Professional Documents
Culture Documents
Neurocomputing
journal homepage: www.elsevier.com/locate/neucom
a r t i c l e i n f o abstract
Available online 9 March 2011 A neuro-fuzzy adaptive control approach for nonlinear dynamical systems, coupled with unknown
Keywords: dynamics, modeling errors, and various sorts of disturbances, is proposed and used to design a wheel
Adaptive neuro-fuzzy control approach slip regulating controller. The implemented control structure consists of a conventional controller and a
Sliding mode incremental learning neuro-fuzzy network-based feedback controller. The former is provided both to guarantee global
algorithm asymptotic stability in compact space and as an inverse reference model of the response of the
Antilock braking system (ABS) controlled system. Its output is used as an error signal by an incremental learning algorithm to update
the parameters of the neuro-fuzzy controller. In this way the latter is able to gradually replace
the conventional controller from the control of the system. The proposed new learning algorithm makes
direct use of the variable structure systems theory and establishes a sliding motion in terms of the
neuro-fuzzy controller parameters, leading the learning error toward zero. In the simulations and in
the experimental studies, it has been tested on the control of antilock breaking system model and the
analytical claims have been justified under the existence of uncertainty and large nonzero initial errors.
& 2011 Elsevier B.V. All rights reserved.
0925-2312/$ - see front matter & 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.neucom.2010.07.035
1884 A.V. Topalov et al. / Neurocomputing 74 (2011) 1883–1893
Besides sliding mode control there are several control schemes effectiveness of the variable structure control (VSC) scheme in
including adaptive, fuzzy logic and neural network control handling imprecisions and uncertainties and the existing possi-
approaches, and hybrid controllers consisting of combinations of bility to analyze rigorously convergence and stability have sti-
these. A self-learning fuzzy controller is combined with a sliding mulated its application in the learning schemes of the intelligent
mode controller in [5]. The stability of the system is guaranteed, structures. The results obtained in this direction have shown that
as the tuning algorithms for the controller are derived in the the convergence of the learning strategies can be significantly
Lyapunov sense. Mauer proposed a fuzzy controller in [6] to improved. Such intelligent systems exhibit the robustness and
identify the condition of the road surface. The controller adjusts invariance properties inherited from VSC technique while still
the braking torque based on the current and past values of wheel maintaining good approximation capability and flexibility.
slip and brake pressure. Layne et al. [7] also adopted the fuzzy The proposed incremental learning algorithm makes direct use
approach and model the plant to be controlled as a first-order of the variable structure systems theory and establishes a sliding
linear dynamic system. motion in terms of the neuro-fuzzy controller parameters, leading
Because of their ability to approximate linear and nonlinear the learning error toward zero. The convergence of the algorithm
mapping, neural network-based (NN-based) controllers are devel- is established and the conditions are given. Its salient character-
oped to improve the stability, convergence, and robustness of the istics are stable online tuning of the parameters in the premise
control system by compensating the effects of nonlinearities and and the consequence parts of the fuzzy rules and fast learning
system uncertainties [8]. Whereas feedforward NNs are designed speed. The zero level set of the learning error variable is
in [9,10], the recurrent neural networks are proposed in [11,12]. considered as a sliding surface in the space of network learning
For instance to the latter approach, in [8], a hybrid control system parameters. It controls the error dynamics. The latter is defined as
with a recurrent neural network observer is developed for the control signal produced by a conventional controller con-
antilock braking systems. The hybrid control system is composed nected in parallel and it is described using a differential equation.
of an ideal controller and a compensation controller. In [13], two Differently from the gradient-based learning methods which aim
parallel NNs are utilized to realize a neuro-sliding mode control, to minimize an error function, here the learning parameters are
where the equivalent control signal and the corrective control tuned by the proposed algorithm in a way to enforce the error to
signal are taken from the outputs of neural network 1 and neural satisfy this stable equation.
network 2, respectively. The so-called feedback-error-learning concept, proposed
An observer-based direct adaptive neuro-fuzzy controller for in [22] and initially applied to control of robot manipulators,
ABS is developed in [14] under the assumption that only the has been used to tune online the controller parameters. It relies
wheel slip is measurable. To track the varying desired slip value, on the parallel work of two feedback controllers—a neuro-fuzzy
an observer-based output feedback control law and a law for feedback controller (NFFC) and a conventional feedback control-
online tuning of the weighting factors of the direct adaptive ler. It is common to assume that a proportional plus derivative
fuzzy-neural controller are derived. Will et al. [15] developed a (PD) controller is used as a conventional controller. It serves both
hybrid nonlinear control system that combines the sliding mode- to guarantee global asymptotic stability in compact space and as
based observer with a PID controller. The proposed controller an inverse reference model of the response of the system under
does not require the information of surface type and can compute control. The output of the PD controller is used as an error signal
the optimal value of wheel slip online based on the data from to update the parameters of the neuro-fuzzy controller and in this
longitudinal accelerometers and wheel speed sensors. In [16] an way the latter is learning online to eliminate the conventional
ABS controller is designed using genetic neural fuzzy control. controller from the control of the system.
While a neural optimizer is employed to find the optimal wheel In the simulations and in the experimental studies, the
slip, the fuzzy component computes the braking torque required proposed control approach has been tested on the control of an
to track the optimal slip. A genetic algorithm is used to adjust the antilock breaking system model and the analytical claims have
parameters of the fuzzy logic component. been justified under the existence of uncertainty and large
In the present paper a new neuro-fuzzy adaptive control nonzero initial errors.
approach is proposed for nonlinear dynamical systems, coupled The paper is organized as follows. Section 2 starts with an
with unknown dynamics, modeling errors, and various sorts of introduction to the quarter vehicle model describing longitudinal
disturbances, and subsequently used to design a wheel slip regulat- motion of the vehicle and angular motion of the wheel under
ing controller. In spite of the existing variety of learning algorithms braking, then continues with the proposed adaptive neuro-fuzzy
used for training of neuro-fuzzy networks, they can be mainly control approach used to design an intelligent controller for
classified into two groups: (i) algorithms relying on the gradient tracking of the varying slip value. Then, the developed new
descent methodology and involving the computation of partial variable structure systems-based method for parametric adapta-
derivatives or sensitivity functions, which can be considered as tion of fuzzy rule-based neural networks with a scalar output is
extensions of the well-known learning algorithm with backpropa- presented. Section 3 is devoted to the obtained results from
gation of the error [17] and (ii) algorithms based on evolutionary simulations and real-time experiments, and the concluding
computations with genetic algorithms (GAs) being the most widely remarks are given in Section 4.
used among them [18]. The algorithms belonging to the first group
inherit some of the major drawbacks of the gradient descent type
learning among which, in particular, are the slow speed of learning 2. The adaptive neuro-fuzzy control approach
because of the entrapment of local minima and the difficulty to
obtain analytical results concerning the convergence and stability of 2.1. The system to be controlled
the learning schemes [19]. The algorithms from the second group
have also well-known disadvantages such as being computationally The free body diagram of the quarter vehicle model describing
intensive and showing slower than gradient-descent algorithms longitudinal motion of the vehicle and angular motion of the
convergence to the optimal point after reaching the near-optimal wheel under braking is presented in Fig. 1 [23]. Although the
area due to their randomness [20]. model is quite simple, it preserves the fundamental character-
Stability of the online learning neuro-fuzzy structures used in istics of an actual system. In deriving the dynamic equations of
control applications is another very important issue [21]. The the system, several assumptions are made. First, only longitudinal
A.V. Topalov et al. / Neurocomputing 74 (2011) 1883–1893 1885
Ft in (1) and (2) stands for the road friction force which is given
by Coulomb Law:
Ft ¼ mðlÞFn ð3Þ
Fn is calculated by the following equation:
d1 o1 þ M10 þTB þ Mg
Fn ¼ ð4Þ
LðsinfmðlÞcosfÞ
where L is the distance between the contact point of the wheels
and the rotational axis of the balance lever and f is the angle
between the normal in the contact point and the line L.
Under normal operating conditions, the rotational velocity of
the wheel would match the forward velocity of the car. When the
brakes are applied, braking forces are generated at the interface
between the wheel and road surface, which causes the wheel
speed to decrease. As the braking force at the wheel increases,
Fig. 1. Schematic view of quarter vehicle model [23].
slippage will occur between the tire and the road surface. The
wheel speed will tend to be lower than vehicle speed. The
parameter used to specify the difference between the above two
Table 1 velocities is called wheel slip (l), and it is defined as
Nomenclature. r2 o2 r1 o1
l¼ ð5Þ
r 2 o2
Name Description
While a wheel slip of 0 indicates that the wheel velocity and
o1 Angular velocity of the upper wheel
the vehicle velocity are the same, a ratio of 1 indicates that the
o2 Angular velocity of the lower wheel
TB Braking torque tire is not rotating and the wheels are skidding on the road
r1 Radius of the upper wheel surface, i.e., the vehicle is no longer steerable.
r2 Radius of the lower wheel The road adhesion coefficient is a nonlinear function of some
J1 Moment of inertia of the upper wheel physical variables including wheel slip and it can be approxi-
J2 Moment of inertia of the lower wheel
d1 Viscous friction coefficient of the upper wheel
mated using the following formula developed by [23]
d2 Viscous friction coefficient of the lower wheel p
c4 l 3 2
Fn Total normal load mðlÞ ¼ p þc3 l þc2 l þ c1 l ð6Þ
m Road adhesion coefficient aþl
l Wheel slip
The corresponding m2l curve for different road conditions can
ld Desired slip
Ft Road friction force
be seen in Fig. 2.
M10 Static friction of the upper wheel
M20 Static friction of the lower wheel 2.2. The control scheme and the neuro-fuzzy network structure
Mg Moment of gravity acting on balance lever
The output signal of the fuzzy-neural network tn ðtÞ is calcu- From (15) to (18) it follows that t and t_ will be bounded
lated as a weighted average of the output of each rule [28]: signals too, i.e.,
PI PJ
i¼1 Wij fij jtðtÞj r Bt , jt_ ðtÞj r Bt_ 8t ð19Þ
tn ðtÞ ¼ PI Pj J¼ 1 ð11Þ
i¼1 j¼1
Wij where Bt and Bt_ are some known positive constants.
After the normalization of (11), the output signal of the neuro- Using the sliding mode control theory principles [29] the zero
fuzzy network will acquire the following form: value of the learning error coordinate tc ðtÞ can be defined as time-
varying sliding surface, i.e.,
J
I X
X
tn ðtÞ ¼ W ij fij ð12Þ Sc ðtn , tÞ ¼ tc ðtÞ ¼ tn ðtÞ þ tðtÞ ¼ 0 ð20Þ
i¼1j¼1
which is the condition that the neuro-fuzzy network is trained to
where W ij is the normalized value of the output signal of the become a nonlinear regulator to obtain the desired response
neuron ij from the second hidden layer of the network: during the tracking-error convergence movement by compensa-
Wij tion for the nonlinearity of the controlled plant.
W ij ¼ PI PJ ð13Þ
Wij The sliding surface for the nonlinear system under control
i¼1 j¼1
_ is defined as
Sp ðe, eÞ
The control signal generated by the wheel slip controller is the
_ ¼ e_ þ we
Sp ðe, eÞ ð21Þ
overall brake torque t determined as follows (see Fig. 3):
t ¼ tc tn ð14Þ with w being a constant determining the slope of the sliding
surface.
where tc and tn are the brake torques generated by the PD
controller and the neuro-fuzzy feedback controller respectively.
Definition. A sliding motion will appear on the sliding manifold
Sc ðtn , tÞ ¼ tc ðtÞ ¼ 0 after a time th, if the condition
2.3. The sliding mode learning algorithm
Sc ðtÞS_ c ðtÞ ¼ tc ðtÞt_ c ðtÞ o 0 is satisfied for all t in some nontrivial
semi-open subinterval of time of the form ½t,th Þ ð1,th Þ.
The following vectors have been specified: XðtÞ ¼ ½x1 ðtÞ x2 ðtÞT :
vector of the time varying input signals; W ðtÞ ¼ ½W 11 ðtÞ W 12 ðtÞ . . . It is desired to devise a dynamical feedback adaptation
W 21 ðtÞ . . . W ij ðtÞ . . . W IJ ðtÞT : vector of the normalized output mechanism, or online learning algorithm for the neuro-fuzzy
signals of the neurons from the second hidden layer; network parameters such that the sliding mode condition of the
sA ¼ ½sA1 . . . sAi . . . sAI T , sB ¼ ½sB1 . . . sBj . . . sBJ T , cA ¼ ½cA1 . . . above definition is enforced.
cAi . . . cAI T and cB ¼ ½cB1 . . . cBj . . . cBJ T :vectors of the tuning
parameters s and c of the Gaussian membership functions Theorem 1. If the adaptation law for the parameters of the
relevant to the fuzzification of the signals supplied to the first considered neuro-fuzzy network is chosen respectively as
and second input of the NFFC respectively; f ðtÞ ¼ ½f11 ðtÞ f12 ðtÞ . . . sAi
f21 ðtÞ f22 ðtÞ . . . fij ðtÞ . . . fIJ ðtÞ: vector of the time variable weight c_ Ai ¼ a signðtc Þ ð22Þ
sTA sA
coefficients of the connections between the neurons from the
second hidden layer and the output neuron of the fuzzy rule- sBj
based neural network. c_ Bj ¼ a signðtc Þ ð23Þ
sTB sB
The following assumptions have been used in this investi-
gation: sAi
Due to the control scheme adopted (Fig. 3), where the s_ Ai ¼ a signðtc Þ ð24Þ
sTA sA
conventional controller serves to guarantee global asymptotic
stability in compact space, the input signals x1 ðtÞ and x2 ðtÞ, and sBj
their time derivatives can be considered bounded: s_ Bj ¼ a signðtc Þ ð25Þ
sTB sB
jx1 ðtÞj r Bx , jx2 ðtÞj r Bx 8t ð15Þ
W ij
jx_ 1 ðtÞj r Bx_ , jx_ 2 ðtÞj rBx_ 8t ð16Þ f_ ij ¼ T a signðtc Þ ð26Þ
W W
where Bx and Bx_ are assumed to be some known positive
constants. where
Based on the same arguments, the vectors defining the tuning sAi ¼ x1 cAi , sA ¼ ½sA1 sA2 . . . sAI T
parameters s and c of the Gaussian membership functions are
considered bounded as follows: sBj ¼ x2 cBj , sB ¼ ½sB1 sB2 . . . sBJ T
JsA J r Bs , JsB J r Bs
and a is a sufficiently large positive design constant satisfying the
JcA J r Bc , JcB J r Bc ð17Þ inequality:
where Bs and Bc are some known positive constants. 4Br Bf þ Bt_
a4 ð27Þ
It will be assumed that, due to physical constraints, the time 18Bq Bf
variable weight coefficients of the connections between the
neurons in the second hidden layer and the output neuron are with the positive constants Br and Bq bounded by the following
also bounded, i.e., inequalities:
Bx þ Bc Bx þBc
jfij jr Bf 8t ð18Þ Br r 2Bx , Bq r2 ð28Þ
B2s B3s
for some positive constant Bf.
From (8)–(13) and (15)–(17) it follows that 0 o W ij o1. In Then for any arbitrary initial condition tc ð0Þ, the learning error
P P
addition, it can be easily seen from (13) that Ii ¼ 1 Jj ¼ 1 W ij ¼ 1. tc ðtÞ will converge to zero during a finite time th, which may be
1888 A.V. Topalov et al. / Neurocomputing 74 (2011) 1883–1893
‘‘chattering’’. Chattering is undesired because it may execute the 0.2 Compact Snow
high-frequency response of the system. The common methods to
eliminate the chattering are usually classified into two 0.15
groups [30]:
0.1 Rough Ice
1. Using a saturation function to replace the signum function.
Reference
2. Inserting a boundary layer so an equivalent control replaces 0.05 NFFC+PD
the corrective one when the system is inside this layer. PD
0
Since when applying the second method a finite steady-state 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
error would always exist, most of the approaches use the satura- Time (s)
tion or the sigmoid function to replace the signum function. In
order to reduce the chattering effect the function in (33) has been Fig. 5. The wheel slip.
A.V. Topalov et al. / Neurocomputing 74 (2011) 1883–1893 1889
the NFFC learns the system dynamics in a finite time duration, Fig. 7 shows the phase space behavior. It figures out that Sp ¼ 0
and makes the system performance better compared to the line is attracting invariant. Clearly the error vector is guided
conventional PD controller alone. As explained in the previous toward the sliding manifold and due to the design presented, it is
sections, to obtain a mathematical model of ABS is very difficult forced to remain in the vicinity of the attracting loci without
and sometimes impossible because of the nonlinearities and explicitly knowing the analytical details of the equations of
noises both coming from inside and outside of the system. In motion of the slip control system.
such cases, the conventional controller cannot remain always Fig 8 presents the best results obtained during the experi-
well-tuned and the adaptive neuro-fuzzy controller structure ments with PD controller implemented alone and with NFFC. It
proposed will help to maintain the desired performance. can be seen that it is not possible to enforce an ABS to the desired
The control signals for both controller structures (the PD trajectory using PD control law only. Although the PD controller
controller alone and the adaptive NFFC coupled with a PD coupled with NFFC is also not capable of tracking the reference
controller) can be seen in Fig. 6. It can be observed that the trajectory accurately, it still exhibits better performance com-
neuro-fuzzy feedback controller has been able to take over the pared to the PD control law only.
control operation, thus becoming the leading controller after a Three membership functions have been used for each of the
short time period. This results in zero output from the conven- two inputs of the NFFC. Since the parameters of the neuro-fuzzy
tional PD controller. The output of the PD controller becomes controller have not been well-tuned initially (initial parameters
nonzero only during the time intervals when the vehicle goes of the fuzzy membership functions m11 , m12 , m13 , m21 , m22 , and
from one road condition to the other. m23 have been selected to be s ¼ ½0:01,0:01,0:01,2,2,2 and
10 10
8 8
6
Derivative of Slip
6
Control Inputs
4 4
2 2
0 0 PD
τc NFFC+PD
τn Reference
−2 −2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Time (s) Slip
10
6
Time Derivative of Error
4 Sp
−2
−4
−0.01 −0.005 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04
Error
0.7
therefore no experimental results could have been carried out
0.6 for changing road conditions. On the other hand, to be able to
0.5
0.4
0.3
0.2
0.1
0
−0.02 −0.01 0 0.01 0.02 0.03 0.04
Error
0.9
μ initial
0.8 μ final
0.9
Membership Function
0.7
0.6 0.8
0.5 0.7
0.4 0.6
0.3 0.5
0.2
0.4
0.1
0.3
0
−5 0 5 10 0.2
Time Derivative of Error Reference
0.1 PD
Fig. 10. Membership functions for the time derivative of error. NFFC+PD
0
20.5 20.6 20.7 20.8 20.9 21 21.1 21.2 21.3 21.4
c¼[0, 0.015, 0.03, 1, 3, 7], respectively) at the beginning the
dominating control signal has been the one coming from the PD Fig. 12. The wheel slip.
controller. After a short time period however, using the control
signal tc as a learning error, the neuro-fuzzy feedback controller
has been able to take over the control, thus becoming the leading
controller (see Fig. 6). The NNFC membership function parameters 20
at the end of the simulation (after 0.7 s) have reached the
following values: s ¼ ½0:008,0:01,0:012,1:78,2:15,2:52 and 10
c¼[ 0.007,0.009,0.025, 1.7,2.25,6.19] for m11 , m12 , m13 , m21 ,
m22 , and m23 , respectively. However, in order to be able to use 0
Control Inputs
the same learning step size for both of the inputs, the error and
the time derivative of error values are normalized. The values −10
given above are for the unnormalized case. Moreover, two
different learning step sizes are used for the weight update law
−20
of the neural network and for the adaptation law for the s and c
values of the input membership functions.
−30
Figs. 9 and 10 show the membership functions of the two NFFC
inputs at the beginning (t ¼0 s) and at the end of the simulation
(at time t ¼0.7 s). −40 τc
τn
3.2. Experimental results −50
20.5 20.6 20.7 20.8 20.9 21 21.1 21.2 21.3 21.4
Time (s)
Number of real time experiments have been performed to
validate the results obtained in the simulations on an ABS Fig. 13. The control signal.
A.V. Topalov et al. / Neurocomputing 74 (2011) 1883–1893 1891
compare the experimental results with the simulation results in Then V_ c can be further expressed as follows:
the previous subsection, the reference wheel slip has been 8 2 0 13 9
changed to an upper value suddenly. Hence, the change in road <X J
I X X J
I X =
V_ ¼ tc 4f_ ij W ij þ fij @W ij K_ ij þW ij _
W ij K ij A5 þ t_
conditions, i.e., from wet to dry asphalt, has been imitated. Once : ;
i¼1j¼1 i¼1j¼1
the velocity threshold has been reached, the braking operation is
started. There is another velocity threshold which states the 2
X J
I X X J
I X
minimum velocity level for applying ABS control algorithms.
¼ tc 4 f_ ij W ij 2 W ij ðAA_ þBBÞf
_ ij
Under this minimum value of the velocity, the system becomes i¼1j¼1 i¼1j¼1
unstable if ABS algorithm is applied. Under such a circumstance, 0 1 3
I X
X J X J
I X
the maximum braking torque should be applied to the wheels þ2 @W ij fij W ij ðAA_ þBBÞ
_ A þ t_ 5
without considering the target value of slip. i¼1j¼1 i¼1j¼1
Similar to the simulation results, the performance of PD 8 "
< X I X J
A
control law has been improved by the hybrid control methodol- ¼ tc a signðtc Þ2 W ij fij 2 ðx_ 1 sAi þ 2a signðtc ÞÞ
ogy consisting of a conventional PD controller coupled to NFFC : sAi
i¼1j¼1
(see Fig. 12). As can be seen from Fig. 13, that the neuro-fuzzy #
X I X J
B
feedback controller is trying to take over the control operation þ 2 ðx_ 2 sBj þ 2a signðtc ÞÞ þ2
after a short time period. The output of the PD controller tends to sBj i¼1j¼1
2
go to zero simultaneously. X J
I X
A
4W ij fij W ij ðx_ 1 sAi
i¼1j¼1
s2Ai
!# )
B
4. Conclusions þ2a signðtc ÞÞ þ ðx_ 2 sBj þ 2a signðtc ÞÞ þ t_
s2Bj
2
A novel approach for generating and maintaining sliding J
I X
X J
I X
X
regime in the behavior of a system with uncertainties in its ¼ tc 4a signðtc Þ2 W ij fij rij 4a signðtc Þ W ij fij qij
dynamics is introduced. The system under control is under a i¼1j¼1 i¼1j¼1
X
I J
X
The work of A.V. Topalov was supported by the Ministry of
rajtc jþ 4ajtc j W ij Bq ðBf þ Bf Þ
Education and Science of Bulgaria Research Fund Project BY-TH- i¼1j¼1
108/2005. The work of O. Kaynak was supported in part by the J
I X
X
Bogazici University Research Fund Project 08A204 and in part by þ2jtc j W ij Br ðBf þ Bf Þ þ Bt_ jtc j
the TUBITAK Project 107E284. i¼1j¼1
¼ jtc j½að18Bq Bf Þ þ 4Br Bf þBt_ o 0 ðA:4Þ
where rij and qij are defined as follows:
Appendix A. Proof of Theorem 1 A B
rij ¼ x_ 1 þ x_ 2 ðA:5Þ
sAi sBj
Consider the following Lyapunov function candidate:
A B
Vc ¼ 12t2c ðtÞ ðA:1Þ qij ¼ þ ðA:6Þ
s2Ai s2Bj
The time derivative of Vc is given by
jrij j rBr , jqij j r Bq ðA:7Þ
V_ c ¼ tc t_ c ¼ tc ðt_ n þ t_ Þ
2 0 1 3 The inequality (A.4) shows that the controlled trajectories of
X I X J
d the learning error tc ðtÞ converge to zero in a stable manner.
¼ tc 4 @ f W A þ t_ 5
dt i ¼ 1 j ¼ 1 ij ij It is possible now to be shown that such a convergence takes
2 3 place in finite time. The differential equation that is satisfied by
X I X J
4 _ _ the controlled error trajectories tc ðtÞ is as follows:
¼ tc ðf ij W ij þ fij W ij Þ þ t_ 5 ðA:2Þ
i¼1j¼1
2 3
X J
I X [2] Q. Zhang, G. Liu, B. Liu, X. Xie, Sensor fusion based estimation technology of
¼ a signðtc Þ41 þ4 W ij qij ðfij tn Þ5 vehicle velocity in anti-lock braking system, in: Proceedings of the 2007
i¼1j¼1 International Conference on Information Acquisition, Jeju City, Korea, 2007,
pp. 106–111.
X
I J
X [3] P. Kachroo, M. Tomizuka, Sliding mode control with chattering reduction and
2 W ij rij ðfij tn Þ þ t_ ðA:8Þ error convergence for a class of discrete nonlinear systems with application
i¼1j¼1 to vehicle control, in: Proceedings of the International Mechanical Engineer-
ing Congress and Expo, vol. 57, Chicago, USA, 1995, pp. 225–233.
For any t r th , the solution, tc ðtÞ, of this equation, with initial [4] C. Unsal, P. Kachroo, Sliding mode measurement feedback control for antilock
condition tc ð0Þ at t¼0, satisfies braking systems, Control Systems Technology 7 (1999) 271–281.
8 [5] C. Lin, C.F. Hsu, Self-learning fuzzy sliding-mode control for antilock braking
Z t Z t< systems, Control Systems Technology 11 (2003) 273–278.
tc ðtÞtc ð0Þ ¼ t_ c ðBÞ dB ¼ a signðtc ðBÞÞ [6] G.F. Mauer, A fuzzy logic controller for an abs braking system, IEEE Transac-
0 0 : tions on Fuzzy System 3 (1995) 381–388.
2 3 [7] J.R. Layne, K.M. Passino, S. Yurkovich, Fuzzy learning control for antiskid braking
X I X J systems, IEEE Transaction on Control System Technology 1 (1993) 122–129.
41þ 4 W ij ðBÞqij ðBÞðfij ðBÞtn ðBÞÞ5 [8] C. Lin, C. Hsu, Neural-network hybrid control for antilock braking systems,
i¼1j¼1 Transactions on Neural Networks 14 (2003) 351–359.
9 [9] J.Y. Choi, J.A. Farrell, Nonlinear adaptive control using networks of piecewise
X
I J
X = linear approximators, Transactions on Neural Networks 11 (2000) 390–401.
2 W ij ðBÞrij ðBÞðfij ðBÞtn ðBÞÞ þ t ðBÞ dB
_ ðA:9Þ [10] R.J. Schilling, J.J. Carroll Jr, A.F. Al-Ajlouni, Approximation of nonlinear
; systems with radial basis function neural network, Transactions on Neural
i¼1j¼1
Networks 12 (2001) 1–15.
At time t ¼th the solution takes zero value and, therefore, [11] T. Chow, Y. Fang, A recurrent neural-network-based real-time learning
8 2 3 control strategy applying to nonlinear systems with unknown dynamics,
Z < th XI X J Transactions on Neural Networks 45 (1998) 151–161.
tc ð0Þ ¼ 4
a signðtc ð0ÞÞ 1 þ 4 W ij ðtÞqij ðtÞðfij ðtÞtn ðtÞÞ5 [12] C. Lee, C. Teng, Identification and control of dynamic systems using recurrent
0 : i¼1j¼1 fuzzy neural networks, Transactions on Neural Networks 8 (2000) 349–366.
9 [13] C. Tsai, H. Chung, F. Yu, Neuro-sliding mode control with its applications to
X I X J = seesaw systems, Transactions on Neural Networks 15 (2004) 124–134.
2 W ij ðtÞrij ðtÞðfij ðtÞtn ðtÞÞ þ t_ ðtÞ dt [14] G. Chen, W. Wang, T. Lee, C.W. Tao, Observer-based direct adaptive fuzzy-
;
i¼1j¼1 neural control for anti-lock braking systems, International Journal of Fuzzy
8 2 3 9
< Z th X I X J = Systems 8 (2006) 208–218.
¼ a signðtc ð0ÞÞ th þ 4 4 W ij ðtÞqij ðtÞðfij ðtÞtn ðtÞÞ5 dt [15] A.B. Will, S. Hui, S.H. Zak, Sliding mode wheel slip controller for an antilock
: 0 ; braking system, International Journal of Vehicle Design 19 (1998) 523–539.
i¼1j¼1
[16] Y. Lee, H.S. Zak, Genetic neural fuzzy control of anti-lock brake systems, in:
8 9 Proceedings of the 2001 American Control Conference, Arlington, Virginia,
Z th < X J
I X =
2001, pp. 671–676.
2 W ij ðtÞrij ðtÞðfij ðtÞtn ðtÞÞt_ ðtÞ dt ðA:10Þ
0 : i¼1j¼1
; [17] D.E. Rumelhart, G.E. Hinton, R.J. Williams, Learning internal representations
by error propagation in: Parallel Distributed Processing: Explorations in the
Microstructure of Cognition, 1986, pp. 318–362.
By multiplying both sides of the equation by signðtc ð0ÞÞ the [18] C. Juang, A tsk-type recurrent fuzzy network for dynamic systems processing
estimate of th can be found using the following inequality: by neural network and genetic algorithm, IEEE Transactions on Fuzzy
2 3 Systems 10 (2002) 155–170.
Z th X J
I X [19] C. Lin, A neural fuzzy control system with structure and parameter learning,
jtc ð0Þj ¼ ath þ 4a 4 W ij ðtÞqij ðtÞðfij ðtÞtn ðtÞÞ5 dt Fuzzy Sets and Systems 70 (1995) 183–212.
0 i¼1j¼1 [20] A. Topalov, K.-C. Kim, J.-H. Kim, B.-K. Lee, Fast genetic on-line learning
2 3 algorithm for neural network and its application to temperature control, in:
Z th J
I X
X International Conference on Evolutionary Computation, vol. 1, Nagoya, Japan,
þsignðtc ð0ÞÞ 42 W ij ðtÞrij ðtÞðfij ðtÞtn ðtÞÞt_ ðtÞ5 1996, pp. 649–654.
0 i¼1j¼1 [21] W. Yu, X. Li, Fuzzy identification using fuzzy neural networks with stable
8 2 3 9 learning algorithms, IEEE Transactions on Fuzzy Systems 12 (2004) 411–420.
< Z th X I X J =
4 [22] H. Gomi, M. Kawato, Neural network control for a closed-loop system using
Z a th þ4 W ij ðtÞqij ðtÞðfij ðtÞtn ðtÞÞ5 dt feedback-error-learning, Neural Networks 6 (1993) 933–946.
: 0 ;
i¼1j¼1 [23] Inteco, User’s manual: the laboratory antilock braking system controlled
2 3
J from pc, Technical Report, Inteco Ltd., Poland, 2006.
XI X
[24] Y. Oniz, E. Kayacan, O. Kaynak, Simulated and experimental study of antilock
42 W ij Br ðBf þ Bf Þ þBt_ 5th
braking system using grey sliding mode control, in: Proceedings of the IEEE
i¼1j¼1
8 9 SMC 2007, Montreal, Canada, 2007, pp. 90–95.
< X J
I X X J
I X = [25] M. Burckhardt, Fahrwerktechnik: Radschlupf-Regelsysteme, Vogel Verlag,
Z a8Bq Bf W ij 4Br Bf W ij Bt_ th 1993.
: ; [26] L.-X. Wang, A Course in Fuzzy Systems and Control, Prentice-Hall, 1997.
i¼1j¼1 i¼1j¼1
[27] D. Rutkowska, Neuro-Fuzzy Architectures and Hybrid Learning, Physica-
¼ ½að8Bq Bf þ 4Br Bf þ Bt_ Þth ðA:11Þ Verlag, 2002.
[28] Y. Jin, Advanced Fuzzy Systems Design and Applications, Physica-Verlag, 2003.
Obviously, for all times t oth , taking into account the chosen [29] V.I. Utkin, Sliding Modes in Control Optimization, Springer-Verlag, 1992.
sliding mode controller gain a in (27) it follows from (A.8) that [30] J.-J. Slotine, W. Li, Applied Nonlinear Control, Prentice-Hall, 1991.
2 3 [31] Y. Oniz, E. Kayacan, O. Kaynak, A dynamic method to forecast the wheel slip
J
I X
X for antilock braking system and its experimental evaluation, IEEE Transac-
4
tc ðtÞt_ c ðtÞ ¼ ajtc ðtÞj 1 þ 4 W ij qij ðfij tn Þ5 tions on Systems, Man and Cybernetics—Part B: Cybernetics 39 (2009)
551–560.
i¼1j¼1
2 3
X
I J
X
42 W ij rij ðfij tn Þt_ 5tc ðtÞ Andon V. Topalov received the M.Sc. degree in control
i¼1j¼1 engineering from the Faculty of Information Systems,
Technologies and Automation, Moscow State Univer-
r ½að18Bq Bf Þ þ4Br Bf þ Bt_ jtc ðtÞj o0 ðA:12Þ sity of Civil Engineering (MGGU), Moscow, Russia, in
1979 and the Ph.D. degree in control engineering from
and a sliding motion exists on tc ðtÞ ¼ 0 for t 4 th .
the Department of Automation and Remote Control,
Moscow State Mining University (MGSU), Moscow,
in 1984.
References
From 1985 to 1986, he was a Research Fellow in the
Research Institute for Electronic Equipment, ZZU AD,
[1] M.R. Akbarzadeh, K.J. Emami, N. Pariz, Adaptive discrete-time fuzzy sliding Plovdiv, Bulgaria. In 1986, he joined the Department of
mode control for anti-lock braking systems, in: Proceedings of the Annual Control Systems, Technical University of Sofia, Plovdiv,
Meeting of the North American, Las Vegas, USA, 2002, pp. 554–559. where he is presently an Associate Professor. He has
A.V. Topalov et al. / Neurocomputing 74 (2011) 1883–1893 1893
held long-term visiting Professor/Scholar positions at various institutions in South Okyay Kaynak received the B.Sc. degree with (First
Korea, Turkey, Mexico, Greece, Belgium, UK, and Germany. He has coauthored one Class Honors) and the Ph.D. degree in electronic and
book and authored or coauthored more than 70 research papers in conference electrical engineering from the University of Birming-
proceedings and journals. His current research interests are in the fields of ham, Birmingham, UK, in 1969 and 1972, respectively.
intelligent control and robotics. From 1972 to 1979, he held various positions within
the industry. In 1979, he joined the Department of
Electrical and Electronics Engineering Bogazici Univer-
sity, Istanbul, Turkey, where he is presently a Full
Professor. He has served as the Chairman of the
Yesim Oniz was born in Istanbul, Turkey, on July 10,
Computer Engineering and the Electrical and Electro-
1981. She received the B.Sc. degree in mechatronics
nic Engineering Departments and as the Director of the
engineering from Sabanci University, Istanbul, in 2004
Biomedical Engineering Institute, Bogazici University.
and the M.Sc. degree in electrical and electronic
Currently, he is the UNESCO Chair on Mechatronics
engineering from Bogazici University, Istanbul, in
and the Director of the Mechatronics Research and Application Centre. He has held
2007, where she is currently working toward the
long-term (near to or more than a year) Visiting Professor/Scholar positions at
Ph.D. degree in electrical and electronic engineering
various institutions in Japan, Germany, the US, and Singapore. His current research
in the Department of Electrical and Electronics Engi-
interests are in the fields of intelligent control and mechatronics. He has authored
neering.
three books and edited five and authored or coauthored more than 200 papers
She is currently a Research Assistant with the
that have appeared in various journals and conference proceedings.
Mechatronics Research and Application Center,
Dr. Kaynak has served as the President of the IEEE Industrial Electronics Society
Department of Electrical and Electronics Engineering,
(2002–2003) and as an Associate Editor of both the IEEE Transactions on Industrial
Bogazici University. Her research interests include
Electronics and the IEEE Transactions on Neural Networks. He is now the Editor-
robotics, variable structure systems, intelligent control, and fuzzy and neurofuzzy
in-Chief of the IEEE Transactions on Industrial Informatics. Additionally, he is on
controls.
the Editorial or Advisory Boards of a number of scholarly journals.