Professional Documents
Culture Documents
Adaptive Control
Adaptive Control
TRANSACTIONS®
ISA Transactions 44 共2005兲 55–68
Abstract
Adaptive nonlinear control is investigated for continuously stirred tank reactor 共CSTR兲 systems using neural net-
works. The CSTR plant under study belongs to a class of nonaffine nonlinear systems, and contains an unknown
parameter that enters the model nonlinearly. Using adaptive backstepping and neural network 共NN兲 approximation
techniques, an alternative adaptive NN controller is developed that achieves asymptotic output tracking control. A
novel integral-type Lyapunov function, which includes both system states and control input as its arguments, is
constructed to solve the difficulty associated with the nonaffine control problem. Numerical simulation is performed to
show the feasibility of the proposed approach for chemical process control. © 2005 ISA—The Instrumentation,
Systems, and Automation Society.
Keywords: Nonlinear systems; Adaptive control; Backstepping design; Lyapunov stability; Neural networks
0019-0578/2005/$ - see front matter © 2005 ISA—The Instrumentation, Systems, and Automation Society.
56 Zhang, Guay / ISA Transactions 44 (2005) 55–68
Table 1
Parameters of the CSTR system.
Parameter Description Nominal value
q process flowrate 100 l/min
C a0 concentration of component A 1 mol/l
Tf feed temperature 350 K
Tcf inlet coolant temperature 350 K
V volume of tank 100 l
ha heat transfer coefficient 7⫻105 J/min K
a0 preexponential factor 7.2⫻1010 min⫺1
E/R activation energy 1⫻104 K
(⫺⌬H) heat of reaction 2⫻104 cal/mol
1 ,c liquid densities 1⫻103 g/l
C p ,C pc heat capacities 1 cal/g K
共⫺⌬H兲a0 ha c C pc
a1⫽ ⫽1.44⫻1013 a 2⫽ ⫽6.987⫻102 a 3⫽ ⫽0.01
1Cp c C pc 1C pV
4
cult to estimate in practice. The approach pro- ẋ 1 ⫽1⫺x 1 ⫺a 0 x 1 e ⫺ 10 /x 2
,
posed circumvents this major limitation while
4
addressing the major technical difficulties associ- ẋ 2 ⫽T f ⫺x 2 ⫹a 1 x 1 e ⫺ 10 /x 2
⫹a 3 u 共 1⫺e ⫺ a 2 /u 兲
ated with this control problem.
In this work, we investigate the application of an ⫻ 共 T c f ⫺x 2 兲 ,
alternative adaptive control method to the CSTR y⫽x 1 , 共2兲
plant 共1兲. The method combines Lyapunov’s sta-
bility design, adaptive backstepping, and neural where the unknown constant parameters are a 0 ,
network approximation to derive the control algo- a 1 , a 2 , and a 3 . The control objective is to design
rithm and adaptive law. The paper is organized as a controller u such that the output y follows a
follows. Section 2 presents some notations and desired signal y d .
background. In Section 3, an adaptive backstep- There are two major difficulties associated with
ping design is performed. Both stability and con- this control problem. First, the unknown param-
trol performance of the closed-loop systems are eter a 2 appears nonlinearly. As a result, it cannot
addressed as well. Numerical simulation results be estimated easily using conventional identifica-
tion technique. Second, the control input appears
are given in Section 4 to show the effectiveness of
nonlinearly. Although it does not constitute a ma-
the proposed approach.
jor drawback for controller design, it requires a
more careful analysis than standard feedback lin-
earization and backstepping-based designs. In this
paper, an alternative design scheme is presented to
solve the adaptive control problem for processes
2. Problem formulation
with similar structural difficulties.
We first state some assumptions that are re-
Let 储•储 denote the 2-norm, i.e., 储 A 储 ⫽ 冑兺 i⫽1 m
a i2 quired for the development detailed in the next
with A⫽ 关 a 1 ,a 2 ,...,a m 兴 苸R . Let 兩 • 兩 1 denote
T m
section.
1-norm, i.e., 兩 A 兩 1 ⫽ 兺 i⫽1 m
兩 a i 兩 . It can be seen that Assumption 1: a 0 ⭐ā 0 , a 2 ⭓aគ 2 and a 3 ⭓aគ 3 with
兩 A 兩 1 ⭐m 储 A 储 .
2 2
known constants ā 0 , aគ 2 , and ā 3 .
In this paper, we assume that plant parameters Assumption 2: The desired signal vector x d
q, C a0 , T f , T c f , and V are at the nominal values ⫽ 关 y d ẏ d ÿ d y (3)
d 兴 苸⍀ d with known compact set
T
where the positive constants c 0 and T c0 are the tems to overcome the restrictions of matching con-
lower bounds of the concentration x 1 and tempera- ditions or growth conditions in nonlinear adaptive
ture x 2 , respectively. The constants T max and u 0 control. This scheme has been further developed
are the maximum values of the tank temperature in recent literature 关19–24兴 for a broad class of
and the coolant flow rate, respectively. nonlinear systems. From the first equation of plant
4
3. Adaptive backstepping design 共2兲, we may consider the function x 1 e ⫺ 10 /x 2 as a
virtual control for x 1 . Then, by controlling x 2
Adaptive backstepping techniques were first de- through input u, we can drive the nonlinearity
4
veloped in 关18兴 for parametric strict-feedback sys- x 1 e ⫺ 10 /x 2 to a desired one which achieves the
58 Zhang, Guay / ISA Transactions 44 (2005) 55–68
冋 册
f 2
x2
z 1 ż 1 1
V̇ 1 ⫽ ⫽z 1 共 1⫺x 1 ⫺ẏ d 兲 ⫺ 共 z 2 ⫹ ␣ 1 兲 . 4
a0 a0 ⫹a 1 x 1 e ⫺ 10 /x 2
兲 ⫺k 1 ż 1 ⫺ ˆ˙ 1 共 1⫺x 1 ⫺ẏ d 兲
Hence, choosing ␣ 1 as
⫹ ˆ 1 共 ẋ 1 ⫹ÿ d 兲 ,
␣ 1 ⫽k 1 z 1 ⫹ ˆ 1 共 1⫺x 1 ⫺ẏ d 兲 , 共7兲 104 x 1 ⫺ 104 /x
g 2共 x 兲 ⫽ e 2 共 x ⫺T 兲 ,
2 cf
where k 1 is a positive constant and ˆ 1 is the esti- x 22
mate of the unknown parameter 1* ⫺1/a 0 , yields
f 共 u 兲 ⫽u 共 1⫺e ⫺ a 2 /u 兲 .
V̇ 1 ⫽⫺k 1 z 21 ⫺ 共 ˆ 1 ⫺ *
1 兲共 1⫺x 1 ⫺ẏ d 兲 z 1 ⫺z 1 z 2 . Since
Now let f共u兲 a2
⫽1⫺e ⫺ a 2 /u ⫺ e ⫺ a 2 /u ,
˜ 21 u u
V s1 ⫽V 1 ⫹ ,
2␥1 2 f 共u兲 a 22
⫽⫺ 3 e ⫺ a 2 /u , 共11兲
u2 u
where ˜ 1 ⫽ ˆ 1 ⫺ 1* and constant ␥ 1 ⬎0. Differen-
tiating with respect to time, that is, it follows that, by considering Assumption 1 and
the operating region 共3兲, we have
˜ 1 8 1
V̇ s1 ⫽⫺k 1 z 21 ⫺˜ 1 共 1⫺x 1 ⫺ẏ d 兲 z 1 ⫺z 1 z 2 ⫹ , f共u兲 aគ 2
␥1 ⭓b 0 ⫽1⫺e ⫺ aគ 2 /u 0 ⫺ e ⫺ aគ 2 /u 0 ⬎0,
u u0
leads to the adaptive law 共12兲
⫽⫺a 3 g 2 共 x 兲关 u⫺  共 兲兴 冕0
1 f 共u兲
u
d,
equation 共17兲 can be written as
ż 2 ⫽⫺a 3 b u g 2 共 x 兲关 u⫺W * T S 共 兲 ⫺ l 兴 . 共20兲
共15兲 Consider the following control structure
where u ⫽u⫹ ( 1⫺ )  ( ) .
u⫽k s 共 z 2 兲 z 2 ⫹Ŵ T S 共 兲 , 共21兲
Defining
冐冕冐
共noting ⫽ 关 ¯ T z 2 兴 T ) . Choose a positive-definite
function w0 z2 b ⫺1
u
⫹ ˙ 储 d
储¯
a3 0 u
V 2 ⫽V s1 ⫹
1
a3
冕 0
z2
b ⫺1
u 共 , ,Ŵ 兲 d .
¯ 共23兲 ⫹
1
a3
冕冐 冐 z2
0
b ⫺1
u
u ˙ T S 共 ¯ , 兲 兩 d .
兩 Ŵ
共27兲
Its time derivative along Eq. 共9兲 is Using Eqs. 共13兲 and 共16兲, we know that
␦
V̇ 2 ⫽⫺k 1 z 21 ⫺z 1 z 2 ⫺ ˜ 1 ˆ 1 ⫹
z2
ż
冏 冏 冏冕
bu
u
⫽
1
0
2 f 共u兲
u 2
d ⭐ 冏 冕 冏 f u共 u 兲 冏d
0
1 2
2
␥1 a 3b u 2
⭐ 冕 b d⫽b .
冋 册
1
冕 b ⫺1 T
1 z2 u u 共 ¯ , ,Ŵ 兲 ˙ 1 1
⫹ ¯ d 0
a3 0 u ¯ 共28兲
⫹
1
冕 z2 b ⫺1
u
u
Ŵ 冋
˙ T u 共 , ,Ŵ 兲 d ,
¯
Ŵ
册 Equations 共12兲 and 共16兲 show that b u ⭓b 0 ⬎0.
Hence
冏 冏冏
a3
冏 冏 冏冏 冏
0
b ⫺1
u 1 bu 1 bu b1
共24兲 ⫽ 2 ⭐ 2 ⭐ 2 . 共29兲
u bu u bu u b0
Using Eqs. 共27兲–共29兲 and error equation 共22兲, in-
˙ is to be assigned in
where the rate of adaptation Ŵ equality 共27兲 can be reexpressed as
the following. Noting controller 共21兲, we have
␦
V̇ 2 ⭐⫺k 1 z 21 ⫺z 1 z 2 ⫺ ˜ ˆ ⫺z g 共 x 兲关 k s 共 z 2 兲 z 2
␥1 1 1 2 2
u 共 ¯ , ,Ŵ 兲
Ŵ
⫽S 共 ¯ , 兲 ,
⫹W̃S 共 兲 ⫺ l 兴 ⫹
w 0b 1
a 3 b 20
冕 z2
0
˙ 储 d
储¯
u 共 ¯ , ,Ŵ 兲
⫽Ŵ Ta
S 共 ¯ , 兲
, 共25兲
⫹
b1
a 3 b 20
冕 0
z2
˙ T S 共 ¯ , 兲 兩 d .
兩 Ŵ
¯ ¯
It is observed that 兩 z 1 z 2 兩 ⭐ ( k 1 z 21 /2) ⫹ ( z 22 /2k 1 )
and a 3 ⭓aគ 3 共by Assumption 1兲. As a result, we
where Ŵ a is the corresponding weight vector for have
input ¯ . It can be seen from Eq. 共19兲 that the first k1 2 ␦
derivative of s ( j ) with respect with i is V̇ 2 ⭐⫺ z ⫺ ˜ ˆ ⫺W̃ T S 共 兲 g 2 共 x 兲 z 2 ⫺g 2 共 x 兲
2 1 ␥1 1 1
d 关 s ( i ) 兴 /d i ⫽ 4/关 ( e i ⫹e ⫺ i ) 2 兴 . Since 4/关 ( e i
⫹e ⫺ i ) 2 兴 ⭐1,᭙ i 苸R, we have
⫻关 k s 共 z 2 兲 z 22 ⫺ l z 2 兴 ⫹z 22 冋 1
2k 1
⫹
w 0b 1C
2aគ 3 b 20 册
冐 u 共 ¯ , ,Ŵ 兲
¯
冐
⭐w 0 ⫽ sup 兵 兩 Ŵ a 兩 1 其 . 共26兲
ˆ 储 W 储 ⭐w m
⫹
b1
aគ 3 b 20
冕 z2
0
˙ T S 共 ¯ , 兲 兩 d ,
兩 Ŵ 共30兲
The following theorem provides the NN learn- Theorem 3.1. Consider the closed-loop adap-
ing law and design function k s ( z ) such that the tive system consisting of plant (2), controller (21),
stability and tracking performance are guaranteed and adaptive law (8). If (i) the following projec-
for the closed-loop system. tion NN learning algorithm is used:
␥ 2S 共 兲 g 2共 x 兲 z 2 ,
冦
if 储 Ŵ 储 ⬍w m or
˙⫽ 储 Ŵ 储 ⫽w m and Ŵ T S 共 兲 g 2 共 x 兲 z 2 ⭐0
Ŵ 共31兲
ŴŴ T S 共 兲 g 2 共 x 兲 z 2
␥ 2S 共 兲 g 2共 x 兲 z 2⫺ ␥ 2 , if 储 Ŵ 储 ⫽w m and Ŵ T S 共 兲 g 2 共 x 兲 z 2 ⬎0
储 Ŵ 储 2
with constant ␥ 2 ⬎0 and initial condition 储 Ŵ ( 0 ) 储 ⭐w m , and (ii) design function k s ( z 2 ) is chosen as
k s 共 z 2 兲 ⫽k 2 ⫹ 冉
1 1
gគ 2 2k 1
⫹
w 0b 1C
2 ⫹
2aគ 3 b 0 冊
b 1␥ 2l
兩z 兩
aគ 3 b 20 2
共32兲
with constant k 2 ⬎0, then (i) all of the signals in the closed-loop system are bounded, and (ii) the mean
square tracking error converges to a small neighborhood of zero whose size can be adjusted by the design
parameters ␦ , ␥ 1 ,k 1 ,k 2 , and NN node number l.
Proof: (i) Consider a Lyapunov function candidate
储 W̃ 储 2
V s ⫽V 2 ⫹ .
2␥2
Its time derivative along Eq. 共30兲 is
V̇ s ⭐⫺
k1 2 ␦
2
z 1 ⫺ ˜ 1 ˆ 1 ⫺W̃ T S 共 兲 g 2 共 x 兲 z 2 ⫺g 2 共 x 兲关 k s 共 z 2 兲 z 22 ⫺ l z 2 兴 ⫹z 22
␥1
1
2k 1
⫹
w 0b 1C
冋
2aគ 3 b 20 册
⫹
b1
aគ 3 b 20
冕 0
z2
˙ T S 共 ¯ , 兲 兩 d ⫹ 1 W̃ T W̃
兩 Ŵ
␥2
˙. 共33兲
The NN weight tuning algorithm 共31兲 is a standard projection learning law, which guarantees that
储 Ŵ ( t ) 储 ⭐w m ,᭙t⬎0 for 储 Ŵ ( 0 ) 储 ⭐w m 共see 关25兴 for the proof兲. Following arguments similar to those used
for the application of the projection algorithm in 关25兴, we obtain
V̇ s ⭐⫺
k1 2 ␦
2
z 1 ⫺ ˜ 1 ˆ 1 ⫺g 2 共 x 兲关 k s 共 z 2 兲 z 22 ⫺ l z 2 兴 ⫹z 22
␥1
1
2k 1
⫹ 冋
w 0b 1C
2 ⫹
2aគ 3 b 0
b1
aគ 3 b 20 册 冕
0
z2
˙ T S 共 ¯ , 兲 兩 d .
兩 Ŵ
共34兲
Considering Eq. 共32兲, ḡ 2 ⭓g 2 ( x ) ⭓gគ 2 , and 兩 l z 2 兩 ⭐ ( k 2 /2) z 22 ⫹ ( l2 /2k 2 ) , we further have
V̇ s ⭐⫺
k1 2 ␦
2 ␥1 冋k2
z 1 ⫺ ˜ 1 ˆ 1 ⫺g 2 共 x 兲 k s 共 z 2 兲 z 22 ⫺ z 22 ⫺
2
2l
2k 2
⫹z 22
1
2k 1
⫹
w 0b 1C
2aគ 3 b 20
册 冋 册
⫹
b1
aគ 3 b 20
冕0
z2
˙ T ¯
k1 2 ␦
兩 Ŵ S 共 , 兲 兩 d ⭐⫺ z 1 ⫺ 1 1 ⫺g 2 共 x 兲
2 ␥1
˜ ˆ
k 2 2 2l
z ⫺
2 2 2k 2
⫺
b1
冉
aគ 3 b 20 2 2
冊
␥ lg 共 x 兲 兩 z 2 兩 3 冋
⫺ 冕 0
z2
˙ T S 共 ¯ , 兲 兩 d .
兩 Ŵ 册 共35兲
62 Zhang, Guay / ISA Transactions 44 (2005) 55–68
冕 0
z2
˙ T S 共 ¯ , 兲 兩 d ⭐2 ␥ lg 共 x 兲 兩 z 兩
兩 Ŵ 2 2 2 冕
0
z2
d ⫽ ␥ 2 lg 2 共 x 兲 兩 z 2 兩 3 .
Therefore,
V̇ s ⭐⫺
k1 2 ␦
2
z 1 ⫺ ˜ 1 ˆ 1 ⫺g 2 共 x 兲
␥1
k 2 2 2l
2
z 2⫺
2k 2
k1
⫽⫺ z 21 ⫺
2
冉
␦
␥1
ˆ 1 ⫺
*
2
1
冊 冉 冊 2
⫹
␦ *1 2
4␥1
冉
⫺g 2 共 x 兲 z ⫺
2 2 2k 2
冊
k 2 2 2l
.
共36兲
Now define a compact set
再
⍀⫽ 共 z 1 ,z 2 , ˆ 1 兲 冏 k1 2 k2
2
z 1 ⫹ គg 2 z 22 ⫹
2
␦
␥1
ˆ 1 ⫺冉*
2
1
冊 2
⭐
␦ *1 2 ḡ 2 2l
4␥1
⫹
2k 2
. 冎 共37兲
We can see that V̇ s ⬍0 as long as z 1 , z 2 , and ˆ 1 Since the NN approximation error l can be re-
remain outside the compact set ⍀. It follows that, duced as the NN node number l increasing, the
applying a standard Lyapunov argument, z 1 , z 2 , tracking performance may be improved if a large
and ˆ 1 are bounded. As the projection algorithm size NN is applied. Inequality 共38兲 implies that the
output tracking error converges to a neighborhood
共31兲 guarantees 储 Ŵ 储 ⭐w m , we conclude that all
of the origin whose size can be reduced by in-
the signals in the closed⫺loop control system are
creasing the controller gains k 1 and k 2 , or the
bounded.
node number of neural networks, or by using a fast
共ii兲 It follows from Eq. 共36兲 and g 2 ( x ) ⭐ḡ 2 that
learning rate ␥ 1 . This completes the proof. Q.E.D.
k 1 2 ␦ 1* 2 ḡ 2 2l
V̇ s ⭐⫺ z 1 ⫹ ⫹ .
2 4␥1 2k 2
k1
2
冕 0
t
z 21 dt⭐V s 共 0 兲 ⫺V s 共 t 兲 ⫹
␦ *1 2
4␥1
t
Table 1 shows the plant parameters used in the
simulation. The initial condition of the CSTR sys-
tem is x ( 0 ) ⫽ 关 0.1,440.0兴 T and the control objec-
⫹
ḡ 2
2k 2
冕 dt.
0
t
2
l
tive is to make the concentration y track the set-
point step change signal r ( t ) . In order to get a
smooth reference signal, a linear reference model
Hence, the mean square tracking error satisfies is used to shape the discontinuous reference signal
r ( t ) for providing the desired signals y d , ẏ d , and
1
t
冕 0
t
z 21 dt⭐
2
V 共 0 兲⫹
k 1t s
␦ *1 2
2 ␥ 1k 1
ÿ d . The following reference model is to be imple-
mented:
y d共 s 兲 w 2n
⫹
ḡ 2
k 1k 2t
冕 dt, t⬎0
t
0
2
l
⫽
r 共 s 兲 s 2 ⫹2 n w n s⫹w 2n
,
Fig. 1. Reference signal y d 共‘‘—’’兲, PI control 共‘‘- -’’兲, and adaptive NN control 共‘‘¯’’兲.
gain proportional plus integral 共PI兲 control law The simulation results given in Fig. 1 indicate
commonly used in chemical process control given that the output response of the adaptive controller
by displays some oscillation during the initial 5 min.
冕 共 y ⫺y 兲d
This behavior is due to the lack of information that
kc t
u pi ⫽k c 共 y d ⫺y 兲 ⫹ d 共39兲 the initial neural networks has about the CSTR
Ti 0 plant dynamics 共note that the initial NN weights
with k c ⫽440, T i ⫽0.8, and u pi ( 0 ) ⫽100.0 关16兴 as were chosen as zero values兲. After several NN
an alternative to the proposed scheme. The param- learning periods, it can be seen from Fig. 1 that the
eters of the PI controller 共39兲 are selected to give tracking error of the adaptive controller decreases
an adequate response for step changes r ( t ) in the significantly. Fig. 2 confirms the boundedness of
set-points of ⫾0.02 mol/l about the nominal prod- the NN weight estimates during the adaptation.
uct concentration of 0.1 mol/l. The numerical experiment shows that better track-
Second-order neural networks were selected for ing performance is achieved with the proposed
the controller 共21兲. The element s i ( ) in Eq. 共19兲 adaptive NN controller.
are chosen as s ( 1 ) , s ( 2 ) ,...,s ( 6 ) and all pos- The value of the manipulated variable required
sible combinations of them 关i.e., s ( i ) s ( j ) with in the NN controller and the PI controller is shown
i, j⫽1,2,3,4,5,6] . The total number of NN nodes in Figs. 3 and 4, respectively. The NN controller is
is 30. The parameters in the adaptive law 共31兲 are considerably more aggressive, but the self-tuning
taken as w m ⫽30.0, ␥ 2 ⫽5.0, and the initial NN nature of the NN controller provides an appropri-
weight Ŵ ( 0 ) ⫽1.0⫻10⫺4 . The other parameters ate control input strategy after a few set-point
of the controller 共21兲 are k 1 ⫽1, k 2 ⫽100, ␥ 1 ⫽1, changes. Note that the PI controller could not be
and ⫽0.1. tuned more aggressively. Any increase in the con-
Zhang, Guay / ISA Transactions 44 (2005) 55–68 65
Fig. 5. Closed-loop response with the NN controller with undesirable controller tuning.
66 Zhang, Guay / ISA Transactions 44 (2005) 55–68
Fig. 6. Manipulated variable for the NN controller with undesirable controller tuning.
troller gain yielded sustained oscillations at the of neural networks remains an active area of in-
high set-point value. The reported performance re- vestigation that warrants further research that is
flects the best possible tuning of the PI controller outside the scope of this study. In most applica-
for this highly nonlinear process. tions, careful tuning of the NN controller remains
In practical applications, high oscillations of the the best strategy to ensure a proper response of the
output are possible. These oscillations are related closed-loop system 关30兴.
to the lack of knowledge about the system nonlin-
earities initially. However, they can also be in-
duced by the specific choice of tuning parameters.
As an example, Fig. 5 shows the tracking response
and the control action of the closed-loop system 5. Conclusion
when the value of ␥ 2 is increased to 10 and the
controller gain k 2 is reduced to 10. The corre- In this paper, we have presented an adaptive NN
sponding manipulated variable is shown in Fig. 6. control scheme for a class of general nonlinear
A very undesirable behavior results. This issue can chemical process control systems. Lyapunov’s sta-
also be resolved by careful a priori training of the bility techniques and an adaptive backstepping de-
neural networks off-line. However, there is no sign have been used to develop the control struc-
guarantee in practice that data with sufficient in- ture and adaptive learning law. It has been proven
formation would be available initially. Therefore, that the proposed controller guarantees the stabil-
the simulation study reflects a typical situation ity and asymptotic tracking performance of the
where only limited knowledge about the system closed-loop adaptive systems. The effectiveness of
nonlinearity is known before a controller is imple- the proposed control method has been illustrated
mented. The issue of iterative and off-line training through a numerical simulation study.
Zhang, Guay / ISA Transactions 44 (2005) 55–68 67
¯˙ 储
Appendix A: Upper Bound of 储 it follows from Eq. 共8兲 that
with
4
f ( x,x d , ˆ 1 ) ⫽1⫺x 1 ⫹ā 0 x 1 e ⫺ 10 /x 2 ⫹ 兩 ẏ d 兩 兩 ˆ 1 共 t 兲 兩 ⭐ 兩 ˆ 1 共 0 兲 兩 e ⫺ t ⫹ 冕e 0
t
⫺ (t⫺ )
␥ 1C 0d
d 兩 ⫹ 兩 ␥ 1 ( 1⫺x 1 ⫺ẏ d ) z 1 ⫺ 1 兩 .
⫹ 兩 ÿ d 兩 ⫹ 兩 y (3) ˆ
We first find the upper bound of 兩 ˆ 1 兩 . Defining ␥ 1C 0
⭐ 兩 ˆ 1 共 0 兲 兩 ⫹ .
C 0⫽ sup 兵 兩 共 1⫺x 1 ⫺ẏ d 兲 z 1 兩 其 共A1兲
x苸⍀ x ,x d 苸⍀ d Defining a compact set
再 冏
⍀ ⫽ 共 x,x d , ˆ 1 兲 x苸⍀ x , x d 苸⍀ d , 兩 ˆ 1 兩 ⭐ 兩 ˆ 1 共 0 兲 兩 ⫹
␥ 1C 0
冎 共A2兲
˙ TS(
Appendix B: Proof of 円Ŵ ¯ ,)円Ï2␥2lg2(x)円z2円
Case 1: 储 Ŵ 储 ⬍w m or 储 Ŵ 储 ⫽w m and Ŵ T S ( ) g 2 ( x ) z 2 ⭐0. It follows from the adaptive law 共31兲 that
˙ T S 共 ¯ , 兲 兩 ⫽ ␥ g 共 x 兲 兩 z 兩 S T 共 兲 S 共 ¯ , 兲 .
兩 Ŵ 2 2 2
We know that S T ( ) S ( ¯
, ) ⭐l, with l being the node number of neural networks because every element
of S ( ) is positive and less than 1. Hence 兩 Ŵ ˙ T S ( ¯ , ) 兩 ⭐ ␥ lg ( x ) 兩 z 兩 .
2 2 2
Case 2: 储 Ŵ 储 ⫽w m and Ŵ T S ( ) g 2 ( x ) z 2 ⬎0. In this case,
兩 Ŵ 2 2 冏冋
˙ T S 共 ¯ , 兲 兩 ⫽ ␥ g 共 x 兲 S T 共 兲 S 共 ¯ , 兲 ⫺ S 共 兲 ŴŴ S 共 , 兲 z
T T
储 Ŵ 储 2
¯
2 册冏
冋
⭐ ␥ 2 g 2 共 x 兲 兩 z 2 兩 S T 共 兲 S 共 ¯ , 兲 ⫹ 冏 S T 共 兲 ŴŴ T S 共 ¯ , 兲
储 Ŵ 储 2
冏册 .
Since S T ( ) Ŵ and Ŵ T S ( ¯
, ) are scalar functions, bounded by 兩 Ŵ 兩 1 , we obtain
˙ T S 共 ¯ , 兲 兩 ⭐ ␥ g 共 x 兲 兩 z 兩 l⫹
兩 Ŵ 2 2 2 冉 兩 Ŵ 兩 21
储 Ŵ 储 2
冊 .
˙ T S ( ¯ , ) 兩 ⭐2 ␥ lg ( x ) 兩 z 兩 .
By the relationship 兩 Ŵ 兩 21 ⭐l 储 Ŵ 储 2 , it is concluded that 兩 Ŵ 2 2 2
68 Zhang, Guay / ISA Transactions 44 (2005) 55–68
References 关17兴 Ge, S. S., Hang, C. C., and Zhang, T., Nonlinear adap-
tive control using neural networks and its application
关1兴 Henson, M. A. and Seborg, D. E., Nonlinear Process to CSTR systems. J. Process Control 9, 313–323
Control. Prentice-Hall, Upper Saddle River, NJ, 1997. 共1999兲.
关2兴 Henson, M. A. and Seborg, D. E., Input-output linear- 关18兴 Kanellakopoulos, I., Kokotovic, P. V., and Morse, A.
ization of general nonlinear processes. AIChE J. 36, S., Systematic design of adaptive controller for feed-
1753–1757 共1990兲. back linearizable systems. IEEE Trans. Autom. Con-
关3兴 Lee, P. L. and Sullivan, G. R., Generic model control trol 36, 1241–1253 共1991兲.
共GMC兲. Comput. Chem. Eng. 12, 573 共1988兲.
关19兴 Krstic, M., Kanellakopoulos, I., and Kokotovic, P.,
关4兴 Kravaris, C. and Chung, C. B., Nonlinear state feed-
Adaptive nonlinear control without overparameteriza-
back synthesis by global input-output linearization.
AIChE J. 33, 592– 603 共1987兲. tion. Syst. Control Lett. 19, 177–185 共1992兲.
关5兴 Bastin, G. and Dochain, D., On-line Estimation and 关20兴 Krstic, M., Kanellakopoulos, I., and Kokotovic, P. V.,
Adaptive Control of Bioreactors. Elsevier, Amster- Nonlinear and Adaptive Control Design. Wiley and
dam, 1990. Sons, New York, 1995.
关6兴 Bastin, G., Adaptive non-linear control of a fed-bath 关21兴 Seto, D., Annaswamy, A. M., and Baillieul, J., Adap-
stirred tank reactors. Int. J. Adapt. Control Signal Pro- tive control of nonlinear systems with a triangular
cess. 6, 273–284 共1992兲. structure. IEEE Trans. Autom. Control 39, 1411–1428
关7兴 Rani, K. Y. and Gangiah, K., Adaptive generic model 共1994兲.
control. AIChE J. 37, 1634 –1644 共1991兲. 关22兴 Yao, B. and Tomizuka, M., Adaptive robust control of
关8兴 Morningred, J. D., Paden, B. E., Seborg, D. E., and SISO nonlinear systems in a semi-strict feedback
Mellichamp, D. A., An adaptive nonlinear predictive form. Automatica 33, 893–900 共1997兲.
controller. Proc. Am. Control Conf. 2, 1614 –1619 关23兴 Polycarpou, M. M., Stable adaptive neural control
共1990兲. scheme for nonlinear systems. IEEE Trans. Autom.
关9兴 Henson, M. A. and Seborg, D. E., Adaptive nonlinear Control 41, 447– 451 共1996兲.
control of a pH neutralization process. IEEE Trans. 关24兴 Zhang, T., Ge, S. S., and Hang, C. C., Adaptive neural
Control Syst. Technol. 2, 169–182 共1994兲. network control for strict-feedback nonlinear systems
关10兴 Isidori, A., Nonlinear Control System, 2nd ed. using backstepping design. Automatica 36, 1835–
Springer-Verlag, Berlin, 1989. 1846 共2000兲.
关11兴 Sastry, S. S. and Isidori, A., Adaptive control of lin- 关25兴 Ioannou, P. A. and Sun, J., Robust Adaptive Control.
earizable systems. IEEE Trans. Autom. Control 34, Prentice-Hall, Englewood Cliffs, NJ, 1995.
1123–1131 共1989兲. 关26兴 Khalil, H. K., Nonlinear Systems. Macmillan, New
关12兴 Teel, A., Kadiyala, R., Kokotovic, P. V., and Sastry, S. York, 1992.
S., Indirect techniques for adaptive input output linear- 关27兴 Sanner, R. M. and Slotine, J. E., Gaussian networks
ization of nonlinear systems. Int. J. Control 53, 193–
for direct adaptive control. IEEE Trans. Neural Netw.
222 共1991兲.
3, 837– 863 共1992兲.
关13兴 Marino, R. and Tomei, P., Nonlinear Adaptive Design:
Geometric, Adaptive, and Robust. Prentice-Hall Inter- 关28兴 Kosmatopoulos, E. B., Polycarpou, M. M.,
national 共UK兲 Limited, London, 1995. Christodoulou, M. A., and Ioannou, P. A., High-order
关14兴 Tao Zhang, K. A. and Kantor, J. C., An exothermic neural network structures for identification of dynami-
continuous stirred tank reactor is feedback equivalent cal systems. IEEE Trans. Neural Netw. 6, 422– 431
to a linear system. Chem. Eng. Commun. 37 共1兲 共1995兲.
共1985兲. 关29兴 Wang, L. X., Adaptive Fuzzy Systems and Control:
关15兴 Aris, R., Elementary Chemical Analysis. Prentice- Design and Analysis. Prentice-Hall, Englewood Cliffs,
Hall, Englewood Cliffs, NJ, 1969. NJ, 1994.
关16兴 Lightbody, G. and Irwin, G. W., Direct neural model 关30兴 Funahashi, K. I., On the approximate realization of
reference adaptive control. IEE Proc.: Control Theory continuous mappings by neural networks. Neural Net-
Appl. 142 共1兲 共1995兲. works 2, 183–192 共1989兲.