You are on page 1of 10

Ind. Eng. Chem. Res.

2006, 45, 8963-8972 8963

PROCESS DESIGN AND CONTROL

Robust Control of a Class of Time-Delay Nonlinear Processes


Yi-Shyong Chou* and Kuen-Tsuen Jih
Department of Chemical Engineering, National Taiwan UniVersity of Science & Technology,
Taipei 106, Taiwan

In this paper, the problem of designing a tracking controller for uncertain nonlinear state-delay systems that
can suppress the effects of both unknown uncertainties and disturbances is investigated. The controller is
designed by using the sliding-mode control concept and the polynomial approximation method. One of the
features in this paper is that model uncertainties and state-delay terms are expressed as the Legendre polynomials
expansion. The expansion coefficients of the Legendre polynomials can furthermore be modified by the update
law derived from the Lyapunov stability theorem. The presented composite nonlinear controller, which consists
of a sliding-mode controller with a coefficient update law and a sliding-mode observer, can achieve offset-
free performance. The major advantage of the proposed control system design is to track the set point without
producing a vigorous control action and requiring the exact knowledge of model uncertainties. The control
scheme is illustrated by an example of a chemical reactor with delayed states. Simulation results indicate that
the proposed method can work for processes with time delays, despite unknown modeling uncertainties.

1. Introduction introduced the concept of flatness to derive the trajectory


tracking control law for nonlinear processes with time delays.
For the past decade, the analysis and control synthesis of time-
In all of the above methods, the admissible nonlinear uncertain-
delay processes has been one of the most active research areas
ties, disturbances, and time delays are usually assumed to be
in process system engineering. This is because time delays are
known a priori or to be satisfied with certain conditions.
frequently encountered in chemical processes. The existence of
However, a time delay in the state is often not measured and
time delays degrades the control performance and sometimes
the bounds of uncertainties and disturbances are difficult to
makes the closed-loop stabilization difficult. Although a number
estimate.
of significant time-delay compensation techniques have been
proposed to improve the control performance of linear time- Control of nonlinear systems with unknown uncertainties has
delay processes, only a few time-delay compensation strategies been a challenging problem. Slotine and Coetsee11 proposed
have been developed for nonlinear state-delay processes involv- an adaptive sliding control to reduce the effect of uncertainties
ing unmodeled dynamics or unknown uncertainties.1 There are on the process and further improve performance by on-line
many chemical processes that are governed by nonlinear state- parameter estimation algorithms. In their study, there is no need
delay equations. Examples are recycled reactors, recycled to know the bounds of unknown parameters. Huang and Kuo12
storage tanks, cold rolling mills, and so forth.2 Shyu and Yan3 presented a sliding control scheme for nonlinear processes with
and Luo and de la Sen4 employed the technique of variable unknown bound time-varying uncertainties. Those uncertainties
structure control to deal with the stabilizing of uncertain were represented in finite-term Fourier functions. Vecchio et
processes with delayed states. Velasco et al.5 proposed a linear al.13 developed a repeatable control design method for a class
approximation of the nonlinear model around an equilibrium of feedback linearizable processes with an unknown nonlinear
point and a causal static-state feedback law to deal with the function. They used the Fourier series to approximate the
disturbance decoupling problem of nonlinear time-delay pro- nonlinear function of the process. In the previously mentioned
cesses. Hu et al.6 introduced the linear matrix inequality papers, all uncertain elements that appeared in the process were
technique and the sliding-mode control method to solve the lumped and expanded in finite-term Fourier functions. This
control design problem of uncertain time-delay processes with requires us to choose sufficiently large terms of Fourier functions
mismatching uncertainties. Oguchi et al.7 presented a coordinate to approximate uncertainties.
transformation involving the delayed states to deal with The aim of this paper is to design a sliding controller for
nonlinear processes with multiple time delays in the state. nonlinear state-delay processes with time-varying uncertainties,
Lehman et al.8 developed the vibrational control technique for despite unknown magnitudes on time delays and model uncer-
nonlinear time-lag processes with arbitrarily large but bounded tainties. The key idea of the paper is to use orthogonal
delay. Niu et al.9 proposed a neural-network-based sliding-mode polynomials to represent model uncertainties and delayed states.
control scheme to solve the control problem of state-delay Because of the fact that the convergence of orthogonal
processes with mismatched parameter uncertainties, unknown polynomials is fast, one only needs to calculate a few terms of
nonlinearities, and external disturbances. Mounier and Rudolph10 coefficients. This allows us to improve the transient perfor-
mance. Once the unknown uncertainties and state-delay terms
* To whom all correspondence should be addressed. Fax: +886- have been approximated and parametrized, a feedback-control
2-27376644. E-mail: yschou@ch.ntust.edu.tw. law and an update law for adjusting the parameters of orthogonal
10.1021/ie0512901 CCC: $33.50 © 2006 American Chemical Society
Published on Web 11/14/2006
8964 Ind. Eng. Chem. Res., Vol. 45, No. 26, 2006

polynomials can be designed by using the Lyapunov stability The Legendre polynomials have the following important recur-
theory. The sliding-mode control structure is the backbone of rence property
the proposed control strategy development. The presented
composite nonlinear controller, which consists of a sliding-mode (n + 1)Pn+1(t) ) (2n + 1)P1(t)Pn(t) - nPn-1(t) (10)
controller with a coefficient update law and a sliding-mode
observer, can achieve offset-free performance. Furthermore, we The first several Legendre polynomials are given by
discuss the stability of the closed-loop process. Finally, the
P0(t) ) 1 (11)
effectiveness of the proposed method is demonstrated by
illustrated simulations.
2t b+a
The scope of this paper covers a class of time-delay nonlinear P1(t) ) - (12)
processes. It is not a general approach to all nonlinear problems.
b-a b-a
The class of problems is frequently encountered in process 3 2t b+a2 1
control of recycled reactors, recycled storage tanks, cold rolling
mills, and so forth.
P2(t) ) [ -
2b-a b-a
-
2 ] (13)

3. Problem Formulation
2. Function Approximation with Orthogonal Polynomials
Consider a class of dynamic processes that may be modeled
Let us suppose that a function f(t) defined on (a, b) is to be by
represented as a series of orthogonal polynomials {Φn(t)}, n )
0, 1, 2, ‚‚‚, that is, x(n) ) ft(X(t), X(t - τ)) + gt(X(t), X(t - τ))u(t) + D(t) (14)

where X ) [x x(1) ‚‚‚ x(n-1)]T is the process state vector, ft(X(t),
f(t) ) ∑ cnΦn(t), a < t < b
n)0
(1) X(t - τ)) and gt(X(t), X(t - τ)) are nonlinear functions, τ > 0
is a delay time in the process state, u(t) is the control input,
The values cn are the coefficients. Define and D(t) is the disturbance. Generally, a large class of chemical
processes can be represented with this type of model structure.
z(t) ) [Φ0(t) Φ1(t) ‚‚‚ Φn(t)]T (2) Examples are recycled reactors, recycled storage tanks, cold
rolling mills, and so on.2 Most of the recycling processes inherit
delays in their state equations. Equation 14 also can be expressed
W ) [c0 c1 ‚‚‚ cn]T (3)
in companion form,16 i.e.,

x̆1 ) x2
E(t) ) ∑ clΦl(t) (4)
l)n+1 x̆2 ) x3
Then eq 1 can be rewritten as l

f(t) ) WTz(t) + E(t) (5) x̆n ) ft(X(t), X(t - τ)) + gt(X(t), X(t - τ))u(t) + D(t) (15)

In applications, f(t) is approximated by a linear combination The control objective is to design a control law u to make x(t)
track xd(t) in the presence of unknown modeling errors in ft-
n (X(t), X(x - τ)) and gt(X(t), X(t - τ)) and unknown disturbances,
fn(t) ) ∑akΦk(t)
k)0
(6) D(t).
The sliding-mode design approach consists of two compo-
nents.16 The first one involves the design of a switching function
and the mean-square error is defined as the form of
so that the sliding motion satisfies design specifications. The
En ) ∫abw(t)[f(t) - fn(t)]2 dt (7)
second one is concerned with the selection of a control law
which will make the switching function attractive to the system
state. Thus, the first step is to define a time-varying surface
where w(t) is the weight function. If the mean-square error s(t) in the state-space as s(X,t) ) 0 that is a differential operator
approaches zero as n becomes infinite, the sequence {fn(t)} acting on some error function
converges in the mean to f(t).14 In other words, as long as n is

(dtd + γ)
large enough, f(t) can be approximated as n-1
s) e(t) (16)
f(t) ≈ W z(t)T
(8)
+∞
e(t) ) x(t) - xd(t) (17)
Note that f(t) must be in L2, that is, means ∫-∞ (f(x)2) dx < ∞.
It is well-known that many orthogonal polynomials are fre- where xd(t) is the desired trajectory and γ is a strictly positive
quently used to approximate the solution of time-varying constant, determining the performance of the system on the
systems because of their computation convergence advantages. sliding surface. The control goal is to force the process state
In this paper, the Legendre polynomials expansion is demon- x(t) to follow the specified trajectory xd(t). The problem is equal
strated to deal with unknown uncertainties and delayed states. to design a control law u(t) to ensure
The Legendre polynomials are defined in a finite interval [a,
b] by15 lim e(t) ) 0 (18)
tf∞
1 dn(t - a)n(t - b)n
Pn(t) ) (9) This condition can be achieved by making the system trajectory
n!(b - a)n dtn converge to the sliding surface. Thus, the problem of tracking
Ind. Eng. Chem. Res., Vol. 45, No. 26, 2006 8965

{ ()
trajectory can be reduced to keep s at zero. s s
if | | < 1
s φ φ
(dtd + γ)
n-1
e(t) ) 0 (19) ()
sat )
φ sgn
s s
if | | g 1
(28)
φ φ

() {
4. Controller Design s
+1 if > 0
s φ
The first step in controller design is to select a feedback sgn ) s (29)
control law u(t). The goal is to reach the sliding surface and φ -1 if < 0
remain on it. The second step is to estimate unknown functions φ
in ft(X(t), X(t - τ)) and gt(X(t), X(t - τ)) by employing
Substituting eq 27 into eq 26 gives
orthogonal polynomials. Classical orthogonal polynomials serve
as an excellent tool for modeling processes in an approximative
way. Here we select the Legendre polynomials in order to
improve the accuracy of approximation of unknown functions.
[
s̆ ) ∆g (f - xd(n) + Ω(e))(∆g-1 - ∆ĝ-1) +

Recall eq 16, (∆H - ∆Ĥ) - ηsat (φs )] (30)


(dtd + γ)
n-1 where
s) e(t) (20)
∆H ) ∆F∆g-1 (31)
The derivative of the sliding surface is calculated as follows,
Assume that ∆H and ∆g-1 are unknown bounded functions and
s̆ ) x (n)
- xd(n)
+ Ω(e) (21) can be approximated as

where ∆H ≈ WHTzF(t) (32)


∆g-1 ≈ WgTzF(t) (33)
n-1 (n - 1)!γn-k
Ω(e) ) ∑
k)1(n - k)!(k - 1)!
e (k)
(22) where

Here, we assume that the functions ft(X(t), X(t - τ)) and gt- zF(t) ) [P0(t) P1(t) ‚‚‚ Pn(t)]T (34)
(X(t), X(t - τ)) can be represented as17
WH ) [wH0 wH1 ‚‚‚ wHn]T (35)
ft(X(t), X(t - τ)) ) f(X(t)) + ∆f(X(t), X(t - τ)) (23)
Wg ) [wg0 wg1 ‚‚‚ wgn]T (36)
gt(X(t), X(t - τ)) ) g(X(t))∆g(X(t), X(t - τ)) (24)
Here, zF is the Legendre polynomials vector. WH and Wg are
This assumption decomposes ft(X(t), X(t - τ)) and gt(X(t), X(t considered as the constant parameter vectors. Generally the
- τ)), respectively, into two parts, where f(X(t)) and g(X(t)) number n should be properly selected so that the approximate
are referred to as the known functions whereas ∆f(X(t), X(t - errors are tolerable. ∆Ĥ and ∆ĝ-1 are the estimates of the
τ)) and ∆g(X(t), X(t - τ)) can be viewed as the time-varying unknown functions ∆H and ∆g-1, respectively, and can be
uncertainties that are unknown functions of time. Here, ∆g(X(t), represented as
X(t - τ)) is, in general, referred to as the matched uncertainty,
which is assumed to be bounded as below: ∆Ĥ ≈ ŴHTzF(t) (37)

0 < ∆gmin e ∆g(X(t), X(t - τ)) e ∆gmax (25) ∆ĝ-1 ≈ Ŵ gTzF(t) (38)

∆f(X(t), X(t - τ)) can be viewed as the mismatched uncertainty. where


With the aid of eqs 23 and 24, by substituting eq 14 into eq
21, one obtains ŴH ) [ŵH0 ŵH1 ‚‚‚ ŵHn]T (39)

s̆ ) (f + ∆F) + g∆gu - xd(n) + Ω(e) (26) Ŵg ) [ŵg0 ŵg1 ‚‚‚ ŵgn]T (40)
where ∆F ) ∆f + D. To make the value of s̆ equal to zero, we The update laws for adjusting the parameters ŴH and Ŵg will
consider the following control input of the form be derived.
Substituting from eqs 32 to 40 into eq 30 yields
1 s
u)
g[-∆Ĥ - ∆ĝ-1(f - xd(n) + Ω(e)) - ηsat
φ ( )] (27)
The thickness of the boundary layer φ is defined to be a positive
[
s̆ ) ∆g (f - xd(n) + Ω(e))(W̃gTzF(t)) +

real scalar. Outside the boundary layer, |s| g φ, the sliding


condition will be used to specify the s dynamics. Inside the
W̃HTzF(t) - ηsat (φs )] (41)
boundary layer, |s| < φ, the control law will be modified to where
impose a smoothing process to the s dynamics. The tuning gain
η is a strictly positive constant, and ∆Ĥ and ∆ĝ-1 are the W̃H ) WH - ŴH (42)
estimates of the unknown functions ∆F∆g-1 and ∆g-1, respec-
tively. The sat is the saturation function W̃g ) Wg - Ŵg (43)
8966 Ind. Eng. Chem. Res., Vol. 45, No. 26, 2006

Consider the Lyapunov function candidate as12

1 1
V(s, W̃H, W̃g) ) s2 + ∆g[W̃HTQHW̃H + W̃gTQgW̃g] (44)
2 2
where QH and Qg are symmetric positive definite matrices. The
derivative of V is

V̇ ) ss̆ + ∆g[W̃HTQHW̃˙H + W̃gTQgW̃˙g] (45)

Since WH and Wg are the constant vectors, W̃˙H ) -Ŵ˙H and W̃˙g
) -Ŵ˙g, the above equation can be further expressed as

V̇ ) ss̆ + ∆g[W̃HTQH(-Ŵ˙H) + W̃gTQg(- Ŵ˙g)] (46)

Substituting eq 41 into 46 and considering within the boundary


layer, we obtain

[
V̇ ) ∆g W̃Tg(zFs(f - xd(n) + Ω(e)) - QgŴ˙g) +

]
2
s
W̃HT(zFs - QHŴ˙H) - η (47)
φ
Choose the update laws as

Ŵ˙g ) Qg-1zFs(f - xd(n) + Ω(e)) (48)

Ŵ˙H ) QH-1zFs (49)

Then the derivative of V becomes

s2
V̇ ) -∆gη
φ
Figure 1. (a) Process responses and (b) controller responses.
Choose η ) η1/∆gmin, η1 > 0. Then

s2 s2 state feedback. The basic sliding observer structure includes


V̇ ) -∆gη e -η1 e 0 (50)
φ φ switching terms added to the design model. Here, we adopt
the basic sliding observer as a state estimator that is added
This implies that s is bounded. The parameter vectors W̃H and to the previous control scheme. A brief review of sliding
W̃g are also bounded. observer has been presented in the literature.19 Here, we consider
To use Barbalat’s lemma,18 let us check the uniform continu- the state-space representation of nonlinear systems of the
ity of V̇. The derivative of V̇ is form
2η x̆1 ) f1(x1, x2,‚‚‚, xn, u)
V̈ ) -∆g ss̆ (51)
φ
x̆2 ) f2(x1, x2,‚‚‚, xn, u)
This shows that V̈ is bounded, since each term in s̆ (eq 41) is
bounded, and s, W̃H, and W̃g were shown above to be bounded. l
Hence, V̇ is uniformly continuous. Application of Barbalat’s x̆n ) fn(x1, x2,‚‚‚, xn, u) (52)
lemma then indicates that s f 0 as t f ∞.
With the aid of the update laws eqs 48 and 49, the unknown where [x1 x2 ‚‚‚ xn]T is the state-space vector and x1 is assumed
functions, ∆H and ∆g, can be estimated and the control law eq to be the only measurable state. [f1 f2 ‚‚‚ fn]T is a nonlinear
27 can be implemented. Notice that ∆H and ∆g can be unknown function. Now, based on the sliding approach, the
approximated by a sum of n terms of orthogonal polynomials; observer dynamics can be written as
they must be bounded in real applications.

5. Sliding Observer
x̂˘ 1 ) f̂ 1(x̂1, x̂2,‚‚‚, x̂n, u) - k1sat ()
x̃1
φ

The previously described design method provides a dynamic


feedback that effectively deals with unknown uncertainties
x̂˘ 2 ) f̂ 2(x̂1, x̂2, ‚‚‚, x̂n, u) - k2sat ()
x̃1
φ
and asymptotically tracks the system if all the states are l

()
available. If they cannot be measured, then a state estimator
x̃1
is necessary. Sliding observer is a high-performance state x̂˘ n ) f̂ n(x̂1, x̂2, ‚‚‚, x̂n, u) - knsat (53)
estimator well-suited for nonlinear uncertain systems with partial φ
Ind. Eng. Chem. Res., Vol. 45, No. 26, 2006 8967

6. Numerical Example

In this section, simulations of the control in an uncertain


nonlinear time-delay system are performed by using the
proposed sliding control method. Consider the first-order,
irreversible, exothermic reaction A f B, carried out in a well-
mixed, continuously stirred tank reactor.8 Suppose that fresh
feed of pure A is mixed with a recycled stream of unreacted A
with recycle flow rate Fr(1 - λ). Notice that no recycled stream
occurs in the limit of λ to 1. If there is a delay, τr, in the recycle
stream, then the material and energy balances become8

dCA(t*) Frλ Fr(1 - λ) Fr


) CAf + CA(t* - τr) - CA(t*) -
dt* Vr Vr Vr
k0 exp -
E
RTr A ( )
C (t*) (55)

dTr(t*) Frλ Fr(1 - λ) Fr


) T + Tr(t* - τr) - Tr(t*) +
dt* Vr rf Vr Vr

( )
(-∆H) E UAr
k exp - C (t*) - (T (t*) - Tj(t*))
FrCpr 0 RTr A FrCprVr r
(56)

dTj(t*) Fj UAr
) (Tjf - Tj(t*)) + (T (t*) - Tj(t*)) (57)
dt* Vj FjCpjVj r

Typically, the above equations are reduced to dimensionless


form using the notation

Figure 2. (a) Process responses and (b) controller responses. E Vr t*


ψ) , τ0 ) , t ) , Da ) k0τ0 exp(-ψ),
RTrf Frλ τ0
where {k1, k2, ‚‚‚, kn} are positive numbers, x̃1 ) x̂1 - x1 is the UAr
estimation error, and f̂(x̂, u) is the estimated value of f(x̂, u). By βr ) τ
FrCprVr 0
using eqs 52 and 53, the resulting error dynamics can be written
as UAr (-∆H)CAf Tjf - Trf
βc ) τ0, B ) ψ, x3f ) ψ,
FjCpjVj FrCprTrf Trf
x̃˘ 1 ) δf1 - k1sat ()
x̃1
φ x1 ) 1 -
CA
CAf

x̃˘ 2 ) δf2 - k2sat ()


x̃1
φ x2 )
Tr - Trf
Trf
ψ, x3 )
Tj - Trf
Trf
Fj
ψ, u ) τ0, τs )
Vj
τr
τ0
l

x̃˘ n ) δfn - knsat ()


x̃1
φ
(54)
Equations 55-57 in dimensionless variables become

1 1
( )
x̆1(t) ) - x1(t) + - 1 x1(t - τs) +

{ }
λ λ
where δf ) f̂(x̂, u) - f(x, u). x2(t)
The more sophisticated estimation techniques can be applied Da(1 - x1(t)) exp (58)
x2(t)
to improve state estimation; however, this is not the main 1+
concern in this paper. Here, we adopt the basic sliding ψ
observation method to estimate the process states without time
1 1
delays. Certainly there exists a discrepancy between the actual
( )
x̆2(t) ) - x2(t) + - 1 x2(t - τs) +

{ }
state and the estimated state. This will degrade the performance λ λ
of tracking. The merit of this paper is that we decompose ft- x2(t)
(X(t), X(t - τ)) and gt(X(t), X(t - τ)), respectively, into two BDa(1 - x1(t)) exp - βr(x2(t) - x3(t)) (59)
x2(t)
parts, where f(X(t)) and g(X(t)) are, in general, referred to as 1+
ψ
the known parts, whereas ∆f(X(t), X(t - τ)) and ∆g(X(t), X(t -
τ)) are viewed as the unknown parts, which can be compensated
x̆3(t) ) βc(x2(t) - x3(t)) + u(x3f - x3(t)) (60)
by the proposed controller.
8968 Ind. Eng. Chem. Res., Vol. 45, No. 26, 2006

Following the proposed method, the sliding surface can be


expressed as

s ) ĕ(t) + γe(t) ) (x̆2(t) - x̆d(t)) + γ(x2(t) - xd(t)) (65)

and then the derivative of s can be computed as

s̆ ) f + ∆f + g∆gu + D - ẍd(t) (66)

where

[ { }
x2(t)
f ) -x2(t) + BDa(1 - x1(t)) exp - βr(x2(t) -
x2(t)
1+

]( { } )
ψ
1 - x1(t) x2(t)

( )
x3(t)) λ + BDa 2
- 1 - βr -
exp
x2(t) x2(t)
1+ 1+

( { })
ψ ψ
x2(t)
-x1(t) + Da(1 - x1(t)) exp ×
x2(t)
1+

( { })
ψ
x2(t)
BDa exp - βrβc(x2(t) - x3(t)) - λx̆d(t) (67)
x2(t)
1+
ψ
∆f + D ) ∆F

( (
1 - x1(t)

)
) (∆x2(t, t - τs) + d) λ + BDa 2
×
x2(t)
1+

{ } )
ψ
x2(t)
exp - 1 - βr +
Figure 3. (a) Process responses and (b) controller responses. x2(t)
1+

{ }
ψ
The main control objective for the reactor would be the
x2(t)
regulation of the product concentration around a desired set- ∆x1(t, t - τs)BDa exp +
point value. The close relationship between the concentration x2(t)
of component A and the temperature enables us to choose the 1+
ψ
regulation of the reactor outlet temperature about a correspond- ∆x̆2(t, t - τs) + ḋ (68)
ing set-point value as our control objective. The coolant flow
rate is considered here as the manipulated variable.
g ) βr(x3f - x3(t)) (69)
6.1. Controller Design. To further conform to the design
forms, the above equations are rearranged as
∆g ) 1 (70)
1 1
(
x̆1(t) ) - x1(t) + - 1 x1(t) +) The unknown function ∆F is assumed as a bounded continuous

{ }
λ λ function of time and is approximated by the Legendre polyno-
x2(t) mials. The control law is
Da(1 - x1(t)) exp + ∆x1(t, t - τs) (61)
x2(t)
1+ 1 s
ψ u)
g[- ŴHTzF - ŴgTzF(f - ẍd) - ηsat
φ ( )] (71)

1 1
( )
x̆2(t) ) - x2(t) + - 1 x2(t) + BDa(1 - x1(t)) exp Ŵ˙g ) Qg-1zFs(f - ẍd)

{ }
λ λ (72)
x2(t)
- βr(x2(t) - x3(t)) + ∆x2(t, t - τs) + d(t) (62) Ŵ˙H ) QH-1zFs (73)
x2(t)
1+
ψ The parameters for all simulations are selected to be λ )
0.5, B ) 8, Da ) 0.135, ψ ) 20, βc ) 1.5, βr ) 1.5, x3f ) -4,
x̆3(t) ) βc(x2(t) - x3(t)) + u(x3f - x3(t)) (63) uin ) 0.97, x1in ) 0.65, x2in ) 3, and x3in ) 0.26, where the
subscript {in} denotes the initial state. The controller parameters
y ) x2(t) (64) are selected to be γ ) 1.5, η ) 1.2, and Qg ) QH ) 0.1 × I,
where I denotes the identity matrix with a proper dimension.
where d(t) is added to the system to represent the process The thickness of boundary layer φ is selected as 0.2 during the
disturbance. course of the simulation.
Ind. Eng. Chem. Res., Vol. 45, No. 26, 2006 8969

Figure 4. (a) Process responses and (b) controller responses. Figure 5. (a) Process responses and (b) controller responses.

6.1.1. Case 1. In this case, all the states are available. Figure Figure 2b shows the input moves produced by the corresponding
1a shows that the proposed controller, the nonlinear controller controllers. Because of the additional oscillatory disturbance,
based on the extension of the Lie derivative developed by the controller moves of nonlinear MPC display the chattering
Oguchi et al.7 (the new input V ) -4z1 - z2 is designed), and behavior. Figure 3a shows the simulation results of the process
the nonlinear model predictive control (MPC)20 are compared having a set-point change from x2 ) 3 to x2 ) 4 and a time
for the process having a time delay τs ) 2 in the state and a delay τs ) 2 in the state for t e 10 and a τs ) 5 in the state for
set-point change from x2 ) 3 to x2 ) 4. The nonlinear MPC is t > 10, with the proposed controller, the nonlinear controller
designed with these parameters: the sampling interval is 0.01, based on the extension of the Lie derivative, and the nonlinear
the prediction horizon is 30, and the control horizon is 10. The MPC. Figure 3b shows the input moves of the corresponding
controller moves are set within the range -2 e u e 1.2. The controllers. Notice that the proposed controller initially needs
nonlinear constrained optimization problem is solved by using some time to update the estimation of model uncertainties. This
the MATLAB Optimization Toolbox routine FMINCON. The results in a sluggish response. Once the work of updating param-
nonlinear controller based on the extension of the Lie derivative eters is finished, the control system possesses satisfactory robust-
produces a large overshoot. The simulation results show that ness. It is evident that the performance retains a nearly similar
nonlinear MPC exhibits a response with the offset to the set response under a time-delay change τs ) 5 for t > 10, with
point within the specified control constraints. By contrast, our respect to that obtained with a time delay τs ) 2 for t e 10.
proposed controller provides a rapid, smooth, and conservative For the sake of demonstrating the choice of polynomial terms,
response. Such a response is much more suitable for chemical Figure 4a compares the proposed controller with different
process control applications. Figure 1b shows the input moves polynomial terms for the process having a time delay τs ) 2 in
produced by the corresponding controllers. This result indicates the state and a set-point change from x2 ) 3 to x2 ) 4. The
that the superior performance of the proposed controller is not controller with three-term polynomials yields a larger deviation
attributable to more aggressive control action but rather to a from the set point, while the controller with five-term polynomi-
more judicious use of the input. Figure 2a shows that the als provides more effective tracking ability. Figure 4b shows
proposed controller, the nonlinear controller based on the the input moves produced by the corresponding controllers.
extension of the Lie derivative, and the nonlinear MPC are Figure 5a shows the simulation results of the process having a
compared for the process having a time delay τs ) 2 in the time delay τs ) 2 in the state, a set-point change from x2 ) 3
state, a set-point change from x2 ) 3 to x2 ) 4, and an oscillatory to x2 ) 4, and an oscillatory disturbance d ) 0.1 sin(t). The
disturbance d ) 0.1 sin(t). As before, the performance of the controller with five-term polynomials improves the performance
proposed controller shows less deviation from the set point. of tracking. Figure 5b shows the input moves produced by the
8970 Ind. Eng. Chem. Res., Vol. 45, No. 26, 2006

and a time delay τs ) 2 in the state for t e 10 and a τs ) 5 in


the state for t > 10, with different polynomial terms. Figure 6b
shows the input moves of the corresponding controllers.
Observing the performance of the proposed design method,
a set-point tracking and disturbance attenuation can be achieved
for bounded uncertainties arising from delayed states and
disturbances.
6.1.2. Case 2. In this case, only the reactor temperature x2 is
available. The observer for this reactor system can be written
of the form

1
( )
1
x̂˘ 1(t) ) - x̂1(t) + - 1 x̂1(t) +

{ }
λ λ
Da(1 - x̂1(t)) exp
x̂2(t)
x̂2(t)
+ k1sat (
(x2 - x̂2)
φ
(74))
1+
ψ
1
( )
1
x̂˘ 2(t) ) - x̂2(t) + - 1 x̂2(t) + BDa(1 - x̂1(t)) exp

{ }
λ λ
x̂2(t)
x̂2(t)
- βr(x̂2(t) - x̂3(t)) + k2sat((x2 - x̂2)
φ )
(75)
1+
ψ

x̂˘ 3(t) ) βc(x̂2(t) - x̂3(t)) + u(x3f - x̂3(t)) + k3sat ( )


(x2 - x̂2)
φ
(76)
Since this state x2(t) is measurable, the computation of x̂1(t) and
x̂3(t) is performed. The choice of {k1, k2, k3} ) {0.1, 1.6, 0.1}
is done by trial and error.21 The initial values of estimate
variables are selected as {0.65, 3, 0.26} during the course of
the simulation. The thickness of boundary layer φ is designed
Figure 6. (a) Process responses and (b) controller responses. as 0.05 to avoid the occurrence of chattering phenomena. Figure
7a shows the dynamic behavior of the process having a time
corresponding controllers. Figure 6a shows the simulation results delay τs ) 2 in the state in which the reference value of the
of the process having a set-point change from x2 ) 3 to x2 ) 4 dimensionless temperature x2 is changed from x2 ) 3 to x2 )

Figure 7. (a) Process responses, (b) controller responses, (c) estimation result of x1, and (d) estimation result of x3.
Ind. Eng. Chem. Res., Vol. 45, No. 26, 2006 8971

Figure 8. (a) Process responses, (b) controller responses, (c) estimation result of x1, and (d) estimation result of x3.

Figure 9. (a) Process responses, (b) controller responses, (c) estimation result of x1, and (d) estimation result of x3.

4. Figure 7b shows the input moves of the controller. The loop performance remains satisfactory despite the extremely
estimated states x1 and x3 of this system are depicted in parts c noisy nature of the time-varying disturbance. Figure 8b shows
and d of Figure 7, respectively. By comparing the estimated the input moves of the controller. Parts c and d of Figure 8
value and the real value, one can see that they present a very present the estimated value and the real value of x1 and x3,
close behavior. Figure 8a presents the simulation results for respectively. Figure 9a shows the simulation results of the
tracking problem with the process having a time delay τs ) 2 process having a set-point change from x2 ) 3 to x2 ) 4 and a
in the state, a set-point change from x2 ) 3 to x2 ) 4, and an time delay τs ) 2 in the state for t e 10 and a τs ) 5 in the
oscillatory disturbance d ) 0.1 sin(t). In this case, the closed- state for t > 10, with the proposed controller. Figure 9b shows
8972 Ind. Eng. Chem. Res., Vol. 45, No. 26, 2006

the input moves of the controller. Parts c and d of Figure 9 (5) Velasco, M.; Alvarez, J. A.; Castro, R. Approximate Disturbance
present the estimated value and the real value of x1 and x3, Decoupling for a Class of Nonlinear Time Delay Systems. Proc. Am. Control
Conf. 1993, 1046-1050.
respectively. As a whole, the simulation results show that
(6) Hu, J.; Chu, J.; Su, H. SMVSC for a Class of Time-delay uncertain
satisfactory control performances can be obtained by applying Systems with Mismatching Uncertainties. IEE Proc. Control Theory Appl.
this proposed strategy. 2000, 147, 687-693.
(7) Oguchi, T.; Watanabe, A.; Nakamizo, T. Input-Output linearization
7. Conclusions of Retarded Nonlinear Systems by Using an Extension of Lie Derivative.
Int. J. Control 2002, 75, 582-590.
A control strategy to track a class of nonlinear state-delay (8) Lehman, B.; Bentsman, J.; Lunel, S. V.; Verriest, E. I. Vibrational
systems with time-varying uncertainties is presented in this Control of Nonlinear Time Lag Systems with Bounded Delay: Averaging
paper. Model uncertainties are represented as the expansion of Theory, Stabilizability, and Transient Behavior. IEEE Trans. Autom. Control
the Legendre polynomials and are estimated by updating the 1994, 39, 898-912.
expansion coefficients. The proposed algorithm takes advantage (9) Niu, Y.; Lam, J.; Wang, X.; Ho, D. W. C. Sliding-mode Control for
Nonlinear State-delayed Systems Using Neural-network Approximation. IEE
of the fast convergence characteristics of the orthogonal Proc. Control Theory Appl. 2003, 150, 233-239.
polynomials. The closed-loop performance of the proposed (10) Mounier, H.; Rudolph, J. Flatness-based Control of Nonlinear Delay
controller is illustrated by numerical simulation results. It is Systems; a Chemical Reactor Example. Int. J. Control 1998, 71, 871-890.
shown that the proposed method is able to track the desired (11) Slotine, J. J. E.; Coetsee, J. A. Adaptive Sliding Controller Synthesis
trajectory despite time-varying modeling uncertainties. for Nonlinear Systems. Int. J. Control 1986, 43, 1631-1651.
The sliding-mode control structure is the backbone of the (12) Huang, A. C.; Kuo, Y. S. Sliding Control of Nonlinear Systems
proposed control strategy development. By using polynomials Containing Time-Varying Uncertainties with Unknown Bounds. Int. J.
to approximate unknowns and adjusting polynomial coefficients Control 2001, 74, 252-264.
by the update law derived from the Lyapunov stability theorem, (13) Vecchio, D. D.; Marino, R.; Tomei, P. Adaptive State Feedback
Control by Orthogonal Approximation Functions. Int. J. Adapt. Control
it provides more robust capability than the traditional sliding- Signal Process. 2002, 16, 635-652.
mode control can offer. It can be viewed as a form of adaptive (14) Johnson, D. E.; Johnson, J. B. Mathematical Methods in Engineer-
control. It is well-known that adaptive control is suitable for ing and Physics; Ronald Press Co.: New York, 1965.
control problems with time-varying uncertainty. (15) Sansone, G.; Diamond, A. H.; Hille, E. Orthogonal Functions;
The presented technique is a general approach. It is applicable Robert E. Krieger Publishing Co.: Huntington, NY, 1977.
to those problems with any combination of uncertain delayed- (16) Slotine, J. J. E. Sliding Controller design for nonlinear systems.
state, unknown disturbance, and uncertainty. It can be used to Int. J. Control 1984, 40, 421-434.
solve problems with or without time delay in the states. (17) Li, M.; Wang, F.; Gao, F. PID-based Sliding Mode Controller for
Therefore, its applicability in chemical engineering is quite Nonlinear Processes. Ind. Eng. Chem. Res. 2001, 40, 2660-2667.
common; even our paper stresses its effectiveness with uncertain (18) Slotine, J. J. E.; Li, W. Applied Nonlinear Control; Prentice Hall
International Inc.: Englewood Cliffs, NJ, 1991.
delayed state.
(19) Wang, G. B.; Peng, S. S.; Huang, H. P. A Sliding Observer for
Nonlinear Process Control. Chem. Eng. Sci. 1997, 52, 787-800.
Literature Cited (20) Henson, M. A. Nonlinear Model Predictive Control: Current Status
(1) Richard, J. P. Time-delay Systems: An Overview of Some Recent and Future Directions. Comput. Chem. Eng. 1998, 23, 187-202.
Advances and Open Problems. Automatica 2003, 39, 1667-1694. (21) Unsal, C.; Kachroo, P. Sliding Mode Measurement Feedback
(2) Malek-Zavarei, M. Time-delay Systems Analysis, Optimization and Control for Antilock Braking Systems. IEEE Trans. Control Syst. Technol.
Applications; Elsevier Science: New York, 1987. 1999, 7, 271-281.
(3) Shyu, K. K.; Yan, J. J. Robust Stability of Uncertain Time-delay
and its Stabilizing by Variable Structure Control. Int. J. Control 1993, 57, ReceiVed for reView November 21, 2005
237-246. ReVised manuscript receiVed September 20, 2006
(4) Luo, N.; de la Sen, M. State Feedback Sliding Mode Control of a Accepted September 29, 2006
Class of Uncertain Time Delay Systems. IEE Proc. Control Theory Appl.
1993, 140, 261-274. IE0512901

You might also like