and M. Di Vaio, P. Falcone, B. Fidan, and S. Santini, “Design and exper-
imental validation of a cooperative driving control architecture for the
linear first order model to each other and the observed data grand cooperative driving challenge 2016,” IEEE Trans. Intell. Transp.
demonstrated that firstly both models were accurate predictors Syst., vol. 19, no. 4, pp. 1290–1301, 2018.
of car velocity, and that secondly the linear model was superior [5] P. Falcone, F. Borrelli, J. Asgari, H. E. Tseng, and D. Hrovat, “Predictive
Active Steering Control for Autonomous Vehicle Systems,” IEEE Trans.
in accuracy to the nonlinear model (Fig. 3). Control Syst. Technol., vol. 15, no. 3, pp. 566–580, May 2007.
The overall VAF for the linear first order model was [6] P. Falcone, F. Borrelli, H. E. Tseng, J. Asgari, and D. Hrovat, “Linear
96.5% and for the nonlinear physical model was 77.4%. There- time-varying model predictive control and its application to active
steering systems: Stability analysis and experimental validation,” Inter-
fore, the linear state-space identified model outperformed the national Journal of Robust and Nonlinear Control, vol. 18, no. 8, pp.
nonlinear model. More importantly, the accurate performance 862–875, May 2008.
of the linear model is combined with the additional benefit of [7] V. Turri, A. Carvalho, H. E. Tseng, K. H. Johansson, and F. Borrelli,
“Linear model predictive control for lane keeping and obstacle avoidance
model simplicity, advantageous in tasks such as control and on low curvature roads,” in 16th International IEEE Conference on
state-estimation. Intelligent Transportation Systems (ITSC 2013), Oct. 2013, pp. 378–
383.
There are few modelling studies that identify car vehicle [8] D. E. Smith and J. M. Starkey, “Effects of Model Complexity on
dynamics from sampled data but one investigation that uses the Performance of Automated Vehicle Steering Controllers: Model
step responses found that a linear second order model was Development, Validation and Comparison,” Vehicle System Dynamics,
vol. 24, no. 2, pp. 163–181, Mar. 1995.
sufficient to accurately describe the car dynamics during single [9] V. Milanés, S. E. Shladover, J. Spring, C. Nowakowski, H. Kawazoe,
maneouvers of about 10 seconds duration [9]. That modelling and M. Nakamura, “Cooperative adaptive cruise control in real traffic
study therefore agrees with the investigation here that linear situations,” IEEE Trans. Intell. Transp. Syst., vol. 15, no. 1, pp. 296–305,
2014.
models can be sufficient to describe vehicle dynamics. [10] L. Ljung, System Identification - Theory for the User. Upper Saddle
A limitation of this investigation is that it only focuses on River, NJ: Prentice Hall, 1999.
longitudinal vehicle dynamics. The lateral dynamics might [11] P. Van Overschee and B. De Moor, Subspace Identification for Linear
Systems: Theory, Implementation, Applications. Norwell, MA: Kluwer
require nonlinear modelling to adequately describe the car Academic Publishers, 1996.
behaviour for full path control. In future work, therefore, we [12] R. Rajamani, Vehicle dynamics and control. Springer Science &
plan to extend the comparison of physical and linear identified Business Media, 2011.
[13] R. A. Waltz, J. L. Morales, J. Nocedal, and D. Orban, “An interior
models to the coupled longitudinal-lateral vehicle dynamics. algorithm for nonlinear optimization that combines line search and trust
region steps,” Mathematical Programming, vol. 107, no. 3, pp. 391–408,
IV. C ONCLUSION 2006.
[14] L. Shampine and M. Reichelt, “The MATLAB ODE Suite,” SIAM
In this paper, we have compared a nonlinear physical model Journal on Scientific Computing, vol. 18, no. 1, pp. 1–22, Jan. 1997.
of longitudinal car vehicle dynamics against a linear model [15] L. Ljung, “Aspects and experiences of user choices in subspace identifi-
obtained by system identification techniques. Experimental cation methods,” IFAC Proceedings Volumes, vol. 36, no. 16, pp. 1765–
1770, 2003.
data used in the study was drawn from normal driving over [16] W. E. Larimore, “Canonical variate analysis in identification, filtering,
53 km of roads. A first order linear model, with torque-brake- and adaptive control,” in Proceedings of the 29th IEEE Conference on
gradient inputs, was found to accurately model vehicle dynam- Decision and Control, 1990, pp. 596–604.
ics (VAF=96.5%), and was superior to the nonlinear physical

978-1-5386-2864-5/18/$31.00 ©2018 IEEE 151

Authorized licensed use limited to: Purdue University. Downloaded on February 18,2025 at 15:43:31 UTC from IEEE Xplore. Restrictions apply.