Upload
ahmadmoaaz
View
216
Download
0
Embed Size (px)
Citation preview
8/21/2019 gw3
1/13
Middle-East Journal of Scientific Research 16 (10): 1348-1360, 2013
ISSN 1990-9233
IDOSI Publications, 2013
DOI: 10.5829/idosi.mejsr.2013.16.10.11842
Corresponding Author: Shahid Qamar, Department of Electrical Engineering,
COMSATS Institute of Information Technology, Abbottabad, Pakistan.1348
Adaptive B-Spline Based Neuro-Fuzzy Control for Full
Car Active Suspension System
Shahid Qamar, Laiq Khan and Saima Ali
Department of Electrical Engineering,
COMSATS Institute of Information Technology, Abbottabad, Pakistan
Abstract: The main role of a car suspension system is to improve the ride comfort and road holding. The
traditional spring-damper suspension is currently being replaced by either a semi-active or active suspension
system, because, passive suspension system cannot give better ride comfort and handling property. In order
to improve the capability of active suspension systems, a robust Adaptive B-spline Neuro-fuzzy (ABNF) based
control strategies are used to give better road handling and passenger comfort. In this paper, different ordersof B-spline membership functions are used in the proposed ABNFs control strategies. The shape of B-spline
membership functions can be adjusted self-adaptively by changing control points during learning process.
The order of the B-spline membership functions gives a structure for choosing the shape of the fuzzy sets.
The update parameters of ABNFs are trained by gradient descent based technique during the learning process.
The ABNFs control techniques are successfully applied to full car suspension model which reduces the seat
heave pitch and roll displacement, suspension travel and wheel displacement of the vehicle. The performance
index of the proposed techniques is taken on the basis of seat, heave, pitch and roll of the vehicle. Simulation
is based on the full car mathematical model by using MATLAB/SIMULINK. The simulation results show that
the ABNFs control techniques give better results than passive and semi-active suspension systems.
Key words: B-spline Fuzzy logic Neural network Full car Suspension system
INTRODUCTION The main objective of an active suspension is to
In vehicle's suspension systems, the suspension road disturbances, to improve the vehicle stability and the
system plays a very important role for the vehicles passenger comfort. The primary requirements of a good
stability or road handling and ride comfort against the vehicle suspension design are possession of good ride
road disturbances. So, the passive system cannot give comfort, road handling and road holding characteristics
better road handling and passengers comfort. The within an acceptable suspension travel range [1-4].
conventional passive suspension system comprises of The conflicting nature of these design objectives makes
combination of springs and dampers. Unlike conventional a necessary trade-off between them.
passive and semi-active suspension systems, an active In the last few years, many researchers applied
vehicle suspension systems external excitation is different linear and nonlinear control methods to thecounteracted with the generation of a control force vehicle suspension models. The linear suspension control
depending on the vehicle response through an actuator strategies range from PID to robust multivariable
driven by an external energy source. Therefore, the use controllers. It is supposed that the systems states reveal
active vibration control mechanism for design of only small oscillations about a stability point so that a
advanced suspension system has attracted considerable linear estimated model can be used. PID and direct
attention in the recent years. adaptive neural control of a nonlinear half car active
reduce the vibrations of the vehicle body induced by the
8/21/2019 gw3
2/13
Middle-East J. Sci. Res., 16 (10): 1348-1360, 2013
1349
suspension system is established in [5-9]. To examine the functions at any arbitrary precision as long as the network
robustness of quarter car suspension system based on is large enough [28]. B-spline neural network is illustrated
stochastic stability robustness presented by [10], but this by a local weight updating scheme with the advantages of
technique needs large feedback gains and an appropriate low computation complexity and fast convergence speed.
phase must be chosen. Adaptive control involves the The B-spline has been identified as very smoothest curveadjustment of feedback gains according to changes in the in the engineering applications. This is, because of the
system parameters and road conditions. In [11] they divergence-moderate property of the B-spline wave.
combined the benefits of adaptive and robust control in These characteristics make it more appropriate for the on-
the design of adaptive vehicle suspension system line adaptive situations. The plus point of the B-spline
controller and focused on the actuator force control. The functions over other radial functions are the local controls
adaptive robust controller for the quarter car active of the curve shape, since the curve changes in the
suspension is designed [12]. In [13], they designed the locality of a few control points that have been changed
feedback linearization control for full car suspension. A [29]. B-spline neural network comprises of the piecewise
variety of simulations showed that the fuzzy logic control polynomials with a set of local basis functions to model
is proficient to give a better ride quality than the other an unknown function for which a finite set of input-output
common control approaches [14-16]. The genetic samples are available. In [30], the authors used B-spline
algorithm-based fuzzy PI and PD controller for the active neurofuzzy technique with different orders of B-splinesuspension system for a quarter car model is designed by functions.
[17]. In [18], designed a hybrid neurofuzzy based sliding In [31], the authors proved that B-spline are
mode control to enhance the convenience of the equivalent to certain fuzzy systems. Therefore, the
passenger and stability of vehicle body. B-spline basis functions may be examined as fuzzy
Fuzzy logic has ability to decrease the complexity and membership functions and can be used to obtain fuzzy
to deal with vagueness and uncertainties of the data rules. B-splines basis functions are better to get the
[19, 20]. Their arrangement permits to build up a system desired properties like, local compact support, low
with fast learning abilities that can explain nonlinear computational and partition of unity. These techniques
structure that are described with uncertainties. Neural are non-adaptive, so can approximate the unknown
networks have self-learning qualities that raise the nonlinear dynamical systems. But the adaptive controllers
precision of the model. The combination of fuzzy logic have the ability to estimate the unknown nonlinear
and wavelets neural network provides the platform for dynamical system, because, the updated parameters hasresolving control problems and signal processing flexibility to update their values according to the system's
problems [21, 22]. There are several probable alternatives requirement. To estimate the nonlinearity of the unknown
for the basis functions to propose local activation system, the adaptive B-spline fuzzy neural network is
functions of neural networks, such as associated memory established. By using the online updating algorithm, both
networks, wavelets, radial basis function [23] and B-spline the B-spline membership functions of the fuzzy sets and
functions [24]. Unlike a global activation function which weighting parameters of the adaptive fuzzy neural
is non zero for all finite inputs, a local activation function network is updated. This online adaptation method will
is nonzero only in a local area of the input space. improves the robustness and the convergence of the
However, fuzzy neural networks with local activation system.
functions are more competent than the global activation In this study, adaptive B-spline Fuzzy Neural
functions in speed and memory, because, only nonzero Network strategies with different orders of B-spline
functions and the weights are calculated [25]. In case of membership function, i.e. B-spline of order 2 and B-spline
local activation functions, the estimators of nonlinear of order 3, are successfully implemented on nonlinear full
mapping can be expressed as a linear combination of basis car active suspension model to improve the ride comfort
functions. An appropriate choice of B-spline membership and road handling. The paper is divided into 6 sections.
and basis functions is the B-spline [26]. Section II describes the full car model. In Section III
The ability of neural networks to approximate various discusses the construction of proposed ABNFs
complex nonlinear functions relating input-output data techniques and section IV gives the online ABFNN
from a nonlinear system has attracted many researchers control algorithm. Finally, simulation results and
[27]. B-spline neural network can also estimate continuous conclusion are given in section V and VI respectively.
8/21/2019 gw3
3/13
M , M , M , Mf, r f, l r, r r, l
.. .
s s s s s s s s s sM Z +k (Z -Z-X -Y f)+c ( Z -Z-X -Y f)=0
.MZ+k (Z-a +Wf-Z )+c (Z-a +Wf-Z ))+k (Z+b +Wf-Z )+f1 f,l f1 f,l r1 r,l
. .c (Z+b +Wf-Z )+k (Z-a -Wf-Z )+c (Z-a -Wf-Z ) +r1 r,l f2 f,r f2 f,r
. .k (Z+b -Wf-Z )+c (Z+b -Wf-Z )-c (Z -Z-X -Y f)-f -f -f -f =0r2 r,r r2 r,r s s s s 1 2 3 4
Middle-East J. Sci. Res., 16 (10): 1348-1360, 2013
1350
Table 1: Simulation variables for full car model
Variable Name Variable Value
Front Left Spring Constant ks1 15000 N/m
Front Right Spring Constant Ks2 15000 N/m
Rear Left Spring Constant ks3 17000 N/m
Rear Right Spring Constant ks4 17000 N/m
Seat Spring Constant ks5 15000 N/m
Front Left Damper Constant cs1 2500 N.s/m
Front Right Damper Constant cs2 2500 N.s/m
Front Left Damper Constant cs3 2500 N.s/m
Front Right Damper Constant cs4 2500 N.s/m
Seat Spring Constant cs5 150 N.s/m
Front Left Tire Suspension kt1 250000 N/m
Front Right Tire Suspension kt2 250000 N/m
Rear Left Tire Suspension kt3 250000 N/m
Rear Right Tire Suspension kt4 250000 N/m
Length of Axle Suspension from C.O.G a 1.2 mLength of Axle Suspension from C.O.G b 1.4 m
Length of Unsprung mass from C.O.G c 0.5 m
Length of Unsprung mass from C.O.G d 1 m
Length of Seat from C.O.G e 0.3 m
Length of Sear from C.O.G f 0.25 m
Front Left Unsprung mass m1 25 kg
Front Right Unsprung mass m2 25 kg
Rear Left Unsprung mass m3 45 kg
Rear Right Unsprung mass m4 45 kg
Seat Mass m5 90 kg
Vehicle Body Mass M 1100 kg
Moment of Inertia for Pitch Ix7 1848 kg.m
2
Moment of Inertia for Roll Ix8 550 kg.m2
Shyhook Damper Constant Csky1 -2500 N.s/m
Full Car Model: There are eight degrees of freedom for a
full car, i.e. Roll, Pitch, Z Heave, Z front left
tyre, Z front right tire, Z rear left tyre, Z rearf,1 f,r r,r right tyre. It comprises only a sprung mass attached to
the four unsprung masses
(front-right, front-left, rear-right and rear-left wheels)
at each corner. The sprung mass is allowed to have pitch,
heave and roll and where the unsprung masses are
allowed only to have heave. For simplicity all other
motions are ignored for this model. This model has
eight degrees of freedom and allocates the body
acceleration and upright body displacement, pitch
and roll motion of the vehicle body. Full car
professionally ensures passenger safety and ride comfort.
This model considers only one seat and this is very
important to take into consideration other fixed with
chassis [32].
Fig 1: Full car model
The eight degrees of freedom consists of (x , x , x , x ,1 2 3 4x , x , x = , x = four wheels displacement, seat5 6 7 8displacement, heave displacement, pitch displacement and
roll displacement. The model of a full-car suspension
system is shown in Fig. 1. The suspensions between thesprung mass and unsprung masses are modeled as
nonlinear viscous dampers and spring components and
the tyres are modeled as simple nonlinear springs without
damping elements. The actuator gives forces that
determine the displacement of the actuator between the
sprung mass and the wheels. The dampers between the
wheels and car body signify sources of conventional
damping like friction among the mechanical components.
The inputs of full car model are four disturbances coming
through the tyres and the four outputs are the heave,
pitch, seat and roll displacement [33].The full car model
will be used as a good approximation of the whole car.
MATERIALS AND METHODS
As full-car model has eight-degrees of freedom, then
equations of motion for this model are explained by the
Newton's 2nd law of motion as:
For the Passenger Seat:
(1)
For the Vehicle's heave motion of the body:
(2)
For the Vehicle's Roll motion of the body:
8/21/2019 gw3
4/13
.I f+Wk (Z-a +Wf-Z )-Wc (Z-a +Wf- Z )+Wk (Z+b +Wf-Z )+x f1 f,l f1 f,l r1 r,l
. .Wc (Z+b +Wf- Z )-Wk (Z-a -Wf-Z )-Wc (Z-a -Wf- Z )-r1 r,l f2 f,r f2 f,r
.Wk (Z+b -Wf-Z )-Wc (Z+b -Wf- Z )+Y k (Z -Z-X -Y f)+r2 r,r r2 r,r s s s s s
.Y c (Z -Zs s s
- X -Y f)-Wf -Wf +Wf +Wf =0s s 1 2 3 4
.I -ak (Z-a +Wf-Z )+ac (Z-a +Wf-Z )+bk (Z+b +Wf-Z )y f1 f,l f1 f,l r1 r,l
.+bc (Z+b +Wf-Z )-r1 r,l
.ak (Z-a -Wf-Z )-ac (Z-a -Wf- Z )+bk (Z+b -Wf-Z )f2 f,r f2 f,r r2 r,r
.+bc (Z+b -Wf- Z ) +r2 r,r
.X k (Z -Z-X -Y f)+X c (Z -Zs s s s s s s s
-X -Y f)+af -bf +af -bf =0s s 1 2 3 4
..
f,l f,l f1 f,l f1
.
f,l t f,l g1 1
M Z -k (Z-a +W -Z )-c
(Z-a +W -Z )+k (Z -z )+f =0
..
r,l r,l r1 r,l r1
.
r,l t r,l g3 2
M Z -k (Z+b +W -Z )-c
(Z+b +W -Z )+k (Z -z )+f =0
..
f,r f,r f2 f,r f2
.f,r t f,r g2 3
M Z -k (Z-a -W -Z )-c
(Z-a ?+ W -Z )+k (Z -z )+f =0
..
r,r r,r r2 r,r r2
.
r,r t r,r g4 4
M Z -k (Z+b +W -Z )-c
(Z+b +W -Z )+k (Z -z )+f =0
p ps s(r)
ij j ij j
i=1 i=1 i=1 i=1
y =A(x)+ B (x)u + G (x)z
p1r -1r -1 T rx=[y ,y ,...,y ,...,y ,y ,...,y ] R 1 1 p p p1
T sr +r +...+r =r, u=[u ,u ,...,u ] R 1 2 p 21 s
T py=[y ,y ,...,y ] R 1 2 p
T sz=[z ,z ,...,z ] R 1 2 s
T1 2 pA=[A (x) A (x) . ..A (x)]
11 1s
p1 ps ps
b (x) b (x)
B(x)= M M
b (x) b (x)
p1 2
11 1s
p1 ps ps
(r )(r ) (r )(r) Tp1 2
g (x) g (x)
G(x)= M M
g (x) g (x)
y =[y ,y , ..., y ]
y=f(x)+B(x).u+G(x).z
(161) (164) (164)f(x) R , B(x) R , G(x) R
x R(161), u R(41) and z R(41)
T1 2
1 2 3
3
6
16
1f(x) = [A (x) A (x) A (x) .
x=[x x x ...
.. A (x)
x ]
]
T
1 2 z 4z=[z z x z ]
Middle-East J. Sci. Res., 16 (10): 1348-1360, 2013
1351
(3)
For the Vehicle's Pitch motion of the body:
(4)
For the Vehicle's Wheels motion:
(5)
(6)
(7)
(8)
The simulation parameters for full car model are
shown in Table 1.
The general class of nonlinear MIMO system is
described by [3,21]:
(9)
Where is the overall
state vector, which is assumed available and
is the control input vector,
is the output vector and
is the disturbance vector. A (x), i =i
1,...,p are continuous nonlinear functions, B (x), i = 1,...,p,ij
j = 1, ...,s are continuous nonlinear control functions and
G (x), i = 1,...,p, j = 1,..., s are continuous nonlinearijdisturbance functions.
Let
(10)
The control matrix is:
(11)
The disturbance matrix is:
(12)
Also, the generic nonlinear car model is,
(13)
, state vector
. The above matrices can be
shown in state-space form, with state vector x that is also
represented in row matrix form.
A (x) to A (x) are velocity states and A (x) to A (x) are1 8 9 16acceleration states of four tires, seat, heave, pitch and roll
[16]. The disturbance inputs for each tire individually are
represented in the form of z matrix.
z are n disturbances applied to full car model. u are nn ncontrollers output to full car model, to regulate the car
model disturbances. y are n states of car. r are nn ndesired outputs of the controller.
Abfnn for Full Car Suspension Control: In this section,
the structure of the fuzzy B-spline neural network is
described. Firstly, the fuzzy B-spline membership function
8/21/2019 gw3
5/13
k-1
ij i q q,2 i
q=0
(x )= c N (x )
i q q+k iq,k i q,k-1 i q+1,k-1 i
q+k-1 q q+k q+1
x -t t -xN (x )= N (x )+ N (x )
t -t t -t
i q q+2 iq,2 i q,1 i q+1,1 i
q+1 q q+2 q+1
x -t t -xN (x )= N (x )+ N (x )
t -t t -t
q q+1q,1 i
1 ; x [t ,t )N (x )=
0 ; otherwise
q+1 q+2q+1,1 i
1 ; x [t ,t )N (x )=
0 ; otherwise
i 00 i 0 1
1 0
2 i i 10 1 i 1 2
2 1 2 1
q+2 i i q+1q q+1 i q+1 q+2ij i
q+2 q+2 q+2 q+1
i q-1k ik-2 z-1 i k-1 k
k k-1 k k-1
k+1 ik-1 i k k+1
k+1 k
x -tc x [t ,t ]
t -t
t -x x -tc +c x [t ,t )
t -t t -t
t -x x -tc +c x [t ,t )t (x )=
t -t t -t
x -tt -xc +c x [t ,t )
t -t t -t
t -xc x [t ,t )
t -t
i ix X q,2 ia=1/sup N (x )
q,2 i q,2 iaN (x )=A (x )
ik ik (x , s(x ))
ikq,2 i
1 ; if round(x )=qA (x )=
0 ; otherwise
m ij ik ik c =t (x )=s(x )
Middle-East J. Sci. Res., 16 (10): 1348-1360, 2013
1352
and B-spline basis function is explained. Secondly, B-
spline fuzzy neural network is described and finally the
update parameters of the B-spline fuzzy neural network
are given.
Abnfs for Full Car Suspension Control: In this section,
the structure of the fuzzy B-spline neural network is
described. As B-spline has localibility properties, fast
convergence and low complexity calculation, so, it can
approximate the nonlinear functions of the dynamical
system. Firstly, the fuzzy B-spline membership function
and B-spline basis function of order 2 and order 3 are
explained. Secondly, B-spline fuzzy Neural Network is
described and finally the update parameters of the
B-spline fuzzy neural networks are given. (19)
Fuzzy B-spline Membership Function of Order 2: Generally, the B-spline basis functions with higherB-spline are the parametric surfaces that give flexibility for order, i.e. more than 1 are not a normal fuzzy set, because,
the visualization and modeling of the numerous types of the maximum value of these basis function does not reach
data [34]. All the B-spline fuzzy membership functions of at 1. This can be resolved by multiplying a positive
fuzzy logic rules are defined here as of order 2 and number with B-spline basis function, so that maximum
calculated by the algorithm explained below. The fuzzy value of basis function reaches at 1. i.e.
B-spline membership function is given by
(14)
Where (x ) is the membership function of the fuzzyij i
variable x to the c fuzzy set, q = 0,1,2,..,k-1, are controli qpoints, is the number of control points. The general form
of the B-spline membership function is ca be expressed as:
(15)
The B-spline basis function for order order (k=2), i.e.
N is cab be written as:q,2
(16)
(17)
(18)
Where t ,t and t are the nodes defined in the intervalq q+1 q+2of x . On substituting (16), (17) and (18) into (14), iti
becomes
Or,
(20)
Fig. 2 shows the shapes of B-spline membershipfunction.
The modeling potentials of the network is depends
on the distribution of the control points, shape and size of
the B-spline membership function. Where the knots and
the order determine the shape and the smoothness of the
B- spline membership function. The fuzzy B-spline
membership function shape can be regulated by changing
the value of the control points c . All the controls pointsqare equally allocated over the fuzzy set and is
the kth targeted value. Thus the values of control points
c can be adjusted asq
(21)
Where round(x ) is the new value $m$ of the closestikcontrol point c from the targeted value s(x ). Now fromm ik(16) and (19), it gives
(22)
8/21/2019 gw3
6/13
s-1
ij i q q,3 i
q=0
t (x )= c N (x )
t (x )ij i
i q q+3 iq,3 i q,2 i q+1,2 i
q+2 q q+3 q+1
x -t t -xN (x )= N (x )+ N (x )
t -t t -t
q q+2q,2 i
1 ; x [t ,t )N (x )=0 ;o therwise
q+1 q+3q+1,2 i
1 ; x [t , t )N (x )=
0 ;o therwise
1 1i 2 2i
m mi i
If x is A and x is A and
... x is A Then u is y
Middle-East J. Sci. Res., 16 (10): 1348-1360, 2013
1353
Fig. 2: (a) B-spline membership (order=2) and (b) Fig. 3: (a) B-spline membership (order=3) and (b)
Adjusted B-spline membership function (order=2) Adjusted B-spline membership function.
The output of the network depends on the B-spline basis
function, because, the B-spline basis function uses the (25)product operator and their smoothness is affected by the
output of the network. For example, the Fig. 2 depicted (26)
that the control points of fuzzy set and the kth target
datum change the shape of the B-spline membership
function. Where t ,t ,t and t , t ,t ,t and t are the nodes
Fuzzy B-Spline Membership Function of Order 3: The The normalized third order basis function can be
fuzzy B-spline membership function of order 3 is given by obtained by the (20). The control points for order 3 are
(23) membership function. Fig. 3 shows the shape of B-spline
Where is the membership function of the fuzzy Fuzzy B-Spline Neural Network Algorithm: A fuzzy
variable x to the jth fuzzy set, c are control points, q=0, 1,i q2,... s-1, s is the number of control points and N is theq,3B-spline basis function of order 3. The B-spline basis
function N is expressed as:q,3
(24)
q q+1 q+2 q+3 q q+1 q+2 q+3
defined in the interval of x .i
changing same as explained in order 2 B-spline
membership function of order 3.
logic system normally accumulates its data in the shape
of a fuzzy algorithm [35], which consists of a fuzzy
linguistic rules relating to the input and output of the
network. Then the ith rule has the form:
8/21/2019 gw3
7/13
m
i i
i=1m
i
i=1
w
u=
z-1
ij q q,2
q=0
(x)= c N (x)
ij i i ij i (x )=? ? (x )
m
i pi
i=1m
ii=1
w
u=
21J= e2
2d i
1J= (y -y )
2
pi pipi
Jw (t+1)=w (t)-
w
ij ijij
J(t+1)= (t)-
?
Middle-East J. Sci. Res., 16 (10): 1348-1360, 2013
1354
Fig. 4: Structure of the ABNF
The output of the system can be expressed as:
(27)
The structure for the fuzzy B-spline neural
network is depicted in Fig. 4. It comprises of four
layers:
Layer I:This layer is the input layer, i.e. introduces the
inputs (x , x ,...x ). This layer accepts the input values and1 2 mtransmits it to the next layer.
Layer II: In this layer the fuzzification process is
performed and neurons represent fuzzy sets used in the
antecedents part of the linguistic fuzzy rules. The
outputs of this layer are the values of the membership
functions, i.e. . The membership of ith input variable toijjth fuzzy set is defined by B-spline function as
(28)
Where shows the membership function of the ithijinput to the jth fuzzy set. Where i=1,2,...m and j=1,2.
Layer III:This layer is the fuzzy inference layer. In this
layer each node represents a fuzzy rule. In order to
compute the firing strength of each rule and $min$
operation is used to estimate the output value of the layer.
i.e.
(29)
where is the meet operation, (x ) are the degrees of thei ij imembership function of the layer II and (x) are the inputij i
values for the next layer.
Layer IV: This layer is the output layer. In this layer, the
defuzzification process is made to calculate the output of
the entire network, i.e. It computes the overall output of
system. Therefore, the output for the fuzzy B-spline neural
network can be expressed as:
(30)
Where u is the output for the entire network. The
training of the network starts after estimating the output
value of the fuzzy B-spline neural network and w are thepiweights between the neurons of III and IV layers and p=1,
2,... n, n is the number of classes.
Parameters Update Rules: The fuzzy B-spline neural
network learning is to minimize a given function or input
and output values by adjusting network parameters. The
gradient descent method is used to adjust the values ofweights w and the control points c of B-spline function.ip qijTo minimize the error between the actual output value of
the system and the desired value. For this purpose,
gradient descent method is can be expressed as:
Then
(31)
Where y is the desired output of fuzzy neural network, yd iis the actual output of system. The updated amount for
the w and can be obtained as:pi ij
(32)
(33)
8/21/2019 gw3
8/13
J
wpi
J
?ij
i
pi i pi
J J y J= . .
w y u w
ji
ij i ij ij
J J y u= . . .
? y u ?
m
i i
i=1
d i mpi 2i
i=1
J
=-(y -y )w
( )
( )
ip
d i jij j
j
w -uJ= -(y -y )
(x)
y
u
m
i i
i=1pi pi d i m
2i
i=1
w (t+1)=w (t)+ (y -y )
( )
( )
ip
ij i d i jj
j
w -u(t+1)= j(t)+ (y -y )
(x)
p
N
i i i
i=1
-15 1 t 2, 3 4
15 7 t 8, 9 10
a (1-cos8pt) 13 t 14,15 16z(t)=
A sin(O s-? ) 19 t 22,23 26
0 otherwise
N
A sin(O s-i i
=1
)i
i
( ; i=(1)Ni
Middle-East J. Sci. Res., 16 (10): 1348-1360, 2013
1355
Where is the learning rate.
By using chain rule the values of and can
be expressed as:
(34)
(35)
By taking the derivative of the above equations, it gives
(36)
(37)
Where the quantity is approximated by a constant
by Frechet derivative (Lin et al. 2000) or and (Khaldi et
al. 1993).
Therefore, updated values of w and can bepi ijobtained by substituting (36) and (37) in (32) and (33) as:
(38)
(39)
Where, = . Therefore, (38) and (39) gives the requiredupdated values for the parameters w , respectively.piFig. 5 depicts the closed loop diagram of the feedback
system.
Online Abnf Adaption Algorithm: Fig.5 shows the closed
loop control structure for the proposed ABNFs control.
This structure is employed to converge and update the
parameters of the ABNFs control. The calculations at any
instant of time can be described in the following steps.
Fig. 5: Control loop structure of ABNF
Step 1:Set the input-output, y and y of the system.d
Step 2: Update the parameters of the proposed ABNFs
scheme, i.e. and w by using (32) and (33) equations.ij pi
Step 3:Calculate the output of proposed ABNFs control
strategies.
Step 4: Finally, output of the proposed controller added
with disturbances is given to the system.
Step 5:Repeat steps (2-4) until solution converges.
Simulation Result: In this study, it is assumed that the
vehicle is moving with uniform velocity unless the road
disturbances create the undesired oscillations in the
vehicle body. To check the performance of the proposed
ABNFs control strategies, a road profile containingbumps, pothole and white noise has the form, i.e.
(40)
Where -0.15 m is the pothole and 0.15 m is the bump.This type of road profile shows the low frequency high
amplitude response. The term a(1-cos8 t) shows the
bell shaped bumps and a =15 m, is the amplitude of thep
bumps on the road. Now the term shows
the road profile like white noise. Where value of A is theiamplitude of the road, is the number of waves and isi ithe phase angle ranging from 0 to 2 .
8/21/2019 gw3
9/13
v|=|v|
6 [1,2,3,4]v=x - x
v
v| |v, t 0 |
Middle-East J. Sci. Res., 16 (10): 1348-1360, 2013
1356
(42)
Fig. 6: Road profile are compared with passive suspension and semi-active
Fig. 7: Seat displacement without controller 1 and ABNF-2 control, the maximum value for seat
Fig. 8: Seat displacement with controller state response is improved as compared to passive and
This road profile is very helpful for observing the heave, increased the passenger comfortability by 10% and 8% as
pitch and roll of the vehicle. This road profile is shown compared to the seat without controllers. The ABNF-1
in Fig. 6. ameliorates the ride comfort than ABNF-2 controlIn order to fulfill the aim of the active suspension strategy.
system, the proposed strategy ABNFs is successfully Fig. 9 to Fig. 11 shows that the response of heave,
implemented on full car suspension system to improve the pitch and roll is improved as compared to passive
vehicles stability and passengers comfort. The comfort suspension and semi-active suspension system. In
is examined by the vertical displacement and acceleration passive suspension and semi-active suspension, the
felt by the passenger. The controller goal is to minimize maximum value of displacement for heave are 0.105 m and
the vertical displacement of the vehicle body with 0.072 m while, for the ABFN-1 and ABFN-2 controls, the
reference to the open-loop, so, as to avoid the suspension maximum value of displacement for heave is 0.022 m and
travel hitting the rattle space limits. The lifetime of 0.025 m, respectively. Here, the ABFN-1 and ABFN-2
elements of the vehicle is conserved by keeping away increase the passenger comfort by 80% and 77% as
from hitting the rattle space limits i.e. to stay away from compared to passive suspension system and 70% and
which is the allowable peak to peak displacement of 67% as compared to semi-active suspension system. Inthe system. passive suspension and semi-active suspension, the
maximum value of displacement for pitch is 0.075 m and
Where 0.053 m while for the ABFN-1 and ABFN-2 controls, the
(41) 0.020 m, respectively. Here, the ABFN-1 and ABFN-2
increase the passenger comfort by 78% and 74% as
and is the maximum limit for bumps amplitudes. The compared to passive suspension system and 68% and
suspension travel can physically be expressed as 64% as compared to semi-active suspension system.
The controller performance is good when it reduces
the vehicle vibrations under road disturbances. In this
section, the simulation results of displacement of theheave, pitch, roll and seat (with and without controller) for
the above mentioned road profile is given. These results
suspension systems. The simulation time of this proposed
work is up to 30 seconds.
Fig. 7 and Fig. 8 show that the response of seat is
improved as compared to passive suspension and semi-
active suspension system. In passive suspension and
semi-active suspension, the maximum value for seat
displacement is 0.1796 m and 0.088 m while, for the ABNF-
displacements are 0.028 m and 0.031 m respectively. Here,
the ABFN-1 and ABFN-2 increase the passenger comfort
by 85% and 82% as compared to passive suspension
system and 69% and 65% as compared to semi-active
suspension system. The response of seat with controller
(wc) is better than seat without controller (woc). Also, the
settling time of ABNFs controllers is reduced and steady
semi-active suspension systems. The seat with controller
maximum value of displacement for pitch are 0.017 m and
8/21/2019 gw3
10/13
TTP P
0
1PI= (Z QZ )dt
2
Middle-East J. Sci. Res., 16 (10): 1348-1360, 2013
1357
Fig. 9: Heave displacement Fig. 12: Suspension travel for front tyre
Fig. 10: Pitch displacement Fig .13: Suspension travel for rear tyre
Fig. 11: Roll displacement Fig. 14: Front tyres displacement
In passive suspension and semi-active suspension, the
maximum value of displacement for roll is 0.029 m and
0.017 m while, for the ABFN-1 and ABFN-2 controls, the
maximum value of displacement for roll are 0.009 m and
0.01 m, respectively. Here, the ABFN-1 and ABFN-2increase the passenger comfort by 69% and 66% as
compared to passive suspension system and 48% and Fig. 15: Rear tyres displacement
44% as compared to semi-active suspension system. This
increased the vehicle stability, ride comfort and passenger Fig. 14 and Fig. 15 show the wheel displacement for
comfortability. Also, the settling time of ABFNN front and rear tyres. It is observed that the ABFN-1 and
controller is reduced and steady state response is ABFN-2 based active suspension system minimizes the
improved as compared to passive suspension. The wheel displacement by 60% and 56% as compared to
proposed active suspension shows better performance passive suspension and 49% and 45% than semi-active
and robustness. [33-35]From the above discussion it is suspension systems. This shows that the ABFN-1 based
observed that the ABFN-1 enhance the vehicle stability active suspension system improves the road handling of
and passenger comfortability than ABFN-2 control the vehicle, robustness and enhances the passenger
strategy. convenience than ABFN-2 control strategy.Fig. 12 and Fig. 13 show the suspension travel for The performance index used for evaluation of the
each front and rear tyres. The results show that the active control is given by,
ABFN-1 and ABFN-2 minimize the suspension travels by
68% and 64% than passive suspension system and 47% (43)
and 44% than semi-active suspension system. It is
observed that the ABFN-1 based active suspension
system provides better road handling and passenger where, Zp is the vector for displacement or
comfort than ABFN-2 control strategy. acceleration, Q i s the identity m atrix. T he R oot
8/21/2019 gw3
11/13
rmszdisp.
mszacc.
[ ]T
2rmsdisp. 6
0
1Z = x (t)
T
[ ]T
2rmsacc. 6
0
1Z = x (t)
T
Middle-East J. Sci. Res., 16 (10): 1348-1360, 2013
1358
Table 2: Performance comparison
DOFs Control Algorithm PI
Seat (woc) Passive 0.0497 4.4034 9.6962
Semi-active 0.0395 3.3742 5.6933
ABNF-1 0.01296 0.8269 0.3419
ABNF-2 0.0110 0.6748 0.2277
Seat (wc) Passive 0.03606 3.211 4.5635
Semi-active 0.03092 2.813 3.9569
ABNF-1 0.01026 0.7045 0.2482
ABNF-2 0.0103 0.5169 0.1336
Heave Passive 0.03606 3.211 4.5635
Semi-active 0.03092 2.813 3.9569
ABNF-1 0.01775 0.1980 0.0197
ABNF-2 0.01382 0.1631 0.0133
Pitch Passive 0.0277 2.4623 3.0318
Semi-active 0.0227 1.9324 1.8673
ABNF-1 0.0072 0.5374 0.1444
ABNF-2 0.0056 0.4831 0.1167
Roll Passive 0.0084 1.5943 1.2709
Semi-active 0.0080 1.2065 0.7278
ABNF-1 0.0029 0.4105 0.00842
ABNF-2 0.0024 0.3619 0.0654
Mean Square (RMS) value for displacement and
acceleration of heave, pitch, roll and seat has been
calculated by,
(44)
(45)
Table 2 shows the performance index of ABNF-1 and
ABNF-2 control strategies as compared to passive and
semi-active suspension systems. The performance of
ABNF-1 and ABNF-2 based active suspension is checked
with the said road profile.
CONCLUSION
In the paper, an ABNF control strategies with
different orders of B-spline membership functions, i.e.
order 2 (ABNF-1) and order 3 (ABNF-2) are successfully
implemented to the full car active suspension. The
objective was to improve the passenger comfort and the
road handling of the vehicle against the road
disturbances. The proposed ABNFs active suspension
system has ability to achieve low steady-state error and
fast error convergence than passive and semi-active
suspension systems. Therefore, ABNF-1 and ABNF-2
based active suspension system show that it improves
the vehicle stability and passenger comfort. The
simulations results depict that heave, pitch, roll, seat
displacement and suspension travel are reduced and road
handling is improved for the active suspension systemsas compared to passive and semi-active suspension
systems. Also, the experimental results show that the
ABNF-1 based active suspension a system performs
better than the ABNF-2 based active suspension system.
REFERENCE
1. Sharkawy, A.B., 2005. Fuzzy and adaptive fuzzy
control for the automobiles active suspension
system. Vehicle Systems Dynamics, 43(11): 795-806.
2. Huang, C., T. Li and C. Chen, 2009. Fuzzy feedback
linearization control for mimo nonlinear system andits application to full-vehicle suspension system.
Circuits, Systems and Signal Processing, 28(6): 959-
991.
3. Kumar, M.S. and S. Vijayarangan, 2007. Analytical
and experimental studies on active suspension
system of light passenger vehicle to improve ride
comfort. Mechanika, 65(3): 34-41.
4. Haiping, Du and Nong Zhang, 2009. Static output
feedback control for electrohydraulic active
suspensions via t-s fuzzy model approach. Asme-
Amer Soc Mechanical Engineering, 131(5).
5. Ekoru, J., O. Dahunsi and J. Pedro, 2011. Pid controlof a nonlinear half-car active suspension system via
force feedback. AFRICON, pp: 1-6.
6. Pedro, J., O. Dahunsi and N. Baloyi, 2011. Direct
adaptive neural control of a quarter-car active
suspension system. AFRICON, pp: 1-6.
7. Ray, L., 1992. Robust linear-optimal control laws for
active suspension systems. Journal of dynamic
systems, Measurement and Control, 114: 592-598.
8. Baranyi, P. and Y. Yam, 2006. Case study of the tp-
model transformation in the control of a complex
dynamic model with structural nonlinearity. IEEE
Transactions on Industrial Electronics, 53(3): 895-904.
9. Chantranuwathana, S. and H. Peng, 2004. Adaptive
robust force control for vehicle active suspensions.
International Journal of Adaptive Control and Signal
Processing, 18(2): 83-102.
10. Huang, S.J. and W.C. Lin, 2009. A neural network
based sliding mode controller for active vehicle
suspension. Proceedings of the Institute of
Mechanical Engineers, Part D: Journal of Automobile
Engineering, Seoul, Korea, 221(1): 1381-1397.
8/21/2019 gw3
12/13
Middle-East J. Sci. Res., 16 (10): 1348-1360, 2013
1359
11. Priyandoko, G., M. Mailah and H. Jamaluddin, 2009. 22. Macnab, C., 2005. Local basis functions in adaptive
Vehicle active suspension system using skyhook control of elastic systems. IEEE International
adaptive neuro active force control. Mechanical Conference on Mechatronics and Automation,
Systems and Signal Processing, 23(3): 855-868. 1: 19-25.
12. Zhang, J., Z. Zhao, Y. Sun and W. Han, 2010. Optimal 23. Wang, C., W. Wang, T. Lee and P. Tseng, 1995.design of fuzzy logic controller for multi-body model Fuzzy bspline membership function (bmf) and its
of semiactive suspension based on genetic algorithm. applications in fuzzy-neural control. IEEE
IEEE Conference on Industrial Electronics and Transactions on Systems, Man and Cybernetics,
Applications, ICIEA, pp: 1478-1483. 25(5): 841-851.
13. Laiq, K., M. Umair and Q. Shahid, 2012. 24. Carrasco, J., J. Quero, F. Ridao, M. Perales and
Comparative Analysis of Suspension Systems Using L. Franquelo, 1997. Sliding mode control of a dc/dc
Adaptive Fuzzy Control. 14th IEEE International pwm converter with pfc implemented by neural
Conference on Intelligent and Advanced Systems, networks. IEEE Transactions on Circuits and
1: 22-27. Systems I: Fundamental Theory and Applications,
14. Li, T. and Y. Kuo, 2004. Evolutionary algorithms for 44(8): 743-749.
passive suspension systems. JSME international 25. Cheng, K., H. Wang and D. Sutanto, 1999.
journal. Series C, Mechanical systems, Machine Adaptive bspline network control for three-
Elements and Manufacturing, 43(3): 537-544. phase pwm ac-dc voltage source converter.
15. Shahid, Q., K. Tariq and K. Laiq, 2012. Adaptive Proceedings of the IEEE 1999 International
Neuro-Fuzzy Sliding Mode Control Based Strategy Conference on Power Electronics and Drive Systems,
for Active Suspension Control. 10th International 1: 467- 472.
Conference on Frontiers of Information Technology, 26. Newman, W. and R. Sproull, 1979. Principles of
pp: 107-115. interactive computer graphics. McGraw-Hill
16. Karim, S. and M.F. Samadi, 2012. Adaptive Fuzzy New York.
Control for Nonlinear Time-Delay Systems with 27. Maier, H., T. Sayed and B. Lence, 2011. Forecasting
Dynamical Uncertainties. Asian Journal of Control, cyanobacterium Anabaena spp, in the River Murray,
14(6): 1589-1598. South Australia, using B-spline neurofuzzy models.
17. Ikenaga, S., F. Lewis, J. Campos and L. Davis, 2012. Ecological Modeling, 146(1): 85- 96.A Novel Fuzzy Approach for State Estimation of 28. Brown, M. and C.J. Harris, 1994.
Nonlinear Hybrid Systems Using Particle Filtering Neurofuzzy adaptive modeling and control. Prentice
Method. Asian Journal of Control, 14(4): 974-990. Hall.
18. Laiq, K., B. Rabiah and Q. Shahid, 2012. Adaptive 29. Rahmi, G., 2003. Active control of seat vibrations of a
Fuzzy Wavelet NN Control Strategy for Full Car vehicle model using various suspension alternatives.
Suspension System. Fuzzy Logic - Emerging Turkish Journal of Engineering and Environmental
Technologies and Applications, Edited by Elmer P. Sciience, 27: 361- 373.
Dadios, ISBN: 978-953-51-0337-0, 147-174. 30. Laiq, K., Q. Shahid and K. Umair, 2012. Adaptive
19. Laiq, K. Q. Shahid and K. Umair, 2012. Neuro-Fuzzy wavelets based fuzzy neural network control for
Wavelets Based Network for Full Car Active active suspension model. Emerging Trends and
Suspension System. 8th IEEE International Applications in Information Communication
Conference on Emerging Technologies, pp: 1- 6. Technologies. pp: 249-260.20. Dang, X. and Y. Tan, 2007. Rbf neural networks 31. Lin, C. and C. Lee, 1991. Neural-network-based fuzzy
hysteresis modelling for piezoceramic actuator using logic control and decision system. IEEE Transactions
hybrid model. Mechanical Systems and Signal on Computers, 40(12): 1320-1336.
Processing, 21(1): 430- 440. 32. Sayed, T., A. Tavakolie and A. Razavi, 2003.
21. Lightbody, G., O'Reilly, G.W. Irwin and J. McCormick, Comparison of adaptive network based fuzzy
1997. Neural modeling of chemical plant using mlp inference systems and B-spline neuro-fuzzy mode
and b-spline networks. Control Engineering Practice, choice models. Journal of Computing in Civil
5(11): 1501-1515. Engineering, 17(2): 123-130.
8/21/2019 gw3
13/13
Middle-East J. Sci. Res., 16 (10): 1348-1360, 2013
1360
33. Muhammad Azam, Sallahuddin Hassan and 35. Mueen Uddin, Asadullah Shah, Raed Alsaqour and
Khairuzzaman, 2013. Corruption, Workers Jamshed Memon, 2013. Measuring Efficiency of Tier
Remittances, Fdi and Economic Growth in Five South Level Data Centers to Implement Green Energy
and South East Asian Countries. A Panel Data Efficient Data Centers, Middle-East Journal of
Approach Middle-East Journal of Scientific Research, Scientific Research, 15(2): 200-207.15(2): 184-190.
34. Sibghatullah Nasir, 2013. Microfinance in India
Contemporary Issues and Challenges. Middle-East
Journal of Scientific Research, 15(2): 191-199.