Adaptive Neural Network Output Feedback Control of

Nonlinear Systems with Actuator Saturation

Wenzhi Gao and Rastko R. Selmic

the magnitude and the rate of actuator inputs. These

Abstract— An indirect adaptive neural network (NN) constraints may be deliberately placed on the actuator to
saturation compensator is presented for a class of nonlinear avoid the damage on the system. When an actuator has
systems. Output feedback control is considered where only the reached such an input limit, it is said to be “saturated”,
system output is assumed to be measurable. The imposed
actuator saturation is assumed to be unknown and treated as
since efforts to further increase the actuator output would
the system input disturbance. A NN-based state observer not result in any variation in the output, which is a
estimates derivatives of the output and another NN-based challenge to the control design engineer. Particularly when
feedback controller is inserted into a feedforward path to integrator is present in the controller, an integrator wind up
capture the nonlinearities of the observed system and to cancel occurs and may generate additional problems.
the effects of the unknown disturbances and the unknown To tackle the problem, many saturation compensation
saturation nonlinearity. The unknown system states identified
by the NN observer are inputs of the NN-based controller. Two schemes have been developed based on a full state
NNs interact together to achieve the desired performance. measurement [1], [2], [3], [4], [6], [8], [9], [11], [12], [14].
Both adaptive, neural network control laws and on line neural However, this assumption is not always realistic in practical
net weights tuning rules are rigorously derived based on control systems. Sometimes, only the output or some of
feedback linearization and Lyapunov approach. The overall system states are known and available.
robust adaptive scheme guarantees that the states estimation
For adaptive control design, usually, the linearity in the
errors, NN weights estimation errors, and output tracking
errors are uniformly ultimately bounded. The simulation system parameters condition is required. Our NN-based
conducted indicates the proposed scheme can effectively observer that estimates the unknown nonlinear function as
estimate the unknown nonlinear system states and well as the unknown actuator saturation nonlinearity does
accommodate the unknown actuator constraints. not require this condition. NNs can successfully deal with
nonlinear systems that are linearly unparameterizable.
I. INTRODUCTION Observers for linear systems were proposed by

A daptive control techniques adjust the controller

parameters using system response to achieve the
desired performance. Adaptive control can accommodate
Luenberger. Recently, many researchers [5], [6], [7], [8],
[11], [13], [15] utilized NNs to synthesize the feedback
linearization based controls and incorporate the Lyapunov
the system modeling, and parametric and environmental theory to ensure the overall system stabilization, command
uncertainties [19]. Due to the non-analytic nature of the following, and disturbance rejection.
actuator nonlinear dynamics and the fact that the exact This paper presents a robust neural net, observer-based
actuator nonlinear functions, namely operation uncertainty, saturation compensator for a class of nonlinear systems in
are unknown, such systems provide an application field for Brunovsky canonical form. The paper is organized as
adaptive control, sliding control and neural network-based follows. Section 2 provides the preliminary remarks.
control. Proportional-derivative (PD) controller has Section 3 presents the nonlinear systems with saturation
observed limit cycles if the actuator nonlinearity exists. To nonlinearity. Section 4 addresses the NN-based observer.
overcome the unavoidable actuator nonlinearities such as Section 5 shows the design process of outer loop tracking
saturation, deadzone, backlash, hysteresis and the actuator controller. In Section 6, a simulation example on
failures, a number of researches [6], [8], [9], [11], [16], “pendulum type” nonlinear system with saturation
[18], [19] have developed techniques for their nonlinearity is given. The conclusion is drawn in Section 7.
compensation.
Saturation is a common nonlinearity in most actuators. II. PRELIMINARY REMARKS
Categories of saturation nonlinearities include constraints of Assuming unknown saturation operation limits τ min and
τ max , the output of the actuator τ (t ) is given by
The first author is working at Rockwell Corp. the second author is with
the Electrical Engineering Department, Louisiana Tech University, USA,
(e-mail: [email protected], [email protected] ).
  τ max : u (t ) ≥ τ max / m states and that only output y is measurable. The objective
  of the controller and compensator designs is to estimate the
τ (t) = mu (t ) : τ min / m < u (t ) < τ max / m , (2.1)
 τ : u (t ) ≤ τ min / m
unknown system states and find a control law u to
 min compensate for the actuator saturation.
where τ max is chosen positive and τ min is negative
saturation limits. The control that cannot be implemented IV. NN-BASED OBSERVER DESIGN
by the actuator δ (t ) , is given by A nonlinear observer that estimates unknown states of
 τ max − u(t ) : u(t ) ≥ τ max / m system (3.4) is given by
 
δ (t) = (m − 1)u(t ) : τ min / m < u(t ) < τ max / m . (2.2) { }
x&ˆ = Axˆ + b fˆ ( xˆ ) + δˆ (t ) + u (t ) + L( y − hT xˆ ) + vo (t )
, (4.1)
 τ − u(t ) : 
u(t ) ≤ τ min / m yˆ = hT xˆ
 min
The nonlinear actuator saturation can equivalently be where x̂ is an estimate of the state vector x , functions fˆ
described using δ (t ) [6] and NN can be used for its
and δˆ are estimates of f and δ respectively. A robust
approximation.
term vo (t ) is added to accommodate the function
III. NONLINEAR SYSTEMS WITH ACTUATOR SATURATION estimation error and unknown disturbances. The observer
Consider the nonlinear systems with state space matrix gain L = [l1 l2 L ln ]T is chosen so that
T
representation in Brunovsky canonical form As = A − Lh is stable i.e., for a symmetric positive
x&1 = x2 definite matrix Q , there exists a symmetric positive definite
x& 2 = x3 matrix P satisfying the following equations
M (3.1) AST P + PAS = −Q
. (4.2)
x& n = f ( x) + d (t ) + τ (t ) Pb = h
y = x1 Dynamic or recurrent NNs are capable to identify or
estimate internal states of an unknown nonlinear dynamical
where x = [ x1 , x2 , L , xn ]T , f : ℜ n → ℜ is an unknown system. Applying NN universal approximation property,
smooth function, d is the unknown bounded disturbance there exist a NN with some ideal weights Vo and Wo that
with a known upper bound D , and τ is the control input.
closely approximates the unknown function f + δ
Define matrices A , b , and h as
f + δ = WoT σ (VoT xoNN ) + εo . (4.3)
0 1 0 L 0  0 1
0 0 1 L 0  0 0  As in [6], a two layer NN has been used. The first layer
      weights V are selected randomly and will not be tuned; the
A= M  , b =  M  , h =  M . (3.2) second layer weights W are tunable. Implemented NN is
     
0 0 L 1 0  0 0  actually an approximation of the ideal NN (4.3), and is
0 0 0 L 0  1 0  given by
     
The equation (3.1) can be written in the following state fˆ + δˆ = Wˆ oT σ (Vo T xˆoNN ) , (4.4)
space form Since not all states are available, input to NN includes the
x& = Ax + b{ f ( x) + d (t ) + τ (t )} known and estimated states.
(3.3) The NN weights approximation error is given by
y = hT x ~
Wo = Wo − Wˆ o . (4.5)
Using the modified saturation nonlinearity (2.2), the above
system is converted to
x& = Ax + b{ f ( x) + d (t ) + u (t ) + δ (t )} Assumption 2 (Bounded Ideal NN Weights): The ideal
(3.4) NN weights Wo are bounded so that Wo ≤ Wo M , with
y = hT x
WoM a known bound.
Define the desired state vector xd (t ) as
Substituting (4.3) and (4.4) into (3.4) and (4.1)
xd (t ) = [ y d , y& d ,L, y d( n−1) ]T . (3.5) respectively yields

Assumption 1 (Bounded Desired Trajectory): The desired

{
x& = Ax + b WoT σ (VoT xoNN ) + ε + d (t ) + u (t ) } (4.6)
trajectory xd (t ) is bounded and continuous, and y = hT x
xd (t ) ≤ Q with Q a known scalar bound. { }
x&ˆ = Axˆ + b Wˆo T σ (VoT xˆoNN ) + u (t ) + L( y − hT xˆ ) + vo (t )
(4.7)
It is assumed that nonlinear function f (x) contains yˆ = hT xˆ
parameter uncertainties including known and unknown
 Subtracting (4.7) from (4.6) yields the estimation error 1 T ~ ~T
L&o = − ~ x Qx + x Pbε + ~ x T Pbd − ~ x T Pvo +
dynamics 2
x& = AS ~
~ {
~
x + b WoT σ (VoT xˆoNN ) + WoT σ~ + ε o + d − vo (t )}(4.8) tr WoT Γ −1Wo + σ (VoT xˆ oNN )( Pb)T ~
~ ~&
x  
~    
y = hT ~
x (4.14)
1 T ~ ~T
where σ~ = σ (VoT xoNN ) − σ (VoT xˆoNN ) . Based on the above =− ~ x Qx + x Pbε + ~ x T Pbd − ~ x T Pvo +
2
ideal bounded weight assumption for Wo and the property
tr WoT Γ −1Wo + σ (VoT xˆ oNN ) ~y  
~ ~&
of the NN activation function, it is obvious that W T σ~ is o  
  
bounded. Applying the NN tuning rules and robust term (4.10),
equation (4.14) is equivalent to
Corollary 1: Term WoT σ~ is bounded such that 1 T ~ ~T
L& o = − ~ x Qx + x Pbε + ~ x T Pbd − K r ~ x T P~ y−
2
WoT σ~ ≤ ε o M , with ε oM a known bound.
KW Wˆ o ~
F
x T P~ y +h ~
~
y tr WoT Wˆ o{ }
(4.15)
Then disturbance term ε = WoT σ~ + ε o is also bounded such 1 ~ T ~ ~T ~ T
= − x Qx + x Pbε + x Pbd − K r P ( Pb) x T ~ 2
2
that ε ≤ ε M where ε M is a known bound. Then, the
estimation error dynamics can be expressed as
− K WˆW o
F
P ( Pb)T ~
x +h ~
2 ~
y tr W T Wˆ { o o }
x& = A ~
~
S
~
{
x + b W T σ (V T xˆ
o ) + ε + d − v (t )
o oNN } o
. (4.9)
Let Pmin , Qmin be the smallest eigenvalues and Pmax , Qmax
~
y =h xT~ the largest eigenvalues of P , Q . Then (4.15) yields
1
L&o ≤ − Qmin ~ x + Pmax (ε N + bD) ~
2
x +
Theorem 1 (Tuning of NN Observer): Given the system in 2 (4.16)
(3.4) with saturation imposed and Assumptions 1 and 2,
choose the dynamic NN estimator equation (4.1), and the
h ( Pb)T ~
~
{
x tr W T Wˆ o o }
robustifying term as Using the inequality,
v =k ~ y + k Wˆ
o r
~y ,
W o
F
(4.10) ~
[ ~ ~
tr X T ( X − X ) ≤ X ] F
X F
~
− X
2

F
, (4.17)
with scalars k r > 0 and kW > 0 . Let the estimated NN (4.16) can be written as
weights be provided by the NN tuning algorithm 1
L&o ≤ − Qmin ~ x + Pmax (ε N + bD) ~
2
x +
& 2
Wˆ = Γσ(V T xˆ
o o)~
y − hΓ ~
y Wˆ ,
oNN (4.11)
o ~ ~ 2
T h ( Pb)T ~
x ( Wo Wo F − Wo ) (4.18)
where Γ = Γ > 0 is a constant matrix representing the F F
learning rates of the NN and h is a small scalar positive  1 ~ T ~ 1 2
( 2 Qmin x + h ( Pb) ( Wo F − 2 WoM ) 
design parameter. By properly selecting the control gains ≤− ~
x  
and the design parameters, the estimation error ~
x and the − P (ε + bD) − 1 h ( Pb)T W 2 
 max N
4
oM 
NN weights Ŵo are uniformly ultimately bounded.
which is guaranteed to remain negative as long as
h 2
Proof: Choose Lyapunov function candidate as ( Pb)T WoM + Pmax (ε N + bD)
1 T ~ 1 ~ ~ ~
x ≥ 4 , (4.19)
Lo = ~ x Px + tr (WoT Γ −1Wo ) (4.12) 1
2 2 Qmin
Differentiating yields 2
or
1 & T ~ ~ T ~& ~ ~&
L&o = ( ~
x Px + x Px ) + tr (WoT Γ −1Wo ) . (4.13) 1
2  2
Whence substitution from error equation (4.9), and using ~ 1  P (ε + bD) 1 2 
Wo ≥ WoM +  max N + WoM  . (4.20)
(4.2) yields F 2  h ( Pb)T 4 
 

V. NN-BASED CONTROLLER DESIGN

Let define the state tracking error vector, eˆ(t ) as
eˆ(t ) = xˆ (t ) − xd (t ) , (5.1)
and a filtered tracking error as
rˆ = K T eˆ , (5.2)
where K = [k1 , k 2 , ...k n −1 ,1]T is appropriately chosen δ = WcT σ (VcT xˆ cNN ) + ε c . (5.8)
coefficient vector [17], so that eˆ → 0 exponentially as Two layer NN is used where the second layer weights Wc
rˆ → 0 . Then, the time derivative of the filtered tracking are tunable and the first layer weights Vc are selected
error can be written as randomly and will not be tuned. Implemented NN is given
rˆ& = f ( x) + u + δ + d + Yd , (5.3) by
n −1 δˆ = WˆcT σ (VcT xˆcNN ) , (5.9)
where Yd = − y d( n ) + ∑ k eˆ i i +1 is a known function of the
where the NN weights approximation error is
i =1 ~
tracking error and the desired state. Wc = Wc − Wˆ c (5.10)
Applying the feedback linearization method choose the Input to the NN saturation compensator is chosen as
tracking control law as xˆcNN ≡ [ xd , eˆ]T .
w = − fˆ ( xˆ ) − Yd + vc − K v rˆ , (5.4)
ˆ . Assumption 4 (Bounded Ideal NN Weights): The ideal
where f ( xˆ ) is the fixed approximation of function f( ). The
NN weights Wc are bounded so that Wc ≤ WcM , with
functional estimation error is given by
~ WcM a known bound.
f = f − fˆ . (5.5)
Using control law (5.7) and (5.4), and substituting into
(5.3) overall closed-loop error dynamics is
Assumption 3 (Bounded Estimation Error): The nonlinear ~ ~
r&ˆ = f + WcT σ (VcT xcNN ) + ν c + d − Kν rˆ + ε c . (5.11)
function f is assumed unknown, but a fixed estimate fˆ is
assumed known, such that the functional estimation error Theorem 2 (Tuning of NN Observer and Compensator):
~
f satisfies Given the system (3.4) with saturation nonlinearity and
~ Assumptions 1-4, choose the dynamic NN estimator (4.1)
f ≤ fM (5.6)
and the robustifying term as
with f M a known bound. v =k ~ y + k Wˆ ~y , (5.12)
o r W o
F
As in [6], [16], approximation fˆ is fixed in this paper
with scalars k r > 0 and kW > 0 . Choose the tracking
and will not be adapted. Robust term vc is chosen for the
control law (5.7), the saturation compensator (5.9), and the
disturbance rejection. The feedback gain K v is a scalar. robustifying term as
The control law u is then a combination of the tracking vc (t ) = −( f M ( x) + D( x))sign(rˆ) , (5.13)
controller and the saturation compensator, as shown in the where the f M ( x), D( x) are bounds on the functional
Figure 1.
estimation error and disturbance, and sign(.) is standard
sign function. Let the estimated NN weights for the
observer and controller be provided by the NN tuning
algorithms
&
Wˆ = Γσ(V T xˆ
o )~
oy − hΓ ~
oNN y Wˆ , (5.14)
o
&
Wˆc = Sσ (VcT xcNN )rˆ − kS rˆ Wˆc , (5.15)
where Γ = Γ T > 0 and S = S T >0 are constant matrices
representing the learning rates of the NN, h and k are small
scalar positive design parameters. By properly selecting the
control gains and the design parameters, the estimation
Figure 1. Saturation compensation based on a NN observer scheme.
error ~x , the filtered error rˆ(t ) , the NN compensator
weights Ŵc , and the NN observer weights Ŵo are
The control signal is given by uniformly ultimately bounded.
u = w − δˆ , (5.7)
ˆ
where δ is the approximation of modified saturation Proof: Choose Lyapunov function candidate as
nonlinear function δ (x) . L = Lo + Lc , (5.16)
Using NN function approximation property, there exists a where Lo is defined in (4.12) and the Lc is given by
NN that closely approximates the modified saturation 1 1 ~ ~
Lc = rˆ 2 + tr (WcT S −1Wc ) . (5.17)
nonlinear function δ (x) 2 2
Differentiating yields
 ~ ~& pendulum” system [6], [17] with actuator saturation
L&c = rˆr&ˆ + tr (WcT S −1Wc ) (5.18)
nonlinearity.
Whence substitution from (5.11) yields, To show and focus on the functionality of adaptive NN
~ ~ ~ ~&
L&c = − Kν rˆ 2 + rˆf + rˆWcT σ (VcT xcNN ) + rˆvc + rˆd + rˆε c + tr (WcT S −1Wc ) saturation compensator, we choose the robust term vo = 0
~ ~ ~&
(5.19)
= − Kν rˆ 2 + rˆ( f + ν c + d + ε c ) + tr (WcT (S −1Wc + σ (VcT xcNN )rˆ)) vc = 0 in the following simulation.
Applying the NN compensator tuning rules, equation (5.19) We consider a nonlinear system of a pendulum type with
is equivalent to bounded disturbance d (t ) = 2 cos(t )
~ ~
L&c = − Kν rˆ 2 + rˆ( f + ν c + d + ε c ) + k rˆ tr (WcT Wˆc ) (5.20) x&1 = x2
Using (5.13) one has x& 2 = −5 x13 − 2 x 2 + d + τ . (6.1)
~ ~
L&c ≤ − Kν rˆ 2 + k rˆ tr (WcT (Wc − Wc )) − rˆ f M
(5.21) y = x1
~
− rˆ D + rˆ d + rˆ f + rˆ ε N The control input τ is constrained by the saturation
Using the inequality (4.17) yields nonlinearity characterized by the parameters
2 ~ ~ 2 τ max = 5, τ min = −5, m = 1 . (6.2)
L&c ≤ − Kν rˆ + k rˆ ( Wc Wc − Wc ) + rˆ ε N
F F F In this paper, the NN observer with sigmoidal activation
 ~ ~ 2  function has six, eleven and one neurons at the input,
≤ rˆ − Kν rˆ + k ( Wc WcM − Wc ) + ε N  . hidden and output layers, respectively. The first-layer
 F F 
weights Vo are selected randomly [10], [16]. They are
 ~ 1 1 
= rˆ − Kν rˆ − k ( Wc − WcM ) 2 + kWcM 2
+ ε N  ,(5.22) uniformly randomly distributed between –1 and +1. These
 F 2 4  weights represent the stiffness of the sigmoid activation
&
Combining terms Lo and Lc , the L becomes function. The threshold weights for the first layer v0 are
L& = L& + L&
o c
uniformly randomly distributed between -20 and +20. The
second layer weights Wo are initialized to zero.
 1 ~ T ~ 1 2
( 2 Qmin x + h ( Pb) ( Wo F − 2 WoM )  The NN compensator with sigmoidal activation function
≤− ~ x   (5.23) has four, fifteen and one neurons at the input, hidden and
− P (ε + bD) − 1 h ( Pb)T W 2  output layers, respectively. The first-layer weights Vc are
 max N 4
oM 
selected randomly. They are uniformly randomly distributed
 ~ 1 1  between –1 and +1. The threshold weights for the first layer
+ rˆ − Kν rˆ − k ( Wc − WcM ) 2 + kWcM 2
+εN 
 F 2 4  vc are uniformly randomly distributed between -20 and
which is guaranteed to remain negative as long as the +20. The second layer weights Wc are initialized to zero.
following inequalities hold
The outer tracking PD controller will make the whole
k 2
WcM + ε N system stable subject to the saturated constraints before the
rˆ ≥ 4 , (5.24) NN saturation compensators start learning and control.
Kν Tracking loop controller parameters are chosen as
1
K v = 4 , K = [2, 1]T . The initial conditions [ x1 (0), x 2 (0) ]T
 k 2 1 2
 WcM + ε N + WcM  are set to [0, 0]T ; the initial conditions [ xˆ1 (0), xˆ 2 (0) ]T are
~
Wc 
≥ 4 2  . (5.25)
F k  [0.05, 0.05]T , and desired trajectory is given by
 
  x1 (t ) = sin(t ), x2 (t ) = cos(t ) .
h 2 The NN observer weights tuning parameters are chosen
( Pb)T WoM + Pmax (ε N + bD)
~ 4 as
x ≥ , (5.26)
1 h = 0.002 , Γ = 2 . (6.3)
Qmin
2 The NN saturation compensator weights tuning
1 parameters are chosen as
 2 k = 0.0001 , S = 5 . (6.4)
~ 1  P (ε + bD) 1 2 
Wo ≥ WoM +  max N + WoM  . (5.27) The performance of the NN observer is shown in Figure
F 2  h ( Pb)T 4 
  2.

VI. SIMULATION OF NN SATURATION COMPENSATOR

The simulation was performed to verify the effectiveness
of the proposed NN compensator using a “generalized
 journal of Control, vol. 69, no. 6, pp. 773-788, Apr.
1998.
[5] F.-C. Chen, and H. K. Khalil, “Adaptive control of
nonlinear systems using neural networks,” Int. J. Contr.,
vol. 55, no. 6, pp. 1299-1317, 1992.
[6] W. Gao and R. Selmic, “Neural network control of a class
of nonlinear systems with actuator saturation,”
Proceedings of American Control Conference, Boston,
Figure 2. Performance of NN observer. Massachusetts, USA, June 2004.
[7] S. S. Ge, J. Zhang and T. H. Lee, “Adaptive neural
The position tracking errors and control signal are shown network control for a class of MIMO nonlinear systems
in the Figure 3 and Figure 4 with and without the saturation with disturbances in discrete-time,” IEEE Trans. System,
compensator. Man, and Cybernetics, vol. 34, pp. 1630-1645, Aug.
2004.
[8] P. He and S. Jagannathan, “Reinforcement learning based
output feedback control of nonlinear systems with input
constraints,” Proceedings of American Control
Conference, Boston, Massachusetts, USA, June 2004.
[9] T. Hu and Z. Lin, Control Systems with Actuator
Saturation: Analysis and Design, Birkhauser, Boston,
2001.
[10] B. Igelnik and Y. H. Pao, “Stochastic Choice of Basis
Figure 3. Tracking errors e1 (solid) and e2 (dotted) without NN saturation Functions in Adaptive Function Approximation and the
compensator (left) and with saturation compensator (right).
Functional-Link Net,” IEEE Trans. Neural Networks, vol.
6, no. 6, pp. 1320-1329, Nov. 1995.
[11] E. N. Johnson, A. J. Calise, “Neural Network Adaptive
Control of Systems with Input Saturation” Proc.
American Control Conference, Arlington, Virginia, June
2001.
[12] R. L. Kosut, “Design of linear systems with saturating
linear control and bounded states” IEEE Trans. Automat.
Contr., no. 28, vol. 1, pp. 121-124, 1983.
[13] F. L. Lewis, A. Yesildirek, and K. Liu, “Multilayer
Figure 4. Control signal τ without NN saturation compensator (left) and
with saturation compensator (right). neural-net robot controller with guaranteed tracking
performance,” IEEE Trans. Neural Networks, vol. 7, no.
2, pp. 1-11, Mar. 1996.
VII. CONCLUSIONS [14] W. Niu and M. Tomizuka, “A Robust Anti-Windup
Controller Design for Asymptotic Tracking of Motion
In this paper, an observer based NN saturation Control System Subjected to Actuator Saturation,” The
compensator for a class of nonlinear systems with unknown 37th IEEE Conference on Decision and Control, Tampa,
nonlinearity has been presented. Further research will be December 1998.
conducted on the more challenging MIMO systems with [15] M. M. Polycarpou, “Stable adaptive neural control
actuator saturation nonlinearity. scheme for nonlinear systems,” IEEE Trans. Automat.
Contr., vol. 41, no. 3, pp. 447-451, March 1996.
VIII. REFERENCES [16] R. R. Selmic and F. L. Lewis, "Deadzone compensation in
[1] K. J. Astrom and B. Wittenmark, Computer-Controlled motion control systems using neural networks," IEEE
Systems: Theory and Design Prentice Hall; 3rd edition, Trans. Automat. Contr., vol. 45, no. 4, pp. 602-613, April
1996. 2000.
[2] A. M. Annaswamy, S. Evesque, S. Niculescu, and A. P. [17] J. E. Slotine and W. Li, Applied Nonlinear Control J.-,
Dowling, “Adaptive Control of a Class of Time-delay Prentice Hall, 1991.
Systems in the Presence of Saturation,” Adaptive Control [18] G. Tao and P. V. Kokotovic, Adaptive Control of Systems
of Nonsmooth Dynamic Systems, Eds. G. Tao and F. With Actuator and Sensor Nonlinearities, John Wiley &
Lewis, Springer-Verlag, New York, NY, 2001. Sons, New York, 1996.
[3] B. C. Atherton, “An Analysis Package Comparing PID [19] G. Tao and S. M. Joshi, and X. Ma, “Adaptive state
Anti-Windup Strategies”, IEEE Control Systems feedback and tracking control of systems with actuator
Magazine, 15 vol. 2, pp. 34 – 40, April, 1995. failures,” Trans. Automat. Control, vol. 46, no. 1, pp. 78-
[4] C. W. Chan and K. Hui, “On the existence of globally 95, Jan 2001.
stable actuator saturation compensators,” International