You are on page 1of 10

Mathematics and Computers in Simulation 60 (2002) 369–378

Design of a non-linear hybrid car suspension system


using neural networks
Konstantinos Spentzas∗ , Stratis A. Kanarachos
Mechanical Engineering Department, National Technical University of Athens, Polytechnioupolis,
Building M, 15780 Athens, Greece

Abstract
A methodology for the design of active/hybrid car suspension systems with the goal to maximize passenger com-
fort (minimization of passenger acceleration) is presented. For this reason, a neural network (NN) controller is pro-
posed, who corresponds to a Taylor series approximation of the (unknown) non-linear control function and the NN is
due to the numerous local minima trained using a semi-stochastic parameter optimization method. Two cases A and B
(continuous and discontinuous operation) are investigated and numerical examples illustrate the design methodology.
© 2002 IMACS. Published by Elsevier Science B.V. All rights reserved.
Keywords: Hybrid car suspension; Neural networks; Semi-stochastic optimization

1. Introduction

The basic idea of an active/hybrid suspension system is to use an actuator (e.g. a hydraulic cylinder)
to increase the dynamic performance of a car, e.g. stability, comfort, etc., beyond the performance of
a passive linear and sometimes non-linear (progressive spring) suspension system. Thus, it is generally
expected that the control of the actuator is non-linear and intelligent (e.g. flexible).
The synthesis of such controllers has been investigated in the past by a number of researchers using
non-linear control, as well as fuzzy-neural approaches. A survey of advanced suspension developments
and related optimal control applications are presented in [4].The most recent papers are listed in [1–6].
Nevertheless, active control is rarely realized. Reasons are the add-on cost and complexity of the active
suspension and the energy consumption.
In the present paper, a new methodological approach using neural networks (NNs) is presented. The
approach proposes a general type NNs non-linear controller based on a Taylor series approximation,
trained by a semi-stochastic optimization algorithm. The remaining part of this paper is organized as
follows. Section 2 contains a description of the car dynamics and the problem statement. The control

Corresponding author.
E-mail address: spentzas@central.ntua.gr (K. Spentzas).

0378-4754/02/$ – see front matter © 2002 IMACS. Published by Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 8 - 4 7 5 4 ( 0 2 ) 0 0 0 2 9 - 0
370 K. Spentzas, S.A. Kanarachos / Mathematics and Computers in Simulation 60 (2002) 369–378

problem is set up in Section 3. Results of numerical calculations are presented in Section 4, while the
conclusions are given in Section 5.

2. Mathematical model

The mathematical model used is the (well known) quarter-car model with passive–active suspension
system shown in Fig. 1. Wheel and axle (unsprung mass m2 ) is connected to the car body through a spring
(k)—damper (c)—actuator (f) suspension system, while the tire is modeled as a spring (k2 ). The car body
is represented by its mass m1 and the road disturbance w(t).
The equations of motion taking into account the dynamics of the actuator (time constant Tact ) are the
following:

m1 ẍ1 + c(ẋ1 − ẋ2 ) + k(x1 − x2 ) = f, m2 ẍ2 − c(ẋ1 − ẋ2 ) − k(x1 − x2 ) + k2 (x2 − w) = −f,
Tact f˙ + f = fc (1)

where fc is the control law to be defined. Other model values are m1 = 289 kg, m2 = 59 kg, k2 =
190,000 N/m, while the values for c, k and Tact are subject to design considerations. The above dynamic
system is restricted mainly by the constraint, with z = x1 − x2 denoting a new variable:
|z| ≤ z̄ (z = x1 − x2 ) (2)

As started, in the present paper we are concerned with the optimization of the ride qualities (acceleration
applied to the passengers) of the car. In this context, it is assumed that the car experiences a sudden
disturbance from the road (see [1]) a bump, described by the equation:

 A 1 − cos 8πt for t ≤ 0.25 s, A = 0–11 cm
w= 2 (3)
 0 for t > 0.25 s

The bump amplitude A varies between A = 0 and 11 cm, while the z̄ restriction, z̄ = 8 cm.

Fig. 1. Quarter-car model and typical bumps.


K. Spentzas, S.A. Kanarachos / Mathematics and Computers in Simulation 60 (2002) 369–378 371

3. Control problem

The control objective is to maximize the passenger comfort, respectively, to minimize the passenger
acceleration, under road disturbances.
The unconstrained problem (|z| not constrained) can be solved for T act = 0 immediately; ẍ1 = 0 is
achieved if
uunc = c(ẋ1 − ẋ2 ) + k(x1 − x2 ) (4)

which means that the resulting passive + active spring and damping constants (kres , cres ) become 0, while
x2 is equal to
 
k2
ẍ2,unc + Ω x2,unc = Ω w → x2,unc = w(1 − cos Ωt)
2 2
= Ω , w = const
2
(5)
m2
Eqs. (4) and (5) indicate now clearly that a control law that can cope with the constraint (2) must be
non-linear.
NNs offer a general powerful tool for designing such a law. The basic artificial neuron can be modeled
as a multi-input non-linear device (zi ) with weighted interconnections wji also called synaptic weights,
while the output (uj ) is described by the equation:
 N 

uj = ψ wji zi + θj (6)
i=1

In (6), ψ is generally a non-linear function, while θ j the external threshold, also called an offset or bias.
The strength of the NN formulation lies in the general structure of the control law and in the introduction
of synaptic weights. On the other hand, the generality of the formulation imposes the problem of defining
an appropriate structure for the controller.
With respect to the controller structure it is proposed in this paper to expand the (unknown) non-linear
law in a Taylor series:

fc ≡ uj (z, ż) = ψ[w1 z + w2 ż + w3 z2 + w4 zż + w5 ż2 + w6 z3 + w7 z2 ż + w8 zż2 + w9 ż3 + · · · ]


(7)

Eq. (7) represents a new methodological approach in applying NN in control problems and is mathemat-
ically accurate and expandable. In the case of a quadratic approximation, the NN is described by N = 9
synaptic weights w1 –w9 , to be determined by an appropriate procedure.
Training of NNs is equivalent with determining an efficient learning algorithm to adjust the weights so
as to minimize a specific objective function. Due to the numerous sub-optima associated with non-linear
control laws, a semi-stochastic parameter optimization procedure for determining the optimum weight
vector and an objective function utilizing penalty approach is used as the quantitative evaluator of the
system performance.
According to the semi-stochastic complex method, a “polyhedron composed of i = 1, . . . , kN + 1
vectors wi ”
wi ≡ { w1 w2 w3 w4 w5 w6 w7 w8 w9 }i (8)
372 K. Spentzas, S.A. Kanarachos / Mathematics and Computers in Simulation 60 (2002) 369–378

with k ≥ 3, is created, forming a regular “complex”. The kN vectors wi of the initial complex are computed
using a Gauss distribution with a pre-selected standard variation σ and the +1 vector is deterministically
selected and defined as ‘base’ vector. The method begins by evaluating the value of the objective function
F(wi ) for each one of the vectors of the initial complex. The vectors with the highest F (ww ) and the
lowest F (wb ) value of the objective function are located, named as “worst” (w) and “best” (b) vectors,
respectively. Then, a “centroid vector”, w̄

1 
kN
w̄ = wi (i = w) (9)
kN i=1

is evaluated excluding ww and a (new) “reflection vector” wR is computed to the opposite direction of
ww considering the centroid (Eq. (9)) as the origin of the space (R = constant):
wR = w̄ + R · (w̄ − ww ) (10)

If the value of the objective function F(wR ) is greater than F(wb ) and less than F(wi ), i = 1, . . . , kN + 1
(i = w), then the worst vector is substituted by the reflection vector wR . A new centroid is then created
and the procedure continues until the termination criterion is fulfilled. A typical termination criterion is
the convergence of the complex to “one” vector.
In case that the value of the objective function of the reflection vector F(wR ) is less than F(wb ), then
an “expansion” step takes place (E = constant)
wE = w̄ + E · (w̄ − ww ) (11)

else the “contraction” step follows (C = constant)


wC = w̄ + C · (ww − w̄) (12)

In the expansion step, a new vector is calculated in the opposite direction of the ww considering the
centroid as the origin of the space and through the contraction a new vector is evaluated in the same
direction of ww . The procedure converges towards a local or global minimum 0 wmin depending on the
initial vector and the properties of the complex (density factor k and standard deviation σ ). The procedure
can be repeated considering 0 wmin , as the new initial vector of a new complex, leading now to a new
1
wmin , etc., until confidence about the result of the optimization procedure is gained. A basic advantage of
the semi-stochastic complex method is that it combines semi-stochastic build up of the kN + 1-complex
and deterministic search.
The problem of minimizing f(w) subject to the constraints gj (w) ≤ 0 (j = 1, . . . , ng ) and hi (w) = 0
(i = 1, . . . , ne ) can now be formulated as an unconstrained optimization problem, where the complex
function F(w) needs to be minimized:
ng ne
 
F (w) = f (w) + λGj · G({gj (w)}+ ) + λHi · H (|hi (w)|) = miminum (13)
j =1 i=1

In the above expression, G is a function of the gj (w)+ , with “+” denoting that only the positive values
of gj (w) are considered (gj (w)+ = 0 for negative values of gj (w)), H a function of hi (w), while λGj and
λH i are positive “penalty” weights.
K. Spentzas, S.A. Kanarachos / Mathematics and Computers in Simulation 60 (2002) 369–378 373

4. Numerical results

Case A (continuous operation). It is assumed that the passive car dynamics are already defined, e.g.
c = 1000 Ns/m, k = 16,812 N/m, and that a non-linear controller, acting all the time (“continuous
operation”) has to be designed to increase the passenger comfort (see [1]).
In order to minimize passenger acceleration with the control law (7), the following goal and constraints
are defined:

f (w) = |ẍ1 |for A=5 cm = minimum, g1 (w) = |z| − 0.08m ≤ 0 for Amax = 11 cm,
g2 (w) = |z| − 0.08m ≤ 0, for Amean = 5 cm, g3 (w) = 5000 − k + w1 ≤ 0 N/m,
g4 (w) = 300 − c + w2 ≤ 0 Ns/m (14)
Eq. (14) indicates that the NN is trained for minimum acceleration for the mean value of A, Amean = 5 cm,
while the constraints g3 and g4 are added to ensure minimum stability. The training of the NN, using
semi-stochastic optimization with R = 1.3, E = 2 and C = 0.5, leads to the following values for the
synaptic weights:

wT = {1.17 × 104 697.25 − 8.75 × 104 − 27.180 415.12


− 4.79 × 106 − 6.28 × 105 5.27 × 103 − 825.86} (15)
Figs. 2 and 3 show the time traces for z, ẍ1 and f (for T act = 0) for two bump amplitudes, Amean = 5 cm
and Amax = 11 cm.
From the above figures it can be concluded, that the NN controller shows a satisfactory performance, as
it succeeds in fulfilling the constraint (2) for Amax = 11 cm with an (acceptable acceleration) of ẍ1,max =

Fig. 2. Continuous operation, road disturbance Amean = 5 cm.


374 K. Spentzas, S.A. Kanarachos / Mathematics and Computers in Simulation 60 (2002) 369–378

Fig. 3. Continuous operation, road disturbance Amax = 11 cm.

18.6 m/s2 , while the maximum acceleration of the passive system for Amean = 5 cm is reduced from
ẍ1,passive,max = 4.1 m/s2 to ẍ1,max = 2.8 m/s2 . Compared with [1], the present NN controller performs
better with respect to ride qualities, as he avoids for Amax = 11 cm high accelerations (ẍ1,max = 24.5 m/s2
in [1]).
The controller Case A is suitable only for manual on–off operation (e.g. to be activated by the driver if
the road is in a bad condition), as it makes no sense to have linear feedbacks w1 z and w2 ż in the control law
(and not change from the beginning the passive feedbacks kz and cż). As a matter of fact, the minimum
stability constraints g3 and g4 are extremely important for the NN training, as the maximum accelerations
depend on them. The NN controller has the trend to minimize static stability (k − w1 , c − w2 ⇒ 0) and to
become even unstable, in order to reduce the maximum passenger acceleration. Therefore, the controller
performance depends heavily on the stability margins g3 and g4 .
Case B (discontinuous operation). In the Case B it is assumed that the passive system is already designed
to satisfactorily fulfill the ride quality requirements for a range of bump amplitudes A = 0, . . . , Amean .
Therefore, the NN controller has to cope essentially with bump amplitudes A between Amean and Amax .
If such a case is given by c = 500 Ns/m and k = 10,000 N/m, the structure of the NN controller results
in
fc ≡ uj (z, ż) = ψ[w3 z2 + w4 zż + w5 ż2 + w6 z3 + w7 z2 ż + w8 zż2 + w9 ż3 ] (16)
as it does not contain linear feedbacks. On the other hand, in order to ensure discontinuous operation, the
ψ-function has to be activated. An appropriate ψ-function in Eq. (16) can be either a sigmoid or a hard
type non-linearity:
ψ =0 for z + %t · ż ≤ 0.08m, ψ =1 for z + %t · ż > 0.08m (17)
with %t being the prediction time.
K. Spentzas, S.A. Kanarachos / Mathematics and Computers in Simulation 60 (2002) 369–378 375

Fig. 4. Discontinuous operation, road disturbance Amax = 11 cm.

In order to minimize passenger acceleration with the control law (16) and (17), the following goal and
constraints are defined:
f (w) = |ẍ|for A=11 cm = minimum, g1 (w) = |z| − 0.08m ≤ 0 for Amax = 11 cm (18)
Eq. (18) indicates that the NN is trained for minimum acceleration for Amax = 11 cm. The training of the
NN, using semi-stochastic optimization with R = 1.3, E = 2 and C = 0.5, leads to the following values
for ψ and the synaptic weights:

wT = {−4.62 × 104 4.28 × 103 − 34.78 − 4.02 × 106 − 1.05 × 105


− 7.21 × 103 − 211.72}, %t = 0.05 s (19)
In Fig. 4 the results for the above case are presented. The controller is not activated for Amean = 5 cm,
while for Amax = 11 cm the maximum passenger acceleration does not exceed ẍ1,max = 12.2 m/s2 and
the maximum actuator forces and the power consumption are kept small (f max = 2438 N).
Corresponding computations are performed also for nonzero time constants (T act = 0.02 s) of the
actuator. The training of the NN, using again semi-stochastic optimization with R = 1.3, E = 2 and
C = 0.5, leads to the following values for ψ and the synaptic weights:

wT = {−2.789 × 105 9.687 × 105 256.8 − 2.564 × 106 − 2.342 × 105


− 6.059 × 103 − 422}, %t = 0.05 s (20)
Figs. 5 and 6 show the time traces for y, ẍ1 and f (for T act = 0.02) for two bump amplitudes, Amean = 5 cm
and Amax = 11 cm. The controller is not activated for Amean = 5 cm. In the second case despite of the small
Tact value, the actuator is rather “slow”, if the 3T act = 0.06 s value is compared to the time needed for the
376 K. Spentzas, S.A. Kanarachos / Mathematics and Computers in Simulation 60 (2002) 369–378

Fig. 5. Discontinuous operation, road disturbance 5 cm, T act = 0.02 s.

Fig. 6. Discontinuous operation, road disturbance 11 cm, T act = 0.02 s.


K. Spentzas, S.A. Kanarachos / Mathematics and Computers in Simulation 60 (2002) 369–378 377

Fig. 7. Discontinuous operation, road disturbance 11 cm, T act = 0.02 s. Evolution of kres , cres .

force bump T forcebump ≈ 0.05 s (Fig. 5): 3T act /T forcebump ≈ 120%. However, the NN controller manages
to control the acceleration, certainly at a cost of the maximum acceleration ẍ1,max (ẍ1,max = 14.2 m/s2 )
and maximum force (f max = 4338 N).
Last, in Fig. 7 the resulting (passive + active) spring and damping coefficients kres and cres
kres ≡ k − w3 z, cres ≡ c − w4 z − w5 ż − w7 z2 − w8 zż − w9 ż2 (21)

are displayed. It is interesting to see, how the spring and damping coefficients kres and cres are increased
at the presence of a greater bump in order to fulfil the z constraint.

5. Conclusions

In this paper the problem of the design of a non-linear hybrid car suspension system for ride qualities
using NNs has been presented. It is shown that the discontinuous operation of the ride qualities controller
has a lot of advantages, e.g. lower power consumption, greater life time, etc., compared to a continuous
operation of the active part of the suspension system.
From the standpoint of methodology, the paper emphasizes a new technique based on NNs, where the
structure of the NN controller is derived from a Taylor series approximation of the (unknown) non-linear
control function. Due to the non-linearity of the controller, the training of the NN is not easy, as it
experiences numerous local minima. For this reason, a semi-stochastic parameter optimization method
is used, which has performance advantages compared to other stochastic methods, e.g. GSA.
The numerical results are compared with recent results in the literature and are found to be very
satisfactory. The applied methodology proves to be very flexible and powerful and is suitable for the
solution of other control problems.
378 K. Spentzas, S.A. Kanarachos / Mathematics and Computers in Simulation 60 (2002) 369–378

References

[1] F.J. D’Amato, D.E. Viassolo, Fuzzy control for active suspensions, Mechatronics 10 (2000) 897–920.
[2] T. Yoshimura, K. Nakaminami, M. Kurimoto, J. Hino, Active suspension of passenger cars using linear and fuzzy-logic
controls, Contr. Eng. Prac. 7 (1999) 41–47.
[3] L. Li, M. Nagai, Control and evaluation of active suspension for MDOF vehicle model, JSAE Rev. 20 (1999) 343–348.
[4] D. Hrovat, Survey of advanced suspension developments and related optimal control applications, Automatica 33 (1997)
1781–1817.
[5] Y. Ando, M. Suzuki, Control of active suspension systems using the singular perturbation method, Contr. Eng. Prac. 4 (1996)
287–2937.
[6] C.-S. Ting, T.-F. Li, F.-C. Kung, Design of fuzzy controller for active suspension system, Mechatronics 5 (1995) 365–383.

You might also like