Professional Documents
Culture Documents
Sliding Mode Control System Design For Chemical Processes
Sliding Mode Control System Design For Chemical Processes
Sliding Mode Control System Design For Chemical Processes
化 學 工 程 學 系
博 士 論 文
化工程序之滑動模式控制系統設計
Sliding Mode Control System Design
for Chemical Processes
指導教授:陳奇中
研 究 生:彭世典
中 華 民 國 九 十 三 年 十 一 月
化工程序之滑動模式控制系統設計
ACKNOWLEDGEMENTS
This thesis would not have been possible without the assistance from many people
who gave their support in different ways. To them I would like to convey my heartfelt
With his enthusiasm, and his great efforts to explain things simply and clearly, he helped
to make process control fun for me. Throughout my thesis-writing period, he provided
encouragement, good teaching, and lots of good ideas. I would have been lost without
him. I would also like to acknowledge the members of my thesis committee: Drs.
Chien, Chyi-Tsong Chen, Shyh-Hong Hwang, Yi-Shyong Chou and Chueh-Ting Chang. I
am grateful to them for their advice, suggestions and critical review of the draft of this
thesis.
I would like to thank my fellow graduate students who offered many valuable
discussions and a helping hand in time of need: Mr. Ching-Kuo Wu, Mr. Cheng-Tien Hsu,
Mr. Nan-Chieh Wen, Mr. Yu-Jung Li, Mr. Ying-Chih Lo, Mrs. Chia-Chi Li and Mrs.
Fang-Yu Hu. I would also like to thank the members of the staff and faculty of the
exceptional friends, Ms. Ling-Yu Chen and Ms. Hsiao-Chun Huang, I extend my warmest
of heartfelt appreciation. Their friendship helped me to cope with the rigors of life as a
graduate student, and made it possible to experience aspects of life other than my work.
Lastly, and most importantly, I wish to thank my parents, Mr. Sung-Tsung Peng and Mrs.
Chun-Yueh Yang, and my brother, Mr. Shih-Chiang Peng, for being there when I needed
中文摘要
大部分的化工程序都存在著多變的動態特性,如非線性、不確定
性、時延與非最小相行為等。這些多變的動態特性使得控制系統的設計
變得更加困難與棘手。因此,如何設計一個優良的控制架構以操控具有
多變動態特性的化工程序,乃是一個非常重要的課題。就理論上而言,
滑動模式的設計能使系統有快速的應答、良好的暫態響應,以及對於系
統的參數變動及外界的擾動有低敏感性的優點。有鑑於此,本論文擬結
合滑動模式及其他實用的技術,提出一個強韌控制策略,以解決化工程
序的控制問題。論文將分四個部分來進行探討與研究,在第一個部分
中,對於線性或非線性程度較低的化工程序,將基於系統識別技術所得
的系統模式來進行控制系統的設計。其中,包含了滑動模式控制器與預
估器的設計與結合,此外,為了使達滑行模式控制後之狀態變動為最
小,吾人也將嘗試導入最佳切換超平面的設計。並將此技術應用於非最
小相的化工程序,進而延伸至多變數化工程序的控制上。另一方面,由
於化工程序大部分皆呈現非線性的動態行為,因此在論文的第二部分
中,將直接基於非線性系統模式進行滑動模式控制器與預估器的設計與
結合,以處理非線性不確定時延之化工程序的控制問題。再者,為了克
服因模式誤差與不可量測擾動所產生的影響,也將嘗試引入預估之狀態
修正項來解決此問題。基於論文第二部分之研究,在第三部分中,將更
進一步地處理高度非線性非最小相不確定時延程序的控制問題。藉由靜
態等效輸出觀念的導入,吾人將得以拓展第二部分的研究成果,來設計
一個結合非線性滑動模式控制器與預估器的架構。由於在滑動模式控制
器的設計中,對於不確定性的種類仍有限制且尚未能涵蓋所有程序的不
確定性,因此為了讓所提的滑動模式控制策略應用更為廣泛。最後,在
論文的第四部分中,吾人將提出一非線性控制策略,其結合了滑動模式
架構與人工智慧中的模糊類神經技術,以處理當系統動態未精確得知時
之非線性化工程序的控制問題。於此,在此論文中吾人將可利用滑動模
式與其他技術的結合,完整地處理化工程序常見的不確定性與時延問
題。甚者,連一向棘手的非最小相系統與多變數系統亦將同時提出解決
的方案。由廣泛的模擬測試結果顯示,本論文所提出的滑動模式控制策
略不但可行且具有優越的控制表現,相信對複雜多變的化工程序之操控
定有十足的助益。
ABSTRACT
This thesis considers the sliding mode control of chemical processes. For handling
and/or time-delay, four sliding mode control schemes are proposed. Firstly, with the
SOPDT model is proposed for uncertain chemical processes with moderate nonlinearity.
scheme for multivariable processes is also explored. The second part develops a
robust stability and performance of the closed-loop system. Subsequently, in the third part,
a sliding mode control scheme is explored for nonlinear, non-minimum phase, uncertain
compensate the process’s input-delay and to circumvent the negative effect of inverse
response. In order to release the constraints and assumptions on process uncertainties for
the sliding mode control system design, in the fourth part a nonlinear control scheme that
incorporates a sliding mode controller and a neural fuzzy strategy is proposed for the
known. Extensive simulation results reveal that the proposed sliding mode control system
design methodologies are applicable and promising for the regulation control of chemical
TABLE OF CONTENTS
1. Introduction ................................................................................................................ 1
................................................................................................................................. 55
4.2. A Case Study: Regulation Control of a Van de Vusse Reactor in the Presence of
5. A Sliding Mode Control Scheme for Imprecisely Known Chemical Processes ... 102
5.4. A Case Study: The Regulation Control of a Nonlinear Uncertain CSTR ..............118
References .......................................................................................................................138
Appendix D. Design of Observer Gains for the Nonlinear CSTR ............................ 150
LIST OF FIGURES
Figure 2.1. A schematic diagram of the proposed sliding mode control system ............. 14
Figure 2.3. Closed-loop system performance of Example 2.1 in the presence of parameter
uncertainties: steady state gain: +50%, each of time constant and dead-time:
-50% .............................................................................................................. 32
Figure 2.8. Closed-loop system performance of the Van de Vusse reactor ...................... 37
Figure 2.9. Closed-loop system performance of the Van de Vusse reactor in the presence
Figure 2.10. Closed-loop system performance of the Van de Vusse reactor in the presence
Figure 2.11. A multi-loop SMC decentralized control system for 2 × 2 process systems
........................................................................................................................ 40
Figure 2.12. Control performance of the proposed approach with and without static
decouplers ...................................................................................................... 41
Figure 2.15. Control performance of the proposed approach in the face of extra
disturbances ................................................................................................... 44
Figure 3.1. Schematic diagram of the proposed sliding mode control system for nonlinear
Figure 3.2. System performance comparison of the proposed strategy with and without
Figure 3.4. Closed-loop system performance in face of process parameter variations ... 64
Figure 3.5. Closed-loop system performance under modeling error in input delays ...... 65
Figure 4.1. Schematic diagram of the proposed predictor-based SMC scheme .............. 68
Figure 4.4. Open-loop simulations using SEOM with designed values of λ s ’s ............ 95
Figure 4.6. Closed-loop system performance in the face with unmodeled side reaction,
........................................................................................................................ 98
Figure 4.8. Schematic diagram of the sliding mode control system along with a sliding
observer ......................................................................................................... 99
Figure 4.9. System response under sliding observer with observer poles of p1 = p 2 = −4
for the case of existing unmodeled side reaction and measuring error ......... 100
Figure 4.10. Closed-loop system performance with time-delay variations .................... 101
Figure 5.1. Schematic diagram of the proposed nonlinear control system ..................... 105
Figure 5.3. System response of the CSTR under the use of a sliding observer with
Figure 5.4. System response of the CSTR under the use of a sliding observer with
Figure 5.5. The effects of boundary layer thickness of the switching function on system
Figure 5.6. Comparison with a PIDSMC control scheme with different tuning constants
....................................................................................................................... 132
Figure 5.7. System performance comparison of the proposed strategy with the PIDSMC
method in the presence of the unmodeled side reaction, measuring error and
Figure 5.8. System performance of the proposed scheme in the face of unmodeled side
LIST OF TABLES
Table 2.1. Controller settings of the proposed approach and comparative methods for
Table 5.1. The initial linking weights for the nonlinear uncertain CSTR (7 segments) 122
NOMENCLATURE
a defined as a = 1 τ
a (⋅), aˆ (⋅) nonlinear functions, defined as Eq. (3.6), Eq. (4.17) and Eq. (5.5)
∆a (⋅), ∆aˆ (⋅) nonlinear functions, defined as Eq. (3.7), Eq. (4.18) and Eq. (5.11)
a~ shape parameter of the activated function in the output layer of the MNN
a1 defined as a1 = 1 τ 2
a2 defined as a 2 = 2ζ τ
b defined as b = K τ
b1 defined as b1 = K τ 2
b(⋅), bˆ(⋅) nonlinear functions, defined as Eq. (3.8), Eq. (4.19) and Eq. (5.6)
∆b(⋅), ∆bˆ(⋅) nonlinear functions, defined as Eq. (3.9), Eq. (4.20) and Eq. (5.12)
Ba heat of reaction
c1 , c 2 , ci sliding coefficients
Da Damökhler number
d1 , d 2 , d 3 unmeasured disturbances
d s (δ ) switching function
)
d ( s) disturbance
d (x ∗ , t ) defined as d (x ∗ , t ) = c 2 d max (x ∗ , t )
1
E objective function, E ≡ (δ (ξ )) 2
2
~ ~ 1 ~
E objective function, E ≡ (δ − δ ) 2
2
e f , eg model uncertainties
∆f (⋅) uncertainty
∆g(⋅) uncertainty
∆H i reaction heat
k , kˆ adaptive gain
)
k sampling number
k1 , k 2 , k 3 , k i 0 rate constants
k̂ 01 , k̂ 02 , k̂ 03 switching gains
k i (T ) rate coefficient
K3 scaling factor
~
P (s) defined in Eq. (4.8)
~
Pd ( s ) defined in Eq. (4.10)
q(⋅), qˆ (⋅) defined as Eq. (3.10), Eq. (4.21) and Eq. (5.7)
⎡q q12 ⎤
Q positive definite, symmetric matrix, defined as Q = ⎢ 11
⎣ q 21 q 22 ⎥⎦
~
Q j (s) defined in Eq. (4.9)
t time
T0 feed temperature
Td steady-state temperature
u control input
um dilution rate
V reactor volume
w characteristic index
~ ,w
w ~ weighting parameters
2 ij 3i
wυ linking weight
y, yˆ system output
~
y model output
z 0j desired zero
z ij zero
Greek Symbols
α1 , α 2 pre-specified parameters
βr , βs momentum parameters
~
δ MNN output
∆i uncertainty
εj constant weight
φ(⋅), φˆ (⋅) defined as Eq. (3.11), Eq. (4.22) and Eq. (5.13)
ϕ activated energy
γ parameter
γ~ tuning parameter
κ characteristic index
λs tuning parameter
Λ gain matrix
θ time-delay
~
θ estimated time-delay
~ ~
θ 2i , θ 3 bias parameters
ρ characteristic index
ρs density
ζ damping coefficient
CHAPTER 1
INTRODUCTION
Conventional control theory is well suited for applications where the process can be
is a tedious task that always requires experts with knowledge both in control theory and
process information. Therefore, in recent years the control of systems with complex,
unknown and uncertain dynamics has become a topic of considerable importance in the
literature and the design of robust control systems has received considerable attention
from control community. Among the developed model-based control strategies for robust
process control, sliding mode control plays an important role because it not only
stabilizes certain and uncertain systems but also provides the capability of disturbance
rejection and insensitivity to parameter variations (Utkin, 1992). Basically, the sliding
mode control (SMC) design is composed of two stages. The first stage is to define a
sliding surface on which the process’s dynamics is restricted. Subsequently, the second
stage is to design a feedback control law such that any process’s trajectory outside the
sliding surface is driven to reach the surface in a finite time and keep on it, which
therefore makes the closed-loop SMC system to be robust to matched uncertainties and
external disturbances. Due to the significant fact that the sliding mode control theory has
the ability of dealing with process uncertainties, much effort has been concentrated on the
design of various sliding mode control systems for handling with diversified process
dynamics.
geometric approach and sliding mode strategy has been proven to be a promising way to
the robust control and many advanced SMC schemes have been developed (Sira-Ramirez,
1992; Slotine and Hendrick, 1993; Colantonio et al., 1995; Li et al., 1995; Chen and Dai,
controller for the regulation of chemical processes, which is designed on the basis of
transforming the original nonlinear system into a Fliess’s local generalized observability
canonical form. Slotine and Hendrick (1993) combined the sliding mode theory with
(1995) proposed the use of nonlinear estimation along with a second-order sliding mode
for an open-loop unstable nonlinear CSTR. With incorporating the feedback linearization
technique and a deterministic approach, Li et al. (1995) presented a sliding mode type
robust controller with parameters being related to the bounds of uncertainties. By means
of the sliding mode control strategy and an adaptive state feedback technique, Chen and
Dai (2001) proposed a hybrid nonlinear controller for the regulation of a nonlinear CSTR
Apart from the sliding mode control design for uncertain processes, some researchers
have conveyed the concepts of SMC to the design of various robust control schemes for
uncertain processes in the presence of time-delay (Shyu and Yan, 1993; Hu et al., 1998;
Roh and Oh, 1999; Camacho and Smith, 2000; Hu et al., 2000; Roh and Oh, 2000; Chou
and Cheng, 2001; Li and Yurkovich, 2001; Camacho et al., 2003). Shyu and Yan (1993)
proposed a specific sliding mode control strategy for the robust stabilizing control of
uncertain linear systems with state-delay. Using the linear matrix inequality technique,
Hu et al. (2000) adopted a sliding mode control method to handle a class of uncertain
time-delay systems. Chou and Cheng (2001) proposed an adaptive variable structure
method does not require the upper bound of perturbations and the performance of the
their work the issue of input-delay was not considered as a whole. To deal with
input-delays, Hu et al. (1998) investigated a transformation that maps the original system
into a system without delay, and based on which a realizable sliding mode control law is
synthesized to assure the asymptotic stability of the control system. Camacho and Smith
(2000) proposed a method for synthesizing a sliding mode controller based on the first
order plus dead time (FOPDT) model. Their approaches resulted in a fixed structure
controller with a set of tuning equations being formulated as a function of the model’s
compensation, Roh and Oh (1999; 2000) presented a sliding mode controller for
convert the delay system into a delay-free one whose spectra embeds all unstable poles of
the original system within the given stability margin, and based on this delay-free system
a sliding mode controller was constructed for uncertain time-delay processes. More
recently, Camacho et al. (2003) developed an internal model sliding mode controller for
It is well known that processes with inverse response, often called the non-minimum
general, an inverse response means that the initial response is in a direction opposite to its
final response during a dynamic testing. It can cause by competing effects of two
dynamics (McMillian, 1983) or from processes with parallel structure (Marlin, 2000).
volatility (Waller and Nygardas, 1975), the bottoms composition response to boil-up
changes in certain distillation columns (Luyben, 1969), and the Van de Vusse reactor
under some special operating conditions (Van de Vusse, 1964), etc. The processes in the
presence of non-minimum phase behavior can cause classical linear feedback controllers
to run into trouble because the correction made initially would be in the wrong direction.
In fact, it also restricts application of some powerful nonlinear control strategies such as
feedback linearization control (Isidori, 1995) and sliding mode control (Utkin et al., 1999)
to achieve high performance and even leads to instability. To deal with this complicated
control problem with sliding mode strategy, Spurgeon and Lu (1997) introduced a higher
order sliding mode and based on which the non-minimum phase characteristics can then
phase process by a FOPDT model, Camacho et al. (1999) proposed a method for
synthesizing a sliding mode controller to the chemical processes with inverse response.
By means of the center manifold theory and on the equivalent control concept, Bonivento
et al. (2001) presented the design of an output feedback sliding mode regulator which is
able to achieve the asymptotic tracking of a reference trajectory for nonlinear systems
having inverse response. Shkolnikov and Shtessel (2002) considered the sliding mode
control strategy for asymptotic output tracking of a class of causal non-minimum phase
plant model is presented in the normal form with internal dynamics being expanded in
power series, which is similar to that used in Gopalswamy and Hedrick (1993). Herrmann
their method, the performance in the sliding mode is determined by a linear stable
This thesis considers the problem of the robust stabilization of chemical processes
inverse response. Alternative but practical sliding mode control schemes are to be
explored and developed in order to achieve the asymptotic stability and the robustness of
the closed-loop system when facing with diversified process dynamics. With the aid of
the state predictor as well as an optimal sliding surface, a simple SMC scheme is
developed for the regulation control of chemical processes which are characterized by a
second order plus dead time (SOPDT) model. Furthermore, we propose the
predictor-based SMC scheme for handling with nonlinear uncertain input-delay processes
and further extend to the non-minimum phase ones by means of using a statically
equivalent output map (SEOM) technique. For releasing the constraints and assumptions
on process uncertainties, the combination of the model-based sliding mode controller (Da
and Song, 1998) with an intelligent neural fuzzy technique (Chen and Peng, 1999) is
proposed to deal with the complex control problem of imprecisely known processes.
identified SOPDT model, an optimal sliding surface and a delay-ahead predictor, Chapter
2 presents a novel and systematic sliding mode control system design methodology. The
satisfying a sliding condition and the control system performance is examined with some
typical chemical processes. In Chapter 3, a sliding mode control strategy with the
sliding mode control scheme is developed for nonlinear, non-minimum phase, uncertain
chemical processes in the presence of input-delay. The proposed scheme, which integrates
input-delay and to circumvent the negative effect of inverse response. Furthermore, the
potential use of a sliding observer along with the proposed scheme is also investigated
therein. Moreover, Chapter 5 will investigate the regulation control of nonlinear chemical
processes whose dynamics are imprecisely known. A nonlinear control scheme that
incorporates a sliding mode controller and a neural fuzzy strategy is proposed to deal with
observer for estimating unmeasurable states will be taken into account in this chapter.
Finally, in Chapter 6, the conclusion remarks and future prospects would be drawn.
CHAPTER 2
DESIGN OF A SLIDING MODE CONTROL SYSTEM
BASED ON AN IDENTIFIED SOPDT MODEL
This chapter develops a simple and novel sliding mode control system for the
predictor is constructed for state estimation and a correction term from the measured
process output is incorporated to enhance the prediction accuracy of the process states.
With the integration of the state predictor and a designed optimal sliding surface, a novel
yet simple sliding mode controller that is able to account for plant uncertainties can be
robust stability as well as the system behavior of the closed-loop system is analyzed
through the satisfaction of a sliding condition. In this chapter the presented scheme is
further extended to one that is able to deal with the process having inverse response.
Moreover, in the same design framework, the control application to the FOPDT system is
included as a special case. For multivariable process control, a decentralized sliding mode
control scheme is proposed and explored in this chapter. The effectiveness and
applicability of the proposed scheme is examined and tested extensively with some
typical chemical processes. Also, the performance comparisons with some existing
sliding mode control techniques are included as a rigorous base for evaluation.
The organization of this chapter follows naturally in four sections. In Section 2.1, the
sliding surface design is presented. For extension to non-minimum phase processes and
FOPDT systems, some simple yet effective strategies for use with the proposed scheme
are introduced. Section 2.2 introduces a decentralized sliding mode control scheme for
and comparisons to verify the control capability and effectiveness of the proposed scheme.
In this section, we will present a SOPDT model-based sliding mode control scheme
for the nonlinear regulation of chemical processes. Additionally, we are going to show
subsequently how the technique developed for an identified SOPDT model can be applied
where K , τ , ζ and θ are, respectively, the steady state gain, natural period of
oscillation, damping coefficient and the input-delay of the process. For handling with the
input-delay and hence facilitating the design of a sliding mode control system, we shall
first consider the development of a delay-ahead predictor on the basis of the SOPDT
model, Eq. (2.1). To begin with, we convert the SOPDT model into an equivalent state
~
x&1 (t ) = ~
x 2 (t ) (2.2a)
~
x& 2 (t ) = −a1 ~
x1 (t ) − a 2 ~
x 2 (t ) + b1u(t − θ ) (2.2b)
~
y (t ) = ~
x1 (t ) (2.2c)
where a1 = 1 τ 2 , a 2 = 2ζ τ and b1 = K τ 2 ; ~
x1 and ~
x 2 are the states, and ~
y and
u represent, respectively, the model output and the control input. By removing the
as follows:
x&1∗ (t ) = x 2∗ (t ) (2.3a)
y ∗ (t ) = x1∗ (t ) (2.3c)
To improve the accuracy of state prediction, especially in face with modeling errors and
xˆ1 (t + θ t ) = x1∗ (t ) + y (t ) − ~
x1 (t ) (2.4a)
and
xˆ 2 (t + θ t ) = x 2∗ (t ) (2.4b)
are suggested for practical implementation, where y (t ) is the actual process output and
(2.4), we arrive at the estimated plant state one time delay ahead, i.e.
xˆ (t + θ t ) = ~
x (t + θ ). This means that the presented prediction model, which is delay free,
can facilitate the design of a sliding mode controller based on the SOPDT model. Also, it
is noted that the present predictor is quite different from that of Roh and Oh (1999) since
in their work merely an open-loop estimator is used which does not include any
design methodology for a delay-ahead sliding mode controller. To begin with, let’s
x&1∗ (t ) = x 2∗ (t ) (2.5a)
for the robust design of a sliding mode controller. It is assumed that the uncertainty
d (x ∗ , u, t ) is bounded and its upper bound function, d max (⋅) can be estimated as
d (x ∗ , u, t ) ≤ d max (x ∗ , t ) (2.6)
δ = c1 x1∗ (t ) + c 2 x 2∗ (t ) (2.7)
where c1 and c2 are the sliding coefficients to be designed later. On the basis of this
sliding function, the following theorem presents a sliding mode controller for the
1 d 2
using the uncertain system of (2.5) to satisfy the sliding condition of δ ≤ −α δ ,
2 dt
where α is the pre-specified positive constant that determines the system performance
and d (x ∗ , t ) = c 2 d max (x ∗ , t ) .
It is worth mentioning that the present sliding mode control law, which is designed to
satisfy the sliding condition, is able to deal with process uncertainties and can hence
provide robust control. In essence, the present controller design methodology is a kind of
deterministic approach of which only the uncertainty bounds are used and the resulting
Basically, the fundamental idea behind the use of the zero level set of the auxiliary output,
outside the manifold, the strategy forces the states toward the designed sliding surface.
Upon reaching Σ fast switching takes place in the immediate vicinity of Σ , which tries
⎧δ β, if δ β <1
sat(δ β ) = ⎨ (2.9)
⎩sign(δ β ), if δ β ≥1
where β > 0 represents the boundary layer thickness. Essentially, the proposed sliding
system performance. The larger value of α , the faster system response while at the
expense of a larger control input. The parameter β is introduced to eliminate the input
chattering. Usually, the selection of β represents the tradeoff between control precision
and the extent of input chattering. The coefficients in the sliding function, c1 and c 2 ,
can be viewed as the weighting factors for the states. Their values can affect the state
trajectories and in turn the system performance. In order to achieve optimal performance,
we shall explore in the following subsection how to design an optimal sliding function for
practical applications.
where x ∗ (t ) ≡ [ x1∗ (t ) x 2∗ (t )]T , t s is the beginning time of the sliding motion, and
⎡q q12 ⎤
Q = ⎢ 11 is a positive definite, symmetric matrix, i.e. q12 = q 21 and
⎣ q 21 q 22 ⎥⎦
where q11∗ = q11 − q122 q22 . Then, with the definition of v , in Eq. (2.11), we have
where a1∗ = q12 q 22 . The optimal control law for the above dynamic equation with the
where p is the positive root of the quadratic polynomial p 2 + 2a1∗ q22 p − q22 q11
∗
= 0, i.e.
p = − q12 + q11 q 22 . With the substitution of Eq. (2.11) into the above optimal solution, it
c1 = p + q12 and c 2 = q22 . Notice that, though the optimal sliding function designed on
the basis of minimizing the quadratic performance index in Eq. (2.10) is only one of
many options (Edwards and Spurgeon, 1998), the present approach is quite simple and is
rather suitable for the design of an effective sliding mode controller for SOPDT
processes.
By means of the output corrections, Eq. (2.4), the proposed sliding mode control law
δˆ = c1 xˆ1 (t + θ | t ) + c2 xˆ 2 (t + θ | t ) (2.16)
and
x&1∗ (t ) = x 2∗ (t ) (2.18a)
It should be noted that the hard input constraint of u (t ) ≤ u has been included in the
derivation of d max in Eq. (2.19). Figure 2.1 depicts the schematic diagram of the
proposed sliding mode control system, where a nominal process model, a prediction
model, an optimal sliding surface and the proposed sliding controller are integrated for
Figure 2.1. A schematic diagram of the proposed sliding mode control system.
In this subsection, we will present a simple strategy such that the proposed sliding
mode control system can be applied directly to deal with the difficult control problem of
non-minimum phase processes. To proceed, one can firstly identify and model this kind
can apply the identification method of Park et al. (1998) to give a model of the form
~ − b s + b1 −θ s
G( s) = 2 2 e (2.20)
s + a2 s + a1
Next, by using the equivalent time-delay concept of Sung and Lee (1996),
one can transform the above non-minimum phase model into the standard SOPDT model
as
b2
b1 − (θ + ) s
G ( s) = 2 e b1 (2.22)
s + a 2 s + a1
This therefore makes the proposed sliding mode control scheme applicable to
equivalence can introduce extra model error, these additional model uncertainties can be
easily handle by the sliding mode control law with an appropriate increment of the upper
This subsection discusses how the proposed scheme can be applied to processes that
x&1∗ (t ) = x 2∗ (t ) (2.24a)
y ∗ (t ) = x 2∗ (t ) (2.24c)
x&1∗ (t ) = x 2∗ (t ) (2.25a)
Using the design technique as stated in subsection 2.1.4, the following sliding mode
control law
with
can be synthesized for FOPDT processes to ensure the sliding condition of δ δ& ≤ −α δ .
x&1∗ (t ) = x 2∗ (t ) (2.29a)
~ ~ ~
x& 2∗ (t ) = −a(1 ± ∆1 ) x2∗ (t ) + b(1 ± ∆ 2 )u (t ) , 0 ≤ ∆ i ≤ 1 , i = 1, 2 (2.29b)
is considered, we have
~ ~
w(x ∗ , u, t ) = m ∆ 1 ax 2∗ (t ) ± ∆ 2 bu (t ) (2.30)
and
~ ~
wmax (xˆ (t + θ | t ), t ) = ∆1 a xˆ 2 (t + θ | t ) + ∆ 2 b u (2.31)
Owing to the complication and interaction between loops, the control of MIMO systems
presents a great challenge. In this section, we shall extend the developed single-loop
sliding mode control system to multivariable case through the use of a decentralized
K ij −θ ij s K ij −θ ij s
or FOPDT model, i.e., Gij ( s ) = e or Gij ( s ) = e . It
τ s + 2ζ ijτ ij s + 1
2
ij
2
τ ij s + 1
should be noted that existing identification techniques such as in Choi et al. (2000) or
Wang et al. (2000) can be applied to model an MIMO process using SOPDT or FOPDT
process, the first thing should be decided is to determine which controlled variable should
answered by relative gain array (RGA) method (Bristol, 1966), which has the advantage
of being easy to calculate and requires only the steady state gains information. In practice,
some alternative guidelines for loop pairing can be found else where, for example Yu and
Luyben (1986), Luyben et al. (1990). For the sake of presentation, we let the loop pairing
sliding mode control system design methodology presented in the previous sections, in
loop i ( i = 1,2,K, n ), the sliding mode controller can be designed mainly based on the
other loops are considered as disturbances that affect this loop. In this sense, the transfer
where d i (x ∗ , u, t ) is the lumped uncertainties arising from the second term of Eq. (2.33)
in the right hand side. Assume that d i (x ∗ , u, t ) is bounded and can be estimated as
d i (x ∗ , u, t ) ≤ d i , max (x ∗ , t ) (2.35)
Then, according to the design technique presented in the previous section, we have the
u~i (t) = b1−i1[a1i xˆ1i (t + θii | t) + (a2i − c2−i1c1i )xˆ2i (t + θii | t)]− (b1i c2i ) −1 (αi + di (xˆ (t + θii | t),t))sat(δˆi βi )
(2.36)
hand, if Gii (s ) is a FOPDT model, the i -th sliding mode control law can be simply
given by
where wi ,max (⋅) is the upper bound of wi (⋅) and wi (⋅) is the lumped uncertainty in the
where ai = 1 τ ii and bi = K ii τ ii .
controllers acts simply as the case of single-loop control since loops are independent. In
such a case, the stability of the closed-loop system is determined by the stability
affect the stability of the entire multivariable control system to some certain extent.
Accordingly, the need has arisen for methods that will provide a decoupling of
disturbances caused by interacting loops and thus enhances the stability. Several
decoupling methods (McAvoy, 1979; Weber and Gaitonde, 1985; Waller and Finnerman,
1987; Perng and Ju, 1994; Wang et al., 2000; Gilbert et al., 2003) are available in the
literature. Among them, a practical solution to multivariable process control without the
need of a perfect dynamic model is the use of the static decoupling technique (McAvoy,
1979). With the static decoupling technique, the decoupling gain Dij ( i ≠ j ) is simply
given as follows:
Gij ( s ) K ij
Dij = − lim =− (2.39)
s →0 Gii ( s ) K ii
Basically, the use of static decouplers gives some substantial advantages: (1) only the
steady state gains rather than complete dynamic models are required; (2) the static
decouplers are merely constants so that they are always physical realizable and easily
implemented. This provides a decentralized sliding mode control system for interacting
example.
with long dead-time, a high order process, a non-minimum phase process and a
multivariable chemical process. The performance comparisons with some existing sliding
mode control strategies are included as a rigorous base for evaluation. In the following
simulation studies, we consider the input constraints of u(t ) ≤ 1 for the illustrated
Examples 2.1-2.3 and 2.5, while for the non-minimum phase nonlinear process of
Example 2.4 the control input is assumed to be constrained by u(t ) ≤ 0.3 . Besides, for
output regulation the parameters of the proposed sliding mode controller are set to be
α = 0.1 and β = 0.4 for the illustrative Examples 2.1-2.4. Those design parameters for
Example 2.5 are listed in Table 2.1. Unless stated otherwise, parameter variations of up to
± 25% are considered for the design of the sliding mode controller.
As many of chemical processes can be modeled by a FOPDT model, the first example
is to examine the output regulating ability of the proposed sliding mode control scheme
In this figure, the sliding mode strategy of Camacho et al. (2003), which is also
developed based on the FOPDT model, as well as the sliding mode scheme of Roh and
Oh (1999) are included therein as a rigorous base for comparison. From Figure 2.2, it can
be seen that the proposed scheme gives excellent control performance. Also observed is
that, though the sliding mode control of Camacho et al. (2003) is able to produce smooth
control input for this large dead-time process, their output regulation process after delay
time ( t = 8 ) is relatively slow as compared with the proposed scheme. This is because
that the method of Camacho et al. (2003) simply applies first-order Taylor series
approximation for the delay term and the tuning of the sliding controller’s parameters is
relatively conservative in order to handle with process uncertainties. On the other hand,
Roh and Oh (1999) has a similar output response as that of Camacho et al. (2003), but
To test the ability of handling with plant/model mismatch, we assume that the plant
1.5 − 4 s
dynamics vary to G p ( s ) = e while the process model remains unchanged. Also,
2s + 1
it is assumed that the plant variations are unknown to the controller. With this
plant/model mismatch, Figure 2.3 shows that the scheme of Camacho et al. (2003) gives
a slower response, while Roh and Oh (1999) is still oscillatory in the control input. In
contrast, the proposed scheme is very robust and the control performance is much better,
even though the actual uncertainties ( ± 50% ) are larger than that ± 25% considered in
⎡0.03 0⎤
Let Q = ⎢ , we arrive at the optimal sliding coefficients as c1 = 0.2449 and
⎣ 0 2⎥⎦
c 2 = 2 . Having these design information, one can easily implement a sliding mode
control system for this process since the SOPDT model can be directly applied to the
controller design. Figure 2.4 depicts the output regulation results and the produced
control inputs. Therein, the performance of the proposed scheme with arbitrary sliding
coefficients is also included for comparison. From this figure, it is shown again that the
proposed scheme can provide a smoother and faster control performance as compared
with both the schemes of Camacho et al. (2003) and Roh and Oh (1999). It is also shown
that the optimal sliding surface for the sliding controller apparently results in a better
performance than the arbitrary one. Also observed is that Roh and Oh (1999) still has
oscillatory control input due to its inherent controller form. To verify the ability of
handling with process uncertainties, we assume that the identified process model remains
1
unchanged, while the dynamics of the actual plant vary to G p ( s ) = e −6 s .
11s + 2 s + 1
2
Figure 2.5 illustrates the system performance in response to this large plant/model
mismatch. The simulation results show clearly that the proposed scheme is very robust in
the presence of plant uncertainties, while both the un-retuned sliding mode controllers of
Camacho et al. (2003) and Roh and Oh (1999) result in relatively large oscillations and
we first convey a system identification technique to get a SOPDT model. By using the
technique of Park et al. (1998), the SOPDT model parameters are identified as
⎡0.3 0 ⎤
a1 = 0.2291 , a 2 = 0.8465 , b1 = 0.2291 and θ = 3.3 . Let Q = ⎢ ⎥ for this
⎣ 0 1.2⎦
process, a set of optimal sliding coefficients of c1 = 0.6 and c 2 = 1.2 is obtained.
From the simulation results shown in Figure 2.6, it is observed that the closed-loop
control performance by the proposed approach is smoother than both the methods of
Camacho et al. (2003) and Roh and Oh (1999). To further evaluate the ability of handling
but the identified model remains unchanged. The control performance in response to this
state estimator, which is able to provide more accurate states for feedback control. Based
on these simulation results, it is again corroborated that the proposed scheme is more
This example considers the control of a Van de Vusse reactor (Van de Vusse, 1964) in
A ⎯⎯→
k1
B ⎯⎯→
k2
C
2A ⎯⎯→
k3
D
steady state, respectively. The state variables x1 (t ) and x 2 (t ) are deviation variables
constants are k1 = 0.8333 min −1 , k 2 = 1.6667 min −1 , and k 3 = 0.1667 L ⋅ mol −1 ⋅ min −1 ,
respectively. It is assumed that the process is at steady state with Fs = 4.0 L min ,
C As = 3.0 mol L , and C Bs = 1.117 mol L initially. The control objective is to regulate
An open loop test shows that this process exhibits the undesirable inverse response.
Therefore, to apply the proposed scheme to this non-minimum phase process, we first
identify the process model as in the form of Eq. (2.20). By applying the identification
technique of Park et al. (1998) around the normal operating condition, we have the model
Now, by considering a larger parameter variation, ± 30% in this case (i.e. ∆ i = 0.3 ,
⎡40 0 ⎤
i = 1,2,3 ), and let Q = ⎢ ⎥ , a sliding mode control system for this non-minimum
⎣ 0 0.01⎦
phase system based on the obtained equivalent model, Eq. (2.45) is constructed. To verify
the regulation ability of the proposed strategy, we assume that the system states are
perturbed to move away from their steady values initially. In this case study, we let
x1 (0) = −2.0 and x2 (0) = −0.3 . From Figure 2.8, it can be seen that the closed-loop
regulation control performance by the proposed scheme is faster than the method of
Camacho et al. (2003). When the obtained equivalent SOPDT model is used, the method
of Roh and Oh (1999) can be applied. As can be examined, the control system
performance of Roh and Oh (1999) is similar to the proposed method, while their control
input is still chattering. For the case that the process parameters C Af , k1 , k 2 , k 3 and
the time-delay vary by +50 % from their nominal values, the method of Camacho et al.
(2003) becomes oscillatory and Roh and Oh (1999) is still chattering, see Figure 2.9. On
the other hand, as -50 % parameter variations occur, the simulation results shown in
Figure 2.10 reveal that the method of Camacho et al. (2003) with un-retuned parameters
fails to control this highly nonlinear process. A reason for this is that in their scheme only
the process output error is used for output feedback control and there is no information
regarding to state variations for use in the controller. In contrast, the proposed sliding
mode control strategy is still very robust despite of the influence of these significant plant
uncertainties.
relative gain array (Bristol, 1966) from the gain matrix as follows:
⎡ 2.01 − 1.01⎤
Λ=⎢ ⎥ (2.48)
⎣− 1.01 2.01 ⎦
This relative gain array indicates that the 2 × 2 system has strong loop interactions and
further suggests that the best control pairing should be: y1 is controlled by u1 ; and y 2
is controlled by u 2 (Luyben and Luyben, 1997). Having determined the variable pairing
issue, the next question to be answered is the setting of the static decouplers. According
to Eq. (2.39), the static decouplers gains are set to be D21 = −0.34 and D12 = −1.48 .
estimated upper bound values of w1,max = 1 and w2,max = 1 instead of using their
accurate formulas. Figure 2.11 depicts the schematic diagram of the 2× 2 decentralized
sliding mode control system. The simulation results shown in Figure 2.12 demonstrate
the benefits of incorporating the static decouplers, where significantly smoother control
inputs and better control performance were obtained. The performance comparisons with
decentralized PI controllers tuned with Sequential (Fang et al., 1994) and BLT methods
(Luyben, 1986) as well as the sliding mode controllers of Camacho and Rojas (2000) are
depicted in Figure 2.13. These comparisons were performed using the parameters listed
in Table 2.1.
Table 2.1. Controller settings of the proposed approach and comparative methods for
Example 2.5.
Sequential
The proposed Camacho and BLT (Luyben,
(Fang et al.,
approach Rojas (2000) 1986)
1994)
loop1 loop2 loop 1 loop 2 loop1 loop2 loop1 loop2
c1 0 0 λ1 1.0599 0.4028 kc 0.61 0.09 0.375 0.075
c2 1.0 1.0 λ0 0.1 0.02 ki 0.065 0.016 0.0452 0.0032
α 0.5 0.5 µ11 2.01
PI controller: Gc ( s ) = k c + k i s
β 5 5 µ 22 2.01
From Figure 2.13, it is clear to observe that the proposed decentralized sliding mode
controllers are able to produce more suitable control inputs such that the regulation time
is relatively short. Figure 2.14 compares the control system performance in response to
the following plant uncertainties: time delays +50%, time constants -25% and the steady
state gains +25%. The existence of steady state gain uncertainties implies that the static
decouplers designed are incapable of fully eliminate the actual loop interactions.
Fortunately, due to the designed decentralized sliding mode controllers having the ability
tuned by Sequential method becomes unstable when facing with such a large plant
variations.
accommodate the extra disturbance, the maximum uncertainties bounds wi ,max of the
order to provide a more aggressive but precise control. With slightly increasing wi ,max ,
i = 1 and 2 , to 1.2 and using the controller parameters as those mentioned above except
for β = 0.02 , we plot the simulation results in Figure 2.15. From this figure, it is evident
that the present decentralized sliding mode control scheme appears to be an effective and
2.4. Summary
This chapter has presented a simple, systematic and novel model-based control
system for the nonlinear regulation of chemical processes. Based on a SOPDT model, a
delay-ahead predictor, and a designed optimal sliding surface, an effective and applicable
sliding mode control scheme has been developed. With satisfying a sliding condition, the
Simulation results reveal that the proposed sliding mode control strategy is effective and
proposed scheme has been shown to be superior in performance than several existing
sliding mode strategies. It is worth mentioning that, from practical point of view, the
presented sliding mode control scheme for use in the field of chemical process control
presents advantages, such as (1) it can make use of a previously available process model
(FOPDT or SOPDT models); (2) the process model used for controller design does not
need to be very precise since it has a great capability in dealing with process uncertainties;
(3) the input hard constraint can be treated directly in the design stage when using a
approach where the only assumption on uncertainties is that they are bounded; (5) the
proposed state estimation which includes process output information has great benefit
especially in the presence of plant/model mismatch; (6) it is simple, robust, and widely
Due to its significant advantages and good performance, the proposed sliding mode
control scheme provides a simple yet robust approach to the nonlinear regulation control
of chemical processes.
Figure 2.3. Closed-loop system performance of Example 2.1 in the presence of parameter
uncertainties: steady state gain: +50%, each of time constant and dead-time: -50%. ─
the proposed approach; ---- Camacho et al. (2003); ······· Roh and Oh (1999).
---
approach with an optimal sliding coefficients ( c1 = 0.2449, c 2 = 2 ); · · the proposed
approach with arbitrary sliding coefficients ( c1 = 0.5, c 2 = 1.5 ); ---- Camacho et al.
plant/model mismatch. ─ the proposed approach; ---- Camacho et al. (2003); ·······
plant/model mismatch. ─ the proposed approach; ---- Camacho et al. (2003); ·······
Figure 2.8. Closed-loop system performance of the Van de Vusse reactor. ─ the
proposed approach; ---- Camacho et al. (2003); ······· Roh and Oh (1999).
Figure 2.9. Closed-loop system performance of the Van de Vusse reactor in the presence
Figure 2.10. Closed-loop system performance of the Van de Vusse reactor in the presence
Figure 2.11. A multi-loop SMC decentralized control system for 2 × 2 process systems.
Figure 2.12. Control performance of the proposed approach with and without static
---
Rojas (2000); ······· Sequential (Fang et al.,1994); · · BLT (Luyben, 1986).
and θ : +50% ). ─ the proposed approach; ---- Camacho and Rojas (2000); ·······
---
Sequential (Fang et al.,1994); · · BLT (Luyben, 1986).
Figure 2.15. Control performance of the proposed approach in the face of extra
disturbances.
CHAPTER 3
DESIGN OF A SLIDING MODE CONTROL SYSTEM FOR
NONLINEAR, UNCERTAIN, INPUT-DELAY
CHEMICAL PROCESSES
and then designed a robust sliding mode controller to handle with the process
sliding mode control strategy of Chen and Dai (2001) is extended to one that has the
ability to deal with nonlinear uncertain time-delay processes. The present methodology
geometric feedback linearization and the sliding mode control strategy. No statistical
properties of the uncertainties are assumed or used, making the present scheme very easy
control scheme for nonlinear uncertain input-delay processes is presented. The design
ideas and concepts of the nonlinear predictor as well as the sliding mode controller are
subsequently Section 3.2 performs a case study on the regulation control of a nonlinear
uncertain input-delay CSTR in the presence of unmodeled side reaction, measuring error,
parameter variations, and extra matched and unmatched disturbances. Finally, in Section
Input-Delay Processes
y (t ) = h( x ) (3.1b)
where x(t ) ∈ R n , u(t ) ∈ R , y (t ) ∈ R and θ ∈ ([0, ∞ ), R ) are state vector, control input,
system output and the time-delay respectively; f (⋅) , g(⋅) , ∆f (⋅) and ∆g(⋅) are smooth
vector fields on an open set U ∈ R n and h(⋅) a smooth function on U . For system
description, we let the Lie derivative of a smooth function h(x) along a vector field
∂h(x) n
∂h(x)
Lg h(x) = g ( x) = ∑ g i ( x) (3.2)
∂x i =1 ∂x i
In terms of Lie derivative, the relative degree of the system (3.1) is defined as
{
Similarly, let κ = min m : L∆f Lmf −1 h(x) ≠ 0 } and w = min{m : L∆g Lmf −1 h(x) ≠ 0} be the
assume that the uncertainties satisfy the so-called generalized matching condition
nominal system and h(x ) vanishes at that equilibrium point. In other words, f (0) = 0
and h(0) = 0 . This means that y represents the tracking error. Notably, a given model
could be easily rewritten in this form by defining appropriate deviation variables. Besides,
we also assume that the system possesses the property of internal stability. In other words,
the nominal system of system (3.1) is assumed to be hyperbolically minimum phase, i.e.,
the corresponding zero dynamics are exponentially stable (Behtash, 1990; Li et al., 1995).
Figure 3.1. Schematic diagram of the proposed sliding mode control system for nonlinear
predictor and a sliding mode controller. In the control scheme, the use of the nonlinear
time-advanced states, the sliding mode controller is designed for attenuating the negative
To introduce the whole sliding mode control system, we shall first present a sliding
mode controller design methodology and then presents a nonlinear predictor design
method in order for compensating the input-delay and to achieve robustness of the whole
control system.
Based on the input-output linearization technique of Isidori (1989), there exists a local
[ξ T
, ηT ]
T
= T( x )
(3.4)
= [h(x), L h(x),L , L ρ −1
h(x),η1 (x),L ,η n − ρ ( x) ]
T
f f
ξ&i = ξ i +1 , i = 1, 2, L , ρ − 1 (3.5a)
y = ξ1 (3.5d)
where a(ξ, η) , ∆a(ξ, η) , b(ξ, η) , ∆b(ξ, η) , q(ξ, η) and φi (ξ, η) are given,
respectively, by
and
x = T −1 (ξ, η) (3.12)
Since the process is of internal stability, the following state feedback control law:
v(t ) − b(ξ, η)
u (t − θ ) = (3.13)
a(ξ, η)
can be applied, where all the quantities in the right-hand side are at time t . To give the
be carefully designed to meet some certain desirable robustness and system performance.
Herein, we modify the robust SMC approach of Chen and Dai (2001) to give v(t + θ ) as
where the adaptive gain k is tuned by k& = γ~(δ (t + θ )) 2 ( γ~ > 0 ); f max , bmin , δ (t + θ )
∆a (ξ (t + θ ), η(t + θ ))
bmin = 1 − sup (3.17)
( ξ , η )∈T (U ) a (ξ (t + θ ), η(t + θ ))
ρ
δ (t + θ ) = c T ξ (t + θ ) = ∑ ci ξ i (t + θ ) , cρ = 1 (3.18)
i =1
and
⎧ δ (t + θ ) β , if δ (t + θ ) β < 1
sat(δ (t + θ ) β ) = ⎨ (3.19)
⎩sign(δ (t + θ ) β ), if δ (t + θ ) β ≥ 1
In the control law, β is the user-specified boundary layer thickness used to eliminate the
input chattering, and coefficients ci in δ (t + θ ) are chosen such that the polynomial
The use of the hybrid control law for v(t + θ ) possesses the following advantages (Chen
and Dai, 2001): (1) it gains benefits from both the adaptive state feedback and sliding
mode strategy; (2) the construction of the hybrid nonlinear controller is very simple and
straightforward; (3) it does not depend on explicit Lyapunov functions and does not make
At this stage, the SMC control law (3.14) for input-delay processes has been
constructed. However, as can be seen from the control law, this controller can not be
directly implemented without having the predictive states. Therefore, in the next
states.
To compensate the time-delay of the process and therefore estimate the process’s
~ ~
x& (t ) = f ( ~
x (t )) + g( ~
x (t ))u(t − θ ) (3.20b)
~
xˆ (t + θ | t ) = x(t ) + x ∗ (t ) − ~
x (t ) (3.20c)
where x ∗ (t ) , ~
x (t ) and x(t ) ∈ R n denote, respectively, the model state vector, the
~
nominal state vector, and the actual plant’s state vector; θ ≥ 0 is the estimated
~
time-delay in the manipulated input and xˆ (t + θ | t ) ∈ R n represents the corrected
time-advanced predictive state vector. By comparing Eqs. (3.20a) with (3.20b), it follows
~ ~
that x ∗ (t ) = ~
x (t + θ ) if the predictor is initialized as x ∗ (0) = ~
x (θ ) . This initialization
~
can be achieved at steady state because in this case ~
x (θ ) = ~
x (0) (Henson and Seborg,
1994). As a result, in the absence of plant/model mismatch the prediction model yields
~ ~
the plant state vector one time delay ahead, i.e. xˆ (t + θ | t ) = x(t + θ ) if θ = θ . Also, it
is worth mentioning that the nonlinear predictor of Eq. (3.20) is quite different from the
open-loop predictor of Kravaris and Wright (1989) since the actual process output is
feeding back for the correction of the estimated states. With the introduction of the
&
ξˆi = ξˆi +1 , i = 1, 2, L , ρ − 1 (3.21a)
& ~
ξˆρ = [bˆ(ξˆ , ηˆ ) + ∆bˆ(ξˆ , ηˆ )] + [aˆ (ξˆ , ηˆ ) + ∆aˆ (ξˆ , ηˆ )] u (t − θ ) (3.21b)
yˆ = ξˆ1 (3.21d)
~ ~ ~
vˆ(t + θ | t ) − bˆ(ξˆ (t + θ | t ), ηˆ (t + θ | t ))
u (t ) = ~ ~ (3.22)
aˆ (ξˆ (t + θ | t ), ηˆ (t + θ | t ))
~
of which the future input vˆ(t + θ | t ) is calculated as
~ ~ ~ ~
vˆ(t + θ | t ) = −kˆδˆ(t + θ | t ) − sat (δˆ(t + θ | t ) β )[bˆmin
−1
( fˆmax + δˆ(t + θ | t ) )] (3.23)
& ~ ~
where the adaptive gain k̂ is tuned by kˆ = γ~(δˆ(t + θ | t )) 2 ( γ~ > 0 ); δˆ(t + θ | t ) , fˆmax ,
With the incorporation of the nonlinear predictor and the insertion of the control law
&
ξˆ = A c ξˆ + B c [vˆ + bˆs−1 ( fˆs + c T ξˆ )] (3.24a)
~ ~
x& = f ( ~
x ) + g( ~
x )u ( t − θ ) (3.24d)
where
⎡ 0 1 0 0 K 0 0⎤
⎢ 0 0 1 0 K 0 0⎥
⎢ ⎥
⎢ M M M M O M M ⎥
Ac = ⎢ ⎥ (3.25)
⎢ 0 0 0 0 K 1 0⎥
⎢ 0 0 0 0 K 0 1⎥
⎢ ⎥
⎢⎣− c1 − c2 − c3 − c4 K − cρ −1 − 1⎥⎦
⎡0⎤
⎢M⎥
Bc = ⎢ ⎥ (3.26)
⎢0⎥
⎢ ⎥
⎣b̂s ⎦
bˆ(ξˆ , ηˆ )
fˆs := fˆs (ξˆ , ηˆ ) = ∆bˆ(ξˆ , ηˆ ) − ∆aˆ (ξˆ , ηˆ ) (3.27)
aˆ (ξˆ , ηˆ )
and
∆aˆ (ξˆ , ηˆ )
bˆs := bˆs (ξˆ , ηˆ ) = 1 + (3.28)
aˆ (ξˆ , ηˆ )
The robust stability and desired behavior of the closed-loop system are described in
Theorem 3.1. Suppose that the uncertain input-delay system (3.1) is subject to the control
law (3.22) and the stable nonlinear predictor (3.20). If bˆmin > 0 , then the closed-loop
(P3.1) Uniform stability: Let η = [ξˆ T , ηˆ T ]T and η0 = [ξˆ T0 , ηˆ T0 ]T . For each d ≥ d , given
any η(⋅) : [t0 , ∞) → R n , and η(t 0 ) = η0 of the closed-loop system (3.24), there
all t ≥ t0 .
(P3.2) Uniform boundedness: Given any r > 0 and any η(⋅) : [t0 , t1 ) → R n and
(P3.3) Uniform ultimate boundedness: For each d ≥ d and r > 0 , given any
t ≥ t0 + t ( d , r ) .
Inherently, the SMC control law in Eq. (3.23) consists of two parts. The first part is
the adaptive term for preserving the nominal performance and the second part containing
the uncertainty measures is to force the system’s trajectory back to the sliding surface
δˆ = 0 despite the influence of the uncertainties. Unlike the classical sliding mode design,
the present sliding mode control law, novelty due to the introduction of an adaptive term
is able to guarantee the stability properties of both the states ξ and η . Besides, the
adaptive version makes the sliding mode controller quite different from that of Li et al.
(1995), where the controller is a combination of a high-gain linear state feedback and a
gain is always unavoidable in order for guaranteeing both the robust stability and
performance. In contrast, based on a simple tuning rule, the feedback gain of the
the magnitude of the sliding function (an index for measuring how far is the current
system output to the desired one). This effort is apparently more practical, without the
need of using a pre-specified, high gain feedback. Therefore, it is more suitable for the
Essentially, the sliding mode control law in Eq. (3.23) consists chiefly of the four
follows. The coefficient vector, c , can be viewed as the weighting factor for the state
vector ξ̂ . The parameter γ~ is a positive constant related to the tuning rate of the
&
adaptive gain. In the light of the tuning rule kˆ = γ~δˆ 2 , a non-negative initial setting of k̂ ,
k̂ (0) , is sufficient to guarantee the negative feedback. Notice again that the parameter β
is introduced to eliminate the input chattering. In practice, one will need to choose β by
simulation experiments. The choice of β represents the trade-off between the extent of
Process
section we apply it to the regulation control of a continuous stirred tank reactor (CSTR)
in the presence of input-delay. The dynamic behavior of this CSTR is described by the
⎛ x 2 (t ) + x 2 d ⎞
x&1 (t ) = − x1 (t ) + Da (1 − x1 (t ) − x1d ) exp⎜⎜ ⎟⎟
⎝ 1 + ( x 2 (t ) + x 2 d ) ϕ ⎠ (3.29a)
⎛ x2d ⎞
− Da (1 − x1d ) exp⎜⎜ ⎟⎟
⎝ 1 + x 2d ϕ ⎠
⎛ x 2 (t ) + x 2 d ⎞
x& 2 (t ) = −(1 + δ a ) x 2 (t ) + Ba Da(1 − x1 (t ) − x1d ) exp⎜⎜ ⎟⎟
⎝ 1 + ( x 2 (t ) + x 2 d ) ϕ ⎠
(3.29b)
⎛ x2d ⎞
− Ba Da(1 − x1d ) exp⎜⎜ ⎟⎟ + δ a u (t − θ )
⎝ 1 + x2d ϕ ⎠
y (t ) = x 2 (t ) (3.29c)
where all variables are in their deviation form. Rewrite the above equation into the
⎡ f (x(t ))⎤
f (x(t )) = ⎢ 1 ⎥
⎣ f 2 (x(t))⎦
⎡ ⎛ x2 (t) + x2d ⎞ ⎛ x2 d ⎞ ⎤
⎢ − x1 (t ) + Da(1 − x1 (t ) − x1d ) exp⎜⎜ ⎟⎟ − Da(1 − x1d ) exp⎜⎜ ⎟⎟ ⎥
⎢ ⎝ 1 + ( x2 (t ) + x2d ) ϕ ⎠ ⎝ 1 + x2 d ϕ ⎠ ⎥
=
⎢ ⎛ x2 (t) + x2d ⎞ ⎛ x2d ⎞⎥
⎢− (1 + δ a ) x2 (t ) + Ba Da(1 − x1 (t ) − x1d ) exp⎜⎜ ⎟⎟ − Ba Da(1 − x1d ) exp⎜⎜ ⎟⎟⎥
⎢⎣ ⎝ 1 + ( x2 (t ) + x2d ) ϕ ⎠ ⎝ 1 + x2d ϕ ⎠⎥⎦
(3.30)
⎡0⎤
g ( x (t )) = ⎢ ⎥ (3.31)
⎣δ a ⎦
and
h( x(t )) = x 2 (t ) (3.32)
Assume that in this reaction system there exist an unmodeled first-order side reaction and
a measuring error. Under such a situation, the model uncertainties can be represented by
⎡ 0 ⎤
∆f ( x (t )) = ⎢ ⎥ (3.33)
⎣ − e f x1 (t )⎦
and
⎡0⎤
∆g( x (t )) = ⎢ ⎥ (3.34)
⎣eg ⎦
ud = 0 . Also, the steady state values of the system states are given as: x1d = 0.144 ,
state value by adjusting the control input u . Furthermore, it is easy to verify that the
proper subset of generalized matching condition. For SMC design, let’s choose a
coordinate transformation of
ˆ = [ξˆ ηˆ ]T = [ xˆ
T xˆ1 ]T (3.35)
1 2
& ~
ξˆ1 = (bˆ(ξˆ , ηˆ ) + ∆bˆ(ξˆ , ηˆ )) + (aˆ (ξˆ , ηˆ ) + ∆aˆ (ξˆ , ηˆ )) u (t − θ ) (3.36)
where
aˆ (ξˆ , ηˆ ) = δ a (3.37)
and
Now, it can be verified that the zero dynamics of the nominal system
⎛ ⎛ x2d ⎞⎞
ηˆ& = qˆ (0, ηˆ ) = −ηˆ ⎜⎜1 + Da exp⎜⎜ ⎟⎟ ⎟
⎟ (3.41)
⎝ ⎝ 1 + x2d ϕ ⎠ ⎠
is exponentially stable. This means that the internal stability is guaranteed and thus the
proposed SMC design methodology can be utilized. Based on Eqs. (3.22) and (3.23), we
have the following control law for this uncertain input-delay CSTR as
~ ~ ~ ~
− kˆδˆ(t + θ | t ) − sat (δˆ (t + θ | t ) β )[bˆmin
−1
( fˆmax + δˆ (t + θ | t ) )] − f 2 (xˆ (t + θ | t ))
u (t ) = (3.42)
δa
To implement the control law, the values of fˆmax and b̂min should be determined. Based
∆aˆ (ξˆ , ηˆ ) eg
bˆmin = 1 − sup = 1− = 0.6667 > 0 (3.43)
( ξˆ , ηˆ )∈T ˆ (ξˆ , ηˆ )
ˆ (U ) a δa
On the other hand, since aˆ (ξˆ , ηˆ ) , ∆aˆ (ξˆ , ηˆ ) , bˆ(ξˆ , ηˆ ) and ∆bˆ(ξˆ , ηˆ ) are state dependent,
one can only estimates possible maximum bound roughly. For this case, we set
fˆmax = 0.5 . The other controller parameters are set as c1 = 1 , kˆ(0) = 1 , γ~ = 1 , and
β = 0.05 . In order to verify the regulation ability of the proposed strategy, we suppose
that the system states are perturbed to move away from their steady states to be
x1 (0) = 0.4 and x 2 (0) = 0.8 initially. Having the previous preparations, we are ready to
The role of the incorporated nonlinear predictor is examined first. In this case, we
~
assume that the estimated input-delay coincides with the actual one, that is θ = θ = 0.5 .
Figure 3.2 depicts the simulation results of using the proposed SMC control law with and
without the incorporation of the nonlinear predictor. As can be seen from this figure, the
sliding mode control can be fail once the nonlinear predictor is not included in the control
system, mainly due to the negative effect of the input-delay on state feedback. Once the
nonlinear predictor has been incorporated as that shown in Figure 3.1, the produced
input-delay and process uncertainties. These simulation results clearly reveal that the
nonlinear predictor plays an important role in the proposed SMC control scheme when
handling with nonlinear uncertain input-delay processes. Also, as shown in the previous
section, the incorporation of the nonlinear predictor with the proposed SMC control law
has been proven to be able to possess certain desired robust performance and stability for
the feeding of the reactant and the reactor temperature. The extra unmeasured
disturbances introduce significantly additional modeling error to process, which leads the
uncertainty vector ∆f to be
⎡ d1 ⎤
∆f ( x (t )) = ⎢ ⎥ (3.44)
⎣ − e f x1 (t ) + d 2 ⎦
It can be verified that, though the uncertainties do not satisfies the matching condition,
the generalized matching condition is still valid and thus the proposed control scheme
remains applicable. For simulation, we let d1 = 0.05 and d 2 = 0.2 , and the unmeasured
disturbance enters into the system after time of 20 min. With increasing fˆmax to 0.7 and
using the controller parameters as those mentioned above, we plot the simulation results
in Figure 3.3. From this figure, it is clearly revealed that the proposed predictor-based
sliding mode control strategy is able to provide robust control without having abrupt
control actions and vigorous control moves in the control loop even though extra
To explore the plant uncertainties on the essential behavior of the control system, we
assume that the values of the Damkhler number, Da , the activated energy, ϕ , and heat
of reaction, Ba , have ± 25% variation from their nominal values while these
parameter values in the model remain unchanged. In designing SMC to deal with these
parameter variations, the value of fˆmax is set as 0.7 for accommodating these
demonstrating that the proposed scheme is robust despite of the presence of the parameter
variations. Furthermore, the simulation results shown in Figure 3.5 reveal that the
proposed observer-based SMC control scheme is robust to modeling error in input delays.
3.3. Summary
In this chapter, we have presented a predictor-based sliding mode control system for
The main idea of the present approach is to combine a prediction scheme for
time-advanced state estimation and a sliding mode control technique for robust control.
The robust stability and system performance of the whole system have been guaranteed
applicability of the proposed nonlinear control strategy has been examined through the
reveal that the proposed sliding mode control scheme appears to be a robust and powerful
Figure 3.2. System performance comparison of the proposed strategy with and without
the use of a nonlinear predictor. ― with the use of a nonlinear predictor; ····· without the
use of a nonlinear predictor.
Figure 3.5. Closed-loop system performance under modeling error in input delays. ―no
CHAPTER 4
A SLIDING MODE CONTROL SCHEME FOR NON-MINIMUM
PHASE, NONLINEAR, UNCERTAIN, INPUT-DELAY
CHEMICAL PROCESSSES
As has been reviewed in Chapter 1, although the established sliding mode control
(SMC) strategies have many successful application in handling with diversified process
dynamics, the sliding mode control of nonlinear processes that possess simultaneously the
dynamics behavior of uncertainties, input-delay and inverse response has not ever been
with this kind of processes. For example, a Van de Vusse reactor operated in a certain
condition could exhibit an inverse response behavior, arising from dynamics competition
(Van de Vusse, 1964; Kravaris et al., 1998; Chen and Lee, 2002). Besides, the
complexities of the consecutive and side reactions inside the reactor can present
there is long mass transport through pipe line. To tackle with these dynamics
simultaneously, this chapter attempts to develop a novel sliding mode control scheme
through making use of a statically equivalent output map (SEOM) and a time-advanced
nonlinear predictor. The main idea is to reap the benefit of both the SEOM and the
nonlinear predictor; the SEOM is used for eliminating the undesirable inverse response
and the time-advanced predictor for curbing the negative effect of input-delay. The
synergistic integration of SEOM and nonlinear predictor provides a new direction toward
the robust control of complex processes, which therefore facilitates the design of an
effective sliding mode control system. Furthermore, the convergence properties of the
proposed sliding mode control scheme are guaranteed by utilizing the Lyapunov stability
theorem.
mode control scheme which integrates an auxiliary output, the input-output feedback
for the robust control of nonlinear, uncertain, non-minimum phase, input-delay processes.
simulations for the SMC control of a Van de Vusse reactor is performed in Section 4.2.
Some important issues regarding the application of the proposed SMC scheme were
investigated. The potential use of a sliding observer for estimating the unmeasurable
states is also included therein. Finally, in Section 4.3 summary remarks are made.
y (t ) = h(x) (4.1b)
where x(t ) ∈ R n , u(t ) ∈ R , y (t ) ∈ R and θ ∈ ([0, ∞), R ) are state vector, control input,
system output and the time-delay respectively; f (⋅) , g(⋅) , ∆f (⋅) and ∆g(⋅) are smooth
vector fields on an open set U ∈ R n and h(⋅) a smooth function on U . Without loss of
equilibrium point of the unforced nominal system and h(x ) vanishes at that equilibrium
point. In other words, f (0) = 0 and h(0) = 0 . This means that y represents the
tracking error. It should be noted that a given model could be easily rewritten in this form
has unstable zero dynamics in the sense defined in Byrnes and Isidori (1985). To control
The proposed nonlinear scheme is mainly constructed on the basis of using the
sliding mode controller to achieve robust control performance. For clear presentation, we
To compensate the undesirable inverse response such that a sliding mode scheme can
be applied, in the proposed scheme the auxiliary output hs (x) should be statically
equivalent to the original output h(x) , i.e., hs (x) should have the same static gain as
the actual process output h(x) and make an u − y s system in the minimum phase
statically equivalent output map (SEOM), hs (x), for nonlinear uncertain processes, we
4.1.2.1. The Design of he (x) : an Auxiliary Output Design Method Using Zero
Assignment Technique
y (t ) = h(x) (4.2b)
which is the nominal model of Eq. (4.1). Apparently, system (4.2) is also a non-minimum
phase one. The following Lemma presents a design methodology for the construction of a
⎡ ∂f (x s ) ∂g (x s ) ⎤
⎢ ∂x + u s ∂x ⎥ and g(x s ) form a controllable pair and g n (x) is nonzero. Then,
⎣ ⎦
there exists an auxiliary locally minimum phase process as follows:
y e (t ) = he (x) (4.3b)
of which
⎡ f ( x) g j ( x) ⎤
Ψ j (x) = det ⎢ j ⎥ , j = 1, 2, K , n − 1 (4.5)
⎣ f n ( x) g n ( x) ⎦
Eq. (4.4) presents a general class of statically equivalent outputs, since the function
ε j (x) is completely arbitrary. To ease the design, we consider the subclass of Eq. (4.4)
n −1
with constant weights ε j , i.e., he (x) = h(x) + ∑ ε j Ψ j (x) . The selection of ε j can be
j =1
completed by using the zero-placement technique as that described in the next lemma.
Lemma 4.2. Let ( x s , u s ) be a reference equilibrium point, and z j be the desired zeros
equations:
n −1
~ ~ ~
P ( s) + ∑ ε j Q j ( s) = P d ( s) (4.6)
j =1
where
~ ∂h(x s ) ⎡ ⎛ ∂f (x s ) ∂g(x s ) ⎞⎤
P ( s) = Adj⎢ sI − ⎜ + us ⎟ g(x s )
∂x ⎣ ⎝ ∂x ∂x ⎠⎥⎦ (4.8)
=~ p +~ p s +L+ ~
0 1p s n −1 n −1
~ ∂Ψ j (x s ) ⎡ ⎛ ∂f (x s ) ∂g(x s ) ⎞⎤
Q j (s) = Adj⎢ sI − ⎜ + us ⎟ g(x s )
∂x ⎣ ⎝ ∂x ∂x ⎠⎥⎦ (4.9)
= q~ j ,1 s + L + q~ j , n −1 s n −1 , j = 1, 2,K, n − 1
and
n −1 ⎛
~ s ⎞
P d (s) = ~
p 0 ∏ ⎜1 − ⎟ = ~p0 + ~
p1d s + L + ~
p nd−1 s n −1 (4.10)
⎜
j =1 ⎝
⎟
zj ⎠
Proof. By considering the zero polynomials of h(x) and Ψ j (x) along with the process
dynamics and by setting appropriate zeros, it can be easily shown that Eq. (4.6) holds.
~ ~
Remark 4.1. In Lemma 4.2, P ( s ) and Q j ( s ) are zero polynomials corresponding to
~
h(x) and Ψ j (x) respectively, and Pd ( s ) for he (x) is the desirable zero polynomial.
can be obtained by zero assignment. If the desired zeros z j ’s are chosen to locate in LHP,
then the obtained ε j can make he (x) minimum phase and statically equivalent to
~
h(x) . It is also noted here that the constant term in Q j ( s ) will have to equal to zero
equivalent minimum phase output is based on the model of the process. If process model
is perfect, the designed auxiliary output, he (x) , can be statically equivalent to the actual
process output, h(x) , and thus makes u − y e system minimum phase. However, as the
process model is not so perfect or the process is subject to model uncertainties, the
minimum phase behavior would not be ensured perfectly. In other words, if the controller
is designed on the basis of the auxiliary output, he (x) , the closed-loop system may
hardly guarantee zero offset performance when facing with process uncertainties. To
overcome this deficiency, in the next subsection we will propose a new algorithm for
ε -redesign and then present a modified synthetic output that is suitable for the robust
4.1.2.2. Synthesis of a Statically Equivalent Output Map (SEOM), hs (x) , for Use
The purpose of this subsection is two folds. The first one is to ensure the minimum
phase behavior under the influence of process uncertainties and the other one is to
guarantee the statically equivalent output property of y s . To meet the first goal, a new
algorithm for the redesign of ε j is proposed. The idea is based on shifting the desired
zeros to make the constructed minimum output invariant despite the influence of the
process uncertainties. The design procedure is depicted in Figure 4.2 and is summarized
as follows:
Initialization: Choose the desired zeros, z 0j ∈ LHP , at the reflections of the RHP
zeros with respect to the imaginary axis. Also, set ∆z j > 0 for zero shifting.
Let i = 1 and z ij = z 0j .
~
Step 1: Set P d ( s ) based on the zeros z ij . Calculate ε ij , j = 1,2, K , n − 1 , from Eq.
Step 2: Check whether hei (x) is minimum phase or not under process uncertainties
Step 3: Shifting the desired zeros by z ij+1 = z ij − ∆z j , then set i = i + 1 and go back
to step 1.
Once the design procedure has converged, a minimum phase output map for the
original uncertain process can be constructed. However, the obtained output map, though
is minimum phase, may not assure that the output is statically equivalent to the actual
equivalent property for the second goal, we suggest the following auxiliary output for
where y e = he (x) , y = h(x) and λ s > 0 is the tuning constant. The role of λ s in this
auxiliary output map is to make a smooth transition from the minimum phase one to
actual process output. Actually, the selection of λ s depends on the process dynamic
characteristics. From Eq. (4.11), it is clear that, for small t , y s is approximately equal
to the minimum phase output map, y e . Notice that y e is simply designed on the basis
process output. This effort ensures that lim y s = y . That is, y s appears to be a statically
t →∞
equivalent output map (SEOM) to the actual process output, which ensures offset-free
and minimum phase behavior despite of the influence of process uncertainties. Having
processes.
y s (t ) = hs (x) (4.12b)
It should be noted here that, based on the constructed SEOM hs (x) , the present system is
minimum phase under uncertainties and the auxiliary output y s is statically equivalent
to the actual process output y . Let the Lie derivative of a smooth function hs (x) along
∂hs (x) n
∂h (x)
Lg hs (x) = g ( x) = ∑ s g i ( x) (4.13)
∂x i =1 ∂xi
In terms of Lie derivative, the relative degree of the system (4.12) is defined as
{ }
Similarly, let κ = min m : L∆f Lmf −1 hs (x) ≠ 0 and w = min{m : L∆g Lmf −1 hs (x) ≠ 0} be the
(Behtash, 1990). Under the above-mentioned assumptions, we shall first present a sliding
mode controller design methodology and then propose a nonlinear predictor design
method in order for compensating the input-delay and to achieve robust stability and
Based on the input-output linearization technique of Isidori (1989), there exists a local
(ξ T
, ηT )
T
= T( x )
(4.15)
(
= hs (x), Lf hs (x),L, Lρf −1 hs (x),η1 (x),L ,η n − ρ (x) )
T
By applying this coordinate transformation, one can transfer the nonlinear uncertain
ξ&i = ξ i +1 , i = 1, 2, L , ρ − 1 (4.16a)
y s = ξ1 (4.16d)
where a(ξ, η) , ∆a(ξ, η) , b(ξ, η) , ∆b(ξ, η) , q(ξ, η) and φ(ξ, η) are given,
respectively, by
and
x = T −1 (ξ, η) (4.23)
Notice that the symbol o denotes a composition function operator. Since the process is
be carefully designed to meet some certain desirable robustness and system performance.
In this chapter, we modify from the SMC approach of Chen and Dai (2001) to give
v(t + θ ) as
where the adaptive gain k is tuned by k& = γ~(δ (t + θ )) 2 ( γ~ > 0 ); f max , bmin , δ (t + θ )
and
⎧ δ (t + θ ) β , if δ (t + θ ) β < 1
sat(δ (t + θ ) β ) = ⎨ (4.30)
⎩sign(δ (t + θ ) β ), if δ (t + θ ) β ≥ 1
In the control law, β is the user-specified boundary layer thickness used to eliminate the
input chattering, and coefficients ci in δ (t + θ ) are chosen such that the polynomial
Γ(λ ) = λ ρ −1 + c ρ −1λ ρ − 2 + L + c 2 λ + c1 has all roots in the open left-half complex plane.
At this stage, the sliding mode control law (4.25) for input-delay processes has been
constructed. However, in the light of the control law, this controller can not be directly
implemented without having the predictive states. Therefore, in the next subsection we
To compensate the time-delay of the process and therefore estimate the process’s
time-advanced states, in this chapter we suggest the use of the following nonlinear
predictor:
~ ~
x& (t ) = f ( ~
x (t )) + g( ~
x (t ))u(t − θ ) (4.31b)
~
xˆ (t + θ | t ) = x(t ) + x ∗ (t ) − ~
x (t ) (4.31c)
where x ∗ (t ) , ~
x (t ) and x(t ) ∈ R n denote, respectively, the model state vector, the
~
nominal state vector, and the actual plant’s state vector; θ ≥ 0 is the estimated
~
time-delay in the manipulated input and xˆ (t + θ | t ) ∈ R n represents the corrected
time-advanced predictive state vector. By comparing Eqs. (4.31a) with (4.31b), it follows
~ ~
that x ∗ (t ) = ~
x (t + θ ) if the predictor is initialized as x ∗ (0) = ~
x (θ ) . This initialization
~
can be achieved at steady state because in this case ~
x (θ ) = ~
x (0) . As a result, in the
absence of plant/model mismatch the prediction model yields the plant state vector one
~ ~
time-delay ahead, i.e. xˆ (t + θ | t ) = x(t + θ ) if θ = θ . With the introduction of the
&
ξˆi = ξˆi +1 , i = 1, 2, L , ρ − 1 (4.32a)
& ~
ξˆρ = [bˆ(ξˆ , ηˆ ) + ∆bˆ(ξˆ , ηˆ )] + [aˆ (ξˆ , ηˆ ) + ∆aˆ (ξˆ , ηˆ )] u (t − θ ) (4.32b)
yˆ s = ξˆ1 (4.32d)
~ ~ ~
vˆ(t + θ | t ) − bˆ(ξˆ (t + θ | t ), ηˆ (t + θ | t ))
u (t ) = ~ ~ (4.33)
aˆ (ξˆ (t + θ | t ), ηˆ (t + θ | t ))
~
where the input vˆ(t + θ | t ) is calculated as
~ ~ ~ ~
vˆ(t + θ | t ) = −kˆδˆ(t + θ | t ) − sat (δˆ(t + θ | t ) β )[bˆmin
−1
( fˆmax + δˆ(t + θ | t ) )] (4.34)
& ~ ~
of which the adaptive gain k̂ is tuned by kˆ = γ~(δˆ(t + θ | t )) 2 ( γ~ > 0 ); δˆ(t + θ | t ) ,
fˆmax , and b̂min are given, similarly as those of δ (t + θ ) , f max and bmin .
With the incorporation of the nonlinear predictor and the insertion of the control law
&
ξˆ = A c ξˆ + B c [vˆ + bˆs−1 ( fˆs + c T ξˆ )] (4.35a)
~ ~
x& = f ( ~
x ) + g( ~
x )u ( t − θ ) (4.35d)
where
⎡ 0 1 0 0 K 0 0⎤
⎢ 0 0 1 0 K 0 0⎥
⎢ ⎥
⎢ M M M M O M M ⎥
Ac = ⎢ ⎥ (4.36)
⎢ 0 0 0 0 K 1 0⎥
⎢ 0 0 0 0 K 0 1⎥
⎢ ⎥
⎢⎣− c1 − c2 − c3 − c4 K − cρ −1 − 1⎥⎦
⎡0⎤
⎢M⎥
Bc = ⎢ ⎥ (4.37)
⎢0⎥
⎢ ⎥
⎣b̂s ⎦
bˆ(ξˆ , ηˆ )
fˆs := fˆs (ξˆ , ηˆ ) = ∆bˆ(ξˆ , ηˆ ) − ∆aˆ (ξˆ , ηˆ ) (4.38)
aˆ (ξˆ , ηˆ )
and
∆aˆ (ξˆ , ηˆ )
bˆs := bˆs (ξˆ , ηˆ ) = 1 + (4.39)
aˆ (ξˆ , ηˆ )
The robust stability and desired behavior of the closed-loop system are described in
Theorem 4.1. Suppose that the uncertain input-delay system (4.12) is subject to the
control law (4.33) and the stable nonlinear predictor (4.31). If bˆmin > 0 , then the
(P4.1) Uniform stability: Let η = [ξˆ T , ηˆ T ]T and η0 = [ξˆ T0 , ηˆ T0 ]T . For each d ≥ d , given
any η(⋅) : [t0 , ∞) → R n , and η(t 0 ) = η0 of the closed-loop system (4.35), there
exists a constant ϑ ( d ) > 0 such that η0 ≤ ϑ (d ) implies that η(t ) ≤ d for all
t ≥ t0 .
(P4.2) Uniform boundedness: Given any r > 0 and any η(⋅) : [t 0 , t1 ) → R n and
(P4.3) Uniform ultimate boundedness: For each d ≥ d and r > 0 , given any
4.2. A Case Study: Regulation Control of a Van de Vusse Reactor in the Presence of
The purpose of this section is to demonstrate the design procedure of the proposed
sliding mode control scheme and evaluate its performance through a chemical
reactor in which the following series/parallel reactions (Van de Vusse, 1964; Kantor, 1986)
A ⎯⎯→
k1
B ⎯⎯→
k2
C (4.40a)
2 A ⎯⎯→
k3
D (4.40b)
where A is the reactant, B the desired product, and C and D are unwanted
(Engell and Klatt, 1993). The reaction rates of A and B are assumed to be
rA = − k1 (T )C A − k 3 (T )C A2 (4.41)
rB = k1 (T )C A − k 2 (T )C B (4.42)
k i (T ) = k i 0 exp( − Ei RT ) (4.43)
species A and B inside the reactor. During operation, the volume of this reactor, V ,
is assumed to be constant. Besides, the feed stream consisting of pure A is fed to this
C A0 = 5 gmol ⋅ L−1 and the feed temperature is T0 = 403.15 K. By means of the material
balances for species A and B and the energy balance for the reactor, the process
the time-delay, ∆H i , ρ s and C P are the reaction heat, density and specific heat of the
reaction mixture, respectively, and − Qs the cooling rate per unit volume. In Table 4.1,
we list the values for the model parameter constants and operation conditions.
E 2 R = 9758.3 K
k10 = 1.287 ⋅ 1012 h −1 C P = 3.01 kJ ( kg ⋅ K ) −1
E3 R = 8560 K
k 20 = 1.287 ⋅ 1012 h −1 ρ s = 0.9342 kg ⋅ L−1
−1
∆H 1 = 4.2 kJ ⋅ mol
The control objective is to maintain the process output C B as close as possible to the set
point (steady-state value) by adjusting the dilution rate, um = F V . In this case, the given
steady-state values are C Ad = 1.25 mol ⋅ L−1 , C Bd = 0.90 mol ⋅ L−1 and Td = 407.15 K ,
model around the reference steady-state, it exhibits locally asymptotically stable and
locally non-minimum phase owing to no RHP pole (-96.518 and − 33.141 ± 9.8118 i )
and the presence of a RHP zero (-11.1673 and +122.6824). Since this chemical reactor
To apply the proposed sliding mode control scheme to this Van de Vusse reactor, we
1
x& 3 (t ) = [ − ∆H 1k1 ( x3 (t ) + x3d ) ⋅ ( x1 (t ) + x1d ) − ∆H 2 k 2 ( x 3 (t ) + x3d ) ⋅ ( x 2 (t ) + x 2 d )
ρ sCP
− ∆H 3 k 3 ( x3 (t ) + x 3d ) ⋅ ( x1 (t ) + x1d ) 2 + ∆H 1k1 ( x3d ) ⋅ x1d + ∆H 2 k 2 ( x 3d ) ⋅ x 2 d
+ ∆H 3 k 3 ( x 3d ) ⋅ x12d ] − x 3 (t ) ⋅ ud + ( x 30 − x 3d − x 3 (t )) ⋅ u(t − θ )
(4.45c)
y (t ) = x2 (t ) (4.45d)
It is further assumed that in this reaction system there exists an unmodeled first-order side
reaction from A and an error in the measurement of the inlet flow rate, F . Under such
y (t ) = h(x(t )) (4.46b)
where
⎡ f1 (x(t )) ⎤
f (x(t )) = ⎢⎢ f 2 (x(t ))⎥⎥ =
⎢⎣ f 3 (x(t )) ⎥⎦
⎡ ⎤
⎢− k ( x (t ) + x ) ⋅ ( x (t ) + x ) − k ( x (t ) + x ) ⋅ ( x (t ) + x ) 2 + k ( x ) ⋅ x ⎥
⎢ 1 3 3d 1 1d 3 3 3d 1 1d 1 3d 1d ⎥
⎢+ k 3 ( x3d ) ⋅ x1d − x1 (t ) ⋅ u d
2 ⎥
⎢ ⎥ (4.47)
⎢k1 ( x3 (t ) + x3d ) ⋅ ( x1 (t ) + x1d ) − k 2 ( x3 (t ) + x3d ) ⋅ ( x 2 (t ) + x 2 d ) − k1 ( x3d ) ⋅ x1d ⎥
⎢+ k ( x ) ⋅ x − x (t ) ⋅ u ⎥
⎢ 2 3d 2d 2 d
⎥
⎢ 1 ⎥
⎢ ρ C [− ∆H 1 k1 ( x3 (t ) + x3d ) ⋅ ( x1 (t ) + x1d ) − ∆H 2 k 2 ( x3 (t ) + x3d ) ⋅ ( x 2 (t ) + x 2 d )⎥
⎢ s P ⎥
⎢− ∆H 3 k 3 ( x3 (t ) + x3d ) ⋅ ( x1 (t ) + x1d ) 2 + ∆H 1 k1 ( x3d ) ⋅ x1d + ∆H 2 k 2 ( x3d ) ⋅ x 2 d ⎥
⎢ ⎥
⎣⎢+ ∆H 3 k 3 ( x3d ) ⋅ x1d ] − x3 (t ) ⋅ u d
2
⎦⎥
⎡ x10 − x1d − x1 (t ) ⎤
g (x(t )) = ⎢⎢ − x 2 d − x 2 (t ) ⎥⎥ (4.48)
⎢⎣ x30 − x3d − x3 (t )⎥⎦
⎡− e f x1 (t )⎤
∆f (x(t )) = ⎢⎢ 0 ⎥
⎥ (4.49)
⎢⎣ 0 ⎥⎦
⎡1⎤
∆g (x(t )) = e g ⎢⎢1⎥⎥ (4.50)
⎢⎣1⎥⎦
and
h(x(t )) = x 2 (t ) (4.51)
of which 0.1 ≤ e f ≤ 0.3 and 0.2 ≤ e g ≤ 0.4 . To construct a statically equivalent output
map (SEOM) for the proposed SMC scheme, we calculate Ψ1 and Ψ2 from the
process model as
Ψ1 (xˆ ) = [−k1 ( xˆ 3 + x3d ) ⋅ ( xˆ1 + x1d ) − k 3 ( xˆ 3 + x3d ) ⋅ ( xˆ1 + x1d ) 2 + k1 ( x3d ) ⋅ x1d + k 3 ( x3d ) ⋅ x12d
1
− xˆ1 ⋅ u d ]( x30 − x3d − xˆ 3 ) − { [−∆H 1 k1 ( xˆ 3 + x3d ) ⋅ ( xˆ1 + x1d ) − ∆H 2 k 2 ( xˆ 3 + x3d )
ρ s CP
⋅ ( xˆ 2 + x2d ) − ∆H 3 k 3 ( xˆ 3 + x3d ) ⋅ ( xˆ1 + x1d ) 2 + ∆H 1 k1 ( x3d ) ⋅ x1d + ∆H 2 k 2 ( x3d ) ⋅ x2d
+ ∆H 3 k 3 ( x3d ) ⋅ x12d ] − xˆ 3 ⋅ u d }( x10 − x1d − xˆ1 )
(4.52a)
(4.52b)
The zeros polynomial for the process output, i.e., the numerator polynomial of the
~
P ( s ) = −0.9s 2 + 100.3636s + 1233.0296 . Subsequently, we also obtain the zeros
~
Q1 ( s ) = −654.7228s 2 − 173130.6147 s (4.53a)
~
Q2 ( s ) = −535.3284s 2 + 33842.4452s (4.53b)
synthesized as
It should be mentioned again that the zeros polynomial around the given steady state, i.e.
~ ~ ~ ~
P ( s ) + ε 1Q1 ( s ) + ε 2 Q2 ( s ) , can be made equal to the given polynomial P d ( s ) by
simulations shown in Figure 4.3 reveal that the obtained weights of ε 1 and ε 2 are able
to make y e the minimum phase under the uncertainty variations. However, this figure
also discloses that the synthesized output y e is still unable to be statically equivalent to
the actual process output, where there exist significant output errors. Therefore, for
statically output equivalence it is desirable to design a SEOM, y s , for this Van de Vusse
reactor. Now, by using the auxiliary output of y s designed based on Eq. (4.11), we
simulate the open-loop system response with various designed values of λ s . From Figure
statically equivalent output for the actual process despite of the influence of process
uncertainties. Also observed is that the parameter λ s controls the transition behavior of
w ≥ ρ = κ . To implement the sliding mode controller, the value of fˆmax and b̂min
fˆmax = 7 and bˆmin = 0.3 for the sliding mode controller. The other parameters are set as
c1 = 1.0 , kˆ(0) = 1.0 , γ~ = 1.0 , λ s = 0.3 and β = 0.01 . In order to verify the
regulation ability of the proposed strategy, we suppose that the system states are
perturbing to move away from their steady states as x1 (0) = −0.7 , x 2 (0) = −0.2 and
x3 (0) = 1.0 initially. Up to this point, we are ready to investigate the following important
The role of the incorporated nonlinear predictor is examined first. In this case, we
~
assume that the estimated input-delay coincides with the actual one, that is θ = θ = 0.01 .
The comparisons of the proposed sliding mode control law with and without the
incorporation of the nonlinear predictor are shown in Figure 4.5. From this figure, it is
evident to observe that the sliding mode control can be fail if the nonlinear predictor is
not included in the control system, mainly due to the negative effect of the input-delay on
state feedback. Once the nonlinear predictor is incorporated as that shown in Figure 4.1,
the produced control input is able to drive the process output to the setpoint, and therefore
process uncertainties. These simulation results obviously reveal that the nonlinear
time-advanced predictor plays an important role in the proposed SMC control scheme.
assume that the process is encountered with not only unmodeled side reaction and
measuring error but also extra unmeasured disturbances. The extra disturbances introduce
⎡− e f x1 ⎤
∆f (x) = ⎢⎢ d1 ⎥⎥ (4.55)
⎢⎣ d 2 ⎥⎦
where d1 = 0.5 and d 2 = 2 . For SMC design, the value of bˆmin = 0.3 is still suitable
for this case, but fˆmax should be increased in order for accommodating this extra
disturbances. With increasing fˆmax to 10 and using the other controller parameters as
those mentioned before, we plot the closed-loop system performance in Figure 4.6. From
this figure, it is clear to observe that the SMC control system simply using y e results in
a small offset on the steady state. The reason for this is that the design of y e is simply
on the base of process’s nominal model and does not take the influence of uncertainties
into consideration, which therefore has insufficient capability in handling with extra
process uncertainties and disturbances. In contrast, with using the modified synthetic
output, y s , for the controller design, the sliding mode control system is capable of
driving the process output gradually to achieve zero offset performance even though
diversified and extra uncertainties are imposed on this chemical reactor. These simulation
results corroborate again that the incorporation of the SEOM, y s , in the proposed
Due to its intrinsic nonlinearity, the static and dynamic behavior of the Van de Vusse
reactor is a strong function of kinetic parameters. When the kinetic parameters change
and/or the model is not so accurate to present the actual process, there is considerable
discrepancy between the actual process and the process model. To explore the plant
uncertainties on the essential behavior of the control system, we assume that each of the
process’s kinetic parameters k i 0 , i = 1, 2, 3 , has ± 25% variation from its nominal value
while these parameter values in the model remain unchanged. In designing SMC to deal
with these parameter variations, the value of fˆmax is set as 10 for accommodating these
control input is shown in Figure 4.7, demonstrating clearly that the proposed scheme is
In practice, the on-line measurement of all states is often not possible. Therefore, for
unavoidably for practical implementation. To deal with this situation, we propose the use
of a sliding observer for estimating the states of this chemical process. In this case, we
assume that only the process output x2 is measured, i.e. the states, x1 and x3 , are
sliding observer is constructed based on the nominal system and the output measurement
x 2 as follows:
(4.56a)
1
x& 3 (t ) = [ − ∆H 1k1 ( x3 (t ) + x3d ) ⋅ ( x1 (t ) + x1d ) − ∆H 2 k 2 ( x3 (t ) + x3d ) ⋅ ( x2 (t ) + x 2 d )
ρ sCP
− ∆H 3k 3 ( x3 (t ) + x3d ) ⋅ ( x1 (t ) + x1d ) 2 + ∆H 1k1 ( x3d ) ⋅ x1d + ∆H 2 k 2 ( x3d ) ⋅ x 2 d
~ (
+ ∆H 3k 3 ( x3d ) ⋅ x12d ] − x3 (t ) ⋅ ud + ( x30 − x3d − x3 (t )) ⋅ u (t − θ ) + kˆ03sat ( x 2 (t ) β ob )
(4.56c)
(
where x2 = x 2 − x2 denotes the error between the measured output x 2 and the
estimated output value x2 . The boundary layer thickness of the observer, β ob , and the
switching gain, k̂ 02 , are set as β ob = 0.05 and kˆ02 = 2 , respectively. With the desired
(1997), we have the remaining gains as kˆ01 = −2.9132 and kˆ03 = 6.1208 . Along with
the incorporation of the sliding observer, the correction term (4.31c) in the nonlinear
~
xˆ (t + θ | t ) = x(t ) + x ∗ (t ) − ~
x (t ) (4.57)
Figure 4.8 depicts the schematic diagram of the proposed sliding mode control scheme
for this Van de Vusse reactor with the sliding observer being incorporated for state
estimation. For simulation studies, we set the initial conditions of the true states as
x(0) = [− 0.7 − 0.2 1]T and the estimated states x(0) = [− 0.5 − 0.2 1.3]T for the
sliding observer. The sliding observer performance is shown in Figure 4.9. Simulation
results reveal that the proposed scheme with sliding observer is an effective and
excellent control performance without the need of full state measurements. Additionally,
therefore of great interest to investigate the capability of the proposed sliding mode
control scheme in handling with time-delay errors. For simulation, we assume that the
~
estimated input-delay, θ , does not coincide with the actual one, θ . From Figure 4.10 it
is clear to observe that the proposed observer-based SMC scheme still provides robust
4.3. Summary
This chapter has presented a novel and systematic sliding mode control scheme for
the robust regulation control of nonlinear uncertain chemical processes in the presence of
simultaneously the non-minimum phase behavior and input-delay. The main idea of the
present approach is to reap the benefits of the statically equivalent output (SEOM),
nonlinear state predictor and a sliding mode control strategy. To overcome the negative
effect of inverse response behavior and eliminate the steady state offset, a new algorithm
has been proposed such that the designed auxiliary output can be statically equivalent to
the actual output and makes the resultant system minimum phase despite the influence of
the process uncertainties. With the incorporation of the constructed statically equivalent
scheme is easily established to tackle with the difficult control problem of uncertain
proposed sliding mode control scheme are guaranteed by the Lyapunov stability theorem.
For demonstration, the effectiveness and applicability of the proposed robust control
strategy has been illustrated through the control of a Van de Vusse reactor in the presence
observer along with the proposed scheme has also been investigated in this work.
Extensive simulation results reveal that the proposed sliding mode control scheme
and input-delay.
Figure 4.4. Open-loop simulations using SEOM with designed values of λ s ’s. ─
y ; ······· y s ( λ s = 5 ); ---- y ( λ
s s ---
= 15 ); · · y s ( λ s = 30 ).
Figure 4.6. Closed-loop system performance in the face with unmodeled side reaction,
Figure 4.8. Schematic diagram of the sliding mode control system along with a sliding
observer.
Figure 4.9. System response under sliding observer with observer poles of p1 = p 2 = −4
for the case of existing unmodeled side reaction and measuring error.
CHAPTER 5
A SLIDING MODE CONTROL SCHEME FOR IMPRECISELY
KNOWN CHEMICAL PROCESSES
This chapter considers the regulation control of nonlinear chemical processes whose
dynamics are imprecisely known. A nonlinear control scheme that incorporates a sliding
mode controller and a neural fuzzy strategy is proposed to deal with this kind of
processes. The sliding mode controller designed on the base of a previously known
process model is implemented to keep system’s trajectory around the desired manifold.
For extra and/or unknown dynamics that can not be captured before the SMC design
stage, an intelligent scheme of utilizing a neural fuzzy strategy is then used to provide an
adaptive ability to accommodate the perturbation, which therefore is able to force the
system output back to and maintain in the desired set point. The combination of the
model-based SMC with the intelligent neural fuzzy technique presents the following
advantages: (1) it can make use of a previously available process model; (2) the nominal
closed-loop control system can be guaranteed theoretically; (3) the imprecisely known
dynamics and extra uncertainties can be handled through the intelligent approach, and
uncertainties; (4) the chattering phenomena of the conventional SMC can be attenuated
with the incorporation of the intelligent neural fuzzy technique; (5) the present approach
The organization of this chapter follows naturally in five sections. In Section 5.1, the
control system and problem have been introduced. Following this, Section 5.2 presents
the control system structure. The related issues on the design of the SMC and an adaptive
neural fuzzy algorithm are explored in Section 5.3. In addition, the potential use of a
sliding observer along with the proposed scheme is also investigated as a whole. To
demonstrate the effectiveness and the applicability of the proposed scheme, Section 5.4
performs a case study on the regulation control of a nonlinear uncertain CSTR with the
presence of unmodeled side reaction, measuring noise, and extra matched and unmatched
imprecisely known but with its nominal system’s dynamics being modeled by the
y = h(x ) (5.1b)
f (⋅) , g(⋅) , ∆f (⋅) and ∆g(⋅) are smooth vector fields on an open set U ∈ R n ; h(⋅) is a
smooth function on U . The process model in Eq. (5.1), which may not include system’s
unknown dynamics and extra uncertainties imposed, is formulated for the purpose of
designing a main sliding mode control loop. Before proceeding to the sliding mode
controller design, let’s consider the nominal system of the uncertain model (5.1) as
follows:
y = h(x ) (5.2b)
asymptotically stable equilibrium point of the unforced nominal system and h(x )
vanishes at that equilibrium point. In other words, f (0) = 0 and h(0) = 0 . This means
that y represents the tracking error. Here, it should be noted that a given model could
important to emphasis again that the process model in Eq. (5.1), though is an uncertain
model, still can not represent all type of processes uncertainties. Therefore, in the
following sections we shall first design a SMC on the basis of the process model and then
incorporate a NFC to deal with extra uncertainties that can not previously modeled in the
The proposed control scheme is shown in Figure 5.1, where the sliding mode
controller (SMC) is recommended for use as the main feedback controller due to that the
SMC has the capability of dealing with some pre-considered process uncertainties. When
the process output is reaching to its desired value or is perturbing to move away from the
set point by the influence of imprecisely known dynamics and/or extra uncertainties to
some certain extent, an intelligent strategy of using an adaptive neural fuzzy controller
(NFC) will gradually join in the control in order to avoid the chattering behavior that may
arise from using the SMC as well as to accommodate extra uncertainties. For identifying
the input-output dynamic behavior of a plant and therefore to provide a reference signal
for on-line tuning of the NFC, a feedforward MNN-based estimator (Chen and Chang,
transition from the SMC to the NFC, a switching function d s (δ ) is introduced to judge
the output weight between the SMC and NFC. For completeness, in what follows we
Based on the previous preparations, we are ready to design a control system for
linearization technique. Second, the intelligent control techniques including neural fuzzy
controller (NFC) and MNN-based estimator are introduced. Third, the overall control
system which integrates above techniques will be explored as a whole. Subsequently, for
Before designing a sliding mode controller, we shall briefly describe the major
aspects of input-output feedback linearization and the properties of zero dynamics, which
are essential to the design of sliding mode controller for the nonlinear uncertain system
Throughout this chapter, we assume the nominal model has strong relative degree
transformation as follows:
[ξ T
, ηT ]
T
= T( x )
(5.3)
[
= h(x), Lf h(x),L , Lρf −1 h(x),η1 (x),L ,η n − ρ (x) ]
T
ξ&i = ξ i +1 , i = 1, 2, L , ρ − 1 (5.4a)
y = ξ1 (5.4d)
where
and
x = T −1 (ξ, η) (5.8)
Notice that the symbol o denotes a composition function operator. The internal
dynamics η can not be observed from the system output. Since we are interest in
achieving stable tracking, it is therefore required that η remains bounded for bounded
dynamics of η whose boundedness entirely depends on the vector field q(ξ, η) . When
play an important role to determine whether the state η remains bounded (Henson and
Seborg, 1991). A sufficient condition for the bounded-input bounded-state stability of the
assume that the zero dynamics of η is exponentially stable and the nominal closed-loop
system is internally stable. Actually, this assumption is very natural because we can not
expect robustness properties except that the nominal system is internally stable.
ξ&i = ξ i +1 , i = 1, 2, L , ρ − 1 (5.10a)
where a(ξ, η) , b(ξ, η) and q(ξ, η) are defined in Eqs. (5.5), (5.6) and (5.7)
and
In the following, we will focus on the design of the sliding mode controller according to
u ≤u (5.14)
and therefore the upper bound functions of ∆b(ξ, η) and ∆a(ξ, η)u can be estimated,
respectively, as
and
Next, let’s choose a sliding function for the transformed system (5.10) as
ρ −1
δ (ξ ) = ∑ ci ξ i + ξ ρ (5.17)
i =1
where the sliding coefficient ci (> 0) should be set to satisfy a Hurwitz polynomial
surface δ = 0 , then all the states ξ i , i = 1, 2, K , ρ , would converge to zero, i.e. there is
no tracking error.
Due to that the internal dynamics η is assumed stable, a sliding mode control law
u SMC = u~ + ∆u
1 ⎧ ⎡ ρ −1 ⎤ ⎫ (5.18)
= ⎨ ⎢− ∑ ci ξ i +1 − b(ξ, η)⎥ + [− ( B (ξ, η) + A(ξ, η) + γ )sign (δ )]⎬
a (ξ, η) ⎩ ⎣ i =1 ⎦ ⎭
⎧1 , δ > 0
⎪
sign (δ ) = ⎨0 , δ = 0 (5.19)
⎪− 1 ,δ < 0
⎩
The control law inherently contains two parts. The first part
1 ⎡ ρ −1 ⎤
u~ = ⎢ − ∑ ci ξ i +1 − b(ξ, η)⎥ (5.20)
a (ξ, η) ⎣ i =1 ⎦
is the steady state one that ensures to preserve the nominal performance. The second part
1
∆u = [− ( B(ξ, η) + A(ξ, η) + γ ) sign (δ )] (5.21)
a (ξ, η)
containing the uncertainty measures is to force the system’s trajectory back to the sliding
analysis is given in Appendix C, showing clearly that the sliding condition of δδ& ≤ −γ δ
is satisfied by the sliding mode control law. That is, with the control law, the system
trajectory will be asymptotically stable and have the desired system performance under
the condition that the process can be represented exactly by model (5.1). However, it
should be mentioned again that we are concerned with the control of imprecisely known
the control scheme. The following subsection describes the design procedure of an NFC
The adaptive neural fuzzy technique (Chen and Peng, 1999), adopted in this article, is
adaptive feedback controller, and the other is an estimator for identifying the input-output
Clearly, the NFC, is a four-layer feedforward connectionist network, which integrates the
connectionist structure. With a learning algorithm (to be designed later), the NFC has
distributed ability to learn the membership functions and fuzzy logic rules. Let each input
has n terms for fuzzy partition, i.e. each input has n membership functions, then the
The input units in this layer are the transformed values of δ and δ& , M 1 and M 2 ,
which are obtained through the mappings of M 1 = (1 − exp(−α 1δ )) /(1 + exp(−α 1δ )) and
M 2 = (1 − exp(−α 2δ&)) /(1 + exp(−α 2δ&)), respectively, where α1 and α2 are the
pre-specified parameters governing the slope of the hyperbolic tangent function. This
effort ensures the universe of discourse to lie in the range of [-1, 1] and therefore can
facilitate the design of the neural fuzzy control scheme. For clear presentation, we depict
This layer transforms the signals from the input layer by Gaussian function as a
membership function to determine the relative contribution of the observed signals. The
(Oij(1) − aij ) 2
Input units: I ij( 2 ) = − , i = 1, 2; j = 1, 2, L , n (5.24)
bij2
where aij and bij are, respectively, the center and the width parameters of the Gaussian
function.
This layer performs centroid defuzzification to obtain the inference output, i.e.
m
Input unit: I ( 4) = ∑O(p3) wp (5.28)
p=1
I ( 4)
Output unit: O ( 4)
=u = *
m
(5.29)
∑O
p =1
( 3)
p
In addition, from Figure 5.2, the output of the NFC is amplified by a scaling factor K 3 ,
between better control performance and larger control input (Chen and Peng, 1999). It is
obvious that NFC presented above is equivalent to a simplified fuzzy inference system,
where layers 1 and 2 correspond to the antecedent part of the fuzzy control rules, and the
Based on the structure of the NFC, a learning algorithm derived on the basis of
1
minimizing the error function of E = (δ (ξ )) 2 is represented as follows:
2
) ) ∂E )
wυ (k + 1) = wυ (k ) − η r + β r ∆wυ (k ) (5.30)
∂ wυ
) ) ∂E )
aij (k + 1) = aij (k ) − η r + β r ∆aij (k ) (5.31)
∂ aij
and
) ) ∂E )
bij (k + 1) = bij (k ) − η r + β r ∆bij (k ) (5.32)
∂ bij
)
for υ = 1, 2, L , m , i = 1, 2 and j = 1, 2, L, n , where ∆χ (k ) is defined by
) ) ) ∂E ∂E ∂E
∆χ (k ) ≡ χ (k ) − χ (k − 1) . In these updating rules, the gradients , and
∂ wυ ∂aij ∂bij
can be obtained by
( 3)
∂E ∂ E ∂δ ∂ u * ∂ δ Oj
= = δ (5.33)
∂ wυ ∂ δ ∂ u * ∂ wυ ∂ u * m ( 3)
∑O
p =1
p
⎧ ∂ δ 2(O1(1j ) − a1 j )O1( 2j ) n ⎛ m m ⎞
⎪δ ∑ O 2 l ⎜ ( j −1) n + l ∑
( 2)
⎜ w O ( 3)
− ∑ O p(3) w p ⎟⎟, i = 1
∂ E ⎪⎪ ∂ u b1 j (∑ p =1 O p )
∗ 2 m ( 3) 2 p
l =1 ⎝ p =1 p =1 ⎠
=⎨ (5.34)
∂ ai j ⎪ ∂ δ 2(O2(1j) − a 2 j )O2( 2j ) n ⎛ m m ⎞
δ
⎪ ∂ u ∗ b 2 ( m O ( 3) ) 2 ∑ O1l ⎜ ( l −1) n + j ∑
( 2)
⎜ w O ( 3)
− ∑ O p(3) w p ⎟⎟, i = 2
2 j ∑ p =1 p
p
⎪⎩ l =1 ⎝ p =1 p =1 ⎠
and
⎧ ∂ δ 2(O1(1j ) − a1 j ) 2 O1( 2j ) n ⎛ m m ⎞
⎪δ ∑ O2 l ⎜ ( j −1) n + l ∑ p
( 2)
⎜ w O ( 3)
− ∑ O p(3) w p ⎟⎟, i = 1
∂ E ⎪ ∂ u b1 j (∑ p =1 O p )
∗ 3 m ( 3) 2
l =1 ⎝ p =1 p =1 ⎠
=⎨ (5.35)
∂ bi j ⎪ ∂ δ 2(O2(1j) − a 2 j ) 2 O2( 2j ) n ⎛ m m ⎞
δ
⎪ ∂ u ∗ b 3 ( m O ( 3) ) 2 ∑ O ⎜
⎜
( 2)
w − + ∑ O ( 3)
− ∑ O ( 3)
w ⎟, i = 2
⎟
2 j ∑ p =1 p
1l( l 1) n j p p p
⎩
l =1 ⎝ p =1 p =1 ⎠
to the paper of Chen and Peng (1999). The only unknown in the learning algorithm of the
NFC is ∂δ / ∂u * , which can be obtained using the technique discussed in the following
subsection.
~ S + θ~ ,
m1
Hidden layer: net 2i = ∑ w 2 ij 1 j 2i i = 1, 2, L, m2 (5.37)
j =1
− net2 i
1− e
S 2i = − net2 i
, i = 1, 2, L, m2 (5.38)
1+ e
~ S + θ~ ,
m2
Output layer: net 3 = ∑ w3i 2 i 3 (5.39)
i =1
~ a~ (1 − e − net3 )
δ = (5.40)
1 + e − net3
where m1 and m2 are, respectively, the numbers of input and hidden layer nodes, and
~ ~ 1 ~
δ is the MNN output. Based on minimizing the error function of E = (δ − δ ) 2 , we
2
have the updating rules for the MNN as follows (Chen and Peng, 1999):
~ (k) + 1) = w
~ (k) ) + η (δ − δ~ )δ w
~ ~ )
w 2 ij 2 ij s 3 3i δ 2 i S1 j + β s ∆w2 ij ( k ) (5.41)
~ (k) + 1) = w
w ~ (k) ) + η (δ − δ~ )δ S + β ∆w
~ (k) ) (5.42)
3i 3i s 3 2i s 3i
~ ) ~ ) ~
~ δ + β ∆θ (k ) ~ )
θ 2i (k + 1) = θ 2i (k ) + η s (δ − δ )δ 3 w3i 2 i s 2i (5.43)
~ ) ~ ) ~ ~ )
θ 3 (k + 1) = θ 3 (k ) + η s (δ − δ )δ 3 + β s ∆θ 3 (k ) (5.44)
and
~
) ) ~ δ )
a (k + 1) = a (k ) + η s (δ − δ ) ~ + β s ∆a~ (k )
~ ~ (5.45)
a
where δ 2i and δ 3 are given by
1
δ 2i = (1 − S 2i )(1 + S 2i ) (5.46)
2
and
~ ~
1 ~⎛ δ ⎞⎛ δ ⎞
δ3 = a ⎜⎜1 − ~ ⎟⎟⎜⎜1 + ~ ⎟⎟ (5.47)
2 ⎝ a ⎠⎝ a⎠
~
Once the MNN has been trained to emulate the sliding function, we have δ ≈ δ . This
~
means that ∂δ / ∂u * can be approximated by ∂δ / ∂u * . Consequently, by using the
input-output relationships of the MNN, we obtain the required gradient information for
It is often not possible to obtain all states of a chemical process via on-line
we propose the use of a sliding observer (Wang et al., 1997) for estimating the states of
nonlinear chemical processes such that the proposed SMC and NFC control strategies can
be implemented without full state measurements. Also, the design of the nonlinear
observer can be done simply on the basis of the nominal system (system model) and the
output measurement. In essence, this sliding observer behaves like a reduced order
observer, and it can guarantee to have a bounded error norm when facing with modeling
errors. In addition, the sliding observer possesses many advantages such as simple
structure and less restrictive in the design and construction, no requirement of canonical
straightforwardly by allocating the observer poles. This chapter will demonstrate the
observer design procedure through illustrative example. For theoretical details, please
where ψ ( > 0 ) is the boundary layer thickness. Based on d s (δ ) , we can generate the
u = (1 − d s (δ )) u NFC + d s (δ ) ∆u + u~ (5.50)
where u is the control input, u NFC is the output of the NFC, u~ and ∆u are the
outputs of the SMC, defined in Eqs. (5.20) and (5.21) respectively. Here, it is emphasized
again that the SMC terms, u~ and ∆u , can be formulated simply on the base of the
previously known process model, Eq. (5.1), while u NFC is obtained through on-line
learning. Also, from Eq. (5.50), it is clear to understand that the SMC can dominantly
govern the control system as the system trajectory is far away from the sliding surface.
When the system trajectory reaches into the boundary layer of the switching function, the
NFC will gradually take over the control in order to avoid the chattering behavior that
may arise from using the SMC (5.18). In addition, with the presented learning algorithms,
the NFC can provide an adaptive ability of eliminating the influences causing by extra
uncertainties. Once the influencing uncertainties give rise to a bad situation where the
system trajectory moves away from the sliding boundary, the SMC will again dominate to
force the system trajectory back to the sliding surface and to ensure the stability of the
overall control system. This integration of techniques from these two fields reaps the
benefits of both model-based and intelligent systems and therefore is able to handle with
In this section, the design procedure and the control performance of the proposed
A B C
is being carried out. The control objective is to maintain the desired concentration of CC
as close as possible to its steady state value by adjusting the molar flow rate of B . The
dynamics of the CSTR is given by the following mathematical model (Chen and Dai,
2001)
) ) ) )
x&1 = 1 − x1 − Da1 x1 + Da 2 x 22 (5.51a)
) ) ) ) ) )
x& 2 = Da1 x1 − x 2 − Da 2 x 22 − Da3 x 22 + u (5.51b)
) ) )
x& 3 = Da3 x 22 − x3 (5.51c)
) )
y = x3 (5.51d)
) )
where xi , i = 1, 2, 3 are dimensionless concentrations defined by x1 = C A C AF ,
) )
x 2 = C B C BF and x3 = C C C AF of which C AF indicates the feed concentration of
feeding rate of species B and F is the volumetric flow rate. Based on the deviation
) ) ) )
variables, x1 = x1 − x1d , x 2 = x 2 − x 2 d , x3 = x3 − x3d and u = u − u d , where xid ,
x& 3 = Da 3 x 22 + 2 Da 3 x 2 d x 2 − x3 (5.52c)
y = x3 (5.52d)
To rewrite the above equation into the standard form as Eq. (5.2), we denote the nominal
y = h(x ) (5.53b)
where
⎡ f1 ⎤ ⎡ − (1 + Da1 ) x1 + 2 Da 2 x 2 d x 2 + Da 2 x22 ⎤
⎢ ⎥ ⎢ 2⎥
f ( x ) = f 2 = ⎢ Da1 x1 − (1 + 2 Da 2 x 2 d + 2 Da3 x 2 d ) x 2 − ( Da 2 + Da3 ) x 2 ⎥ (5.54)
⎢ ⎥
⎢⎣ f 3 ⎥⎦ ⎢⎣ Da3 x 22 + 2 Da3 x 2 d x2 − x3 ⎥
⎦
⎡ 0⎤
g ( x ) = ⎢1⎥ (5.55)
⎢ ⎥
⎢⎣0⎥⎦
and
h( x ) = x3 (5.56)
Assume that in this reaction system there exist first-order side reaction from B and an
error in measuring the molar feeding rate of B . Under such a situation, the process
y = h(x ) (5.57b)
⎡ 0 ⎤
∆f (x) = ⎢⎢− e f x 2 ⎥⎥ (5.58)
⎢⎣ 0 ⎥⎦
and
⎡0⎤
∆g ( x ) = ⎢ e g ⎥ (5.59)
⎢ ⎥
⎢⎣ 0 ⎥⎦
The values for the various constants are as follows (Kravaris and Palanki, 1988):
The steady state values of the states are: x1d = 0.3467 , x2 d = 0.8796 and
x 3d = 0.8796. For SMC design, we first apply the following coordinate transformation,
to process model (5.57) and then obtain the transformed system as follows:
ξ&1 = ξ 2 (5.61a)
where
and
Here, it is easy to verify that the zero dynamics of the nominal system
is exponentially stable. This means that the internal stability is guaranteed and the
proposed SMC design methodology can be utilized. To implement the SMC, the upper
and
For the later simulation studies, the parameters of the SMC for this case are set as
c = [100 1] , γ = 50 and the hard input constraint for this CSTR is − 5 ≤ u ≤ 5 , i.e.
u = 5 . The structure of the NFC is arbitrarily chosen to be of 2-14-49-1, and the initial
linking weights w j (0) suggested are those listed in Table 5.1 (Chen and Peng, 1999).
Table 5.1. The initial linking weights for the nonlinear uncertain CSTR (7 segments).
M1 partitions
negative ← → positive
NB NM NS ZO PS PM PB
M2
NB −1 −1 2 2 1 1
− − − − 0
( w1 ) ( w8 ) 3 3 3 3
−1
→ negative
NM 2 2 1 1 1
− − − − 0
( w2 ) 3 3 3 3 3
NS 2 2 1 1 1 1
− − − − 0
3 3 3 3 3 3
partitions
ZO 2 1 1 1 1 2
− − − 0
3 3 3 3 3 3
PS 1 1 1 1 2 2
− −
positive ←
0
3 3 3 3 3 3
PM 1 1 1 2 2 1
− 0
3 3 3 3 3 ( w48 )
PB 1 1 2 2 1 1
0
3 3 3 3 ( w42 ) ( w49 )
j = 1, 2, ⋅ ⋅⋅, 7 to cover the universe of discourse [-1, 1] uniformly. The structure of the
~ , w
of the MNN parameters, w ~ , θ~ and θ~ , are given randomly in the range
2 ij 3i 2i 3
of –0.01 and 0.01, and the initial shape parameter of the activated function in the output
layer is chosen as a~(0) = 1 , which is of the same shape as the activated function used in
the hidden layer. The learning rate η s of 0.3 is set for the MNN-based estimator and the
momentum parameters are set to β r = 0 and β s = 0.01 for the NFC and the
(α1 , α2 ) = [1, 0.01] , K 3 = 0.5 and ηr = 0.5. For sliding observer design, we assume that
only the process output x3 is measured, i.e. the states x1 and x 2 are unmeasurable.
Following the design methodology of Wang et al. (1997), a sliding observer for this
nonlinear process is constructed based on the nominal model and the output measurement
x3 as follows:
(
x&1 = −(1 + Da1 ) x1 + 2 Da 2 x 2 d x 2 + Da 2 x 22 + kˆ01sat ( x3 β ob ) (5.69a)
(
x& 2 = Da1 x1 − (1 + 2 Da 2 x 2 d + 2 Da3 x 2 d ) x 2 − ( Da 2 + Da3 ) x 22 + u + kˆ02 sat ( x3 β ob ) (5.69b)
(
x& 3 = Da3 x 22 + 2 Da3 x 2 d x 2 − x3 + kˆ03 sat ( x3 β ob ) (5.69c)
(
where x3 = x3 − x3 denotes the error between the measured output x3 and the
estimated output value x3 . The observer boundary layer thickness, β ob , and the
switching gain, k̂ 03 , are set as β ob = 0.01 and kˆ03 = 2 respectively. The design of the
With the incorporation of the sliding observer, the whole control law can be expressed
as follows:
where u NFC is generated by NFC with δ in the updating rules being replaced by the
estimated δ of
and ∆u and u~ , from SMC are, respectively, given based on process model as
∆u =
[ ]
− 2 Da3 (e 0f + 0.01)( x 22 + x 2 d x 2 + 2e g Da3 ( x 2 + x 2 d )u + γ sign (δ )
(5.72)
2 Da3 ( x 2 + x 2 d )
and
u~ =
[
2 Da3 ( x 2 + x 2 d ) Da1 x1 − (1 + 2 Da 2 x 2 d + 2 Da3 x 2 d ) x 2 − ( Da 2 + Da3 ) x 22 ] (5.73)
2 Da3 ( x 2 + x 2 d )
Having the previous preparations, we are ready to investigate the following issues.
For the later simulation studies, we set the initial conditions of the true states
x(0) = [0.5 0.5 0.5]T and the estimated states of x(0) = [0.55 0.45 0.5]T . Figures
5.3 and 5.4 depict the closed-loop system responses when the sliding observer with
designed poles is implemented. From the comparison of these two figures, it is apparent
to recognize that the location of the poles affects the accuracy of the state estimation and,
in turn, the produced control input. However, the system performance by using the
proposed nonlinear scheme is still very robust. Based on the above simulation results, it is
evident that the proposed scheme with sliding observer appears to be an effective and
previously, the effect of the boundary layer thickness ψ of the switching function on
system performance is depicted in Figure 5.5. From this figure, it can be seen that the
produced control input appears chattering as the boundary thickness ψ is set too small,
for example ψ = 0.1 . This is because that the transition from SMC to NFC is too slow,
causing chattering control input around the sliding surface. As the value of ψ is further
set too large, for example ψ = 15 , a relative slow transient response was obtained due to
that the model-free intelligent strategy with on-line tuning rule dominantly governs the
control system. Furthermore, it is observed that the control system can be considerably
slow by merely applying the adaptive NFC control scheme since it is started up randomly.
Based on the above simulation results, it is evident that with a suitable boundary
thickness value the incorporation of the SMC and NFC present significant advantage for
nonlinear process control. Also, from the above simulation results, it is concluded that the
boundary layer thickness ψ = 3 of the switching function is appropriate for the control
(PIDSMC) for chemical processes. The PIDSMC control system consists of three
errors, and a PID control of the sliding surface for the purpose of eliminating the
parameters online in order to control the process states onto a sliding surface. The
applicability of the PIDSMC control scheme is mainly based on the assumption of the
existence of a set of PID parameters that satisfy the reachability condition. With the initial
system performance is shown in Figure 5.6. From this figure, it is clearly to observe that
the PID tuning constants, γ 1 and γ 2 , can affect the system performance significantly.
The increment of the constants gives rise to a fast initial response while at the expense of
high oscillatory control input. To obtain smooth control input, the slow tuning of the PID
parameters should be applied. However, this can result in a relative slow regulation
process. In contrast, the proposed scheme which combines the techniques of the SMC and
NFC can have a good balance and high performance during the whole regulation control
process. Apparently, the produced control input by the proposed scheme is more
extra disturbances on the feeding of the reactants. The extra unmeasured disturbances
introduce significantly additional modeling errors, which induces the uncertainty vector
∆f to be
⎡ d1 ⎤
∆f = − e f x 2 + d 2 ⎥
⎢ (5.74)
⎢ ⎥
⎢⎣ d3 ⎥⎦
These uncertainties impose extra model discrepancy on control system and can make
many of the existing algorithms fail to be applied, for example (Kravaris and Palanki,
1988). In the later simulations, we assume that d 1 = 0.1 , d 2 = 0.02 and d 3 = 0.05 are
entering into system after time of 5. It is also noted that these extra disturbances are
assumed unmeasurable and unknown to the control system. With the controller
shown in Figure 5.7. From this figure, it is clear to show that the control input is soon
results that the proposed control strategy is able to provide robust control without having
abrupt control actions and vigorous control moves in the control loop, even though extra
unknown dynamics are imposed. On the contrary, the PIDSMC scheme seems unable to
handle with the extra uncertainties well, which obviously presents a significant offset.
The obtained simulation result implies that the assumption of the existence of a set of
PIDSMC parameters that satisfy the reachability condition is not assured for this case.
further suppose that the CSTR exists not only the unmodeled side reaction and measuring
error but also the unmatched uncertainties including measurement disturbance and noise.
The magnitude of the measurement step disturbance is 0.1 entering into the control
system after time of 2.5. Also, the measurement noise in the process output is simulated
by adding zero mean random numbers with standard deviation of 0.003. Figure 5.8
depicts the simulation results with different initial conditions of the system state when
facing with measurement disturbance and noise. From this figure, it is evident that the
5.5. Summary
In this chapter, a nonlinear control strategy that combines a model-based SMC and an
adaptive NFC technique has been proposed for the control of an imprecisely known
process. The convergence property of the model-based SMC is guaranteed to satisfy the
sliding condition, which enhances the SMC to provide the great ability in dealing with
some certain model uncertainties. For extra uncertainties and/or unknown process
dynamics that cannot be captured before the SMC design stage, the NFC with on-line
achieve high control performance. The present approach that reaps the benefits of both
the SMC and the NFC is widely applicable for imprecisely known process since it is not
subject to any type of process uncertainties. Furthermore, the smooth transition from
conventional SMC, which thus avoids producing rigorous control action and, in turn,
leads to excellent control performance. Besides, for practical implementation the potential
use of a sliding observer along with the SMC and NFC has been investigated as a whole.
Extensive simulation results reveal that the proposed nonlinear control strategy is
effective and applicable for the control of imprecisely known processes, which provides
robust control without having abrupt control moves and without the need of full state
measurements.
Figure 5.3. System response of the CSTR under the use of a sliding observer with
observer poles of p1 = p 2 = −2 for the case of existing unmodeled side reaction and
measuring error.
Figure 5.4. System response of the CSTR under the use of a sliding observer with
observer poles of p1 = p 2 = −4 for the case of existing unmodeled side reaction and
measuring error.
Figure 5.5. The effects of boundary layer thickness of the switching function on system
performance.
Figure 5.6. Comparison with a PIDSMC control scheme with different tuning constants.
Figure 5.7. System performance comparison of the proposed strategy with the PIDSMC
method in the presence of the unmodeled side reaction, measuring error and extra
Figure 5.8. System performance of the proposed scheme in the face of unmodeled side
CHAPTER 6
CONCLUSIONS AND FUTURE PROSPECTS
This thesis has devoted to the development of sliding mode control schemes for the
regulation control of chemical processes having diversified process dynamics. For the
control of linear or simple nonlinear chemical processes, Chapter 2 has presented a novel
and systematic sliding mode control system design methodology, which integrates an
identified SOPDT model, an optimal sliding surface and a delay-ahead predictor. The
satisfying a sliding condition and the control system performance is examined with some
typical chemical processes. Besides, with the concept of delay equivalent, the proposed
sliding mode control scheme can be utilized directly to the regulation control of a
non-minimum phase process. As a special case the proposed scheme is further extended
to the control of chemical processes whose dynamics are simply described by a FOPDT
model. In addition, the decentralized sliding mode control scheme for multivariable
processes in the presence of simultaneous uncertainties and input-delay. The idea is based
designed a robust sliding mode controller to handle with the process uncertainties.
difficulty in the control system design. To tackle with this difficulty, Chapter 4 develops a
novel sliding mode control scheme by the integration of a statically equivalent output
map (SEOM) and a time-advanced nonlinear predictor; the SEOM is used for eliminating
the undesirable inverse response, while a predictor is adopted for curbing the negative
effect of input-delay. This integration reaps the benefits of both techniques and therefore
facilitates the design of a sliding mode controller. Besides, to guarantee the convergence
properties of the whole SMC control scheme, a Lyapunov-based technique is utilized for
analysis.
For handling with chemical processes having imprecisely known dynamics, Chapter 5
combines the model-based sliding mode controller with an intelligent neural fuzzy
technique to deal with this complex control problem. A previously developed process
model is used to design a nominal SMC loop, while those dynamics that are not captured
in the model is suggested to be handled with the enhancement of an adaptive neural fuzzy
technique. The control system performances of the proposed approaches have been
results reveal that the proposed design methodologies appear to be effective and
promising for the regulation control of chemical processes whose process dynamics are
imprecisely known.
Finally, it should be mentioned that, though this thesis has presented several practical
SMC schemes for chemical processes with diversified dynamic behaviors, it is still
unable to encompass all kind of problems arose from chemical process control. Therefore,
some suggestion and recommendations for future work are summarized as follows. The
mode control system design methodology presented in Chapter 4 gives rise to a nonlinear
model-based controller but the design specifications are local, around a given operation
weights ε j (x) for the SEOM should be used. The use of state-dependent weights will
also allow constructing statically equivalent outputs of higher relative order. Therefore,
how to design the state-dependent weights ε j (x) becomes an essential problem that
the sliding mode control using diversified configurations is worth further exploring since
the process model is not always perfect. Besides, the practical implementation of the
proposed SMC schemes in actual chemical processes would be an interesting issue that
REFERENCES
Behtash, S., “Robust Output Tracking for Nonlinear Systems,” Int. J. Contr., 51,
1381-1407 (1990).
Bristol, E., “On a New Measure of Interactions for Multivariable Process Control,” IEEE
Camacho, O. and R. Rojas, “A General Sliding Mode Controller for Nonlinear Chemical
(2000).
Camacho, O., R. Rojas and W. Garcia, “Variable Structure Control Applied to Chemical
Chen, C. T. and C. S. Dai, “Robust Controller Design for a Class of Nonlinear Uncertain
(2002).
Choi, J. Y., J. Lee, J. H. Jung, M. Lee and C. Han, “Sequential Loop Closing
Perturbed Time-Varying State Delay Systems,” Journal of the Franklin Institute, 338,
35-46 (2001).
of a CSTR: Disturbance Rejection Using Sliding Mode Control,” Ind. Eng. Chem.
Ultimate Boundedness for Uncertain Dynamic Systems,” IEEE Trans. Auto. Contr.,
Da, F. P. and W. Z. Song, “Sliding Mode Adaptive Control Based on Fuzzy Neural
Edwards, C. and S. Spurgeon, Sliding Mode Control, Taylor and Francis, Bristol, PA
(1998).
(1993).
Henson, M. A. and D. E. Seborg, “An Internal Model Control Strategy for Nonlinear
(2003).
Hu, J., J. Chu and H. Su, “SMVSC for a Class of Time-Delay Uncertain Systems with
Isidori, A., Nonlinear Control Systems, Springer-Verlag, New York, U.S.A. (1989).
Isidori, A., Nonlinear Control Systems, 3rd ed., London: Springer (1995).
Kantor, J. C., “Stability of state feedback transformations for nonlinear systems- Some
Li, Z. H., T. Y. Chai and C. Wen, “Systematic Design of Robust Controllers for Nonlinear
Li, M., F. Wang and F. Gao, “PID-Based Sliding Mode Controller for Nonlinear
Li, X. and S. Yurkovich, “Sliding Mode Control of Delayed Systems with Application to
802-810 (2001).
Luyben, W. L., “Feedback and Feedforward Control of Distillation Column with Inverse
Luyben, W. L., “Simple Method for Tuning SISO Controllers in Multivariable Systems,”
(1997).
Control: Part I-Structure and Design Methodology,” IEC Res., 29, 382-388 (1990).
Marlin, T. E., Process Control: Designing Processes and Control Systems for Dynamic
McAvoy, T. J., “Steady-State Decoupling of Distillation,” IEC Fund., 18, 269-273 (1979).
McMillian, G. K., Tuning and Control Loop Performance, Instrument Society of America:
Park, J. H., H. I. Park and I. B. Lee, “Closed-Loop On-Line Process Identification Using
Perng, M. H. and J. S. Ju, “Optimally Decoupled Robust Control of MIMO Plants with
Sliding Mode Control with Delay Compensation,” Automatica, 35, 1861-1865 (1999).
Roh, Y. H. and J. H. Oh, “Sliding Mode Control with Uncertainty Adaptation for
(2000).
Systems with Nonlinear Internal Dynamics via Sliding Mode Control Using Method
Shyu, K. K. and J. J. Yan, “Robust Stability of Uncertain Time-Delay and its Stabilizing
(1997).
Utkin, V. I., Sliding Modes in Control and Optimization, Berlin: Springer (1992).
Utkin, V. I., J. Guldner and J. Shi, Sliding Modes in Electromechanical Systems, London:
Van de Vusse, J. G., “Plug-Flow-Type Reactor versus Tank Reactor,” Chem. Eng. Sci., 19,
994-998 (1964).
Waller, K. and D. Finnerman, “On Using Sum and Differences to Control Distillation,”
Waller, K. V. T. and C. G. Nygardas, “On Inverse Response in Process Control,” Ind. Eng.
Wang, Q. G., B. Huang and X. Guo, “Auto-Tuning of TITO Decoupling Controllers from
Wang, G. B., S. S. Peng and H. P. Huang, “A Sliding Observer for Nonlinear Process
APPENDIX A
PROOF OF THE SLIDING CONDITION
By taking time derivative of the sliding function (2.7) and inserting the control law of
(2.8), we have
1 d
δ 2
= δ ⋅ δ& = − α δ − d ( x ∗ , t ) δ + δ c 2 d ( x ∗ , t )
2 dt
⎛ δ c 2 d ( x ∗ , t ) ⎞⎟
= − α δ − d ( x ∗ , t ) δ ⎜⎜ 1 −
⎝ δ d ( x ∗ , t ) ⎟⎠ (A2)
⎛ δ c 2 d ( x ∗ , t ) ⎞⎟
= − α δ − d ( x ∗ , t ) δ ⎜⎜ 1 −
⎝ δ c 2 d max ( x ∗ , t ) ⎟⎠
≤ −α δ
APPENDIX B
Since the system is assumed to be internally stable, the nominal system of Eq. (3.1)
possesses the property of hyperbolically minimum phase, i.e., the corresponding zero
dynamics are exponentially stable (Behtash, 1990; Li et al., 1995). Besides, for internal
where L is called a Lipschitz constant of qˆ (ξˆ , ηˆ ) . Under the condition of Eq. (B1) and
using a converse theorem of Lyapunov (Hahn, 1967), there exists a Lyapunov function
2 2
σ 1 ηˆ ≤ V0 ( ηˆ ) ≤ σ 2 ηˆ (B2)
∂V0 2
qˆ (0, ηˆ ) ≤ −λ1 ηˆ (B3)
∂ηˆ T
∂V0
≤ λ 2 ηˆ (B4)
∂ηˆ
φˆ (ξˆ , ηˆ ) ≤ l1 ( ξˆ + ηˆ ) + l 2 (B5)
for all (ξˆ , ηˆ ) ∈ Tˆ (U ) , where l1 and l2 are positive constants. Now, let’s consider the
Lyapunov candidate as
~ ~
where V1 (ξˆ ) = 1 2 δˆ 2 + γ 2k 2 and k ≡ kˆ − k ∗ with k ∗ being the desired steady state
noted that the first part in the right hand side of V1 (ξˆ ) corresponds to the classic sliding
condition and the second part is introduced for constructing a tuning rule for the feedback
gain. Furthermore, the term V0 ( ηˆ ) in Eq. (B6) is added for ensuring the stability of the
2 1 ˆ2
≤ −µ1λmin (Q) ξˆ − (e1 − e22 ) ηˆ +
2
ξ + µ 0 λ2 l 2 ηˆ
4
1 2
= −[µ1λmin (Q) − ] ξˆ − (e1 − e22 ) ηˆ + µ 0 λ2 l 2 ηˆ
2
4
(B7)
& bˆ
where the gain tuning rule kˆ = s δˆ 2 = γ~δˆ 2 with γ~ > 0 and a non-negative value of
γ
k̂ (0) has been applied in the above derivation. In Eq. (B7), Q ≡ −cc T A c is a positive
1 2
e2 ηˆ ξˆ ≤ ξˆ + e22 ηˆ . Since l1 < λ1 λ2 , we can appropriately chose µ 0 and µ1
2
such that all the square terms in Eq. (B7) are negative. Finally, by letting η = [ξˆ T , ηˆ T ]T ,
2
V& ( η) ≤ −c1 η + c 2 η (B8)
1
where c1 = min{µ1λmin (Q) − , e1 − e22 } and c2 = µ 0 λ2 l2 . The time derivative V& is
4
strictly negative for a sufficient large η . Noticing that from Eqs. (B2)-(B4), and (B6)
and following a similar procedure of Li et al. (1995), the closed-loop system properties
(P3.1), (P3.2) and (P3.3) can follow upon using standard arguments in the literature
⎧ 0, if r ≤ ( r2 r1 )1 2 d
⎪
t (d , r ) = ⎨ r2 r 2 − r12 r2−1d 2 (B9)
, if otherwise
⎪ c r r −1d 2 − c r 1 2 r −1 2 d
⎩ 112 2 2 1 2
⎧( r r )1 2 R, if r ≤ R
d (r) = ⎨ 2 1 1 2 (B10)
⎩ ( r2 r1 ) r, if r > R
d = ( r2 r1 )1 2 R (B11)
and
ϑ (d ) = R (B12)
λmax (Q) denotes the maximum eigenvalue of Q . This completes the proof.
APPENDIX C
STABILITY ANALYSIS OF USING THE SMC
Then, by inserting the control law (5.18) and using the above equation, we have
⎧ ρ −1 ⎫
δ ⋅ δ& = δ ⋅ ⎨∑ ci ξ i +1 + [b(ξ, η) + ∆b(ξ, η)] + [a(ξ, η) + ∆a(ξ, η)]uSMC ⎬
⎩ i =1 ⎭
ρ −1
⎧ 1 ⎫
⎪∑ ci ξ i +1 + [b(ξ, η) + ∆b(ξ, η)] + a(ξ, η) ⋅ a (ξ, η) ⎪
⎪ i =1 ⎪
= δ ⋅ ⎨ ρ -1 ⎬
⎪⎡(− c ξ − b(ξ, η)) + (−( B(ξ, η) + A(ξ, η) + γ )sign (δ ))⎤ + ∆a(ξ, η)u ⎪
⎪⎩⎢⎣ ∑ i =1
i i +1 ⎥
⎦
SMC
⎪⎭
= δ ⋅ [∆b(ξ, η) − ( B(ξ, η) + A(ξ, η) + γ )sign (δ ) + ∆a(ξ, η)uSMC ]
≤ δ ⋅ [B(ξ, η) + A(ξ, η) − ( B(ξ, η) + A(ξ, η) + γ )]
= −γ δ < 0
(C3)
Obviously, with the proposed sliding mode control scheme the sliding condition is
satisfied.
APPENDIX D
DESIGN OF OBSERVER GAINS FOR THE NONLINEAR CSTR
Following the procedure of Wang et al. (1997), the switching gains of the observer
Step 2: Calculate
⎡ ∂f1 ∂f1 ⎤
⎢ ∂x ∂x2 ⎥ ⎡ − 4 0.8796 ⎤
Ar ≡ ⎢ 1 ⎥ =⎢ ⎥
⎢ ∂f 2 ∂f 2 ⎥ ⎣ 3 − 3.6388⎦
⎢⎣ ∂x1 ∂x2 ⎥⎦ x =0
⎡ ∂f ∂f 3 ⎤
Cr ≡ ⎢ 3 = [0 1.7592]
⎣ ∂x1 ∂x2 ⎥⎦ x =0
⎡0.6895 0.5684⎤
[
P1 ≡ CTr A Tr CTr ]
−1
=⎢
0 ⎥⎦
⎣0.1895
and
⎡1 0⎤ ⎡ 1 0⎤
R r ≡ ⎢~ ⎥ =⎢
⎣a 2 1⎦ ⎣7.6388 1⎥⎦
Step 3: Let the observer poles be p1 and p2 . Then, calculate α~2 and α~3 from
Step 4: Denote a r ≡ [a~2 a~3 ] and b r ≡ [α~2 α~3 ] , then the remaining gains, k̂ 01
⎡ kˆ01 ⎤ ˆ T −1
⎢ ˆ ⎥ = k 03 P1 R r (b r − a r )
T
⎣⎢k 02 ⎦⎥