IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308

[Link] Received: 22-09-2018, Accepted: 01-11-2018, Published: 24-11-2018

INFLUENCE OF MODEL PARAMETERS ON VEHICLE SUSPENSION

CONTROL

Sérgio Junichi Idehara1, Matheus Rogério Roesler Sabka2

1
Campus of Joinville, Federal University of Santa Catarina, Santa Catarina, Brazil
2
Campus of Joinville, Federal University of Santa Catarina, Santa Catarina, Brazil

Abstract
Numerical models are widely used to characterize the vehicle dynamics in order to control the active suspension process.
However, little information is available on the evaluation of performance when the model parameters do not match real vehicle
configurations. Obtaining estimates of the influence of these factors on the system control requires statistical analysis, which
generates stochastic data on the issue under consideration. A sensitivity analysis of the test data is one the most successful
approaches to this type of problem. A Monte Carlo simulation with uncertainty parameters for mass, front and rear stiffness and
damping was used with design of experiments analysis to evaluate the performance of three methods of active suspension control
(PID, MPC and LQR). In this study a sensitivity analysis was developed to determine the relevant factors and the cross-
correlation effects of their features. The methodology is applied to a model of a passenger car, which is excited by an asymmetric
speed bump and uneven road profile. The changes in the behavior of the main parameters of each controller were observed and
evaluated as improved for the PID and MPC controllers and worsened for the LQR controller when compared to the designed
condition.

Keywords: Vehicle dynamic, suspension, active control, and sensitivity analysis

---------------------------------------------------------------------***---------------------------------------------------------------------
1. INTRODUCTION performance of the rear suspension based on information
obtained from the front wheels. Suspension control was
Suspension control has been the focus of many studies to simulated based on the optimal control theory for a half-car
improve the dynamic behavior of vehicles aimed at model. The results indicate that better performance of the
increasing automotive safety and comfort. The application active suspension can be obtained with a combination of the
of active or semi-active suspension control has expanded two methods, wheelbase and look-ahead previews. Thus,
and it is used to improve ride comfort, by isolating the suspension control is important to vehicle engineering and
chassis from road excitations, and to guarantee the road some of the techniques available in this regard are listed in
holding characteristics. With the aim of reducing the motion Table 1, [8-14].
of the sprung mass, active control is used to store, dissipate
or even generate energy, whereas semi-active control only Table 1: Suspension control methods described in the
dissipates energy [1]. As examples, some researchers [2, 3] literature
have used suspension control to reduce the vibration of the Literature Control Model Control Year
vehicle passengers. The model considers the dynamics of concept
the person with reference to the head acceleration. The
[8] LQ theory Full Semi- 2013
method employed was linear H∞ control, which was more
vehicle active
effective than a single controlled vehicle model. Also, the
[9] LPV theory Quarter Semi- 2008
application of adaptive strategies has been used to improve
car active
the control methods. Specifically, [4] compares different
[10] MPC Half car Semi- 2006
controllers, such as proportional-integral-derivative (PID)
active
and linear quadratic regulator (LQR), with a proposed
method using feedback error learning (FEL) on a quarter [11] Mixed H2 Full Semi- 2010
vehicle model. This method employs a fuzzy neural network and H∞ vehicle active
with LQR control to optimize the suspension system. The [12] Fault- Full Active 2015
results indicate better conditions for ride comfort, steering tolerant vehicle
and stability of the vehicle dynamics. In models applied in robust H∞
previous research, preview active suspension control was [13] PID Quarter Active 2012
used to improve performance based on measurements of car
road irregularities [5, 6]. For instance, [7] evaluates two [14] Modal non- Model Semi- 2007
types of prediction in the controller: a look-ahead preview linear free active
with information on the irregularities of the track ahead of skyhook
the vehicle; and a wheelbase preview, to improve the

_______________________________________________________________________________________
Volume: 07 Issue: 12 | Dec-2018, Available @ [Link] 1
IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308

[Link] Received: 22-09-2018, Accepted: 01-11-2018, Published: 24-11-2018

The simulations in the cited studies use a quarter-car (single

corner), half-car or full-car model as a base model, with 2, 4
and 7 degrees of freedom (dof), respectively. As another
example of this type of modelling, on analyzing a complete
heavy commercial vehicle, in [15] a semi-active suspension
control was simulated using a multibody model with the
cabin and trailer models from modal skyhook control.
Therefore, there are several approaches to controlling the
vehicle suspension; however, most of them require an
accurate plant model. Thus, one of the challenges in
suspension control is tuning the model parameters. In many
studies today, the main challenge is dealing with the model
uncertainties, seeking online compensation for unknown Fig 1: Illustration of vehicle model with seven dof and its
nonlinearities through adaptive control [16, 17]. dimensions
Nevertheless, in the commercial vehicle context, traditional
methods based on a vehicle model are still frequently 2.1 Vehicle Dynamics
applied. Thus, when developing a control technique, an
understanding of the model parameters and their sensitivity The equation of motion, derived from this seven dof model,
is required to obtain a robust controller setup. represents the vehicle motions of pitch, bounce and roll and
the vertical movement of the wheels. The differential
In this context, this article reports a study on the influence of equation solution generates the time response for the vertical
the model parameters on the suspension control movement of the sprung mass (z), angular displacement
performance, where numerical experiments were conducted around the x-axis (θx ) and the y-axis (θy ) and the vertical
to obtain a statistical understanding of the system behavior. movement of the unsprung mass (x1 , x2 , x3 and x4 ).
A Monte Carlo simulation and design of experiments (DoE) According to this model, the control force of the active
method were employed in this analysis. The vehicle model suspension is attributed to input excitation u(t) and the
has seven degrees of freedom and is controlled by three other force, f(t) , is the road excitation force. Thus, the
types of controllers: proportional-integral-derivative (PID), equation of motion is written as:
model prediction control (MPC) and linear quadratic
regulator (LQR) methods. 𝑀 . 𝑥 + 𝐶 . 𝑥 + 𝐾 . 𝑥 = 𝑓(𝑡) + 𝑢(𝑡) , (1)

2. MATHEMATICAL FORMULATION where,

In the area of numerical simulation, models with seven dof x = z θy θx x1 x2 x3 x4 , and x and x are the
have proven valuable for predicting the dynamics of first and second time derivates of x ,
vehicles. For example, Fig. 1 illustrates a vehicle with a M = diag ms Jy Jx m1 m2 m3 m4 is the mass
sprung mass (ms ) and four unsprung masses (m1 , m2 , m3 matrix,
and m4 ) from the wheels. The scheme also represents the
suspension system with stiffness (k i ) and damping (ci ) for
corners i = 1, 2, 3, 4. Tire stiffness and damping of the
wheels are described by k t and ct , respectively.

c11 c12 c13 −c1 −c2 −c3 −c4

c21 c22 c23 c1 . d3 c2 . d3 −c3 . d4 −c4 . d4
c31 c32 c33 −c1 . d1 c2 . d2 −c3 . d1 c4 . d2
C = −c1 c1 . d3 −c1 . d1 ct + c1 0 0 0 is the damping matrix,
−c2 c2 . d3 c2 . d2 0 ct + c2 0 0
−c3 −c3 . d4 −c3 . d1 0 0 ct + c3 0
−c4 −c4 . d4 c4 . d2 0 0 0 ct + c4

k11 k12 k13 −k1 −k 2 −k 3 −k 4

k 21 k 22 k 23 k1 . d3 k 2 . d3 −k 3 . d4 −k 4 . d4
k 31 k 32 k 33 −k1 . d1 k 2 . d2 −k 3 . d1 k 4 . d2
K = −k1 k1 . d3 −k1 . d1 k t + k1 0 0 0 is the stiffness matrix,
−k 2 k 2 . d3 k 2 . d2 0 kt + k2 0 0
−k 3 −k 3 . d4 −k 3 . d1 0 0 kt + k3 0
−k 4 −k 4 . d4 k 4 . d2 0 0 0 kt + k4

_______________________________________________________________________________________
Volume: 07 Issue: 12 | Dec-2018, Available @ [Link] 2
IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308

[Link] Received: 22-09-2018, Accepted: 01-11-2018, Published: 24-11-2018

f(t) = Rear suspension 𝑘3 , 𝑘4 1.90x104N/m

0 0 0 k t . xt1 + ct . x t1 k t . xt2 + ct . xt2 k t . xt3 + stiffness
ct.xt3 kt.xt4+ct.xt4T is the vector of excitation forces on Rear suspension 𝑐3 , 𝑐4 3.14x103N.s/m
the wheels and u(t) is the control force of the suspension, damping coefficient
with xti being the base excitation from road irregularities Tire stiffness 𝑘𝑡 1.90x105N/m
for suspensions 1, 2, 3 and 4. Tire damping 𝑐𝑡 0.00
coefficient
The matrix elements are:
2.2 Vehicle Decoupled Controller
𝑐11 = 41 𝑐𝑖 , (2)
𝑐12 = 𝑐21 = −𝑐1 . 𝑑3 − 𝑐2 . 𝑑3 + 𝑐3 . 𝑑4 + 𝑐4 . 𝑑4 , A decoupled controller is based on a quarter car model,
𝑐13 = 𝑐31 = 𝑐1 . 𝑑1 − 𝑐2 . 𝑑2 + 𝑐3 . 𝑑1 − 𝑐4 . 𝑑2 , which considers two vibration dof and it sets out a sprung
mass (ms ) and unsprung mass (mus ) motion (Fig. 2). The
𝑐22 = 𝑐1 . 𝑑32 + 𝑐2 . 𝑑32 + 𝑐3 . 𝑑42 + 𝑐4 . 𝑑42 ,
idea is to control independently each suspension with this
𝑐23 = 𝑐32 = −𝑐1 . 𝑑1 . 𝑑3 + 𝑐2 . 𝑑2 . 𝑑3 + 𝑐3 . 𝑑1 . 𝑑4 −
model and select the optimal simulated parameters to find
𝑐4 . 𝑑2 . 𝑑4 ,
ways to reduce the vibration in the whole vehicle. To
𝑐33 = 𝑐1 . 𝑑12 + 𝑐2 . 𝑑22 + 𝑐3 . 𝑑12 + 𝑐4 . 𝑑22 . simulate the results for the control of the 7-dof vehicle, four
2-dof controls are used simultaneously (one for each
𝑘11 = 41 𝑘𝑖 , (3) suspension). This strategy is a good way to reduce the
𝑘12 = 𝑘21 = −𝑘1 . 𝑑3 − 𝑘2 . 𝑑3 + 𝑘3 . 𝑑4 + 𝑘4 . 𝑑4 , processing time and model complexity of the controller.
𝑘13 = 𝑘31 = 𝑘1 . 𝑑1 − 𝑘2 . 𝑑2 + 𝑘3 . 𝑑1 − 𝑘4 . 𝑑2 ,
𝑘22 = 𝑘1 . 𝑑32 + 𝑘2 . 𝑑32 + 𝑘3 . 𝑑42 + 𝑘4 . 𝑑42 ,
𝑘23 = 𝑘32 = −𝑘1 . 𝑑1 . 𝑑3 + 𝑘2 . 𝑑2 . 𝑑3 + 𝑘3 . 𝑑1 . 𝑑4 −
𝑘4 . 𝑑2 . 𝑑4 ,
𝑘33 = 𝑘1 . 𝑑12 + 𝑘2 . 𝑑22 + 𝑘3 . 𝑑12 + 𝑘4 . 𝑑22 .

The magnitude of the torque on the x-axis (Mx ) and y-axis

(My ) of the sprung mass is found by a geometric
transformation matrix (T) from suspension forces:

𝑓1 𝑓1 (4)
𝑀𝑥 𝑑1 −𝑑2 𝑑1 −𝑑2 𝑓2 𝑓
𝑀𝑦 = −𝑑3 = 𝑇. 2 ,
−𝑑3 𝑑4 𝑑4 𝑓3 𝑓3 Fig 2: Quarter car model.
𝑓4 𝑓4
Similarly to the 7-dof model, the motion equation of the 2-
where, f1 , f2 , f3 , f4 are the suspension forces at the vehicle dof model is written as
corners (sprung mass).
𝑚𝑠 0 𝑥 𝑐 −𝑐 𝑥 (5)
. 𝑠 + −𝑐 𝑐 + 𝑐 . 𝑠 +
The values used in the numerical simulations of a full 0 𝑚𝑢𝑠 𝑥𝑢𝑠 𝑡 𝑥𝑢𝑠
vehicle excited by a transient excitation from vehicle tire 𝑘 −𝑘 𝑥𝑠 𝑢(𝑡)
. = ,
contact to evaluate the influence of the active suspension on −𝑘 𝑘 + 𝑘𝑡 𝑥𝑢𝑠 −𝑢 𝑡 + 𝑓(𝑡)
the vibration are shown in the Table 2.
where c and ct are the suspension and tire damping
Table 2: Seven-dof car parameters coefficients, respectively, k and k t are the suspension and
Variable Symbol Value tire stiffness, respectively, u t is the controller force, and
Sprung mass 𝑚𝑠 678.00kg f t is the external excitation force on the tire. The numbers
Front vehicle unsprung 𝑚1 , 𝑚2 31.50kg of dof are xs and xus for the displacement of the sprung
mass mass and unsprung mass, respectively. The quarter car
Rear vehicle unsprung 𝑚3 , 𝑚4 44.50kg parameters are shown in Table 3.
mass
Mass moment of 𝐽𝑥 850kg.m² Table 3: quarter car parameters
inertia x-axis Variable Symbol Front vehicle Rear vehicle
Mass moment of 𝐽𝑦 2.40x103kg.m² Sprung mass 𝑚𝑠 169.50kg 169.50kg
inertia y-axis Unsprung 𝑚𝑢𝑠 31.50kg 44.50kg
Front suspension 𝑘1 , 𝑘2 1.69x104N/m mass
stiffness Suspension 𝑐 1.55x103Ns/m 3.14x103Ns/m
Front suspension 𝑐1 , 𝑐2 1.55x103N.s/m damping
damping coefficient coefficient

_______________________________________________________________________________________
Volume: 07 Issue: 12 | Dec-2018, Available @ [Link] 3
IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308

[Link] Received: 22-09-2018, Accepted: 01-11-2018, Published: 24-11-2018

Tire damping 𝑐𝑡 0.00 0.00 02𝑥1

−1 . 𝑢1𝑥1 ,
coefficient 𝑀2𝑥2 . 𝐿2𝑥1
Suspension 𝑘 1.69x104N/m 1.90x104 𝑑 = 𝐶𝑠𝑠 . 𝜙 + 𝐷𝑠𝑠 . 𝑢 , (10)
stiffness
Tire stiffness 𝑘𝑡 1.90x105N/m 1.90x105N/m 0 0 0
where, O2x2 = and O2x1 = are the zero matrix
0 0 0
3. CONTROLLER METHODS and vector, respectively, L = 1 0 T is the actuator
placement, Css is the output matrix and Dss is the direct
The control force, u t , is calculated using one of the transition matrix. By representing as a function of state and
controller algorithms. The system response from the input matrix, respectively:
external excitation force f t and control force u t is the
input to the controller. In this paper, the three algorithms 𝑂2𝑥2 𝐼2𝑥2 (11)
used are (a) a PID control, (b) an explicit model predictive 𝐴4𝑥4 = −1 −1 and
−𝑀2𝑥2 . 𝐾2𝑥2 −𝑀2𝑥2 . 𝐶2𝑥2
control (MPC) and (c) a linear quadratic regulator (LQR), 02𝑥1
which are written in the Matlab program based on a desired 𝐵4𝑥1 = −1 ,
𝑀2𝑥2 . 𝐿2𝑥1
signal (d∗ t ), according to the formulations and conditions
described in the following sections. In this case, the desired the first equation is written as:
signal, sprung mass vibration, is d∗ t = 0.
𝑞4𝑥1 = 𝐴4𝑥4 . 𝑞4𝑥1 + 𝐵4𝑥1 . 𝑢1𝑥1 . (12)
3.1 PID control
The proportional-integral-derivative (PID) control is The discrete time equation can then be expressed, sampled
described by a force determined as at equal intervals ∆t, with Dss = [0], as:

𝑑𝑒 𝑡 1 𝑡 (6) 𝑞[𝑘 + 1|𝑘] = 𝛷 . 𝑞[𝑘|𝑘 − 1] + 𝛤𝑢 . 𝑢[𝑘|𝑘 − (13)

𝑢 𝑡 = 𝐾. 𝑒 𝑡 + 𝑇𝑑 . + . 𝑒 𝜏 . 𝑑𝜏 .
𝑑𝑡 𝑇𝑖 0 1] +𝛤𝑒.𝑒[𝑘|𝑘],
𝑑[𝑘|𝑘 − 1] = 𝐶𝑠𝑠 . 𝑞[𝑘|𝑘 − 1] . (14)
where, e t is the control error between the system response
(d) and the reference signal or set point (d∗ ). In relation to
And
the controller constants, K is a proportional gain, Td a time
derivative and Ti a time integral. In the discrete form we A.∆t n
have the following: Φ4x4 = eA .∆t = ∞ n=0 n! is the matrix exponential. For
A.∆t
small time steps, e ≅ I4x4 + A4x4 . ∆t, where I4x4 is the
𝑢𝑘 = (7) identity matrix;
𝑒 𝑘 −𝑒 𝑘−1
𝐾𝑝 . 𝑒 𝑘 + 𝐾𝑑 . +
∆𝑡
𝑘 Γu = A−1 . eA .∆t − 1 . B is a constant (4x1) vector;
+𝐾𝑖 . 𝑛=1 𝑒
𝑛 . ∆𝑡,
Γe is a constant estimator gain vector (4x1); and
𝑒 𝑘 = 𝑑 ∗ 𝑘 − 𝑑[𝑘]. (8)
e k k = d∗ k − d k k − 1], which d∗ [k] is the desired
output signal.
K
where, K p = K, K d = K. Td and K i = are the gains of the
Ti
proportional, derivative and integral terms. In the To define the force control for each time instant, a control
simulations, the numerical values for the constants are objective function is written as
K p = 10000, K d = 8000 and K i = 4000. 1
𝐽𝑀𝑃𝐶 = 𝜓 𝑘 − 𝜓 ∗ [𝑘] 𝑇 . 𝑄. 𝜓 𝑘 − 𝜓 ∗ 𝑘 + (15)
2
1
3.2 Explicit Model Predictive Control 𝑢𝑇 𝑘 . 𝑅. 𝑢 𝑘 ,
2
The model predictive control (MPC) is a control method
commonly employed by researchers and engineers for where,
academic and industrial research and development, mainly
T
in advanced process control. The unconstrained MPC ψ k = dT k + 1 k dT k + 2 k … dT k + p k
method presented here is based on [18]. Rewriting the
motion equation of the 2-dof system in a state space form and
gives the state equation (9) and output equation (10):
ψ∗ k =
𝑥 (9) [d∗T k + 1 k d∗T k + 2 k … d∗T k + p k ]T are
𝑞4𝑥1 = 2𝑥1 =
𝑥2𝑥1 (px1) vectors. p is the number of step-ahead predictions of
𝑂2𝑥2 𝐼2𝑥2 𝑥2𝑥1 the output signal, known as the prediction horizon.
. +
−1
−𝑀2𝑥2 . 𝐾2𝑥2 −1
−𝑀2𝑥2 . 𝐶2𝑥2 𝑥2𝑥1

_______________________________________________________________________________________
Volume: 07 Issue: 12 | Dec-2018, Available @ [Link] 4
IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308

[Link] Received: 22-09-2018, Accepted: 01-11-2018, Published: 24-11-2018

Based on minimizing equation (15), the optimal predictive where,

control force is given by 𝑐ℎ𝑜𝑜𝑠𝑒 𝐾𝐿𝑄𝑅 𝑡𝑜 𝑚𝑖𝑛𝑖𝑚𝑖𝑧𝑒 𝐽𝐿𝑄𝑅 = (20)
𝑞 𝑇 . 𝑄𝐿𝑄𝑅 . 𝑞 + 𝑢 𝑇 . 𝑅𝐿𝑄𝑅 . 𝑢 𝑑𝑡.
𝑢𝑘 = (16)
𝐻𝑇 . 𝑄𝑀𝑃𝐶 . 𝐻 + The optimal controller force, for any initial state, is given as
𝑅𝑀𝑃𝐶−1.𝐻𝑇.𝑄𝑀𝑃𝐶.𝑌𝑧.𝑞𝑘𝑘−1+𝑌𝑒.𝑒[𝑘|𝑘],
𝑢 𝑡 = − 𝐾𝐿𝑄𝑅 . 𝑞 . (21)
The elements of matrix H (pxm) are given by hi =
C. Φi−1 . Γu (1x1) and written as equation (17), where m is
the control horizon, step-ahead prediction of the force. For the optimization problem the following expression is
chosen, where P is a symmetric and positive-definite matrix.
ℎ1 0 ⋯ 0 (17) 𝑇
ℎ2 ℎ1 ⋯ 0 𝑞 𝑇 . 𝑄 + 𝐾𝐿𝑄𝑅 . 𝑅 . 𝐾𝐿𝑄𝑅 . 𝑞 = (22)
𝐻= .
⋮ ⋮ = −2. 𝑞 𝑇 . 𝑃 . 𝑞 .
ℎ𝑝 ℎ𝑝−1 ⋯ ℎ𝑝−𝑚 +1
For the negative real parts of the eigenvalues of A −
And B . K LQR , the system is asymptotically stable and thus the
first term is zero, q ∞ → 0. Therefore,
𝑌𝑧 = 𝐶. 𝛷 𝑇 𝐶. 𝛷 2 𝑇
⋯ 𝐶. 𝛷 𝑝 𝑇 𝑇
a matrix
(𝑝𝑥4),
𝐽𝐿𝑄𝑅 = 𝑞(0) 𝑇 . 𝑃 . 𝑞(0) . (23)
𝑌𝑒 =
𝐶. 𝛤𝑒 𝑇 𝐶. 𝐼 +
In order to minimize JLQR , where q(0) is a constant vector,
𝛷.𝛤𝑒𝑇 ⋯ 𝐶.𝑘=1𝑝𝛷𝑘−1.𝛤𝑒𝑇 𝑇 a vector (𝑝𝑥1), the matrix P should be minimized with respect to K LQR :
∂P
= 0. Hence,
∂K LQR
𝑞 0⋯ −1
𝑄𝑀𝑃𝐶 = ⋮ a matrix (𝑝𝑥𝑝), for 2-dof 𝐾𝐿𝑄𝑅 = 𝑅 . 𝐵 𝑇. 𝑃 . (24)
⋮
0 ⋯ 𝑞
1 0 The expression in equation (24) gives the static LQR gain
problem, 𝑞 = ,
0 0 matrix and the matrix P is obtained by a backward
1 ⋯ 0
numerical solution of the algebraic Riccati equation, using
𝑅𝑀𝑃𝐶 = 𝜆. ⋮ ⋮ a matrix ( 𝑝 − 𝑚 + 1 𝑥 𝑝 −
standard numerical tools in linear algebra:
0 ⋯ 1
𝑚+1).
𝐴𝑇 . 𝑃 + 𝑃. 𝐴 − 𝑃. 𝐵. 𝑅 −1 . 𝐵𝑇 . 𝑃 + 𝑄 = 0. (25)
λ is a weighting on the rate of change of the inputs. For
small penalties, the controller tends to give higher values 1 0 0 0
and is less robust, while for large penalties it becomes a 0 5 1 0 0
The input matrix is assumed as Q = 1x10 .
more robust controller but with a slow response [19]. The 0 0 1 0
configuration values of the controller for the simulations 0 0 0 1
1 0
consist of a state prediction horizon of p = 100, control and R = 2.5x10−4 . .
0 1
horizon of m = 20 and weighting of λ = 5x10−4 .
4. ROAD EXCITATION FORCES
3.3 Linear Quadratic Regulator Control
The roughness of a road is considered a source of external
The linear quadratic regulator (LQR) controller is one of the excitation of the vehicle with, for instance, forces acting on
optimal control techniques used in engineering design and the tire. In this simulation, an asymmetrical excitation is
applications. The algorithm chooses weighting factors as a used for the passage of the right wheels of a vehicle over a
matrix of a linear state-feedback gain [20]. Mathematically, speed bump (case 1) and a vehicle passing over an uneven
a LQR is defined as a cost function by road profile (case 2), as seen in Fig. 3. The longitudinal
speed of the vehicle is constant at 20km/h. The bump model,
𝐽𝐿𝑄𝑅 = 𝑞 𝑇 . 𝑄𝐿𝑄𝑅 . 𝑞 + 𝑢 𝑇 . 𝑅𝐿𝑄𝑅 . 𝑢 𝑑𝑡, (18) case 1, is a sinusoidal profile with a wavelength of 2. l =
3.0m and height of h = 0.080m. The (right) frontal wheel
where, Q LQR (n x n) and R LQR (r x r) are a positive-definite travels a distance of 0.5m before the excitation starts. To
Hermitian matrix, and a performance index is described by represent the general road excitation in case 2, a white noise
signal is filtered through a low-pass first-order filter. The
𝐽𝐿𝑄𝑅 = 𝐾𝐿𝑄𝑅 𝑞, 𝑢 . 𝑑𝑡, (19) cutoff frequency is fc = fo′ . v, where fo′ [cycle/m] is a cutoff
spatial frequency and v [m/s] is the constant velocity of the

_______________________________________________________________________________________
Volume: 07 Issue: 12 | Dec-2018, Available @ [Link] 5
IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308

[Link] Received: 22-09-2018, Accepted: 01-11-2018, Published: 24-11-2018

vehicle. The power spectral density (PSD) of the road by half. In general, the response with the LQR control was
displacement (Sr ) is given by equation (26), [21]: of lower amplitude, but it had a low decay ratio.

1/𝑓 ′ 2 (26)
𝑆𝑟 𝑓 ′ = 𝑆𝑜 . 2 .
1+ 𝑓0′ /𝑓′

The parameter So is a roughness magnitude and it is

calculated using the signal standard deviation σ as So =
σ. fo′ 2 . The f ′ [cycles/m] is the spatial frequency, a
division between frequency and velocity, f ′ = f/v.

The road excitation frequency in the range of 0.11Hz is

given by So = 1.25x10−2 and fo ′ = 0.02cycle/m at a speed
of 20km/h.

Fig 4: Displacement of vehicle center of gravity or bounce

during passage over bump for different control methods

In order to analyze a random excitation response for each

controller, an uneven road profile was applied to the 7-dof
model. The road profile spectrum (power spectrum density)
is provided in Fig. 5 and the road classification provided by
ISO 8608 lies between letters “A” and “H”. For this
simulation, the PSD profile operates at levels “C” and “D”,
between an average and poorly paved road profile [22].
Fig 3: Profile of the speed bump (case 1) and uneven road
excitation (case 2)
At the end of this simulation of the random profile
excitation, the vehicle CG displacements in Fig. 6,
5. RESULTS AND DISCUSSIONS controlled by these methods, are noticeably smoother and
This section reports the vehicle response in cases 1 and 2 in have lower amplitude compared with the passive
the dynamic simulation of 7 dof and using three different suspension. Once again, the vibration attenuations of the
control methods. A delay time of 10ms in the control force PID and MPC controllers were found to be similar.
was employed in the numerical integration using the
Newmark integrator with a sample rate of 1000Hz. The data Interestingly, the performance of the controllers is analyzed
collected were analyzed using the vibration displacement of with the integral squared error (ISE) as an index given by
the vehicle center of gravity (CG) and the integral squared
𝑇 2 (27)
error (ISE) of the controllers. Additionally, a sensitivity 𝐼𝑆𝐸 = 𝑒 𝑡 . 𝑑𝑡,
0
analysis was carried out to observe the interdependence and
correlation between the controller parameters. For these
evaluations, a design of experiments was employed with the in which e t is the error between the desired and actual
Monte Carlo simulation. output values.

5.1 Vehicle Response from 7-dof Model The ISE observed as a result of cases 1 and 2 in Table 4
shows attenuations of over 70%, when compared with the
The focus of the active suspension, in this case, is to provide passive suspension. Based on these results, the three control
vibration control and improve the internal comfort by techniques investigated in this study achieved valuable
reducing the oscillatory movements of the sprung mass. Fig. dynamic control, attenuating the sprung mass vibration.
4 shows the displacement of the vehicle center of gravity
(CG) – bounce (heave) movement – on passing over the It is important to note that these results are specific to the
right side bump (case 1) and compares how the different quarter car model used in these simulations. Thus,
controllers act. This illustrates the variation that can be application to cases other than the values of the 2-dof model
generated in the controlled system, where the vehicle parameters employed here may lead to different responses,
response with passive suspension has a higher vibration resulting from variations in the more sensitive parameters,
peak, while the controllers reduce this amplitude. The PID limiting the effectiveness of the controllers.
and MPC controls have similar behaviors, decreasing the Table 4: Comparisons between different control methods
overshoot to almost a third and shortening the settling time considering the integral squared error (ISE)
_______________________________________________________________________________________
Volume: 07 Issue: 12 | Dec-2018, Available @ [Link] 6
IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308

[Link] Received: 22-09-2018, Accepted: 01-11-2018, Published: 24-11-2018

ISE Speed Attenuati Uneven Attenuati sprung mass and the position of the CG are far from being
bump on road on constant values. Under such conditions, the controller
(𝒙𝟏𝟎−𝟔 ) (𝒙𝟏𝟎−𝟑 ) performance can be inconsistent, as observed in simulation
W/o 79.16 – 21.97 – tests. Thus, the aim of this study was to apply a generic
control methodology to evaluate the controller, in an environment
PID 9.35 88.2% 3.71 83.1% with a high level of uncertainty, via a numerical simulation
control using the Monte Carlo method and design of experiment
LQR 9.81 87.6% 5.57 74.6% (DoE) tools, sometimes called the numerical design of
control experiments (NDoE), instead of a combination of costly
MPC 6.51 91.8% 2.88 86.9% numerical modelling and numerous experiments [23]. In
control recent years, DoE has been recognized as an important tool
for validation and system characterization, applying
5.2 Sensitivity of the Model Parameters statistical approaches to experimental design and analysis.
The behavior of an input parameter and its effect on a
Since the uncertainties associated with the vehicle are real, a response can be investigated separately, and interactions
non-adaptive controller method may be sensitive to between different factors that affect the system output can
variations in the system characteristics. For instance, the be highlighted.

Fig 5: Power spectrum density for the uneven road profile

Fig 6: Displacement of the vehicle center of gravity or heave for the random road profile with different control methods

The Monte Carlo method is applied to compute a stochastic better understanding of the effect of variations in the system
simulation with seven dof of vehicle vibration with PID, characteristics on the performance of an active suspension
LQR and MPC controllers. The simulation employs the controller, the design of experiments provides an important
application of random numbers, which are normally method to identify the critical input parameters. In this case,
distributed values for the front and rear stiffness, front and a 25 full-factorial design was applied with 10 replications
rear damping coefficients and sprung mass (Table 5). In this for the parameters described in Table 5.
way, sampling experiments, associated with the uncertain Table 5: Mean and standard deviation of two levels of
behavior of the system, can be generated. In order to gain a Monte Carlo inputs

_______________________________________________________________________________________
Volume: 07 Issue: 12 | Dec-2018, Available @ [Link] 7
IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308

[Link] Received: 22-09-2018, Accepted: 01-11-2018, Published: 24-11-2018

Factor Variable Mean Mean Standard 18 +1 -1 -1 -1 +1

Level 1 Level 2 Deviation 19 +1 -1 -1 +1 -1
(a) Sprung 678.0 1078.0 100.0 20 +1 -1 -1 +1 +1
mass 21 +1 -1 +1 -1 -1
(kg) 22 +1 -1 +1 -1 +1
(b) Front 13,700.0 16,878.0 1,500.0 23 +1 -1 +1 +1 -1
stiffness 24 +1 -1 +1 +1 +1
(N/m) 25 +1 +1 -1 -1 -1
(c) Rear 15,810.0 19,000.0 1,500.0 26 +1 +1 -1 -1 +1
stiffness 27 +1 +1 -1 +1 -1
(N/m) 28 +1 +1 -1 +1 +1
(d) Front 1,300.0 1,554.0 150.0 29 +1 +1 +1 -1 -1
damping 30 +1 +1 +1 -1 +1
(Ns/m) 31 +1 +1 +1 +1 -1
(e) Rear 2,700.0 3,144.0 150.0 32 +1 +1 +1 +1 +1
damping
(Ns/m) Testing with every combination of the five factors shown in
Table 6 required simulation with 32 conditions. The total
The DoE evaluated with two levels receives lower (-1) and number of simulations for 10 replications is 320 runs. The
higher (+1) values for each parameter and the statistical results of the t-statistical analysis for a two-tail distribution,
correlation is calculated. The parameter sensitivity can be the beta values determined (linear regression coefficients)
obtained by linear regression in the form of: and the p-values for each parameter and its cross-
correlation, are given in Table 7 (see Appendix for table of
𝑌 = 𝑋. 𝛽 + 𝜖, (28) results). The controller parameters were found to have a
very significant impact on the performance, with a p-value
where, X is the DoE matrix from Table 6 (see Appendix for of less than α = 5% (95% of the confidence interval for βj ).
table of results), β are the coefficients of the surface lines, ϵ This sensitivity parameter can be defined as the average
is the regression error and Y is the ISE value from the response to the factor, changing between low and high
controlled suspension output signals. levels. This is summarized by the statement:

Table 6: DoE matrix (-1) represents the lower level and  null hypothesis 𝐻𝑜 : 𝛽𝑗 = 0 (29)
(+1) the higher level of the parameters  and alternative hypothesis, 𝐻𝑎 : 𝛽𝑗 ≠ 0
Sprun Front Rear Front Rear
g stiffnes stiffnes dampin dampin where, Ho is rejected if t o > t α ,n−k−1 , otherwise the null
mass s s g g 2

(𝒎𝒔 ) (𝒌𝟏 = (𝒌𝟑 = (𝒄𝟏 = (𝒄𝟑 = hypothesis is accepted. n is the number of observations and
𝒌𝟐 ) 𝒌𝟒 ) 𝒄𝟐 ) 𝒄𝟒 ) k the number of variables.
Test 𝒂 𝒃 𝒄 𝒅 𝒆
s It is clear from Table 7 that for each controller the most
sensitive parameters differ. Front stiffness (b) and rear
1 -1 -1 -1 -1 -1
damping (e) had significant effects in terms of the ISE value
2 -1 -1 -1 -1 +1
for the PID control, while sprung mass (a), front stiffness (b)
3 -1 -1 -1 +1 -1
and rear damping (e) were important for the MPC control,
4 -1 -1 -1 +1 +1 and sprung mass (a) for the LQR control. In addition, this
5 -1 -1 +1 -1 -1 study permits the quantification of parameter interactions in
6 -1 -1 +1 -1 +1 cross product terms, as found for the sprung mass and front
7 -1 -1 +1 +1 -1 stiffness of the PID and LQR controls. Other parameters
8 -1 -1 +1 +1 +1 show no or little interaction.
9 -1 +1 -1 -1 -1
10 -1 +1 -1 -1 +1
11 -1 +1 -1 +1 -1
12 -1 +1 -1 +1 +1
13 -1 +1 +1 -1 -1
14 -1 +1 +1 -1 +1
15 -1 +1 +1 +1 -1
16 -1 +1 +1 +1 +1
17 +1 -1 -1 -1 -1
Table 7: Linear regression t-tests for PID, MPC and LQR controllers in case 2. Significance for p-value < 0.05
_______________________________________________________________________________________
Volume: 07 Issue: 12 | Dec-2018, Available @ [Link] 8
IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308

[Link] Received: 22-09-2018, Accepted: 01-11-2018, Published: 24-11-2018

PID MPC LQR

Variable 𝟑 p-value 𝟑 p-value 𝟑 p-value
𝜷𝒋 x 𝟏𝟎 𝜷𝒋 x 𝟏𝟎 𝜷𝒋 x 𝟏𝟎
Mean 2.5486 0 2.1904 0 7.1459 0
𝑎 -0.1073 0.1021 -0.1804 0.0006 -0.4952 0.0116
𝑏 0.2036 0.0020 0.1512 0.0040 0.3001 0.1248
𝑐 0.0803 0.2207 0.0777 0.1367 0.1157 0.5532
𝑑 -0.0134 0.8379 -0.0233 0.6544 -0.1620 0.4067
𝑒 0.1550 0.0184 0.1143 0.0290 0.3258 0.0957
𝑎. 𝑏 -0.1550 0.0184 -0.0975 0.0621 -0.3927 0.0449
𝑎. 𝑐 -0.0092 0.8877 0.0264 0.6123 0.0196 0.9200
𝑎. 𝑑 0.0134 0.8376 0.0238 0.6483 0.1426 0.4652
𝑎. 𝑒 0.0290 0.6579 0.0380 0.4663 0.1948 0.3186
𝑏. 𝑐 0.0140 0.8302 0.0061 0.9063 0.0673 0.7302
𝑏. 𝑑 -0.0976 0.1366 -0.0690 0.1861 -0.2554 0.1911
𝑏. 𝑒 0.1070 0.1028 0.0799 0.1263 0.3765 0.0544
𝑐. 𝑑 0.0450 0.4921 0.0216 0.6789 -0.0053 0.9784
𝑐. 𝑒 0.0793 0.2261 0.0473 0.3645 0.1439 0.4611
𝑑. 𝑒 -0.0366 0.5758 -0.0347 0.5057 -0.0782 0.6888
𝑎. 𝑏. 𝑐 -0.0957 0.1446 -0.0513 0.3255 -0.2062 0.2910
𝑎. 𝑏. 𝑑 0.0400 0.5414 0.0380 0.4663 0.1986 0.3093
𝑎. 𝑏. 𝑒 -0.0063 0.9232 -0.0014 0.9791 -0.1858 0.3413
𝑏. 𝑐. 𝑑 0.0293 0.6549 0.0487 0.3503 0.2127 0.2762
𝑏. 𝑐. 𝑒 0.0986 0.1329 0.0619 0.2356 0.3058 0.1178
𝑐. 𝑑. 𝑒 0.0357 0.5854 0.0385 0.4602 0.1222 0.5314
𝑑. 𝑒. 𝑎 0.0455 0.4877 0.0019 0.9709 -0.0255 0.8961
𝑑. 𝑒. 𝑏 -0.0067 0.9186 0.0028 0.9570 0.0204 0.9168
𝑐. 𝑑. 𝑎 0.0599 0.3608 0.0510 0.3283 0.2071 0.2890
𝑒. 𝑎. 𝑐 -0.0302 0.6451 -0.0251 0.6297 -0.1153 0.5545
𝑎. 𝑏. 𝑐. 𝑑 -0.0268 0.6819 -0.0244 0.6396 -0.0775 0.6911
𝑎. 𝑏. 𝑐. 𝑒 -0.0700 0.2856 -0.0343 0.5103 -0.1196 0.5399
𝑏. 𝑐. 𝑑. 𝑒 -0.0610 0.3517 -0.0255 0.6247 -0.0644 0.7413
𝑐. 𝑑. 𝑒. 𝑎 -0.0297 0.6503 -0.0409 0.4332 -0.2541 0.1935
𝑑. 𝑒. 𝑎. 𝑏 0.0208 0.7508 0.0451 0.3870 0.0365 0.8515
𝑎. 𝑏. 𝑐. 𝑑. 𝑒 0.0221 0.7352 0.0176 0.7363 0.0871 0.6554

In the interaction plots in Fig. 7 to 9, the factor Conversely, for the PID and MPC controllers the
interdependence can be observed. The main diagonal distribution generally lies below that of the control without
represents the factor influence on the ISE values with a ±σ the random parameters, so that the vibration attenuation was
curve (one standard deviation). For the PID controller (Fig. significantly higher than the original controller design. In
7), the main factor effects can be observed for (b) and (e), contrast to the LQR controller, the ISE distribution changes
both showing an increase in ISE for higher values of the scale significantly between the minimum value and the
parameter. Similarly, the highest gradients found for MPC uncontrolled condition. Therefore, the higher vibration
(Fig. 8) are for the factors (a) with a negative slope and (b) amplitude from some samples controlled by the LQR
and (e) with positive slopes, and for LQR (Fig. 9) for factor method indicates that it did not improve the suspension
(a) with a negative slope. The main interaction between performance.
factors occurs for variables (a) and (b), in the case of the
PID and LQR controllers, and no apparent interactions are
observed for the MPC controller.

Fig. 10 shows the factor level plots comparing the

distributions of the ISE between two controller parameters.
The column distribution (cyan square) illustrates the ISE
variation from minimum and maximum values for each
combination of factors. In all cases, there is a large variation
in the values, indicating that the parameters with high
uncertainty provoke changes in the controller performance.
_______________________________________________________________________________________
Volume: 07 Issue: 12 | Dec-2018, Available @ [Link] 9
IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308

[Link] Received: 22-09-2018, Accepted: 01-11-2018, Published: 24-11-2018

Fig 7: Interaction plot for ISE for the PID controller

Fig 10: ISE distribution for interaction between parameters

Based on these results, the maximum attenuation of the

sprung mass vibration using different controllers can be
reduced or increased because of the uncertain and variable
values attributed to the controller model parameters.
However, the changes in the performance of the suspension
controller are more pronounced only for specific parameters,
indicating the need for further engineering research to
investigate them.

6. CONCLUSION
When using a conventional control method to control the
active suspension of a vehicle, considerable caution needs to
Fig 8: Interaction plot for ISE for the MPC controller be taken due to the effects associated with the model
parameters. This paper presented a methodology and
analyzes the relation between active suspension control and
DoE, in order to provide a better understanding of the
impact and interactions of the controller parameters with
regard to the performance dynamics of the vehicle. The
methodology is applied in a Monte Carlo simulation, which
allows random factors to be established for statistical
analysis. The method was applied on the uneven road profile
(case 2), and the numerical simulations regarding the ISE
values of the CG displacement were carried out to evaluate
and compare three types of controllers: PID, MPC and LQR.
As a result, the main effects and their interactions were
identified. The sensitivity study using linear regression
identified groups of factors for each controller, highlighting
parameters with the strongest influence on the performance
variation. These variations were found to have a random
distribution, increasing or reducing the ISE value of the
designed condition. In the extreme cases, the associated
Fig 9: Interaction plot for ISE for the LQR controller vibration attenuation in the sprung mass is greater than that
of a passive suspension. This quantitative procedure for
evaluating the controller provides a way for model analysis

_______________________________________________________________________________________
Volume: 07 Issue: 12 | Dec-2018, Available @ [Link] 10
IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308

[Link] Received: 22-09-2018, Accepted: 01-11-2018, Published: 24-11-2018

to be carried out and supports predictions regarding the systems with finite-frequency constraint', Mech. Syst.
behavior of the system dynamics. However, as reported Signal Process., 2015, 62, (63), pp. 341-355
herein, this is dependent on the control method chosen, as [13] Choudhury, S.F., Sarkar, M.A.R.: 'An approach on
the variables of a 2-dof model are more sensitive to changes. performance comparison between automotive passive
It emerged, in general, that the sprung mass, front stiffness suspension and active suspension system (PID
and rear damping have the greatest influence on the controller) using matlab/Simulink', Journal of
controllers investigated. However, it is significant that for Theoretical and Applied Information Technology,
the LQR method the attenuation observed is sometimes poor 2012, 43, (2), pp. 295-300
due to the model parameter uncertainty. [14] Sweversa, J., Lauwerysa, C., Vandersmissenb, B., et
al.: 'A model-free control structure for the on-line
REFERENCES tuning of the semi-active suspension of a passenger
car', Mech. Syst. Signal Process., 2007, 21, pp. 1422-
[1] Poussot-Vassal, C., Spelta, C., Sename, O., Savaresi, 1436
S.M., Dugard, L.: 'Survey and performance [15] Ieluzzi, M., Turco, P., Montiglio, M.: 'Development
evaluation on some automotive semi-active of a heavy truck semi-active suspension control',
suspension control methods: A comparative study on Control Eng. Pract., 2006, 14, pp. 305-312
a single-corner model', Annu. Rev. in Control, 2012, [16] Huang, Y., Na, J., Wu, X., et al.: 'Adaptive control of
36, pp. 148-160 nonlinear uncertain active suspension systems with
[2] Suzuki, T., Takahashi, M.: 'Semi-Active Suspension prescribed performance', ISA T., 2015, 54, pp. 145-
Control System Design for Vibration Reduction of 155
Passenger's Body Based on Lissajous Figure of [17] Jastrzebski, R.P., Hynynen, K.M., Smirnov, A.: 'H∞
Damping Force', Journal of System Design and control of active magnetic suspension', Mech. Syst.
Dynamics, 2011, 5, (2), pp. 279-295 Signal Process., 2010, 24, pp. 995-1006
[3] Du, H., Li, W., Zhang, N.: 'Integrated seat and [18] Mei, G., Kareem, A., Kantor, J.C.: 'Real-time model
suspension control for a quarter car with driver predictive control', Earthq. Eng. Struct. D., 2001, 30,
model', IEEE T. Veh. Technol., 2012, 61,(9), pp. pp. 995-1019
3893-3908 [19] Garriga, J.L., Soroush, M.: 'Model Predictive Control
[4] Marofi, [Link]., Seyedalian, S.J., Akram, L.: 'Improve Tuning Methods: A Review', Ind. Eng. Chem. Res.,
active suspension system by FEL controller design', 2010, 49, pp. 3505-3515
Mechanika, 2013, 19, (6), pp. 681-687 [20] Nagarkar, M.P., Vikhe, G.J., Borole, K.R., et al.:
[5] Göhrle, C., Schindler, A., Wagner, A., et al.: 'Design 'Active control of quarter-car suspension system
and vehicle implementation of preview active using linear quadratic regulator', International Journal
suspension controllers', IEEE T Contr Syst T, 2014, of Automotive and Mechanical Engineering, 2011, 3,
22, (3), pp. 1135-1142 pp. 364-372
[6] Li, P., Lam, J., Cheung, K.C.: 'Multi-objective [21] Ulsoy, A.G., Peng, H., Çakmakci, M.: 'Automotive
control for active vehicle suspension with wheelbase Control Systems' (University Press, Cambridge, UK,
preview', J. Sound Vib., 2014, 333, pp. 5269-5282 2012)
[7] ElMadany, M.M., Al Bassam, B.A., Fayed, A.A.: [22] Loprencipe, G., Zoccali, P.: 'Use of generated
'Preview control of slow-active suspension systems', artificial road profiles in road roughness evaluation',
J. Vib. Control, 2011, 17, (2), pp. 245-258 J. Mod. Transport., 2017, 25, (1), pp. 24-33
[8] Unger, A., Schimmack, F., Lohmann, B., et al.: [23] Blondet, G. Duigou, J. L., Boudaoud, N., et al.: 'An
'Application of LQ-based semi-active suspension ontology for numerical design of experiments
control in a vehicle', Control Eng. Pract., 2013,21, processes', Comput. Ind., 2018, 94, pp. 26-40.
pp. 1841-1850
[9] Poussot-Vassal, C., Sename, C., Dugard, L., et al.: 'A
new semi-active suspension control strategy through
LPV technique', Control Eng. Pract., 2008, 16, pp.
1519-1534
[10] Canale, M., Milanese, M., Novara, C.: 'Semi-active
suspension control using “fast” model-predictive
techniques', IEEE T. Contr. Syst. T., 2006, 14, (6),
pp. 1034-1046
[11] Zheng, L., Li, Y., Chen, B.: 'A new semi-active
suspension control strategy through mixed H2/H∞
robust technique', J. Cent. South Univ. T., 2010, 17,
pp. 332-339
[12] Wang, R., Jing, H., Karimi, H.R., et al.: 'Robust
fault-tolerant H∞ control of active suspension

_______________________________________________________________________________________
Volume: 07 Issue: 12 | Dec-2018, Available @ [Link] 11