1

Active Suspension Control of Full-car Systems without

Function Approximation
Jing Na, Member, IEEE, Yingbo Huang*, Member IEEE, Qianqian Pei, Xing Wu, Guanbin Gao and Guang Li,
Member, IEEE

(e.g. springs and dampers) and thus can effectively dismiss
Abstract—This paper proposes a new control approach for
full-car active suspension systems with unknown nonlinearities. forces from the road excitation [9]. However, active suspension
The main advantage of this approach is that the uncertainties and has not been widely used in the commercial products yet
nonlinearities in the system can be handled without using any because of its high energy demand and increased cost. In this
function approximator (e.g. neural networks (NNs), fuzzy logic respect, the development of advanced control strategies for
systems (FLSs)), and the associated online adaptation. Hence, the active suspension systems has been recognized as one of the
heavy computational cost and sluggish learning phase to achieve
most promising pathways to achieve cost-reduction and safety
convergence can be remedied. To maintain the transient and
steady-state suspension responses, a coordinate suspension error enhancements, and thus has attracted significant attentions in
transformation with prescribed performance functions (PPF) is both academics and engineering fields [9-13].
adopted. Then an approximation-free control (AFC) is developed In the active suspension system designs, several performance
to achieve stabilization of the transformed system so as to retain requirements should be considered. Apart from eliminating the
predefined suspension response. Extreme Value Theorem is used uninterrupted disturbance from the road roughness, the ride
together with Lyapunov theorem to prove the stability and comfort, road holding and suspension deflection limitation
convergence of the closed-loop control system. To validate the
proposed method and show its practical applicability, a dynamic should be considered. To manage the potential contradictions
simulator is built by using a commercial vehicle software, Carsim, between these requirements, some advanced control methods
where an E-SUV type vehicle is configured to describe realistic have been tailored for active suspension applications, e.g.
vehicle dynamics. Simulation results reveal that the proposed multi-objective control[14], adaptive control[15], backstepping
control can achieve better suspension performance and require control [16] and preview control [17]. In the aforementioned
less model information compared with some existing approaches. studies, a critical assumption is that all dynamics of the studied
Index Terms—Active suspension control, Full-car system,
Prescribed performance control, Nonlinear dynamics. suspension systems should be known accurately and even
should be linear. This assumption may not be true in the
I. INTRODUCTION realistic vehicle suspension systems (e.g. the springs, dampers
and actuators used for suspension have nonlinear behaviors),
S USPENSION systems have always played an essential role in
the automotive products due to its ability to guarantee the
passengers’ ride comfort and driving safety (i.e. road holding
which makes these model based approaches less effective.
To accommodate unknown nonlinearities in the control
systems, function approximators, such as neural networks (NNs)
and suspension stroke limitation) [1, 2]. According to different
and fuzzy logic systems (FLSs) [7, 13, 18-21], have been
mechanical configures, vehicle suspension systems can be
incorporated into adaptive control designs, where the unknown
classified into three types: passive suspension, semi-active
weights of NNs or FLSs can be online updated via adaptive
suspension [3-5] and active suspension [6, 7]. Although passive
laws designed to minimize the control errors [22-24]. This
suspension has been widely used in commercial vehicles due to
methodology allows proving the closed-loop stability by means
its low cost and simple structure, its ability to absorb excitations
of the Lyapunov theorem. Although it is mathematically
induced by the road conditions is restricted by the fixed spring
elegant, it is found that practical implementation of function
and damper dynamics. Hence, recent studies have been made
approximation based adaptive control schemes is still not fully
toward the developments of semi-active suspension systems
mature for commercial products due to their complex structure
[3-5] and active suspension systems [6, 7], which are able to
and demanding computational cost. Specifically, it is still a
achieve considerably improved suspension performance over
nontrivial task to tune the parameters used in these adaptive
passive suspensions [8]. Specifically, active suspension system
control methods. Moreover, the online learning of the function
uses extra actuators together with passive suspension elements
approximators needs fairly long time to achieve convergence,
This work was supported by the National Natural Science Foundation of and the potentially sluggish convergence rate and/or large
China under Grants 61573174, and a joint PhD Scholarship between Chinese overshoot during the transient stage may create risks for
Scholarship Council (CSC) and Queen Mary University of London under Grant operation safety of adaptive control systems [25].
201708530252. (Corresponding author: Yingbo Huang).
Jing Na, Yingbo Huang, Xing Wu and Guanbin Gao are with the Faculty of
It is also noted that quantitatively study transient response of
Mechanical and Electrical Engineering, Kunming University of Science and nonlinear control systems is generally difficult. Recently,
technology, Kunming, 650500, China. (e-mail: [email protected]; Bechlioulis and Rovithaks introduced a new control framework
[email protected]; [email protected]; [email protected]).
Qianqian Pei is with the Faculty of Mechanical and Electrical Engineering,
[26, 27] to guarantee both the transient and steady-state control
Yunnan Land And Resources Vocational College, Kunming, 652501, China. convergence responses within a predefined boundary. The key
(e-mail: [email protected]) idea is to incorporate a prescribed performance function (PPF)
Guang Li is with the School of Engineering and Materials Science, Queen
and the associate error transformation into the control designs.
Mary University of London E1 4NS, UK (e-mail: [email protected]).
 2

This idea has been subsequently tailored for other control In this paper, a nonlinear uncertain full-car model with four
system designs [28-30]. Nevertheless, in these PPF based independent active suspension systems is considered, as shown
control designs, function approximators are still needed to in Fig.1. This full-car suspension system has seven degrees of
accommodate the unknown system dynamics. To address this freedom (DOFs), e.g. vertical, roll and pitch motions of the
issue, an approximation-free control (AFC) [31] was further
sprung mass, and the vertical motions of the four unsprung
introduced to address tracking control of nonlinear systems. In
masses connected to the wheels (front-right, front-left,
this method, NNs and FLSs are not required, while both the
rear-right and rear-left) [10]. The variables of this suspension
transient and steady-state control performances are guaranteed.
Hence, this technique provides a potentially new constructive model as shown in Fig.1 and model (1) are defined as: M is
methodology for nonlinear control system designs, which is the sprung mass and mi , i  1 4 is the unsprung mass,
also useful for active suspensions [21]. representing the car body and the wheel assembly, respectively;
On the other hand, it is also found from active suspension I and I denote the mass moment of inertia for the roll and
control literatures that most of existing results focus on quarter- pitch motions. Fsi and Fdi are the forces produced by the
car or half-car dynamics only, while only few results have been springs and dampers. ksi and kdi are the spring stiffening
reported for full-car dynamics [10, 32]. In fact, the full-car coefficients and damper damping coefficients. kti is the
suspension system has multi-inputs-multi-outputs and certain stiffness of the four wheels. zs ,  and  define the vertical, roll
couplings between the vertical, roll and pitch motions, which and pitch motions of the vehicle body. zui and yi denote the
makes the corresponding control design difficult. Hence, the displacement of the unsprung mass and road displacement
aim of this paper is to introduce a new control design for
input. a, b, c and d are the distances of the suspensions to the
full-car active suspension systems with unknown dynamics.
center of the vehicle body mass. The control inputs ui are the
The main merit is to further tailor the concept of AFC [9, 31]
and extend this idea to full-car suspension applications, where forces produced by the four active suspension systems to
the vertical, roll and pitch motions are all considered. This is a eliminate the effect of the road roughness. V is the driving
nontrivial advancement compared with the recent work [10, 32]. velocity of the car.
b a
Accommodation of unknown dynamics without using any
function approximators can improve the computational M
d u4 zs u2
efficiency and eliminate the effort required for the control Fs4 Fd4 Fs2 Fd2
m4 zu4 ϕ zu2
m2
parameter tuning. Furthermore, the suspension response (e.g. Iϕ
x
kt4 Iθ θ kt2
overshoot, convergence rate and ultimate displacement) can be c y4 y2 V
y
strictly guaranteed, which directly contributes to the
enhancement of operation safety for active suspension systems.
This control design is model independent, and the derived Fs3 u3 Fs1 u1
Fd3 Fd1
control actions are with a proportional-like form. Hence, it is m3 zu3 m1 zu1
easy to implement, and may be more preferable in practice. kt3 kt1
The contributions of this paper are summarized as follows: y3 y1

a) This paper addresses active suspension control for full-car

systems, where the nonlinear dynamics of springs and dampers Fig.1 Schematic of full-car active suspension systems.
in the systems are not necessarily known. Hence, the modeling The detailed mathematical model of active suspension
efforts required for the control designs can be reduced. system shown in Fig.1 can be developed as [10, 32]:
b) An approximation-free control is obtained for active M z   F  F  F  F  F  F  F  F  u
suspension by tailoring the idea of [9, 31]. The predefined  s s1 s2 s3 s4 d1 d2 d3 d4 z

suspension response of vehicle motions is retained, while the m1 zu1  Fs1  Fd 1  kt1 ( zu1  y1 )  u1

function approximators are not used. Thus, the computational m2 zu 2  Fs 2  Fd 2  kt 2 ( zu 2  y2 )  u2

efficiency and system operation safety can be enhanced. m3 zu 3  Fs 3  Fd 3  kt 3 ( zu 3  y3 )  u3 (1)
c) A dynamic simulator with realistic vehicle dynamics is m z  F  F  k ( z  y )  u
built in a commercial vehicle simulation software Carsim 8.1.  4 u4 s4 d4 t4 u4 4 4

Comparative simulation results are given to exemplify the  I   c( Fs1  Fs 3  Fd 1  Fd 3 )  d ( Fs 2  Fs 4  Fd 2  Fd 4 )  u

proposed control strategy and illustrate its superior responses. 
 I   b( Fs 3  Fs 4  Fd 3  Fd 4 )  a ( Fs1  Fs 2  Fd 1  Fd 2 )  u
The paper is structured as follows. The modeling of full-car
suspension systems and preliminaries are given in Section II. In (1), uz , u and u denote the lumped control forces,
Section III presents the AFC design and analysis of the which can be calculated in terms of the realistic control actions
closed-loop system stability. Simulation results are provided in ui . Hence, the formulation can be expressed as follows:
Section IV and Section V gives some conclusions. u z  u1  u2  u3  u4

II. MODELLING OF FULL-CAR AND PRELIMINARIES u  d (u2  u4 )  c(u1  u3 ) (2)

A. Modeling of Full-car Suspension System u  a(u1  u2 )  b(u3  u4 )
Then based on (2) and the assumption cu3  du4  0 used in
 3

[10, 32, 34, 35], the real control inputs ui , i  1 4 for active Hence, the dynamics of Fsi and Fdi can be denoted [10, 32] as
Fsi  Fsi  x1 x6 , x  , Fdi  Fdi  x1 x6 , x  , where x  [ x7 x14 ] .
suspension systems can be calculated based on the derived
three control actions uz , u and u as:
Therefore, the forces Fsi and Fdi are all bounded in a sufficient
u1  bdu z  du  (a  b)u   a  b  c  d   compact set in practice. In the proposed control, we do not

u2  cbu z  cu  (a  b)u   a  b  c  d   require accurate system dynamics, e.g. dynamics of springs and
 (3) dampers as required in [10, 32, 34, 35]. Instead, only the
u3   d  au z  u    a  b  c  d   vehicle motions x1 , , x6 and wheel base (i.e. the distance

u4  c  au z  u    a  b  c  d   between the wheels a, b, c, d ) are used in this paper, while the
unknown forces and system dynamics are all handled.
Hence, the objective is to design control for uz , u and u
based on (1), and then calculate the realistic control actions ui Remark 1: In most of existing active suspension control
based on (3). It should be noted that the full-car suspension designs, the forces of the springs and dampers in the suspension
model described by (1) is used for the stability analysis only, models (4) and (5) are usually assumed as linear functions,
and the exact force dynamics are not required for the control and/or their generation dynamics are fully known [10], which
implementation. Nevertheless, the idea proposed in this paper are stringent and unrealistic in practical applications. To
can be extended to other suspension systems with different address this issue, the unknown nonlinear dynamics generated
DOFs provided that a similar relationship as (3) can be found by the springs and dampers are considered in this paper. More
for force allocations. specifically, realistic forces embedded in Carsim are adopted in
To facilitate subsequent control designs, we first reformulate our case studies, whose generation models are unknown. This
full-car system (1) as a state-space model. Define the system implies that the proposed control can cover more realistic
applications, whilst requiring less information and reduced
states as x1  zs , x2  zs , x3   , x4   , x5   , x6   , x7  zu1 ,
modeling effort in the control synthesis.
x8  zu1 , x9  zu 2 , x10  zu 2 , x11  zu 3 , x12  zu 3 , x13  zu 4 , x14  zu 4 .
Remark 2: In the control design for systems with unknown
Then the system (1) can be rewritten as: nonlinearities and uncertainties, the unknown dynamics can be
 x1  x2 compensated by incorporating function approximators (e.g.
 NNs and FLSs) into the adaptive control implementation [20].
 x2  1   Fs1  Fs 2  Fs 3  Fs 4  Fd 1  Fd 2  Fd 3  Fd 4   1 u z
 M M However, the function approximation based control methods
x  x usually have complex structures and impose demanding
 3 4
computational costs. Specifically, it is still a nontrivial task to
 x  c  F  F  F  F  d  F  F  F  F    1 u
1
 4 I  d4 
s1 s3 d1 d3 s2 s4 d2  tune large amount of parameters used in the adaptive control
I
 methods [20]. Moreover, the potentially sluggish transient
 x5  x6 response (online training of NNs and FLSs requires fairly long
 time to reach convergence) also limits the practical application.
 x6  1 b  Fs 3  Fs 4  Fd 3  Fd 4  a  Fs1  Fs 2  Fd 1  Fd 2    1 u
 I   I

Remark 3: It is noted that the transient suspension response
(e.g. convergence rate, overshoot) during the first few seconds
(4)
is a critical issue, since too sluggish convergence or large
 x7  x8 overshoot may degrade the ride comfort, or even cause damage
 of suspension components. However, most of existing active
 x8  1  Fs1  Fd 1  kt1  x7  y1    1 u1
 m1   m
1
control designs (e.g. [9-12, 21]) can merely guarantee the
 steady-state motion behaviors, i.e. x j  0, j  1, 3, 5 , but they
x
 9  x10
cannot address the transient suspension behaviors explicitly.
 1 1
 x10   Fs 2  Fd 2  kt 2  x9  y2    u2 Remark 4: It is noted that the realistic active suspension
 m 2 m 2 systems are operated by current or pressure applied to actuators
 (5)
 x11  x12 (e.g. hydraulic or pneumatic actuators) to produce the required
 1 1 forces ui , i  1 4 . Since this paper mainly focuses on
 x12   Fs 3  Fd 3  kt 3  x11  y3    u3 presenting and validating a novel active suspensions control
 m3 m3
x  x method to deal with uncertain dynamics and unknown
 13 14 nonlinearities without using function approximators, the
 1 1
 x14  m  Fs 4  Fd 4  kt 4  x13  y4    m u4
actuator dynamics are not considered explicitly.
 4 4
Inspired by the above discussions, this paper will propose
where the parameters 1 M ,1 I and 1 I in (4) are bounded control designs for obtaining uz , u and u to accommodate
by positive constants 1min 3min and 1max 3max as the road excitations so as to maintain the vertical, roll and pitch
1min  1 M  1max , 2 min  1 I   2 max and 3min  1 I  3max . motions x j  0, j  1, 3, 5 (i.e. zs ,  and  ) of the vehicle body,
In practical vehicle suspension systems, the applied forces while both the predefined transient and steady-state suspension
are the functions of sprung/unsprung masses and tires motions. responses can be retained without using function approximator.
 4

B. Preliminaries The key idea of PPF control is to represent the condition (8)
To accomplish the stability analysis of the closed-loop into an equivalent “unconstrained” one as [26, 28] by
control system, the following preliminaries of the initial value introducing a coordinate transformation on the control error
problem [33] are briefly introduced: dynamics. For this purpose, we define  k  R as the
  t     t ,   ,   0    0   (6) normalized error and S  k   R is a smooth and strictly
with  : R    R being a continuous function and
n increasing function of  k , which satisfies the following
conditions:
  R n is a nonempty open set.
1)   S  k    ,  k  L
Definition 1 [33]: The solution of the initial value problem (6)
is maximal, if its solution   t  has no proper right extension. 2) lim S ( k )   , lim S ( k )  
 k   k 
Theorem 1 [33]: For the initial value problem (6), if   t ,   From the properties of S  k  , it can be verified that the
fulfills: a) locally Lipschitz on   t  for t  0 ; b) piecewise condition (8) equals to:
continuous on t for each   t    ; and c) locally integrable xk  t   k  t  S  k  (9)
on t for each   t    . Then, there exists a unique maximal
Hence, one can use the function S  k  given by:
solution   t  on the time interval 0, max  with
 max  {  , } such that   t    for t  0, max  .  e   e
k k

S ( k )  k  k
(10)
Proposition 1 [33]: If the conditions of Theorem 1 are true, e e
then for a maximal solution   t  on the time interval 0, max  Based on the facts that k 0  k   0 and S  k  is strictly
with  max   and any compact set    , there exists a increasing, we can calculate its inverse function as:
time instant t1  0, max  such that   t1    . 1  k 
 k  S 1  k   ln   (11)
III. ACTIVE SUSPENSION CONTROL AND ANALYSIS 2   k 

For completeness, we first briefly present the concept of the which defines a mapping from the unconstrained error xk to
PPF and error transformation to characterize the suspension the intermediate variable  k with the transformation (9). Then,
response of xk including the convergence rate and overshoot. the stabilization of the transformed error  k is sufficient to
Then, we will present the AFC schemes to regulate the vehicle guarantee the suspension of vehicle motions xk given in (4)
motions xk with the predefined error constraints. with the prescribed bound given in (8). In this sense, as stated in
[26, 28], the tracking control with a predefined error bound (8)
A. Prescribed Performance Function can be reduced to retain the boundedness of the transformed
To guarantee the suspension of vehicle motions x j within a error (11). Hence, in the next section, the transformed variables
predefined bounded region, we choose the following positive (11) will be used in the AFC design.
decreasing function k  t  : R   R  as the PPF [26-28]:
C. AFC Controller Design for Vertical Motion
k  t   k 0  k   e ak t  k ，k  1,,6 (7)
The control design objective is to regulate the vertical, roll
where k 0  k   0 and ak  0 are positive constants set to and pitch motions x1 , x3 , x5 (i.e. zs ,  and  ) governed by (4)
predefine the initial error, ultimate error and convergence speed, into the predefined boundary (8). For this purpose, we will
respectively. Clearly, the following facts can be verified: present a new control design, where the PPF (7) and error
1) k  0   k 0  k   e ak 0  k   k 0 ; transformation (11) will be utilized throughout the following
developments and analyses.
2) lim  k  0    k 0 , lim  k  t    k  .
t 0 t  We first present the control design for vertical motion x1 ,
Then we can use k  t  to construct predefined boundaries, while the controllers for the roll and pitch motions x3 , x5 can be
within which the system states xk can be retained. This can be obtained similarly.
formulated as: Step 1: For the vertical motion dynamics given in (4), we first
k  t   xk  t   k  t  , t  0 (8) define the suspension error as x1 , and then the normalized error
where  is a positive constant chosen by the designers to fulfill is derived as  1 (t )  x1 (t ) 1 (t ) with a PPF defined in (7) as:
the initial condition k 0  xk  0  . 1  t   10  1  e a1t  1 (12)
Remark 5: In the PPF formulation given in (7) and (8), ak where 10 , 1 and a1 are set as positive constants such that
represents the convergence rate; k  defines the ultimate the initial condition x1 (0)  10 is fulfilled.
steady-state error; k  0  and k  0  account for the lower Then, one can construct the first virtual control as:
bound of the undershoot and upper bound of the overshoot [26, k1     1 
28]. In this sense, both the transient and steady-state u1  k11  ln   (13)
2    1 
performance can be a priori designed by tuning the parameters
ak ,  , k 0 ,  k  . where k1  0 is a control gain, and the transformed error is
derived along with (11) as:
B. Error Transformation
 5

1    1  conditions x5 (0)  50 and e3 (0)  60 are fulfilled.

1  S 1  1   ln   (14) Hence, as explained in (3), we can calculate the realistic
2    1  control inputs ui , i  1, 2,3, 4 applied on the practical
Step 2: By using the obtained virtual control action, one can suspension system as long as the control actions uz , u and u
denote the virtual control error as: are obtained by using (17), (19) and (20).
e1  t   x2  t   u1  t  (15)
Remark 6: It is noted in the above control designs that the
Then the corresponding nominalized virtual control error is intermediate control laws u1 , u3 , u5 are similar to the virtual
given by  2 (t )  e1 (t ) 2 (t ) , where the second PPF 2  t  is control variables in the backstepping [10, 16]. However, unlike
given as: backstepping methods where the derivatives of the virtual
2  t   20  2  e a2t  2 (16)
controls needs to be obtained, the proposed AFC actions (17),
with 20 , 2 and a2 being properly selected positive (19) and (20) are with a proportional-like form of the transform
constants to guarantee the initial condition e1 (0)  20 . errors  k  S 1[ k ] for the normalized errors  k . Clearly, the
Then, the required control action u z for the vertical motion proposed controller has a simpler structure, making its practical
can be described by: implementation easier. Moreover, the ‘explosion of complexity’
k  2  issue encountered in the backstepping methods is also avoided.
u z   k2 2   2 ln   (17) Remark 7: As shown in (17), (19) and (20), the AFC does not
2   2 
require any function approximators, while the unknown
with k2  0 being a constant control gain, and the transformed dynamics in the suspension system (4) can be accommodated
error  2 of  2  t  is calculated by: effectively. Hence, this control has reduced computational
1  2  costs, and avoids the sluggish online learning procedure. This is
 2  S 1  2   ln   (18) clearly different to the well-known function approximation
2   2 
based adaptive control methods, where NNs or FLSs are used to
D. AFC Controller Design for Roll and Pitch Motions estimate and compensate the unknown nonlinearities.
Remark 8: In the proposed AFC method, the parameters to be
The control actions u and u for the roll and pitch motions
tuned can be classified into two categories, i.e. PPF parameters
x3 and x5 can be designed following similar manipulations as k 0 , k  , ak ,  and control gains kk . The parameter tuning of
that for vertical motion x1 , which can be briefly given as: the AFC is more straightforward than adaptive control schemes
1) The control action u of roll motion x3 is designed as: and thus easy to implement for practitioners. The required
k4     4  initial conditions x1 (0)  10 , e1 (0)  20 , x3 (0)  30 ,
u   k4 4   ln   e2 (0)  40 , x5 (0)  50 , e3 (0)  60 can be satisfied by
2   4 
choosing large PPF parameters  k 0 and  . The convergence
 4 (t )  e2 (t )  4 (t ) rate ak can be set small initially and then adjusted large to
e2  t   x4  t   u3  t  , (19) achieve fast error convergence. The ultimate error bounds k 
can be set large in the initial phase and then reduced to obtain
k3     3 
u3   k3 3   ln   smaller errors. On the other hand, the control gains kk are
2   3  related to the control error and the smoothness and amplitude of
 3 (t )  x3 (t ) 3 (t ) control actions, thus they can be set small initially and then
increased gradually to seek for a tradeoff between the control
where k3 , k4  0 are positive control gains, 3  t  and 4  t  response and the required control actions.
are PPFs defined in (7) with 30 , 3 ,  3 and 40 , 4 ,  4 The practical implementation of the proposed control
being properly selected positive constants to fulfill the initial method can be described as the following algorithm:
conditions x3 (0)  30 and e2 (0)  40 .
2) The control action u of the pitch motion x5 is designed as: Algorithm: Online Implementation of the proposed AFC
1: Initialize the PPF parameters k , k  1, , 6 for vertical,
k6     6 
u   k6 6   ln   roll and pitch motions;
2   6  2: Calculate the suspension errors ek (t ) according to the
 6 (t )  e3 (t ) 6 (t ) measured variables x1 , , x6 ;
e3  t   x6  t   u5  t  (20) 3: Define the normalized errors  k as:
k5     5   k  ek  t  k (21)
u5   k5 5   ln  
2   5  where k is the PPF defined in (7).
 5 (t )  x5 (t ) 5 (t ) 4: Calculate the transformed errors  k as:
where k5 , k6  0 are positive control gains, the PPFs 5  t  1  k 
 k  S 1  k   ln   (22)
and 6  t  are defined in (7) with properly selected positive 2   k 
constants 50 , 5 ,  5 and 60 , 6 ,  6 such that the initial
 6

5: Design the control signals as: functions with respect to time and its coordinates, the function
kk     k   (t ,  ) given in (28) is piecewise continuous on time t , locally
uk   k k  k  
ln   (23) Lipschitz on  within a nonempty set   ( ,  )  ( ,  ) .
2   k 
Hence, the main results of this paper can be given as:
6: Allocate the control signals uz  k2 2 , u  k4 4 ,
Theorem 2: For active suspension system (4), consider the
u  k6 6 based on (3) and obtain the real control actions
AFC (17) with (13) is designed for the vertical motion x1 under
u1 , , u4 to be applied on the system.
the initial conditions ek (0)  k 0 , then all signals in the
7: Go back to step 2.
vertical dynamic system are bounded, and x1 is retained within
E. Stability and Convergence Analysis the prescribed bound defined in (8).
The stability of the closed-loop control system and the Proof: The proof is shown in the Appendix ◇
convergence of suspension displacements will be examined in It is noted that similar analysis results can also be claimed for
this section. As shown in (4), the vertical, roll and pitch motions the roll motion x3 and pitch motion x5 , which implies that the
zs ,  and  and the proposed corresponding controllers have suspension responses of x3 and x5 can be retained within the
similar structures. Hence, we will provide detailed analysis for bound 3 (t )  x3 (t )  3 (t ), 5 (t )  x5 (t )  5 (t ) .
the vertical motion x1 , x2 only. Similar analysis can be
conducted for the roll motion x3 , x4 and pitch motion x5 , x6 , IV. SIMULATIONS
which will not be detailed due to the page limit. In this section, numerical simulations with two different road
Before presenting the convergence property of the proposed profiles are carried out to validate the effectiveness of the
control system, we first derive the dynamics of the adopted proposed AFC method. To cover more realistic vehicle
transformed errors. Consider the definition of virtual control dynamics, we have built a dynamic simulator by using
errors and the corresponding normalized errors  k  t  , one can professional vehicle simulation software Carsim○R and Matlab○R .
obtain from (13)-(20): For the purpose of demonstration, an E-SUV type vehicle
x1   11 , x2   2 2  u1， model embedded in Carsim○R is used. The full-car suspension
x3   33 , x4   4 4  u3 , (24) dynamics are generated by Carsim○R based on the embedded
realistic vehicle data, which are all unknown and thus not
x5   55 , x6   66  u5 . required in the control implementation, and the proposed AFC
By using (24) and the explanations below (4), the governing method is implemented in Simulink/Matlab. The control
dynamics of Fsi and Fdi in the suspension system (4) can be signals and the suspension system states are communicated
reformulated as: between these two softwares in real time. The structure of the
developed simulator can be found in Fig. 2.
Fsi  x1 x6 , x   Fsi  11 ,  2 2  u1 ,  33 ,  4 4  u3 ,  55 ,  66  u5 , x 
Fdi  x1 x6 , x   Fdi  11 ,  2 2  u1 ,  33 ,  4 4  u3 ,  55 ,  66  u5 , x  2013MATLABb
SIMULINK
(25) Control Strategy
Based on the definition of the normalized errors  1 ,  2 , their U1 U4
Vehicle Dynamics

derivatives are calculated along (24) and (13)-(17) with (4) as: Import

d ( x1 1 ) 1
Channels
1
1    x1   11  =  2 2  u1   11 
dt 1 1 (26)
  1 (t ,  1 ,  2 ) Road Condition Signal Response Model Type

d (e2  2 ) 1 1
2    e2   2 2    x2  u1   2 2 
dt 2 2
Export

1 1  4
Channels

=    Fsi  11 ,  2 2  u1 ,  33 ,  4 4  u3 ,  55 , x1 x6

2  M  i 1 Fig.2 Diagram of the closed-loop dynamic simulator.
4
 66  u5 , x    Fdi  11 ,  22  u1 ,  33 ,  4 4  u3 , (27) It is noted that in this developed simulator, the dynamics of
i 1
full-car suspension system (1) are generated from Carsim○R and
 55 ,  66  u5 , x   u z   u1   2 2  their accurate models are unknown. Only the system states
  2 (t ,  1 ,  2 ,  3 ,  4 ,  5 ,  6 ) x1 , x2 , , x6 that can be measured by using the configured
From (26)-(27), the dynamics of the normalized error vector sensors are required in this paper. In this sense, the plant to be
controlled can be regarded as a “black box”. This requirement
  [ 1 ,  2 ,  3 ,  4 ,  5 ,  6 ]T are represented in a compact form:
is clearly less stringent than most of existing active suspension
 (t ,  1 ,  2 ) 
 (t )   (t ,  )   1  (28) control results, which assume fully known suspension system

 2 (t ,  1 ,  2 ,  3 ,  4 ,  5 ,  6 
) dynamics. Hence, this control method can reduce the modeling
Since the induced forces in (4) are assumed to be continuous effort as it is a model-independent control structure.
 7

The major parameters of the studied E-SUV type vehicle are may be larger than the passive scheme as shown in Fig.5.
listed in TABLE I, which are taken from Carsim○R . It is worth Moreover, there are significant couplings between the vertical,
mentioning that only parts of these parameters are used in the pitch and roll dynamics, which lead to asymmetric behaviors in
AFC strategy in comparison to the backstepping control (BSC) the deflections as shown in Fig.5, which may also stem from
the asymmetric forces applied on the left and right tires.
[10, 16], i.e. only the distances between the wheels a, b, c, d are
used to allocate control signals in (3). For comparison, passive
suspension scheme embedded in the Carsim○R is also tested,
where the corresponding parameters, i.e., stiffness coefficients
of spring ksi (i  1, , 4) and damper kdi (i  1, ,4) , are also
listed in Table I.
TABLE I PARAMETERS FOR E-SUV TYPE CAR MODEL
Symbol Value Symbol Value
M 1590 kg c 0.7875 m
I 894.4 kgm2 d 0.7875 m
I 2687.1 kgm2 V 40 km/h
mi 120 kg ksi 46 N/mm
a 1.18 m kdi 30000 Ns/m
Fig.3 Ground road elevations of the combined bump road.
b 1.77 m

In the following, two driving road profiles embedded in

Carsim○R are chosen to verify the active suspension system.
Scenario 1 (Bump road with example roughness): Unlike
most of existing suspension results that only use bump road
condition to test the effectiveness of active suspension control
system, this paper incorporates the example road roughness
into a bump road to conduct a more realistic test driving road
condition ( 3.5 cm high and 40 cm long). The generated road
profile is shown in Fig.3.
In the simulations, the initial suspension system states are
xk  0   0, k  1, ,14 . Following the aforementioned parameter
tuning guidelines in Remark 8, the PPFs’ parameters are set as:
1  t    0.1  0.04  e10t  0.04, 2  t   10000  6000  e10t  4000, Fig.4 Vehicle motions and accelerations under bump road.
3  t  =  5  2  e15t  2, 4  t    5000  3000  e20t  3000, 5  t 
=  6  3 e22t  3 and 6  t   10000  6500  e20t  6500 . The
feedback control gains are set as k1  7000, k2  9000, k3 
7500, k4  12000, k5  5100 and k6  7400 . Simulation results
are given in Figs.4-5. In Fig.4, the responses of displacements
and accelerations of the vehicle motions in the vertical, roll and
pitch directions are provided. One can find from Fig.4 that the
proposed active suspension control system has lower peaks and
less fluctuations for both the vehicle motion displacements and
accelerations in comparison to the passive suspension. This fact
indicates that the proposed AFC method can isolate the
vibrations transmitted to the vehicle body from irregular road
roughness effectively to maintain the stability of vehicle body,
which contributes to improving the ride comfort. The profiles
of suspension deflections for four active suspension systems
are shown in Fig.5, where all the deflections are within the Fig.5 Profiles of suspension deflections.
allowable range 0.1 m, which help to retain the driving safety.
It is noted that there are inherent conflicts between the ride Scenario 2 (Bounce sine sweep road): To further validate the
comfort and suspension deflections, which means that better efficiency and robustness of the AFC method, a more oscillated
ride comfort may result in larger suspension deflection. In the road condition, bounce sine sweep road, is selected, which is
control design presented in this paper, we choose the ride given in Fig.6. Considering the fact that the bounce sine sweep
comfort as the primary target. Hence, the suspension deflection road is more aggressive (i.e. with fast varying frequency and
 8

amplitude) than bump road, the corresponding PPF boundaries

can be further modified to enhance the control response as:
1  t   1.2  0.5 e2.5t  0.5, 2  t  = 8000  5000  e15t  5000,
3  t   15  10  e3t  10, 4  t    8000  5000  e 1.5t  5000,
5 (t )  15  10  e2.5t  10 and 6  t   8000  6000  e2t  6000 .
The feedback control gains are set as k1  4000, k2  7880,
k3  1700, k4  4830, k5  1482 and k6  4790 .
Figs.7-10 give comparative simulation results. It is shown in
Fig.7 that the introduced AFC scheme achieves fairly good
suspension performance under the bounce sine sweep road
condition compared with the BSC method and passive
suspension. Specifically, the vibrations of the vehicle body in
the vertical, roll and pitch directions are significantly mitigated
with the AFC method, and both the transient and steady-state
suspension errors are the smallest among the three tested Fig.8 Acceleration responses of vehicle motions.
controllers. To evaluate the ride comfort of different suspension To further quantitatively exemplify the performance of the
schemes, the acceleration signals that have been well proposed AFC method, three commonly used performance
recognized as a ride comfort performance index are also indices: Integral Absolute Error (IAE), Root Mean Square
depicted in Fig.8. One can find from Fig.8 that both the AFC (RMS) and Maximum (MAX), are calculated for the three
method and BSC method can reduce the amplitude of the different controllers. Comparative results for the three control
accelerations of the vehicle body in the vertical, roll and pitch strategies under the sine sweep road condition are shown in
directions compared with passive suspension. In particular, the Fig.9. From Fig.9, it can be clearly found that the proposed
proposed AFC yields lower peaks than the BSC method. This AFC method yields the smallest index values among the three
fact indicates that the AFC scheme provides improved ride given performance indices, which means that the AFC provides
comfort over the BSC method. superior performs over the BSC method and passive suspension
in terms of suspension motion behaviors. This again implies
that the AFC has better capability to isolate the vehicle motions
as much as possible from the road-induced shocks.
Performance Indices
Passive BSC AFC
36% 40 26%
1.4 50%
1 41%
39 39
77% 77%
0.9 20 29
0.5 23
0.7
8.9 8.9
0 0
Vertical Roll Pitch
40

20% 35% 39% 51% 36%

0.2 45%
20 29
0.2 22
0.16 18
0.13 14 14 12
0 0
Vertical Acceleration Roll Acceleration Pitch Acceleration
300
2 35% 45% 34%
200 44% 61% 61%
2 215
193
1.3 100
Fig.6 Ground road elevations of sinusoid road. 1.1 120
83
127
127
0 0
Vertical Acceleration Roll Acceleration Pitch Acceleration

Fig.9 Performance evaluation of different methods.

Moreover, since the suspension strokes have certain effects
on the driving safety, it must be limited within a resaonalbe
bound especially in the presense of aggressive riding road
conditon. Simulation results of the suspension deflections with
sine sweep road excitation can be observed from Fig.10. As
shown in Fig.10, the suspension stroke responses with both the
AFC and BSC appraoches are all guaranteed within the
allowable bounded set. Furthermore, the proposed AFC again
can obtain less fluctuationss and lower peaks than BSC method
and passive suspension, in particular in the high-frequency
regimes. Comparative results of control inputs with the
proposed AFC method and the BSC method are plotted in
Fig.11. One may find from Fig.11 that although the control
Fig.7 Vehicle motions under sinusoid road condition. inputs of both methods are bounded, the AFC requires less
forces and provides smoother control signals (i.e. less
 9

fluctuations) compared to the BSC method, i.e. it requires less is carried out, where the simulation parameters used in Scenario
control power due to the use of the prescribed performance 2 are also adopted for Scenario 1 since the required initial
function that leads to better suspension responses. This conditions can be fulfilled for both road conditions. It can be
advantage is preferable in practical suspension system designs.
found from Fig.12 that both the transient and steady-state
convergence can be retained for both road conditions, though
the well-tuned parameters can contribute to better control
performance as shown in the above simulations.
Vertical Motion
2
Scenario1 PPF Boundary Scenario2

Zs(m)
0

-2
20
Roll Motion

10

 (rad)
0

-10
-20
Pitch Motion
20

10
Fig.10 Comparative results of suspension deflections.

 (rad)
0

-10

-20
0 2 4 Time (s) 6 8 10

Fig.12 Control performance with same parameters for different

road conditions.
To justify the implementation of the AFC in practical
applications, the required variables of the proposed AFC
method and the BSC method are summarized in Table III. For
the BSC method, the vehicle forces Fdi , Fsi , inertial variables
M , I , I , vehicle motions x1 , , x6 and wheel base (i.e. the
distance between the wheels a, b, c, d ) are all required to be
measureable and available, while for the proposed AFC, only
the vehicle motions and wheel base are required in the control
Fig.11 Control inputs of the proposed AFC and BSC method. implementation. In this respect, the proposed AFC may be
Moreover, to illustrate the computational efficiency of the more suited for practical active suspension systems due to its
proposed method, the computational time of simulations for the reduced modeling effort and less sensors.
three control methods under two different road conditions TABLE III REQUIRED VARIABLES FOR TWO CONTROLLERS
(Scenario 1 and Scenario 2) is provided in Table II. One can
AFC Method BSC Method
find from Table II that the computational time of the AFC is
Forces — Fdi , Fsi
dramatically decreased for both road conditions compared with
adaptive control given in [36] and the BSC method. This is Inertial Variables — M , I  , I
reasonable since the online training of the NNs weights Vehicle Motions xi , i  1, ,6 xi , i  1, ,6
involved in the adaptive control and the repeated calculation of Wheel Base a, b, c, d a, b, c, d
the derivatives of virtual control actions in the BSC method are
all avoided in the proposed AFC method. It should be noted that the control designs for the vertical,
Table II COMPARATIVE RESULTS OF COMPUTATIONAL TIME pitch and roll motions of the full-car suspension system
(SCENARIO 1/SCENARIO 2) presented in this paper have been derived separately as shown
Computational in (17), (19) and (20), and then the obtained control actions are
Methods Decreasing ratio allocated based on (3). The motion dynamics of each DOF of
time (s)
the sprung mass are indeed the same as those of quarter-car
Adaptive control 15.83 s/19.1 s Benchmark suspension systems studied in the literature, e.g. [8-9] and
BSC method 10.76 s/11.5 s 68.1%/66.1%↓ references therein. In this sense, the proposed AFC scheme can
be applied to quarter-car suspension systems, where its validity
AFC method 4.46 s/4.3 s 71.8%/77.5%↓ and effectiveness can also be demonstrated.
Finally, to demonstrate the effectiveness of the proposed V. CONCLUSION
AFC method and the generality of using same control This paper presents a new control design for full-car active
parameters under different road conditions, an extra simulation suspension systems with unknown nonlinear dynamics. Unlike
 10

the existing results, the proposed control does not use any We select a Lyapunov function as V1  12 2 , and calculate
function approximation, whilst the unavoidable uncertainties its time derivative along (29) as:
and nonlinearities in the suspension systems can be tackled V1  1  1r1 ( 22   11  k11 ) (30)
effectively. This control has a simple proportional-like
Consider  k (t )   , t  [0, max ) and the facts that 1 , 2
structure with less requirements on the system and improved
computation efficiency. Moreover, both the predefined are all bounded, then by recalling the Extreme Value Theorem
transient and steady-state suspension motion bounds are strictly [33], we can verify that
guaranteed by incorporating PPFs into the control design.  22   11  F1 , for t   0, max  (31)
These features make it more suitable for practical application. is true for a positive constant F1  0 . Thus, it follows from (30)
The stability of the closed-loop control system is rigorously -(31):
V1  1 rM 1  F1  k1 1 
proved by using Extreme Value Theorem and Lyapunov
Theorem. A dynamic simulator consisting of commercial (32)
software Carsim ○R and Matlab/Simulink has been built to Consequently, one can conclude that V1 is negative when
conduct comparative simulations, where a realistic E-SUV type 1  F1 / k1 . Then based on the Lyapunov theorem, the
full-car model embedded in Carsim ○R is utilized. Simulation variable 1 will ultimately converge to a set defined by
results show the efficacy and superior performance of the 1 : 1 | 1   M 1  max{ 1 (0) , F1 / k1 } for t  [0, max ) .
proposed control over several other suspension methods. Moreover, the boundedness of 1 can be verified from (29),
Future work will focus on the validation of the proposed AFC
and the derivative of the first virtual control u1 is also
schemes on practical active suspension systems including the
bounded. Then it can be claimed that 1 , 1 , u1 , u1  L for
actuator dynamics.
t  [0, max ) .
APPENDIX On the other hand, consider the relationship between 1 and
Proof: The proof follows a similar procedure as [31], which  1 given by (11), we have
consists of three steps. We first prove the existence of the   1
e 2  1
(33)
maximal solution  (t ) of (28) over the set  for a time   1
interval [0, max ) . Then, we prove the boundedness of the Consequently, it can be derived that
closed-loop system signals with the proposed controls (13) and e2 M 1    e 2 M 1   
  2 M 1   1 (t )  2 M 1  (34)
(17) for all t  [0, max ) . We finally prove that  max   holds e 1 e 1
for all  (t )   , t  [0, max ) . Therefore, we know   x1 / 1   is true, which indicates
Step 1. It has been stated that the PPF parameters can be set to 1 (t )  x1 (t )  1 (t ) , i.e. the vertical displacement x1 is
guarantee the initial conditions x1 (0)  10 , e1 (0)  20 .
retained within the bound (8) for t  [0, max ) .
Then the fact  k (0)   , k  1, 2 is true, i.e.  (0)   .
Following the above analysis, we set a Lyapunov function
Moreover, since the suspension dynamics in (4) and the PPFs
V2   22 2 , then calculate V2 from (29) as:
k (t ) are continuous with respect to their coordinates, the
1  4
function  in (28) is piecewise continuous, locally Lipschitz V2   2 r2    Fsi  11 ,  2 2  u1 ,  33 ,  4 4  u3 ,  55 ,
on  over the set  , and piecewise continuous on t for each  M  i 1
fixed  (t )   . Hence, based on Theorem 1, there exists a 4

maximal solution  (t ) of (28) on a time interval t  [0, max ) ,  66  u5 , xi    Fdi  11 ,  2 2  u1 ,  33 ,  4 4  u3 ,
i 1 (35)
such that  i (t )  ( ,  ), i  1, 2, t  [0, max ) .
1 
Step 2. To prove the boundedness of all closed-loop control  55 ,  66  u5 , xi   k2 2  u1   22 
system signals for all t  [0, max ) with the proposed control,
M 
we will calculate the time derivative of  k , k  1, 2 along (14)   2 rM 2  F2  k21min  2 
and (18) with (26) and (27) as: Similar to the arguments given in (31), we know that
S 1  i , i , i , u3 , u1 , i  1, 2 are all bounded for t  [0, max ) .
1   1  r1 ( x1   11 )=r1  2 2  k11   11  Thus, there exists a positive constant F2  0 such that
 1
S 1 1  4 1  4
2   2 =r2    Fsi  11 ,  2 2  u1 ,  33 ,  4 4  u3 ,  55 ,   Fsi  11 ,  2 2  u1 ,  33 ,  4 4  u3 ,  55 ,  6 6  u5 , xi  
 2  i 1 M  i 1
M
4

 F   ,    u ,   ,    u ,   ,    u , x   u     F
4
 66  u5 , x    Fdi  11 ,  2 2  u1 ,  33 ,  4 4  u3 , i 1
di 1 1 2 2 1 3 3 4 4 3 5 5 6 6 5 i 1 2 2 2

For t  0, max  , which can be validated by recalling the

i 1

 55 ,  66  u5 , x   u z   u1   2 2 
Extreme Value Theorem and the fact  2 (t )   . Therefore, we
(29) conclude from (35) and the facts 0  r2  rM 2 and
where the variables r1  1/ [1 (   12 )] and r2  1/ [2 (   22 )] are
1min  1 M  1max that V   r  F  k    , which
bounded by positive constants rMk  0, k  1, 2 , i.e. 2 2 M2 2 2 1 min 2

implies that  2 ultimately converges to the set defined by

0  rk  rMk is true [28].
   |     max   (0) , F  k   for t  [0,  max ) .
2 2 2 M2 2 2 2 1 min
 11

Thus, the control signal u z and the state variable x2 are all [14] H. Chen, P.-Y. Sun, and K.-H. Guo, "A multi-objective control design for
bounded for t  [0, max ) . Moreover, the boundedness of  2 active suspensions with hard constraints," in Proceedings of the 2003
American Control Conference (ACC), 2003, pp. 4371-4376.
and u z can be easily verified from (29) and (17). Again, [15] L. Zuo, J.-J. Slotine, and S. A. Nayfeh, "Model reaching adaptive control
similar to (33), we know e2  (1   2 ) / (1   2 ) , and thus it
2 for vibration isolation," IEEE Trans. Control Syst. Technol, vol. 13, pp.
can be derived that 611-617, 2005.
[16] J.-S. Lin and C.-J. Huang, "Nonlinear backstepping active suspension
e2 M 2    e 2 M 2    design applied to a half-car model," Veh. Syst. Dyn, vol. 42, pp. 373-393,
  2 M 2   2 (t )  2 M 2  (36)
e 1 e 1 2004.
[17] C. Gohrle, A. Schindler, A. Wagner, and O. Sawodny, "Design and
Therefore, the fact 2 (t )  e1 (t )  2 (t ) is also true for vehicle implementation of preview active suspension controllers," IEEE
t  [0, max ) . Trans. Control Syst. Technol, vol. 22, pp. 1135-1142, 2014.
[18] H. Li, J. Yu, C. Hilton, and H. Liu, "Adaptive sliding-mode control for
Step 3. The last step will validate that  max   is true. nonlinear active suspension vehicle systems using T–S fuzzy approach,"
Equations. (34) and (36) show that the condition IEEE Trans. Ind. Electron, vol. 60, pp. 3328-3338, 2013.
 (t )   , t  [0, max ) holds for a nonempty open set [19] H. Li, X. Jing, H.-K. Lam, and P. Shi, "Fuzzy sampled-data control for
uncertain vehicle suspension systems," IEEE Trans. Cybern, vol. 44, pp.
2
 e  Mk    e Mk     1111-1126, 2014.
     Mk ,  . Consequently, we can easily [20] M. Wang, B. Chen, and P. Shi, "Adaptive neural control for a class of
k 1  e  1 e Mk  1  perturbed strict-feedback nonlinear time-delay systems," IEEE Trans.
Syst., Man, Cybern. B, Cybern, vol. 38, pp. 721-730, 2008.
verify that    for   ( ,  )  ( ,  ) . If we assume [21] Y.-J. Liu and S. Tong, "Adaptive fuzzy control for a class of unknown
 max   is true, Proposition 1 implies that there exists a nonlinear dynamical systems," Fuzzy Sets and Systems, vol. 263, pp. 49-70,
2015.
finite time t1  [0, max ) such that  (t1 )   . This claim leads [22] J. Na, X. Ren, and D. Zheng, "Adaptive control for nonlinear
to a contradiction. Therefore, we can conclude that  max   is pure-feedback systems with high-order sliding mode observer," IEEE
Trans. Neural. Netw, Learn. Syst, vol. 24, pp. 370-382, 2013.
true. Thus all the closed-loop signals in control system are [23] R. Cui, C. Yang, Y. Li, and S. Sharma, "Adaptive neural network control
bounded for t  0 . Specifically, the vertical motion x1 is of AUVs with control input nonlinearities using reinforcement learning,"
retained within the predefined bound (8) for t  0 . This IEEE Trans. Syst., Man, Cybern., Syst., vol. 47, pp. 1019-1029, 2017.
[24] B. Xu, C. Yang, and Z. Shi, "Reinforcement learning output feedback NN
completes the proof. ◇ control using deterministic learning technique," IEEE Trans. Neural. Netw,
Learn. Syst, vol. 25, pp. 635-641, 2014.
REFERENCES [25] J. Na, M. N. Mahyuddin, G. Herrmann, X. Ren, and P. Barber, "Robust
adaptive finite ‐ time parameter estimation and control for robotic
[1] D. Hrovat, "Optimal active suspension structures for quarter-car vehicle systems," Int. J. Robust Nonlin. Control, vol. 25, pp. 3045-3071, 2015.
models," Automatica, vol. 26, pp. 845-860, 1990. [26] C. P. Bechlioulis and G. A. Rovithakis, "Robust adaptive control of
[2] T. D. Le, M. T. N. Bui, and K. K. Ahn, "Improvement of Vibration feedback linearizable MIMO nonlinear systems with prescribed
Isolation Performance of Isolation System Using Negative Stiffness performance," IEEE Trans. Autom. Control, vol. 53, pp. 2090-2099, 2008.
Structure," IEEE/ASME Trans. Mechatron, vol. 21, pp. 1561-1571, 2016. [27] C. P. Bechlioulis and G. A. Rovithakis, "Adaptive control with guaranteed
[3] X. Tang, H. Du, S. Sun, D. Ning, Z. Xing, and W. Li, "Takagi–Sugeno transient and steady state tracking error bounds for strict feedback
fuzzy control for semi-active vehicle suspension with a systems," Automatica, vol. 45, pp. 532-538, 2009.
magnetorheological damper and experimental validation," IEEE/ASME [28] J. Na, Q. Chen, X. Ren, and Y. Guo, "Adaptive prescribed performance
Trans. Mechatron, vol. 22, pp. 291-300, 2017. motion control of servo mechanisms with friction compensation," IEEE
[4] H. Zhang, X. Zheng, H. Yan, C. Peng, Z. Wang, and Q. Chen, "Codesign Trans. Ind. Electron, vol. 61, pp. 486-494, 2014.
of Event-Triggered and Distributed H∞ Filtering for Active Semi-Vehicle [29] S. Wang, J. Na, and X. Ren, "RISE-based asymptotic prescribed
Suspension Systems," IEEE/ASME Trans. Mechatron, vol. 22, pp. performance tracking control of nonlinear servo mechanisms," IEEE
1047-1058, 2017. Trans. Syst., Man, Cybern., Syst., vol. 48, pp. 2359 - 2370, 2018.
[5] H. Li, H. Liu, H. Gao, and P. Shi, "Reliable fuzzy control for active [30] J. Na, Y. Huang, X. Wu, G. Gao, G. Herrmann, and J. Z. Jiang, "Active
suspension systems with actuator delay and fault," IEEE Trans. Fuzzy Syst, adaptive estimation and control for vehicle suspensions with prescribed
vol. 20, pp. 342-357, 2012. performance," IEEE Trans. Control Syst. Technol, pp. 1-15, 2017.
[6] Y. Huang, J. Na, X. Wu, X. Liu, and Y. Guo, "Adaptive control of [31] C. P. Bechlioulis and G. A. Rovithakis, "A low-complexity global
nonlinear uncertain active suspension systems with prescribed approximation-free control scheme with prescribed performance for
performance," ISA Trans, vol. 54, pp. 145-155, 2015. unknown pure feedback systems," Automatica, vol. 50, pp. 1217-1226,
[7] H. Li, X. Jing, and H. R. Karimi, "Output-feedback-based H∞ control for 2014.
vehicle suspension systems with control delay," IEEE Trans. Ind. Electron, [32] W. Sun, H. Gao, and B. Yao, "Adaptive Robust Vibration Control of
vol. 61, pp. 436-446, 2014. Full-Car Active Suspensions With Electrohydraulic Actuators," IEEE
[8] W. Sun, H. Gao, and O. Kaynak, "Adaptive Backstepping Control for Trans. Control Syst. Technol, vol. 21, pp. 2417-2422, 2013.
Active Suspension Systems With Hard Constraints," IEEE/ASME Trans. [33] E. D. Sontag, Mathematical control theory. London, U.K.: Springer, 1998.
Mechatron, vol. 18, pp. 1072-1079, 2013. [34] M. Moradi and A. Fekih, "Adaptive PID-sliding-mode fault-tolerant
[9] Y. Huang, J. Na, X. Wu, and G. Gao, "Approximation-free control for control approach for vehicle suspension systems subject to actuator
vehicle active suspensions with hydraulic actuator," IEEE Trans. Ind. faults," IEEE Trans. Veh. Technol, vol. 63, pp. 1041-1054, 2014.
Electron, vol. 65, pp. 7258-7267, 2018. [35] S. Ikenaga, F. L. Lewis, J. Campos, and L. Davis, "Active suspension
[10] N. Yagiz and Y. Hacioglu, "Backstepping control of a vehicle with active control of ground vehicle based on a full-vehicle model," in Proceedings
suspensions," Control Eng. Pract, vol. 16, pp. 1457-1467, 2008. of the 2000 American Control Conference (ACC), 2000, pp. 4019-4024.
[11] S.-J. Huang and H.-Y. Chen, "Adaptive sliding controller with self-tuning [36] F. Zhao, S. S. Ge, F. Tu, Y. Qin, and M. Dong, "Adaptive neural network
fuzzy compensation for vehicle suspension control," Mechatronics, vol. control for active suspension system with actuator saturation," IET
16, pp. 607-622, 2006. Control Theory Appl, vol. 10, pp. 1696-1705, 2016.
[12] I. Fialho and G. J. Balas, "Road adaptive active suspension design using
linear parameter-varying gain-scheduling," IEEE Trans. Control Syst.
Technol, vol. 10, pp. 43-54, 2002.
[13] Y. Jia, "Robust control with decoupling performance for steering and
traction of 4WS vehicles under velocity-varying motion," IEEE Trans.
Control Syst. Technol, vol. 8, pp. 554-569, 2000.
 12

Jing Na (M’15) received the B.Eng. and Faculty of Mechanical and Electrical Engineering, Kunming
Ph.D. degrees from the School of University of Science and Technology. His current research
Automation, Beijing Institute of interests include modern signal processing theory and their
Technology, Beijing, China, in 2004 and applications on fault feature extracting, and internet-based
2010, respectively. mechanical fault diagnosis technology and expert systems.
From 2011 to 2013, he was a
Monaco/ITER Postdoctoral Fellow at the Guanbin Gao received the B.Sc. and M.Sc.
ITER Organization, degrees in mechanical engineering and
Saint-Paul-lès-Durance, France. From 2015 to 2017, he was a automation from Northeastern University,
Marie Curie Intra-European Fellow with the Department of Shenyang, China, in 2001 and 2004,
Mechanical Engineering, University of Bristol, U.K. Since respectively, and the Ph.D. degree in
2010, he has been with the Faculty of Mechanical and Electrical mechanical manufacturing and automation
Engineering, Kunming University of Science and Technology, from Zhejiang University, Hangzhou, China,
Kunming, China, where he became a Professor in 2013. He has in 2010.
coauthored one monograph and more than 100 international His current research interests include precision measuring &
journal and conference papers. His current research interests control, kinematics of industrial robots, and NNs.
include intelligent control, adaptive parameter estimation,
nonlinear control and applications for robotics, vehicle systems Guang Li (M’09) received the Ph.D.
and wave energy convertor, etc. degree in Electrical and Electronics
He is currently an Associate Editor of the IEEE Transactions Engineering, specialized in control systems,
on Industrial Electronics, the Neurocomputing, and has served from the University of Manchester, in
as the Organization Committee Chair of DDCLS 2019, 2007.
international program committee Chair of ICMIC 2017. Dr Na He is currently a senior lecturer of
has been awarded the Best Application Paper Award of the 3rd dynamics modelling and control with
IFAC International Conference on Intelligent Control and Queen Mary University of London, UK.
Automation Science (IFAC ICONS 2013), and the 2017 His current research interests include constrained optimal
Hsue-shen Tsien Paper Award. control, model predictive control, adaptive robust control and
control applications including renewable energies and energy
Yingbo Huang (M’19) received the B.Sc. storage, etc.
degree from Lanzhou City University,
Lanzhou, China, and the Ph.D degree from
the Faculty of Mechanical and Electrical
Engineering, Kunming University of
Science and Technology, Kunming, China,
in 2013 and 2019, respextively.
He is currently a lecture with Faculty of
Mechanical and Electrical Engineering, Kunming University of
Science and Technology, Kunming, China. His current
research interests include adaptive control and transient
performance improvement of nonlinear systesms with
application to vehicle suspension systems.

Qianqian Pei received the M.Sc. degree

from the Faculty of Mechanical and
Electrical Engineering, Kunming University
of Science and Technology, Kunming, China
in 2017.
She is currently a lecture with Faculty of
Mechanical and Electrical Engineering,
Yunnan Land And Resources Vocational
College, Kunming, China. Her current research interests
include adaptive control, parameter estimation and application.

Xing Wu received the B.Sc. and M.Sc.

degrees from Kunming University of
Science and Technology, China, in 1994 and
1997, respectively, and the Ph.D. degree in
mechanical engineering from Shanghai Jiao
Tong University, China, in 2005.
He is currently a Professor with the