This paper proposes a new adaptive neural network based control scheme for switched linear systems with parametric uncertainty and external disturbance. A key feature of this scheme is that the prior information of the possible upper bound of the uncertainty is not required. A feedforward neural network is employed to learn this upper bound. The adaptive learning algorithm is derived from Lyapunov stability analysis so that the system response under arbitrary switching laws is guaranteed uniformly ultimately bounded. A comparative simulation study with robust controller given in [Zhang L, Lu Y, Chen Y, Mastorakis NE. Robust uniformly ultimate boundedness control for uncertain switched linear systems. Computers and Mathematics with Applications 2008; 56: 1709–14] is presented.
2. 106 H.P. Singh, N. Sukavanam / ISA Transactions 51 (2012) 105–110
on the Stimulation Initiative for European Neural Applications
(SIENA project) indicates ANN is used in 39% of the production or
manufacturing sectors, with 35% of the usage related to control,
monitoring, modeling and optimization [15]. In [16], the possibility
to apply neural network modeling for simulation and prediction
of the EU-27 global index of production and domestic output
price index behavior is investigated. The most useful property
of neural network in control is their ability to approximate
arbitrary linear or nonlinear functions through learning. Due to
this property neural networks have proven to be a suitable tool
for controlling complex nonlinear dynamical systems. The basic
idea behind neural network based control is to learn unknown
nonlinear dynamics and compensate for uncertainties existing in
the dynamic model. There are many NN based control schemes
available for a class of switched systems [17–19]. But the case of
uncertainty bound estimation using neural network is not found
in the literature for switched systems.
In this paper, we propose a new adaptive neural network
based control scheme for switched linear systems with parametric
uncertainties and external disturbances. The proposed control
scheme has the following salient features:
(i) prior information of the upper bound of the uncertainty is not
required,
(ii) the feedforward neural network is able to learn the upper
bound of uncertainty, and
(iii) uniformly ultimately bounded (UUB) stability of system re-
sponse and NN weight error are guaranteed under arbitrary
switching laws.
This paper is organized as follows. Some basics of stability analysis
are given in Section 2. In Section 3, a review of feedforward neural
network and a new robust adaptive controller is proposed. A
common Lyapunov function approach is used to show that system
response and NN weight error are all UUB. A simulation example is
provided in Section 4 to illustrate the effectiveness of the proposed
control scheme. Section 5 gives conclusions.
2. Preliminaries
Consider a switched linear system represented by a differential
equation of the form
˙x(t) = Aσ(t)(ω)x(t) + Bσ(t)(ω)u(t), (1)
σ(t) : R+
−→ S = {1, . . . , N}
where t ≥ 0, x = (x1, x2, . . . , xn) ∈ Rn
denotes the state vector of
the system. u(t) ∈ Rm
is the control input and R+
denotes the set
of nonnegative real numbers. σ(t) is a piecewise constant function
called a switching law, which indicates the active subsystem at
each instant. Ai(ω), Bi(ω), i = 1, . . . , N are matrices whose
elements are continuous functions of a time-varying vector ω on
a compact set Ω ⊂ Rq
. When switched system (1) has parametric
uncertainty and external disturbance then it can be written in the
following form
˙x(t) = ( ¯Ai + Ai(ω))x(t) + ( ¯Bi + Bi(ω))u(t) + d(t), (2)
where d(t) is the bounded external disturbance, ¯Ai, i =
1, . . . , N are commuting Hurwitz matrices and Ai(ω), Bi(ω) are
uncertainty terms which satisfy the following conditions [20,21]
Ai(ω) = ¯BDi(ω) (3)
Bi(ω) = ¯BEi(ω) (4)
I +
1
2
(Ei(ω) + ET
i (ω)) ≥ δI (5)
where δ is a positive constant, Di : Ω −→ Rm×n
and Ei : Ω −→
Rm×m
are continuous matrix functions.
Now consider the nominal case with u(t) = 0 and d(t) = 0
˙x(t) = Aσ(t)(ω)x(t). (6)
Then stability conditions for system (6) are given by the following
theorems [22].
Theorem 2.1. If {Ai : i = 1, . . . , N} is a set of commuting Hurwitz
matrices, then the switched linear system (6) is globally uniform
asymptotic stable for any arbitrary switching sequence between Ai.
Theorem 2.2. For a given symmetric positive definite matrix Q ,
let P1, P2, . . . , PN > 0 be the unique symmetric positive definite
solutions of the following equations
AT
1 P1 + P1A1 = −Q , (7)
AT
i Pi + PiAi = −Pi−1, (8)
with the condition of Theorem 2.1, then the function L = xT
PN x is a
common Lyapunov function for the switched linear system (6).
Because of the parametric uncertainty and external disturbance in
switched linear system (2), we cannot derive the state response
x(t) which exactly converges to the equilibrium point. Therefore
it is only reasonable to expect that the system response converges
into a neighborhood of the equilibrium point and remains within it
thereafter, which is the so-called uniformly ultimate boundedness.
In [23], the following uniformly ultimate boundedness is proposed
for dynamical systems.
Definition 2.1. The solution of a dynamical system is uniformly
ultimately bounded (UUB) if there exists a compact set S ⊂ Rn
such that for all x(t0) = x0 ∈ S, there exists a λ > 0 and a number
T(λ, x0) such that ‖x(t)‖ < λ for all t ≥ t0 + T.
Theorem 2.3. If there exists a function L with continuous partial
derivatives such that for x in a compact set S ⊂ Rn
, L is positive
definite and ˙L < 0 for ‖x‖ > R, R > 0 such that the ball of radius R
is contained in S, then the system is UUB and the norm of the state is
bounded to within some neighborhood of R.
3. Neural network based control design
In this section, our aim is to design a control input u(t) in such
a way that the switched linear system (2) is uniformly ultimately
bounded (UUB). The following robust control scheme
u(t) =
−
w
‖w‖
ρ, if ‖ω‖ > ϵ
−
w
ϵ
ρ, if ‖ω‖ ≤ ϵ
(9)
is proposed in [14], where w = ¯BT
PN ρx, ϵ > 0 is any constant and
ρ denotes the possible upper bound of the parametric uncertainty
defined as
ρ =
1
δ
max
i
max
ωϵΩ
‖Di(ω)‖ ‖x‖ (10)
which is assumed to be known.
However, there are some potential difficulties associated with
the robust controller (9). (i) Implementation requires a precise
knowledge of the uncertainty bound, (ii) if an uncertain switched
system has many subsystems, the computation of the uncertainty
bound will be a complex and time consuming task, and (iii) if a
new switched system comes in, the whole procedure has to start
again. Therefore, we introduce a new adaptive control approach to
avoid the requirement of a prior knowledge of the upper bound
ρ in the expression (10) by using a feedforward neural network
(FFNN). Mathematically, a two-layer feedforward neural network
(Fig. 1) with n input units, m output units and L units in the hidden
layer is given as
3. H.P. Singh, N. Sukavanam / ISA Transactions 51 (2012) 105–110 107
zl =
L−
j=1
wljσ
n−
k=1
vjkxk + θvj
+ θwl
l = 1, 2, . . . , m (11)
where σ(.) are the activation functions of the neurons of the
hidden-layer [23]. The inputs-to-hidden-layer interconnection
weights are denoted by vjk and the hidden-layer-to-outputs
interconnection weights by wlj. The bias weights are denoted by
θvj, θwl. There are many classes of activation functions e.g. sigmoid,
hyperbolic tangent and Gaussian. The usual choice is the sigmoid
activation function, defined as
σ(x) =
1
1 + e−x
. (12)
By collecting all the NN weights vjk and wlj into matrices of weights
VT
and WT
, we can write the NN equation in terms of vectors as
z = WT
σ(VT
x) (13)
with the vector of activation functions defined by σ(y) = [σ(y1)
· · · σ(yn)]T
for a vector y ∈ Rn
. Without losing generality first-layer
weights are fixed and only the second-layer weights are tuned,
then the NN has only one layer of tunable weights. Then according
to [23–26], ρ can be written in neural network form as
ρ = WT
φ(x) + ρ
and an estimate of ρ can be given by
ˆρ = ˆWT
φ(x)
where ρ denotes NN approximation error, φ(.) is a basis matrix
and ˆW is the estimation of the weight matrix W provided by
some on-line weight tuning algorithms. For further analysis, the
following assumptions are selected as in [24,26,27].
Assumption 3.1. Given an arbitrary small positive constant ξ and
a continuous function ρ on a compact set Ω, there exist an optimal
vector W∗
such that
‖ ρ‖ = ‖W∗T
φ(x) − ρ‖ < ξ.
Assumption 3.2. The norm of parametric uncertainty µ defined as
µ = Di(ω)x and its upper bound satisfy the following relationship
ρ − ‖µ‖ > ξ.
For the design of a neural network based control input u(t) and the
weight vector ˆW, we have the following theorem.
Theorem 3.1. Consider the switched linear systems (2). If u(t)
in (2) is designed as
−
ˆρ2
w
ˆρ‖w‖ + ϵ
, (14)
Singh et al. [24], where ˙ϵ = −γ ϵ, ϵ(0) > 0, γ is a positive constant,
w = ¯BT
PN x and the adaptive neural network weight law is given as
˙ˆW = F‖w‖(φ(x) − κ ˆW),
with a positive definite diagonal gain matrix F, and a positive constant
κ, then the system response x and ˜W are uniformly ultimately
bounded (UUB), and x can be made small by a suitable choice of F
and κ.
Proof. Consider the following Lyapunov function candidate
L =
1
2
xT
PN x +
1
2
tr( ˜WT
F−1 ˜W) + k−1
ϵ (15)
where ˜W = W∗
− ˆW,
˙˜W = − ˙ˆW. Then, the time derivative of the
Lyapunov function is given as
˙L =
1
2
˙xT
PN x +
1
2
xT
PN ˙x + tr( ˜WT
F−1 ˙˜W) + k−1
˙ϵ
˙L =
1
2
[xT
( ¯Ai
T
+ AT
i ) + uT
(¯BT
+ BT
i )]PN x
+
1
2
xT
PN [( ¯Ai + Ai)x + (¯B + Bi)u] +
1
2
dT
PN x
+
1
2
xT
PN d + tr( ˜WT
F−1 ˙˜W) + k−1
˙ϵ.
Now using condition (3) and (4), we get
˙L =
1
2
[xT
( ¯Ai
T
PN + PN
¯Ai) + (¯BDi)T
PN + PN (¯BDi)]x
+
1
2
[uT
(¯BT
+ Ei
T ¯BT
)PN x + xT
PN (¯B + ¯BEi)u]
+
1
2
dT
PN x +
1
2
xT
PN d + tr( ˜WT
F−1 ˙˜W) + k−1
˙ϵ
˙L =
1
2
xT
( ¯Ai
T
PN + PN
¯Ai)x
+
1
2
xT
[(¯BDi)T
PN + PN (¯BDi)]x
−
ˆρ2
wT
2(ˆρ‖w‖ + ϵ)
(I + Ei)T ¯BT
PN x
−
ˆρ2
2(ˆρ‖w‖ + ϵ)
xT
PN
¯B(I + Ei)w +
1
2
dT
PN x
+
1
2
xT
PN d + tr( ˜WT
F−1 ˙˜W) + k−1
˙ϵ
using condition (5) with δ = 1 and condition (8) with ¯Ai
T
PN +
PN
¯Ai = −Ri, where Ri is a positive definite symmetric matrix, then
we have
˙L ≤ −
1
2
λmin(Ri)‖x‖2
−
ˆρ2
‖w‖2
ˆρ‖w‖ + ϵ
+ ‖w‖(‖µ‖ − ρ) + ‖w‖(W∗T
φ(x) + ρ)
+ tr( ˜WT
F−1 ˙˜W) + η‖PN ‖ ‖x‖ − ϵ
where η is the disturbance bound.
˙L ≤ −
1
2
λmin(Ri)‖x‖2
−
ˆρ2
‖w‖2
ˆρ‖w‖ + ϵ
+ ‖w‖(−ξ + ρ)
+ ‖w‖W∗T
φ(x) + tr( ˜WT
F−1 ˙˜W) + η‖PN ‖ ‖x‖ − ϵ
˙L ≤ −
1
2
λmin(Ri)‖x‖2
−
ˆρ2
‖w‖2
ˆρ‖w‖ + ϵ
+ ˆρ‖w‖
+ tr(‖w‖W∗T
φ(x) − ‖w‖W∗T
φ(x))
+ κ‖w‖tr ˜WT
(W∗
− ˜W) + η‖PN ‖ ‖x‖ − ϵ
˙L ≤ −
1
2
λmin(Ri)‖x‖2
+ κ‖w‖tr ˜WT
(W∗
− ˜W) + η‖PN ‖ ‖x‖.
Let the optimal weight W∗
is bounded by ‖W∗
‖F < Wmax on a
compact set Ω, where Wmax > 0. Then the following identity is
used in the derivation of the above inequality
tr ˜WT
(W∗
− ˜W) = ⟨ ˜W, W∗
⟩F − ‖ ˜W‖2
F
≤ ‖ ˜W‖F ‖W∗
‖F − ‖ ˜W‖2
F .
Then
˙L ≤ −
1
2
λmin(Ri)‖x‖2
+ κ‖¯BT
PN x‖ ‖ ˜W‖F (‖W∗
‖F − ‖ ˜W‖F )
+ η‖PN ‖ ‖x‖
˙L ≤ −‖x‖
[
1
2
λmin(Ri)‖x‖ + κ‖¯B‖ ‖PN ‖ ‖ ˜W‖F (‖ ˜W‖F
− ‖W∗
‖F ) − η‖PN ‖
]
.
4. 108 H.P. Singh, N. Sukavanam / ISA Transactions 51 (2012) 105–110
Fig. 1. Feedforward neural network.
Thus ˙L is negative as long as the term in brackets is positive.
Completing square terms, it is obtained that
1
2
λmin(Ri)‖x‖ + κ‖¯B‖ ‖PN ‖ ‖ ˜W‖F (‖ ˜W‖F − ‖W∗
‖F ) − η‖PN ‖
=
1
2
λmin(Ri)‖x‖ + κ‖¯B‖ ‖PN ‖
‖ ˜W‖F −
Wmax
2
2
−
κ
4
‖¯B‖ ‖PN ‖W2
max − η‖PN ‖.
Then RHS is positive if
‖x‖ >
κ‖¯B‖ ‖PN ‖W2
max + 4η‖PN ‖
2λmin(Ri)
(16)
or
‖ ˜W‖F >
Wmax
2
+
η
‖¯B‖κ
+
W2
max
4
. (17)
Thus ˙L is negative outside a compact set. According to the
Theorem 2.3, x and ˜W are UUB. Moreover, x can be made small by
a suitable choice of F and κ.
4. Simulation results
In this section, we consider a numerical example to demon-
strate the utility of the proposed control scheme for the switched
linear system (2) with the following two subsystems similar to [14]
A1(ω) =
[
0 1
−6 + ω2(t) −3 + ω1(t)
]
,
B1(ω) =
[
0
1.4387 + ω3(t)
]
,
A2(ω) =
[
0 1
−6.225 + ω2(t) −3 + ω1(t)
]
,
B2(ω) =
[
0
0.5613 + ω3(t)
]
,
where ω1(t) = 0.5 cos(t), ω2(t) = sin(t), ω3(t) = 0.25 sin(t).
¯Ai =
[
0 1
−6 −3
]
, ¯B =
[
0
1
]
,
d(t) =
[
0.1 sin(2t)
0.1 cos(2t)
]
,
D1(w) = [w2(t) w1(t)],
D2(w) = [−0.225 + w2(t) w1(t)],
Fig. 2. Response of state variable x1 with control design (9).
Fig. 3. Response of state variable x2 with control design (9).
E1(w) = 0.4387 + w3(t), E2(w) = −0.4387 + w3(t),
A1(ω) =
[
0 0
ω2(t) ω1(t)
]
,
B1(ω) =
[
0
0.4387 + ω3(t)
]
,
A2(ω) =
[
0 0
−0.225 + ω2(t) ω1(t)
]
,
B2(ω) =
[
0
−0.4387 + ω3(t)
]
which satisfies the condition (3) and (4). We choose a positive
definite matrix Q as
Q =
[
36 3
3 24
]
(18)
then the common Lyapunov solution matrix
PN =
[
17 3
3 5
]
5. H.P. Singh, N. Sukavanam / ISA Transactions 51 (2012) 105–110 109
Fig. 4. Response of state variable x1 with control design (14) for F = I10 and κ = 1.
Fig. 5. Response of state variable x2 with control design (14) for F = I10 and κ = 1.
can be easily obtained using the Lyapunov equations (7) and (8).
ρ = 1.15‖x‖ is selected as in [14] for comparison between the
robust controller (9) and our proposed controller (14).
The architecture of the FFNN is composed of 2 input units, 10
sigmoid units and a single output unit. The controller parameters
F, κ and γ in the proposed scheme are selected as I10, 0.01 and 0.01,
respectively. For simulation, we take a particular switched law σ(t)
defined as σ(t) = 1 for t < 30 and σ(t) = 2 for 30 ≤ t ≤ 70. The
simulation of the whole system is shown for 70 s. The effectiveness
of the FFNN to learn the upper bound of the switched linear system
uncertainty for producing the desired system response is shown
in the following figs. Response of state variables x1 and x2 are
shown in Figs. 2 and 3, respectively with known uncertainty bound
(i.e. with robust control design (9)). Figs. 4–5, Figs. 6–7 and Figs. 8–9
depict the response of the state variable x with neural network
based control design (14) for (F = I10, κ = 1), (F = 0.1I10, κ =
0.01) and (F = I10, κ = 0.01), respectively.
Fig. 6. Response of state variable x1 with control design (14) for F = 0.1I10 and
κ = 0.01.
Fig. 7. Response of state variable x2 with control design (14) for F = 0.1I10 and
κ = 0.01.
From Figs. 4–5 and Figs. 6–7, we see that the UUB region of the
state x is increasing for (F = I10, κ = 1) and (F = 0.1I10, κ =
0.01). In Figs. 8 and 9, the UUB region of the state x is small for
F = I10, κ = 0.01 in comparison to Figs. 2–7. In Figs. 4–9, for
the fixed γ , we observe that the state x is improved (decreased) by
increasing F and decreasing κ. Overall, we can say that the FFNN
based controller is quite effective without prior knowledge of the
upper bound of uncertainty, in comparison to the controller using
known upper bound of uncertainty as in [14].
5. Conclusion
In this paper, we investigated the uniformly ultimately bound-
edness of a switched linear system with parametric uncertainty
and external disturbance under arbitrary switched signals using a
Lyapunov function generated by the weighting matrices. A feed-
forward neural network is employed to learn the upper bound of
6. 110 H.P. Singh, N. Sukavanam / ISA Transactions 51 (2012) 105–110
Fig. 8. Response of state variable x1 with control design (14) for F = I10 and
κ = 0.01.
Fig. 9. Response of state variable x2 with control design (14) for F = I10 and
κ = 0.01.
uncertainty. The NN weights may be simply initialized to zero or
randomized. Finally, the numerical simulation is carried out for
two subsystems. The simulation results show that the feedforward
neural network with the on-line updating law can compensate the
switched linear system efficiently.
In our proposed controller (14) any prior knowledge of the
uncertainty bound is not required. This is the sharp contrast to
the control approach in [14], which has some potential difficulties
such as (i) implementation requires a prior knowledge of the
uncertainty bound, (ii) if an uncertain switched system has many
subsystems, the computation of the uncertainty bound will be a
complex and time consuming task, and (iii) if a new switched
system comes in, the whole procedure has to start again.
Acknowledgment
This work is financially supported by the Council of Scientific
and Industrial Research (CSIR), New Delhi, India and Department of
Science and Technology (DST), Government of India through grant
No. DST-347-MTD.
References
[1] Liberzon D, Morse AS. Basic problems in stability and design of switched
systems. IEEE Control Systems Magazine 1999;19:59–70.
[2] Asarin E, Bournez O, Dang T, Maler O, Pnueli A. Effective synthesis of switching
controllers for linear systems. Proceedings of the IEEE 2000;88:1011–25.
[3] Hespanha JP, Morse AS. Switching between stabilizing controllers. Automatica
2002;38:1905–17.
[4] Shorten R, Wirth F, Mason O, Wulff K, King C. Stability criteria for switched and
hybrid systems. SIAM Review 2005;49:545–92.
[5] Sun Z, Ge SS. Analysis and synthesis of switched linear control systems.
Automatica 2005;41:181–95.
[6] Zhang L, Chen Y, Cui P. Stabilization of a class of second-order switched
systems. Nonlinear Analysis 2005;62:1527–35.
[7] Zhang Z. Robust H∞ control of a class of discrete impulsive switched systems.
Nonlinear Analysis: Theory, Methods and Applications 2009;71:2790–6.
[8] Ji Z, Guo X, Wang L, Xie G. Robust H∞ control and stabilization of uncertain
switched linear systems: a multiple Lyapunov functions approach. Journal of
Dynamic Systems, Measurement, and Control 2006;128:696–700.
[9] Oya H, Hagino K. Adaptive robust stabilizing controllers for a class of uncertain
switched linear systems. In: Proceedings of the 2006 international workshop
on variable structure systems. 2006.
[10] Yang H, Zhao J. Robust control for a class of uncertain switched fuzzy systems.
Journal of Control Theory and Applications 2007;5:184–8.
[11] Rifai KE, Toumi KY. Robust adaptive control of a class of switched systems. In:
American control conference. 2008.
[12] Zhang X, Zhaoab J, Dimirovskicd GM. L2-gain analysis and control synthesis
problem for a class of uncertain switched linear systems subject to actuator
saturation. International Journal of Systems Science 2010;1–10.
[13] Fang Z, Xiang Z, Chen Q, Hua J. Optimal switching sequence of a class
of switched systems with parameter uncertainty. International Journal of
Nonlinear Science 2010;9:233–40.
[14] Zhang L, Lu Y, Chen Y, Mastorakis NE. Robust uniformly ultimate boundedness
control for uncertain switched linear systems. Computers and Mathematics
with Applications 2008;56:1709–14.
[15] Heikkonen J, Lampinen J. Building industrial applications with neural
networks. In: Proceedings of the European symposium on intelligent
techniques. 1999.
[16] Liana E, Nicolae T. The neural model of the European Union (EU-27) annual
growth rates. Ovidius University Annals, Economic Sciences Series, Ovidius
University of Constantza, Faculty of Economic Sciences 2010; 638–44.
[17] Long F, Fei S, Fu Z, Zheng S. Adaptive neural network control for switched
system with unknown nonlinear part by using backstepping approach: SISO
case. Lecture notes in computer science, Berlin (Heidelberg): Springer; 2006.
p. 842–8.
[18] Han TT, Ge SS, Lee TH. Adaptive neural control for a class of switched nonlinear
systems. Systems and Control Letters 2009;58:109–18.
[19] Yu L, Fei S, Li X. Robust adaptive neural tracking control for a class of switched
affine nonlinear systems. Neurocomputing 2010;73:2274–9.
[20] Schmitendorf WE, Barmish BR. Robust asymptotic tracking for linear systems
with unknown parameters. Automatica 1986;22:355–60.
[21] Pai MC. Observer-based adaptive sliding mode control for nonlinear uncertain
state-delayed systems. International Journal of Control, Automation, and
Systems 2009;7:536–44.
[22] Narendra KS, Balakrishnan J. A common Lyapunov function for stable LTI
systems with commuting A-matrices. IEEE Transactions on Automatic Control
1994;39:2469–71.
[23] Lewis FL, Jagannathan S, Yesildirek A. Neural network control of robot
manipulators and nonlinear systems. Taylor and Francis; 1999.
[24] Singh HP, Sukavanam N, Panwar V. Neural network based compensator for
robustness to the robot manipulators with uncertainties. In: Proceeding of
IEEE international conference on mechanical and electrical technology. 2010.
p. 444–8.
[25] Debbache G, Bennia A, Goléa N. Neural networks-based adaptive state
feedback control of robot manipulators. International Journal of Computers,
Communications and Control 2007;2:328–39.
[26] Zuo Y, Wang Y, Huang L, Li C. Intelligent hybrid control strategy for trajectories
tracking of robot manipulators. Journal of Control Science and Engineering
2008; doi:10.1155/2008/520591. Article ID-520591.
[27] Singh HP, Sukavanam N. Control of robot manipulators in task-space under
uncertainties using neural network. International Journal of Intelligent
Engineering Informatics 2011;1:142–55.