This action might not be possible to undo. Are you sure you want to continue?

Welcome to Scribd! Start your free trial and access books, documents and more.

Find out moreIOS Press

Temperature control using neuro-fuzzy

controllers with compensatory operations and

wavelet neural networks

Cheng-Jian Lin

a,∗

, Chi-Yung Lee

b

and Cheng-Chung Chin

c

a

Department of Computer Science and Information Engineering, Chaoyang University of Technology, Taichung

County, 413 Taiwan

b

Department of Computer Science and Information Engineering, Nankai Institute of Technology, Nantou County

542, Taiwan

Abstract. This paper addresses a Compensatory Wavelet Neuro-Fuzzy System (CWNFS) for temperature control. The proposed

CWNFS model is ﬁve-layer structure, which combines the traditional Takagi-Sugeno-Kang (TSK) fuzzy model and the wavelet

neural networks (WNN). We adopt the non-orthogonal and compactly supported functions as wavelet neural network bases.

Besides, the compensatory fuzzy reasoning method is used in adaptive fuzzy operations that can make the fuzzy logic system

more adaptive and effective. An on-line learning algorithm, which consists of structure learning and parameter learning, is

presented. The structure learning is based on the degree measure to determine the number of fuzzy rules and wavelet functions.

The parameter learning is based on the gradient descent method to adjust the shape of membership function, compensatory

operations and the connection weights of WNN. Simulation results have been given to illustrate the performance and effectiveness

of the proposed model.

Keywords: Temperature control, TSK-type fuzzy model, wavelet neural networks, on-line learning, gradient descent, compen-

satory operation

1. Introduction

In recent years, the concept of fuzzy logic and artiﬁcial neural network for control problem has been grown into

a popular research topic [9,10,17]. The reason is that the classical control theory usually requires a mathematical

model for designing the controller. The inaccuracy of mathematical modeling of the plants usually degrades the

performance of the controller, especially for nonlinear and complex control problems [1]. On the contrary, the

fuzzy logic controller (FLC’s) and the artiﬁcial neural network controller, they offer a key advantage over traditional

adaptive control systems. That is, they do not require mathematical models of the plants. The traditional neural

networks can learn from data and feedback, but the meaning associated with each neuron and each weight in the

network is not easily understood. Alternatively, the fuzzy logical models are easy to appreciate, because it uses

linguistic terms and the structure of if-then rules. However, as compared with the neural networks, learning ability is

lack of fuzzy logic. In contrast to the pure neural network or fuzzy system, the fuzzy neural network representations

have emerged as a powerful approach to the solution of many problems [5,11,13,15].

In this paper, the compensatory wavelet neuro-fuzzy system (CWNFS) is proposed to overcome the disadvantages

of the FLC and the artiﬁcial neural network. Each fuzzy rule corresponding to a WNN consists of single-scaling

wavelets. The non-orthogonal and compactly supported functions are adopted as wavelet neural network bases. The

∗

Corresponding author. Fax: +886 4 23742375; E-mail: cjlin@mail.cyut.edu.tw.

1064-1246/06/$17.00 © 2006 – IOS Press and the authors. All rights reserved

146 C.-J. Lin et al. / Temperature control using neuro-fuzzy controllers with compensatory operations

1 0

φ

φ φ

φ φ φ

φ φ φ φ

φ φ φ

. .

1

1

1

1

0

0

0

0

Fig. 1. Wavelet bases are over-complete and compactly supported.

compensatoryfuzzy reasoningmethod is used in adaptive fuzzy operations that can make the fuzzy logic systemmore

adaptive and effective. Therefore, an effective neural fuzzy system should be able not only to adaptively adjust fuzzy

membership functions but also to dynamically optimize adaptive fuzzy operators. An on-line structure/parameter

learning algorithm is performed concurrently in the CWNFS. The CWNFS model cannot only ﬁnd itself optimal

or almost optimal network size, but the parameters of the CWNFS are adjusted via a proposed dynamic training

algorithm. Finally, the encouraging results are obtained via series simulations of a water bath temperature control

system. We also compare our approach with other methods in the literature early.

2. The structure of a compensatory wavelet neuro-fuzzy system

2.1. Description of wavelet neural network

To generate the novel form of the TSK model, the CWNFS is integrates the traditional TSK-type fuzzy model and

the WNN [4]. Each fuzzy rule corresponding to a WNN consists of single-scaling wavelets. The non-orthogonal

and compactly supported functions are adopted in ﬁnitely range as wavelet bases [3]. The shape and position of

wavelet bases are shown in Fig. 1.

Neural networks employing wavelet neurons are refereed to wavelet neural networks. According to Yamakawa et

al. [18], we propose a new type of wavelet neural network model that is shown in Fig. 2. Consider n inputs vectors

{x

1

, x

2

, . . . , x

n

}∈ R

n

and single-output Y ∈ R, respectively. This model is obtained by replacing a sigmoidal

activation function with single-scaling wavelets. The wavelet neural networks are characterized by weighted and

wavelet base. Each linear synaptic weight of wavelet basis is adjustable by learning. Noted that, the ordinary wavelet

neural network model applications, it is often useful to normalize the input vectors into the interval [0, 1]. When the

input signal ﬁre up the interval of wavelet neurons, and the φ

a.b

(x

i

) function be calculated by

_

φ(x

i

) = cos(x

i

) −0.5 x

i

0.5

0 (otherwise)

, φ

a.b

(x

i

) = cos(ax

i

−b) (1)

Above equation is formulating the non-orthogonal wavelet neurons in ﬁnitely range, the symbol b is a shifting

parameter, the maximumvalue of witch equals the corresponding scaling parameter a. Obviously, a crisp value Ψ

a.b

obtained as follows:

C.-J. Lin et al. / Temperature control using neuro-fuzzy controllers with compensatory operations 147

Fig. 2. Schematic diagram of the WNN.

ψ

a.b

=

n

i=1

φ

a.b

(x

i

)

|X|

(2)

where |X| means the number of input dimension. The ﬁnal output of the wavelet neural networks is:

ˆ y

j

=

M

k=1

w

jk

ψ

a.b

(3)

where ˆ y

j

is the local output of the WNN for output Y and jth rule, and the link weight w

jk

is the output action

strength associated with in the jth rule and kth Ψ

a.b

. The symbol M denotes the number of wavelets, which are

equal the number of existing fuzzy rules in CWNFS.

2.2. Description of the CWNFS model

According to [12], a novel CWNFS model, which combines fuzzy model, compensatory operations, and wavelet

neural networks, can be written in the following general form:

R

j

: [IF x

1

is A

1j

and . . . and x

n

is A

1−γj+rj/n

nj

THEN ˆ y

j

=

M

k=1

w

jk

ψ

a.b

(4)

where x

i

is input variable, ˆ y

j

is output variable, A

nj

is linguistic term of the precondition part, and n is number of

input variables.

The structure of the CWNFS is shown in Fig. 3. It is a ﬁve-layer structure. Each node in layer 1 is an input node;

these nodes only pass the input signal to next layer. Each node in layer 2 acts as membership function representing

the term of the respective input-linguistic variables. The Gaussian function is adopted as the membership function.

Layer 3 is a rule node representing the precondition part of one fuzzy logic rule. We use a compensatory fuzzy

operator mentioned in [12] to perform IF-condition matching of fuzzy rules. As a result, the output function of each

inference nodes is

O

(3)

j

=

_

i

I

(2)

ij

_

1−γj+

γ

j

n

(5)

where γ

j

∈ [0, 1] is called the compensatory degree. Nodes in layer 4 only receive the signal, which is ˆ y

j

fromoutput

of wavelet neural network model. The node in layer 5 computes the output signal Y . The output node together with

links connected to it act as a defuzziﬁer. The mathematical function is

148 C.-J. Lin et al. / Temperature control using neuro-fuzzy controllers with compensatory operations

Fig. 3. Schematic diagram of CWNFS model.

O

(5)

= Y =

M

j=1

O

(4)

j

O

(3)

j

M

j=1

O

(3)

j

=

M

j=1

(w

j1

ψ

0.0

+w

j2

ψ

1.0

+w

j3

ψ

1.1

· · ·)O

(3)

j

M

j=1

O

(3)

j

(6)

where O

(4)

j

is the output of the local model of the WNN model for an output Y and the jth rule, O

(3)

j

is the output

of layer 3, and O

(5)

is the output of the CWNFS. The structure of the proposed CWNFS model is different from

the CNFN model [12]. In [12], it is four-layer structure and the consequent part of the rules is singleton. In the

proposed CWNFS model, the consequent part of the rules is a nonlinear function of input linguistic variables. This

study adopts the wavelet neural network to the consequent part of the rules. The local properties of wavelets in the

CWNFS model enable arbitrary functions to be approximated more effectively.

3. An on-line learning algorithm for CWNFS controller

In this paper, an on-line learning algorithm, which consists of structure learning and parameter learning, is used

concurrently for constructing the CWNFS. The structure learning scheme is used to decide proper fuzzy partition,

C.-J. Lin et al. / Temperature control using neuro-fuzzy controllers with compensatory operations 149

the number of the rule nodes and the wavelet neural networks while the parameter learning scheme is used to tune

the adjustable parameters. The detail of the learning algorithm is presented below:

3.1. The structure learning scheme

Initially, there are no rules in CWNFS model; the ﬁrst task in structure learning is to decide when to generate a

new rule. Geometrically, a rule is corresponding to a cluster in the input space with m

ij

and σ

ij

representing the

mean and variance of that cluster. For each incoming pattern x

i

the strength a rule is ﬁred can be regard as the degree

of the incoming pattern belongs to the corresponding cluster. An input data x

i

with higher ﬁring strength means

its spatial location is nearer the center of cluster than those with smaller strength. Based on this concept, the ﬁring

strength obtained from Eq. (5) is used as the degree measure. The criterion of generating a new fuzzy rule for new

incoming data is

F

max

= max

1jM

F

j

= max

1jM

O

(3)

j

(7)

If F

max

F, then a new rule is generated where F ∈ (0, 1) is a pre-speciﬁed threshold that should be decayed

during the learning process limiting the size of CWNFS. Once the new rule is generated, the next step is to assign

initial value of the free parameters. The structure learning method is similar to [8,12].

3.2. The parameter learning scheme

After the network structure is adjusted according to the current training pattern, the network then enters the

parameter learning scheme to turn the adjustable parameters of the network optimally based on the same training

pattern. Notice that the following parameter learning is performed on the whole networks after structure learning,

no matter whether the nodes (links) are newly added or are existent originally. Since the learning process involves

the determination of the vector which minimize a given energy function. The gradient of the energy function with

respect to the vector is computed and the vector is adjusted along the negative gradient. The energy function E is

deﬁned as

E =

1

2

(Y −Y

des

)

2

(8)

where Y is the model output and Y

des

is the desired output. Assuming that w is the adjustable parameter in a

nodded, the general learning rule used is

w(t + 1) = w(t) +η

_

−

∂E

∂w

_

(9)

where η is the learning rate. To show the learning rules, we derive the rules layer by layer. For clarity, we consider

the single output case.

Layer 5: The error to be propagated to the preceding layer is

δ

(5)

= −

∂E

∂O

(5)

=

−∂

1

2

(Y −Y

des

)

2

∂O

(5)

= Y

des

−Y (10)

Layer 4: The link weight of wavelet neural network is update by

∆w

jk

= −η

w

∂E

∂w

jk

=

_

−η

w

∂E

∂O

(5)

_

·

_

∂O

(5)

∂O

(4)

lj

_

·

_

∂O

(4)

lj

∂w

jk

_

= η

w

· δ

(5)

·

O

(3)

j

ψ

a.b

(x)

j

O

(3)

j

(11)

where η

w

is the learning rate.

Layer 3: In this layer only the error term needs to be calculated and propagated

δ

(3)

= −

∂E

∂O

(3)

j

=

_

−

∂E

∂O

(5)

_

·

_

∂O

(5)

∂O

(3)

j

_

= δ

(5)

·

⎛

⎝

O

(4)

lj

j

O

(3)

j

−

j

O

(4)

lj

O

(3)

j

⎞

⎠

1

_

j

O

(3)

j

_

2

(12)

150 C.-J. Lin et al. / Temperature control using neuro-fuzzy controllers with compensatory operations

To eliminate the constraint γ

j

∈ [0, 1], we redeﬁne γ

j

as follows:

γ

j

=

c

2

j

c

2

j

+d

2

j

(13)

Then we have

∆c

j

= η

c

_

2c

j

d

2

j

[c

2

j

+d

2

j

]

2

_

δ

(3)

_

1

n

− 1

_

ln

_

Π

i

O

(2)

ij

_

O

(3)

j

(14)

∆d

j

= −η

d

_

2c

2

j

d

j

[c

2

j

+d

2

j

]

2

_

δ

(3)

_

1

n

− 1

_

ln

_

Π

i

O

(2)

ij

_

O

(3)

j

(15)

In all above formulas, η

c

and η

d

are the learning rate of the parameter c

j

and the parameter d

j

.

Layer 2 : In this layer, the error term is computed as follows:

δ

(2)

= −

∂E

∂O

(2)

ij

=

_

−

∂E

∂O

(5)

_

·

_

∂O

(5)

∂O

(3)

j

_

·

_

∂O

(3)

j

∂O

(2)

ij

_

= δ

(3)

·

i=j

O

(3)

j

(16)

The updated law of m

ij

is

∆m

ij

= −η

m

∂E

∂m

ij

=

_

−η

m

∂E

∂O

(2)

ij

_

·

_

∂O

(2)

ij

)

∂m

ij

_

= η

m

δ

(2)

·

∂O

(2)

ij

∂m

ij

(17)

The updated law of σ

ij

is

∆σ

ij

= −η

m

∂E

∂σ

ij

=

_

−η

σ

∂E

∂O

(2)

ij

_

·

_

∂O

(2)

ij

∂σ

ij

_

= η

σ

δ

(2)

·

∂O

(2)

ij

∂σ

ij

(18)

where η

m

and η

σ

are the learning rate parameter of the mean and the standard deviation of the Gaussian function,

respectively.

4. Control of water bath temperature system

The goal of this section is to control the temperature of a water bath system given by

dy(t)

dt

=

u(t)

C

+

Y

0

−y(t)

RC

(19)

where y(t) is system output temperature in

◦

C; u(t) is heating ﬂowing inward the system; Y

0

is room temperature;

C is the equivalent system thermal capacity; and R is the equivalent thermal resistance between the system borders

and surroundings.

Assuming that R and C are essentially constant, we rewrite the system in Eq. (19) into discrete-time form with

some reasonable approximation. The system

y(t + 1) = e

−αTs

y(k) +

β

α

(1 −e

−αTs

)

1 +e

0.5y(k)−40

u(k) + [1 −e

−αTs

]y

0

. (20)

is obtained, where α and β are some constant values describing R and C. The system parameters used in this

example are α = 1.0015e

−4

, β = 8.67973e

−3

and Y

0

= 25.0 (

◦

C), which were obtained from a real water bath

plant in [17]. The inputu(k) is limited to 0 and 5 V represent voltage unit. The sampling period is Ts = 30. The

system conﬁguration is shown in Fig. 4, where y

ref

is the desired temperature of the controlled plant. Recently,

many researchers [2,6,7,14] use various different methods for solving the temperature control problems. The control

approach in this paper is different from [2,6]. Chen and Pao [2] compute the derivative of the model’s output with

respect to its input by means of the back-propagationprocess, which evaluates the transpose of the network Jacobian

C.-J. Lin et al. / Temperature control using neuro-fuzzy controllers with compensatory operations 151

y(k+1)

On-Line Learning

Algorithm

Water

Bath System

Z

CWNFS

Controller

u(k)

y(k)

y

ref

(k+1)

Fig. 4. Flow diagram of using CWNFS controller for solving the temperature control problem.

at the network’s current input vector. This usually implies that we need a model for the plant and the Jacobian

matrix obtained from the model, which could be a neural network, a neuro-fuzzy system, or another appropriate

mathematical description of the plant. As a result, propagating errors between actual and desired plant outputs back

through the forward model produces error in the control signal, which can be used to train another network to be a

controller [6].

By implement the on-line training scheme for CWNFS, a sequence of randominput signals u

rd

(k) limited to 0 and

5 V is injected directly into the simulated system described in Eq. (20). The 120 training patterns are chosen from

the input-outputs characteristic in order to cover the entire reference output. The initial temperature of the water is

25

◦

C, and the temperature rises progressively when randominput signals are injected. For the CWNFS, the learning

rate η

w

= η

m

= η

σ

= η

c

= η

d

= 0.05, the initial variance σ

init

= 6 and the prespeciﬁed threshold F = 0.05 are

chosen. After training, there are 12 fuzzy rules generated. The obtained fuzzy rules with a compensatory degree are

shown as follows:

R

1

: If [I

1

is A

1,1

(24.6024, 3.6861) and I

2

is A

2,1

(26.9721, 7.1613)]

0.37

Then ˆ y

1

1

= 0.7056ϕ

0.0

+ 0.0029ϕ

1.0

−0.0422ϕ

1.1

+ 0.0592ϕ

2.0

+ 0.0739ϕ

2.1

−0.0336ϕ

2.2

+ 0.0164ϕ

3.0

+ 0.0669ϕ

3.1

+ 0.0432ϕ

3.2

−0.0497ϕ

3.3

+0.0605ϕ

4.0

−0.0241ϕ

4.1

R

2

: If [I

1

is A

1,2

(33.0827, 3.0011) and I

2

is A

2,2

(31.5810, 7.7886)]

0.93

Then ˆ y

1

2

= 0.0608ϕ

0.0

+ 2.9094ϕ

1.0

+ 0.0806ϕ

1.1

−0.0338ϕ

2.0

−0.0027ϕ

2.1

−0.0344ϕ

2.2

−0.0550ϕ

3.0

+ 0.0669ϕ

3.1

−0.0736ϕ

3.2

−0.0551ϕ

3.3

−0.0105ϕ

4.0

+ 0.0384ϕ

4.1

R

3

: If [I

1

is A

1,3

(45.9519, 5.0566) and I

2

is A

2,3

(37.5630, 5.1166)]

0.73

Then ˆ y

1

3

= −0.0573ϕ

0.0

+ 0.0874ϕ

1.0

+ 17.1185ϕ

1.1

−0.0910ϕ

2.0

−0.0918ϕ

2.1

+0.0962ϕ

2.2

−0.0323ϕ

3.0

+ 0.0463ϕ

3.1

−0.0734ϕ

3.2

−0.0047ϕ

3.3

−0.0321ϕ

4.0

+ 0.0250ϕ

4.1

R

4

: If [I

1

is A

1,4

(38.7137, 4.7003) and I

2

is A

2,4

(45.9058, 5.1329)]

0.94

Then ˆ y

1

4

= −0.0996ϕ

0.0

−0.0251ϕ

1.0

−0.0464ϕ

1.1

−12.5963ϕ

2.0

+ 0.0857ϕ

2.1

+0.0823ϕ

2.2

−0.0022ϕ

3.0

+ 0.0176ϕ

3.1

−0.0320ϕ

3.2

+ 0.0014ϕ

3.3

−0.0036ϕ

4.0

+ 0.0027ϕ

4.1

R

5

: If [I

1

is A

1,5

(56.3778, 5.1274) and I

2

is A

2,5

(48.0105, 4.7223)]

0.25

Then ˆ y

1

5

= −0.0042ϕ

0.0

+ 0.0153ϕ

1.0

+ 0.0179ϕ

1.1

+ 0.0053ϕ

2.0

+ 13.7761ϕ

2.1

+0.0632ϕ

2.2

−0.0280ϕ

3.0

−0.0370ϕ

3.1

−0.0241ϕ

3.2

−0.0865ϕ

3.3

−0.0471ϕ

4.0

+ 0.0842ϕ

4.1

152 C.-J. Lin et al. / Temperature control using neuro-fuzzy controllers with compensatory operations

R

6

: If [I

1

is A

1,6

(62.8080, 8.3051) and I

2

is A

2,6

(66.1094, 6.6011)]

0.41

Then ˆ y

1

6

= −0.0763ϕ

0.0

+ 0.0380ϕ

1.0

−0.0943ϕ

1.1

+ 0.0567ϕ

2.0

+ 0.0369ϕ

2.1

+2.2305ϕ

2.2

−0.0607ϕ

3.0

−0.0893ϕ

3.1

+ 0.0720ϕ

3.2

−0.0015ϕ

3.3

+0.0747ϕ

4.0

−0.0416ϕ

4.1

R

7

: If [I

1

is A

1,7

(74.4883, 5.5007) and I

2

is A

2,7

(68.1826, 7.6617)]

0.95

Then ˆ y

1

7

= 0.0028ϕ

0.0

+ 0.0722ϕ

1.0

+ 0.0858ϕ

1.1

−0.0837ϕ

2.0

+ 0.0739ϕ

2.1

−0.0315ϕ

2.2

+ 3.5460ϕ

3.0

−0.0142ϕ

3.1

+ 0.0327ϕ

3.2

+ 0.0699ϕ

3.3

+0.0614ϕ

4.0

+ 0.0655ϕ

4.1

R

8

: If [I

1

is A

1,8

(77.4191, 6.6648) and I

2

is A

2,8

(78.0521, 5.7519)]

0.35

Then ˆ y

1

8

= 0.0553ϕ

0.0

−0.0123ϕ

1.0

−0.0703ϕ

1.1

−0.0936ϕ

2.0

+ 0.0774ϕ

2.1

−0.0907ϕ

2.2

−0.0672ϕ

3.0

+ 2.6022ϕ

3.1

−0.0951ϕ

3.2

−0.0450ϕ

3.3

−0.0597ϕ

4.0

−0.0852ϕ

4.1

R

9

: If [I

1

is A

1,9

(49.3033, 4.1272) and I

2

is A

2,9

(54.9042, 3.9521)]

0.12

Then ˆ y

1

9

= −0.0126ϕ

0.0

+ 0.0186ϕ

1.0

+ 0.0649ϕ

1.1

−0.0651ϕ

2.0

+ 0.0903ϕ

2.1

−0.0978ϕ

2.2

+ 0.0797ϕ

3.0

+ 0.0506ϕ

3.1

−7.3473ϕ

3.2

−0.0747ϕ

3.3

+0.0500ϕ

4.0

−0.0143ϕ

4.1

R

10

: If [I

1.10

is µ(29.5346, 5.6278) and I

2.10

is µ(28.9791, 6.5138)]

0.72

Then ˆ y

1

10

= −0.0133ϕ

0.0

−0.0522ϕ

1.0

+ 0.0096ϕ

1.1

+ 0.0757ϕ

2.0

+ 0.0594ϕ

2.1

−0.0107ϕ

2.2

+ 0.0405ϕ

3.0

−0.0732ϕ

3.1

+ 0.0019ϕ

3.2

+ 1.7451ϕ

3.3

+0.0476ϕ

4.0

+ 0.0710ϕ

4.1

R

11

: If [I

1

is A

1,11

(44.0306, 2.8736) and I

2

is A

2,11

(48.5759, 1.2315)]

0.96

Then ˆ y

1

11

= 0.0830ϕ

0.0

+ 0.0382ϕ

1.0

+ 0.0861ϕ

1.1

−0.0607ϕ

2.0

−0.0223ϕ

2.1

−0.0926ϕ

2.2

+ 0.0707ϕ

3.0

−0.0246ϕ

3.1

−0.0520ϕ

3.2

+ 0.0209ϕ

3.3

−0.9817ϕ

4.0

−0.0653ϕ

4.1

R

12

: If [I

1

is A

1,12

(34.8565, 3.4774) and I

2

is A

2,12

(35.2144, 4.4854)]

0.5

Then ˆ y

1

12

= −0.0324ϕ

0.0

−0.0283ϕ

1.0

+ 0.0880ϕ

1.1

+ 0.0404ϕ

2.0

+ 0.0155ϕ

2.1

−0.0193ϕ

2.2

+ 0.0520ϕ

3.0

+ 0.0252ϕ

3.1

+ 0.0809ϕ

3.2

+ 0.0851ϕ

3.3

−0.0171ϕ

4.0

+ 2.1611ϕ

4.1

In this paper, we compare the CWNFS controller to the PIDcontroller [16], the manually designed fuzzy controller

and the self-constructing fuzzy neural network (SCFNN) [15]. Each of the three controllers is applied to the water

bath temperature control system. The comparison performance measures include set-points regulation, the inﬂuence

of impulse noise, and a large parameter variation in the system.

For the PID control, a velocity-form discrete PID controller [16] is used and is described by

∆u(k) = K

_

e(k) −e(k − 1) +

T

s

2T

i

[e(k) +e(k − 1)] +

T

d

T

s

[e(k) − 2e(k − 1) +e(k − 2)]

_

(21)

= K

P

[e(k) −e(k − 1)] +K

I

e(k) +K

D

[e(k) − 2e(k − 1) +e(k − 2)]

where K

P

= K −

1

2

K

I

, K

I

=

KTs

Ti

, K

D

=

KT

d

Ts

. The parameter ∆u(k) is the increment of the control input, e(k)

is the performance error at the sampling instant k, and, K

P

, K

I

and K

D

are the proportional, integral, and derivative

C.-J. Lin et al. / Temperature control using neuro-fuzzy controllers with compensatory operations 153

Table 1

Fuzzy rule table formulated for the water bath temperature control system

Error, e(t)

NL NM NS ZE PS PM PL

Change error, ce(t) PL PL PL PL PL

PM PM PM PM PL

PS PS PS PS PM PL

ZE NL NM NS ZE PS PM PL

NS NS NS NS

NM NM

NL NL

parameters, respectively. In order not to aggravate noise in the plant, only a two-termPID controller is used, i.e., K

D

is set to zero in the water bath system. The other two parameters K

P

and K

I

are chosen as 80 and 70, respectively.

For the above designed PID controller, we have tried our best to achieve their respective best performance through

several trial-and-error experiments.

For the manually designed fuzzy controller, the input variables are chosen as e(t) and ce(t), where e(t) is the

performance error indicating the error between the desired water temperature and the actual measured temperature

and ce(t) is the rate of change in the performance error e(t). The output or the controlled linguistic variable is the

voltage signal u(t) to the heater. Seven fuzzy terms are deﬁned for each linguistic variable. These fuzzy terms consist

of Negative Large (NL), Negative Medium (NM), Negative Small (NS), Zero (ZE), Positive Small (PS), Positive

Medium(PM), and Positive Large (PL). Each fuzzy termis speciﬁed by a Gaussian membership function. According

to common sense and engineering judgment, 25 fuzzy rules are speciﬁed in Table 1. Like other controllers, a fuzzy

controller has some scaling parameters to be speciﬁed. They are GE, GCE, and GU, corresponding to the process

error, the change in error, and the controller’s output, respectively. We choose these parameters as follows: GE =

1/15, GCE = 1/15, GU = 450.

Recently, Lin et al. [15] presented a self-constructing fuzzy neural network (SCFNN) for control problems. The

SCFNN controller is a standard four-layer structure. Each node in layer 3 performs the product operation. The

consequence of each fuzzy rule is a singleton value. The output node sums all incoming signals to obtain inferred

result. An on-line learning algorithm was proposed to decide the structure of fuzzy rules and turn the adjustable

parameters through the backpropagation algorithm. The structure of the proposed CWNFS controller is difference

from [15]. Our model is ﬁve-layer structure, using the compensatory operation in layer 3, and adopting the wavelet

neural network as consequent part of each fuzzy rule.

For the aforementioned controllers (CWNFS controller, PID controller, manually designed fuzzy controller and

SCFNN controller), three groups of computer simulations are conducted on the water bath temperature control

system. Each simulation is performed over 120 sampling time steps.

The ﬁrst task is to control the simulated system to follow three set-points.

y

ref

(k) =

⎧

⎨

⎩

35

◦

C,

55

◦

C,

75

◦

C,

for

for

for

k 40

40 < k 80

80 < k 120.

(22)

The regulation performance of the CWNFS model is shown in Fig. 5(a). We also test the regulation performance

by using SCNNF controller [15]. The error curves of CWNFS controller and SCFNN controller between k = 80

and k = 100 are shown in Fig. 5(b). In this ﬁgure, the CWNFS controller obtains smaller errors than the SCNNF

controller. To test their regulation performance, a performance index, sum of absolute error (SAE), is deﬁned by

SAE =

k

|y

ref

(k) −y(k)| (23)

where y

ref

(k) and y(k) are the reference output and the actual output of the simulated system, respectively. The

SAE values of the CWNFS controller, the PID controller, the fuzzy controller and SCFNN controller are 352.95,

418.5, 401.5, and 356.41, which are shown in the ﬁrst row of Table 2.

The second set of simulations is carried out for the purpose of studying the noise-rejection ability of the four

controllers when some unknown impulse noise is imposed on the process. One impulse noise value −5

◦

C is added

154 C.-J. Lin et al. / Temperature control using neuro-fuzzy controllers with compensatory operations

(a)

(b)

Fig. 5. (a) Final regulation performance of the CWNFS controller for water bath system. (b) The error curves of CWNFS controller and SCFNN

controller between k = 80 and k = 100.

to the plant output at the sixtieth sampling instant. A set-point of 50

◦

C is performed in this set of simulations. For

the CWNFS controller, the same training scheme, training data and learning parameters are used as those used in

the ﬁrst set of simulations. The behaviors of the CWNFS controller under the inﬂuence of impulse noise and the

corresponding errors are shown in Fig. 6(a)–(b). The SAE values of the CWNFS controller, the PID controller, the

fuzzy controller, and SCFNN are 273.25, 311.5, 275.8, and 280.5, which are shown in the second rowof Table 2. It is

observed that the CWNFS controller performs quite well. It recovers very quickly and steadily after the presentation

of the impulse noise.

One common characteristic of many industrial-control processes is that their parameters tend to change in an

unpredictable way. To test the robustness of the four controllers, a value of 0.7 ∗ u(k −2) is added to the plant input

after the sixtieth sample in the fourth set of simulations. A set-point of 50

◦

C is used in this set of simulations. For

C.-J. Lin et al. / Temperature control using neuro-fuzzy controllers with compensatory operations 155

Table 2

Performance comparison of various controllers

SAE =

120

k=1

|y

ref

(k) − y(k)| CWNFS controller PID controller [16] manually designed SCFNN controller [15]

fuzzy controller

Regulation performance 352.95 418.5 401.5 356.41

Inﬂuence of impulse noise 273.25 311.5 275.8 280.50

Effect of change in plant dynamics 262.51 322.2 273.5 270.21

(a)

(b)

Fig. 6. (a) Behavior of the CWNFS controller under the impulse noise for water bath system. (b) The error curves of CWNFS controller and

SCFNN controller.

the GA-NFS controller, the same training scheme, training data and learning parameters are used as those used in

the ﬁrst set of simulations. The behaviors of the CWNFS controller when there is a change in the plant dynamics

are shown in Fig. 7(a). The corresponding errors of the CWNFS and SCNNF controllers are shown in Fig. 7(b).

The SAE values of the CWNFS controller, the PID controller, the fuzzy controller, and SCFNN are 262.51, 322.2,

156 C.-J. Lin et al. / Temperature control using neuro-fuzzy controllers with compensatory operations

(a)

(b)

Fig. 7. (a) Behavior of the CWNFS controller when a change occurs in the water bath system. (b) The error curves of CWNFS controller and

SCFNN controller.

273.5, and 270.21, which are shown in the third row of Table 2. The results show the good control and disturbance

rejection capabilities of the trained CWNFS controller in the water bath system.

For the aforementioned simulation results, Table 2 has shown that the proposed CWNFS controller has better

performance than that of other methods. For the fuzzy controller, the numbers of rules and membership functions

have to be decided and tuned by hand. As for the PID controller, the parameters K

p

, K

I

, and K

D

also have to

be decided properly. For the fuzzy and PID controllers, therefore, they usually require a long time in design for

achieving good performance. In the CWNFS controller, however, no controller parameters have to be decided

in advance. We only need to choose propose training patterns of the CWNFS controller. Although the structure

of CWNFS controller is more complicated than the fuzzy and PID controllers, in general, the CWNFS controller

usually spends a relatively short time in design for achieving good performance. This study attempts to emphasize

C.-J. Lin et al. / Temperature control using neuro-fuzzy controllers with compensatory operations 157

the methodology and control abilities of the proposed CWNFS model. In the future, we will apply the proposed

CWNFS controller on a real water bath temperature control system.

5. Conclusion

In this paper, the new CWNFS controller, which combines TSK-type fuzzy model and wavelet neural networks, is

proposed and applied to the water bath temperature system. The CWNFS controller can automatically construct and

adjust free parameters itself by performing online supervised structure/parameter learning schemes concurrently.

Finally, computer simulation results have shown that the proposed CWNFS controller has better performance than

that of other methods.

Acknowledgement

This research is supported by the National Science Council of R.O.C. under grant NSC 93-2213-E-324-008.

References

[1] K.J. Åstrom and B. Wittenmark, Adaptive Control, Reading, MA: Addison-Wesley, 1989.

[2] V.C. Chen and Y.H. Pao, Learning Control with Neural Networks, Proc. of International Conf. on Robotics and Automation, 1989,

1448–1453.

[3] I. Daubechies, Orthonormal Bases of Compactly Supported Wavelets, Comm. Pur. Appl. Math. 41 (1998).

[4] D.W.C. Ho, P.A. Zhang and J. Xu, Fuzzy Wavelet Networks for Function Learning, IEEE Trans. on Fuzzy Systems 9 (2001), 200–211.

[5] J.-S.R. Jang, ANFIS: Adaptive-Network-Based Fuzzy Inference System, IEEE Trans. on Syst., Man, and Cybern. 23 (1993), 665–685.

[6] J.-S.R. Jang, C.T. Sun and E. Mizutani, Neuro-Fuzzy and Soft Computing, Ch. 17, Prentice-Hall, 1997.

[7] C.F. Juang, J.Y. Lin and C.T. Lin, Genetic reinforcement learning through symbiotic evolution for fuzzy controller design, IEEE Trans. on

Systems, Man, and Cybernetics B30 (2000), 290–302.

[8] C.F. Juang, A TSK-type recurrent fuzzy network for dynamic systems processing by neural network and genetic algorithms, IEEE Trans.

on Fuzzy Systems 10 (2002), 155–170.

[9] C.L. Karr and E.J. Gentry, Fuzzy Control of pH Using Genetic Algorithms, IEEE Trans. on Fuzzy Syst. 1 (1993), 46–53.

[10] C.C. Lee, Fuzzy Logic in Control Systems: Fuzzy Logic Controllers – Parts I, II, IEEE Trans. on Syst., Man, Cybern. 20 (1990), 404–435.

[11] C.J. Lin and C.C. Chin, A Wavelet-Based Neuro-Fuzzy System and Its Applications, Proc. IEEE Int. Joint Conference on Neural Networks,

Oregon, Portland, 2003, 20–24.

[12] C.J. Lin and C.H. Chen, Nonlinear System Control Using Compensatory Neuro-Fuzzy Networks, IEICE Transactions on Fundamentals

of Electronics, Communications and Computer Sciences E86-A (2003), 2309–2316.

[13] C.T. Lin and C.S.G. Lee, Neural Fuzzy Systems: A Neuro-Fuzzy Synergism to Intelligent System, NJ: Prentice-Hall, 1996.

[14] C.T. Lin, C.F. Juang and C.P. Li, Temperature control with a neural fuzzy inference network, IEEETrans. on Systems, Man, and Cybernetics

C29 (1999), 440–451.

[15] F.J. Lin, C.H. Lin and P.H. Shen, Self-Constructing Fuzzy Neural Network Speed Controller for Permanent-Magnet Synchronous Motor

Drive, IEEE Trans. Fuzzy Syst. 9 (2001), 751–759.

[16] C.L. Phillips and H.T. Nagle, Digital Control System Analysis and Design, Prentice Hall, 1995.

[17] J. Tanomaru and S. Omatu, Process Control by On-Line Trained Neural Controllers, IEEE Trans. on Ind. Electron. 39 (1992), 511–521.

[18] T. Yamakawa, E. Uchino and T. Samatsu, Wavelet Neural Networks Employing Over-Complete Number of Compactly Supported Non-

Orthogonal Wavelets and Their Applications, Proc. of IEEE Conf. on Neural Networks 3 (1994), 1391–1396.

[19] Y.Q. Zhang and A. Kandel, Compensatory Neurofuzzy Systems with Fast Learning Algorithms, IEEE Transactions on Neural Networks

9 (1998), 83–105.

According to Yamakawa et al. 1. Noted that. . Neural networks employing wavelet neurons are refereed to wavelet neural networks. Finally. 1 Fig. An on-line structure/parameter learning algorithm is performed concurrently in the CWNFS. Consider n inputs vectors {x1 . it is often useful to normalize the input vectors into the interval [0. [18]. Lin et al. and the φ a. . respectively. The shape and position of wavelet bases are shown in Fig. Obviously. xn }∈ Rn and single-output Y ∈ R. . 1]. The wavelet neural networks are characterized by weighted and wavelet base.-J. The CWNFS model cannot only ﬁnd itself optimal or almost optimal network size.146 C. we propose a new type of wavelet neural network model that is shown in Fig. . Each linear synaptic weight of wavelet basis is adjustable by learning. Each fuzzy rule corresponding to a WNN consists of single-scaling wavelets. the symbol b is a shifting parameter.1. the CWNFS is integrates the traditional TSK-type fuzzy model and the WNN [4].b obtained as follows: . The non-orthogonal and compactly supported functions are adopted in ﬁnitely range as wavelet bases [3]. a crisp value Ψ a. 1. 0 (otherwise) φa. 2.5 xi 0. When the input signal ﬁre up the interval of wavelet neurons. an effective neural fuzzy system should be able not only to adaptively adjust fuzzy membership functions but also to dynamically optimize adaptive fuzzy operators.b (xi ) = cos(axi − b) (1) Above equation is formulating the non-orthogonal wavelet neurons in ﬁnitely range. Description of wavelet neural network To generate the novel form of the TSK model. the maximum value of witch equals the corresponding scaling parameter a. We also compare our approach with other methods in the literature early.. the encouraging results are obtained via series simulations of a water bath temperature control system. 2. / Temperature control using neuro-fuzzy controllers with compensatory operations φ 0 0 0 0 φ 0 φ φ φ φ φ φ φ φ φ φ 1 1 1 1 φ . This model is obtained by replacing a sigmoidal activation function with single-scaling wavelets. compensatory fuzzy reasoning method is used in adaptive fuzzy operations that can make the fuzzy logic system more adaptive and effective. Wavelet bases are over-complete and compactly supported. the ordinary wavelet neural network model applications.b (xi ) function be calculated by φ(xi ) = cos(xi ) −0. x2 .5 . The structure of a compensatory wavelet neuro-fuzzy system 2. but the parameters of the CWNFS are adjusted via a proposed dynamic training algorithm. Therefore.

The structure of the CWNFS is shown in Fig. n φa. It is a ﬁve-layer structure. The output node together with links connected to it act as a defuzziﬁer. can be written in the following general form: Rj : [IF x1 is A1j and . Each node in layer 2 acts as membership function representing the term of the respective input-linguistic variables. Nodes in layer 4 only receive the signal. the output function of each inference nodes is 1−γj + Oj (3) = i Iij (2) γj n (5) where γj ∈ [0. and the link weight w jk is the output action ˆ strength associated with in the jth rule and kth Ψ a. The ﬁnal output of the wavelet neural networks is: M yj = ˆ k=1 wjk ψa. The symbol M denotes the number of wavelets. The Gaussian function is adopted as the membership function. 3. The mathematical function is . which are equal the number of existing fuzzy rules in CWNFS. We use a compensatory fuzzy operator mentioned in [12] to perform IF-condition matching of fuzzy rules. these nodes only pass the input signal to next layer. and n is number of ˆ input variables. Schematic diagram of the WNN. Lin et al.b . which is y j from output ˆ of wavelet neural network model. 2. a novel CWNFS model.-J. Each node in layer 1 is an input node. which combines fuzzy model. .b (3) where yj is the local output of the WNN for output Y and jth rule. .2.C.b (xi ) |X| (2) ψa. and xn is Anj 1−γj +rj /n M THEN yj = ˆ k=1 wjk ψa. 1] is called the compensatory degree.b = i=1 where |X| means the number of input dimension. compensatory operations. 2. and wavelet neural networks. / Temperature control using neuro-fuzzy controllers with compensatory operations 147 Fig. Layer 3 is a rule node representing the precondition part of one fuzzy logic rule. y j is output variable.b (4) where xi is input variable. A nj is linguistic term of the precondition part. Description of the CWNFS model According to [12]. As a result. The node in layer 5 computes the output signal Y .

-J. . Lin et al. The structure of the proposed CWNFS model is different from the CNFN model [12]. / Temperature control using neuro-fuzzy controllers with compensatory operations Fig. In [12]. M O(5) = Y = j=1 Oj Oj Oj (3) (4) (3) M M j=1 = j=1 (wj1 ψ0.0 + wj2 ψ1. and O (5) is the output of the CWNFS. an on-line learning algorithm. The structure learning scheme is used to decide proper fuzzy partition.0 + wj3 ψ1. the consequent part of the rules is a nonlinear function of input linguistic variables. 3. is used concurrently for constructing the CWNFS.1 · · ·)Oj M j=1 (3) (6) Oj (3) where Oj is the output of the local model of the WNN model for an output Y and the jth rule. O j is the output of layer 3. This study adopts the wavelet neural network to the consequent part of the rules. An on-line learning algorithm for CWNFS controller In this paper. it is four-layer structure and the consequent part of the rules is singleton.148 C. In the proposed CWNFS model. The local properties of wavelets in the CWNFS model enable arbitrary functions to be approximated more effectively. (4) (3) 3. Schematic diagram of CWNFS model. which consists of structure learning and parameter learning.

Layer 5: The error to be propagated to the preceding layer is −∂ 1 (Y − Y des )2 ∂E 2 = = Y des − Y ∂O(5) ∂O(5) Layer 4: The link weight of wavelet neural network is update by δ (5) = − ∆wjk = −ηw ∂Olj ∂E ∂E ∂O(5) = −ηw · · (5) (4) ∂wjk ∂wjk ∂O ∂Olj (4) (10) (3) = ηw · δ (5) · Oj ψa. Since the learning process involves the determination of the vector which minimize a given energy function. the next step is to assign initial value of the free parameters. (5) is used as the degree measure.b (x) j Oj (3) (11) where ηw is the learning rate. The structure learning scheme Initially. Geometrically. The parameter learning scheme After the network structure is adjusted according to the current training pattern. a rule is corresponding to a cluster in the input space with m ij and σ ij representing the mean and variance of that cluster.C.2. we consider the single output case. we derive the rules layer by layer. For each incoming pattern x i the strength a rule is ﬁred can be regard as the degree of the incoming pattern belongs to the corresponding cluster.-J. Based on this concept. / Temperature control using neuro-fuzzy controllers with compensatory operations 149 the number of the rule nodes and the wavelet neural networks while the parameter learning scheme is used to tune the adjustable parameters. then a new rule is generated where F ∈ (0. The detail of the learning algorithm is presented below: 3. 1) is a pre-speciﬁed threshold that should be decayed during the learning process limiting the size of CWNFS. 3. the general learning rule used is w(t + 1) = w(t) + η − ∂E ∂w (9) where η is the learning rate. Lin et al. Notice that the following parameter learning is performed on the whole networks after structure learning. Once the new rule is generated. the ﬁring strength obtained from Eq. The criterion of generating a new fuzzy rule for new incoming data is Fmax = max Fj = max Oj 1 j M 1 j M (3) (7) If Fmax F . The energy function E is deﬁned as 1 E = (Y − Y des )2 (8) 2 where Y is the model output and Y des is the desired output. no matter whether the nodes (links) are newly added or are existent originally. there are no rules in CWNFS model.12]. An input data x i with higher ﬁring strength means its spatial location is nearer the center of cluster than those with smaller strength.1. The structure learning method is similar to [8. For clarity. To show the learning rules. The gradient of the energy function with respect to the vector is computed and the vector is adjusted along the negative gradient. Assuming that w is the adjustable parameter in a nodded. the network then enters the parameter learning scheme to turn the adjustable parameters of the network optimally based on the same training pattern. Layer 3: In this layer only the error term needs to be calculated and propagated ⎞ ⎛ ∂E ∂E ∂O(5) (4) (3) (4) (3) ⎠ δ (3) = − (3) = − (5) · Oj − Olj Oj = δ (5) · ⎝Olj (3) ∂O ∂O ∂O j j j j j 1 (3) Oj 2 (12) . the ﬁrst task in structure learning is to decide when to generate a new rule.

where α and β are some constant values describing R and C. (19) into discrete-time form with some reasonable approximation.5y(k)−40 is obtained. Recently. The system parameters used in this example are α = 1. Assuming that R and C are essentially constant. β = 8.7. 4. where y ref is the desired temperature of the controlled plant. Y 0 is room temperature. / Temperature control using neuro-fuzzy controllers with compensatory operations To eliminate the constraint γ j ∈ [0. we rewrite the system in Eq.6].6. (20) 1 + e0. C is the equivalent system thermal capacity. many researchers [2. The system conﬁguration is shown in Fig. The sampling period is T s = 30. The control approach in this paper is different from [2. which evaluates the transpose of the network Jacobian y(t + 1) = e−αT s y(k) + β α (1 .0 (◦ C). The system − e−αT s ) u(k) + [1 − e−αT s ]y0 . Layer 2 : In this layer. 1].67973e−3 and Y0 = 25.-J. The inputu(k) is limited to 0 and 5 V represent voltage unit. and R is the equivalent thermal resistance between the system borders and surroundings. the error term is computed as follows: δ (2) = − ∂E ∂Oij (2) = − ∂Oj ∂E ∂O(5) · · (5) (3) (2) ∂O ∂Oj ∂Oij (2) (3) = δ (3) · i=j Oj (3) (16) The updated law of m ij is ∆mij = −ηm ∂Oij ∂Oij ) ∂E ∂E = −ηm · = ηm δ (2) · (2) ∂mij ∂mij ∂mij ∂Oij (2) (2) (2) (17) The updated law of σ ij is ∂Oij ∂E ∂E · ∆σij = −ηm = −ησ (2) ∂σij ∂σij ∂Oij = ησ δ (2) ∂Oij · ∂σij (18) where ηm and ησ are the learning rate parameter of the mean and the standard deviation of the Gaussian function. 4. we redeﬁne γj as follows: γj = c2 j c2 + d2 j j 2cj d2 j [c2 + d2 ]2 j j 2c2 dj j [c2 + d2 ]2 j j 1 (2) (3) − 1 ln Π Oij Oj n i 1 (2) (3) − 1 ln Π Oij Oj n i (13) Then we have ∆cj = ηc δ (3) (14) ∆dj = −ηd δ (3) (15) In all above formulas. Lin et al. η c and ηd are the learning rate of the parameter c j and the parameter d j . u(t) is heating ﬂowing inward the system.14] use various different methods for solving the temperature control problems.0015e −4. Chen and Pao [2] compute the derivative of the model’s output with respect to its input by means of the back-propagation process. Control of water bath temperature system The goal of this section is to control the temperature of a water bath system given by dy(t) u(t) Y0 − y(t) = + (19) dt C RC where y(t) is system output temperature in ◦ C. which were obtained from a real water bath plant in [17].150 C. respectively.

By implement the on-line training scheme for CWNFS.0321ϕ4.0164ϕ3.3778.0566) and I2 is A2. As a result. 3.0823ϕ2.2 (33.0996ϕ0.0053ϕ2.0734ϕ3.3 +0.9058.1 ˆ1 +0. the initial variance σ init = 6 and the prespeciﬁed threshold F = 0.5810.1 + 0.4 (45.0153ϕ1.0739ϕ2.0179ϕ1. which could be a neural network.1329)]0.0 + 17.0 + 0.0827.5 (56. 4.3 −0.3 −0.0036ϕ4.1 R5 : If [I1 is A1.0608ϕ0.3 (37.0463ϕ3.2 −0.0323ϕ3.0 −0.1 −0.0 + 0.0471ϕ4.1 ˆ1 +0.C.0 + 0.0842ϕ4. 5.2 + 0.1 −0.0736ϕ3.0344ϕ2.25 Then y5 = −0.5630.0027ϕ2.0592ϕ2.1 ˆ1 −0. 5.1 (26.0251ϕ1.0 + 0.0338ϕ2.0573ϕ0.5 (48.0042ϕ0. 7.2 + 0.0011) and I2 is A2. the learning rate η w = η m = η σ = ηc = ηd = 0. or another appropriate mathematical description of the plant.7886)] 0.0 −0.0806ϕ1.1 R4 : If [I1 is A1.0 + 0.1613)]0.0047ϕ3. 3.0 + 0.05.7761ϕ2.1166)] 0.0384ϕ4.1 −12.0 −0.0422ϕ1.9721.1274) and I2 is A2.0 + 0. at the network’s current input vector. After training. 7.0 + 0.2 −0.2 −0.0 −0.4 (38.0551ϕ3.2 −0. a sequence of random input signals u rd (k) limited to 0 and 5 V is injected directly into the simulated system described in Eq.2 (31.0669ϕ3.0865ϕ3.05 are chosen. 5. a neuro-fuzzy system. there are 12 fuzzy rules generated.1 −0.0497ϕ3.5963ϕ2. and the temperature rises progressively when random input signals are injected.2 −0.0 + 0.0669ϕ3.0105ϕ4.0320ϕ3.1 ˆ1 +0.2 −0.0027ϕ4. The 120 training patterns are chosen from the input-outputs characteristic in order to cover the entire reference output.1 −0.1 (24.0250ϕ4.37 Then y1 = 0.0432ϕ3. Lin et al.0962ϕ2.1 R3 : If [I1 is A1.7056ϕ0.-J.7137. which can be used to train another network to be a controller [6].93 Then y2 = 0.0 + 13.0605ϕ4. / Temperature control using neuro-fuzzy controllers with compensatory operations 151 On-Line Learning Algorithm yref (k+1) y(k) CWNFS Controlle r u(k) Water Bath System y(k+1) Z Fig.73 Then y3 = −0.0632ϕ2.2 −0.1 −0. For the CWNFS.0241ϕ3.0 −0.1 ˆ1 −0.7003) and I2 is A2.0 + 0.6024.0105.0857ϕ2.1 + 0.1 −0.0 + 0.3 (45.0022ϕ3.1 + 0. The initial temperature of the water is 25◦ C. The obtained fuzzy rules with a compensatory degree are shown as follows: R1 : If [I1 is A1.0014ϕ3.1 .9094ϕ1.0176ϕ3.0241ϕ4. 5. (20). 4.0280ϕ3.3 −0.0 −0.2 −0.0918ϕ2.9519.1185ϕ1. propagating errors between actual and desired plant outputs back through the forward model produces error in the control signal. 4. This usually implies that we need a model for the plant and the Jacobian matrix obtained from the model.7223)] 0.0 + 0.6861) and I2 is A2.0874ϕ1.0464ϕ1.0910ϕ2.0 + 0.0370ϕ3. Flow diagram of using CWNFS controller for solving the temperature control problem.0550ϕ3.0 + 2.0029ϕ1.0 + 0.94 Then y4 = −0.0 −0.1 R2 : If [I1 is A1.0336ϕ2.3 −0.0 + 0.

1 ˆ1 −0.0 −0.5759.0506ϕ3.0450ϕ3.0653ϕ4.2 −0.5346.0797ϕ3.-J.0903ϕ2. 4.8736) and I2 is A2.0246ϕ3.0926ϕ2.0747ϕ3.2 −0.2 + 0.4883.1 −0.0830ϕ0.3033.2315)] 0.6617)] 0.8 (77.0283ϕ1.0327ϕ3.0607ϕ3.7 (74.1 ˆ1 +2.1272) and I2 is A2. 5.2 + 0.0 −0.1 ˆ1 −0.6022ϕ3.1 R12 : If [I1 is A1.0 + 2.0672ϕ3. The comparison performance measures include set-points regulation.0 −0.1 R9 : If [I1 is A1.0722ϕ1.0607ϕ2.12 (34.9521)] 0.96 Then y11 = 0.2305ϕ2. KI = KTs . 5. we compare the CWNFS controller to the PID controller [16].10 is µ(29.2 + 3.0651ϕ2.9 (49.1 −0. For the PID control.0861ϕ1.2 + 0.0732ϕ3.1 −0.0597ϕ4.3 +0.2 + 1.0 + 0. K P .12 Then y9 = −0. The parameter ∆u(k) is the increment of the control input.0852ϕ4.11 (48. 3. and.6278) and I2.0 + 0. integral.11 (44.0951ϕ3.0380ϕ1.0809ϕ3. 4.0 + 0.0837ϕ2.1 + 0. 2.0553ϕ0.1611ϕ4.0252ϕ3.0 + 0.35 Then y8 = 0.0143ϕ4.3473ϕ3.0015ϕ3.2 −0.1826.0936ϕ2.3 +0.3 +0. 1.0 −0.0142ϕ3.0774ϕ2.0171ϕ4.4191.0649ϕ1. the inﬂuence of impulse noise.0567ϕ2.0614ϕ4.0907ϕ2.0763ϕ0.1 −0.0703ϕ1.0720ϕ3.6 (66.0520ϕ3.0 + 0.0 + 0.1 R7 : If [I1 is A1.1 + 0.95 Then y7 = 0.0 −0.0123ϕ1. 7.4854)] 0. and derivative .0747ϕ4.0 + 0.7 (68. 8.8 (78. / Temperature control using neuro-fuzzy controllers with compensatory operations R6 : If [I1 is A1.1 ˆ1 −0.0 −0.0 + 0.1 ˆ1 −0.3 −0.72 Then y10 = −0.8080.0500ϕ4. Lin et al.1 + 0.152 C. e(k) 2 Ti Ts is the performance error at the sampling instant k. 6.0405ϕ3.0028ϕ0.0 + 0.1 + 0. 6. a velocity-form discrete PID controller [16] is used and is described by ∆u(k) = K e(k) − e(k − 1) + Td Ts [e(k) + e(k − 1)] + [e(k) − 2e(k − 1) + e(k − 2)] 2Ti Ts = KP [e(k) − e(k − 1)] + KI e(k) + KD [e(k) − 2e(k − 1) + e(k − 2)] (21) where K P = K − 1 KI .8565.0324ϕ0. the manually designed fuzzy controller and the self-constructing fuzzy neural network (SCFNN) [15].1 + 0.2 −0.5 Then y12 = −0.0757ϕ2.5138)] 0.0 + 0.0 + 0.1 −7.1 −0.0 −0.0521.0 −0. 3.0306.3 +0.2 + 0.0476ϕ4.2 + 0. KD = KTd .0893ϕ3.0193ϕ2.0404ϕ2.0 + 0.0 −0.0096ϕ1.10 is µ(28.12 (35.7451ϕ3.0155ϕ2.0655ϕ4.1 R8 : If [I1 is A1.3 −0.0 −0.6648) and I2 is A2.5007) and I2 is A2.0 + 0.0209ϕ3.0 + 2. Each of the three controllers is applied to the water bath temperature control system.1 ˆ1 −0.1 In this paper.1 + 0.0880ϕ1.0369ϕ2.0710ϕ4.6011)] 0.0851ϕ3.1 + 0.2 + 0.1 R10 : If [I1.0019ϕ3.41 Then y6 = −0.0 −0.1 −0.0 −0.0133ϕ0.0382ϕ1.6 (62.9817ϕ4.0522ϕ1.9791.0126ϕ0.9042.0 + 0.3 −0.0707ϕ3.0978ϕ2.5460ϕ3.0416ϕ4.0943ϕ1.1094.0315ϕ2. 5.2 −0.2144. and a large parameter variation in the system.3051) and I2 is A2.0 −0.0 + 0.0186ϕ1.0 + 0.1 ˆ1 −0.0739ϕ2.0223ϕ2. KI and KD are the proportional.0 −0. 6.0520ϕ3.7519)] 0.0 + 0.0699ϕ3.2 + 0.0858ϕ1.0107ϕ2.0 + 0.9 (54.4774) and I2 is A2.0594ϕ2.1 R11 : If [I1 is A1.0 + 0.

The second set of simulations is carried out for the purpose of studying the noise-rejection ability of the four controllers when some unknown impulse noise is imposed on the process. Lin et al. a fuzzy controller has some scaling parameters to be speciﬁed. i. Each fuzzy term is speciﬁed by a Gaussian membership function. The SAE values of the CWNFS controller. Seven fuzzy terms are deﬁned for each linguistic variable. One impulse noise value −5 ◦ C is added . Recently. the PID controller. respectively. and adopting the wavelet neural network as consequent part of each fuzzy rule. respectively. The structure of the proposed CWNFS controller is difference from [15]. An on-line learning algorithm was proposed to decide the structure of fuzzy rules and turn the adjustable parameters through the backpropagation algorithm. For the above designed PID controller. The ﬁrst task is to control the simulated system to follow three set-points. In this ﬁgure. / Temperature control using neuro-fuzzy controllers with compensatory operations Table 1 Fuzzy rule table formulated for the water bath temperature control system NL Change error. is deﬁned by SAE = k |yref (k) − y(k)| (23) where yref (k) and y(k) are the reference output and the actual output of the simulated system. 401. the CWNFS controller obtains smaller errors than the SCNNF controller. The consequence of each fuzzy rule is a singleton value. GCE. The output or the controlled linguistic variable is the voltage signal u(t) to the heater. using the compensatory operation in layer 3.-J. These fuzzy terms consist of Negative Large (NL). the change in error. Each node in layer 3 performs the product operation. respectively. We choose these parameters as follows: GE = 1/15. The other two parameters K P and KI are chosen as 80 and 70. Our model is ﬁve-layer structure. GCE = 1/15. In order not to aggravate noise in the plant. PID controller. Each simulation is performed over 120 sampling time steps. and GU. manually designed fuzzy controller and SCFNN controller). The regulation performance of the CWNFS model is shown in Fig. corresponding to the process error. For the manually designed fuzzy controller. the fuzzy controller and SCFNN controller are 352. we have tried our best to achieve their respective best performance through several trial-and-error experiments. for k 40 (22) yref (k) = 55◦ C. and the controller’s output.41.. for 40 < k 80 ⎩ ◦ 75 C. Negative Small (NS). and 356. The SCFNN controller is a standard four-layer structure. 418. K D is set to zero in the water bath system. We also test the regulation performance by using SCNNF controller [15].C. the input variables are chosen as e(t) and ce(t). Zero (ZE). The output node sums all incoming signals to obtain inferred result. Like other controllers. respectively. Positive Medium (PM). They are GE.5. and Positive Large (PL). For the aforementioned controllers (CWNFS controller. ⎧ ◦ ⎨ 35 C. a performance index. 5(a). e(t) NS ZE PL PM PS PS NS ZE NS NS NM NL PS PL PM PS PS NS PM PL PM PM PM PL PL PL PL PL 153 NL NM parameters. for 80 < k 120.e. Lin et al. [15] presented a self-constructing fuzzy neural network (SCFNN) for control problems. Negative Medium (NM). which are shown in the ﬁrst row of Table 2. 25 fuzzy rules are speciﬁed in Table 1. ce(t) PL PM PS ZE NS NM NL NM Error. sum of absolute error (SAE). According to common sense and engineering judgment. To test their regulation performance. The error curves of CWNFS controller and SCFNN controller between k = 80 and k = 100 are shown in Fig.95. only a two-term PID controller is used. three groups of computer simulations are conducted on the water bath temperature control system.5. where e(t) is the performance error indicating the error between the desired water temperature and the actual measured temperature and ce(t) is the rate of change in the performance error e(t). GU = 450. 5(b). Positive Small (PS).

(a) Final regulation performance of the CWNFS controller for water bath system.5. One common characteristic of many industrial-control processes is that their parameters tend to change in an unpredictable way. A set-point of 50 ◦ C is used in this set of simulations. a value of 0. training data and learning parameters are used as those used in the ﬁrst set of simulations. (b) The error curves of CWNFS controller and SCFNN controller between k = 80 and k = 100. The SAE values of the CWNFS controller. For . For the CWNFS controller. / Temperature control using neuro-fuzzy controllers with compensatory operations (a) (b) Fig.154 C. to the plant output at the sixtieth sampling instant. the fuzzy controller. 275. 6(a)–(b). It recovers very quickly and steadily after the presentation of the impulse noise. the PID controller.7 ∗ u(k − 2) is added to the plant input after the sixtieth sample in the fourth set of simulations. Lin et al. and SCFNN are 273. To test the robustness of the four controllers. 5. and 280. A set-point of 50 ◦ C is performed in this set of simulations. the same training scheme.-J. which are shown in the second row of Table 2.8.25. 311. The behaviors of the CWNFS controller under the inﬂuence of impulse noise and the corresponding errors are shown in Fig. It is observed that the CWNFS controller performs quite well.5.

5 SCFNN controller [15] 356. 7(b). the same training scheme.25 262.51.5 322.-J.5 275. / Temperature control using neuro-fuzzy controllers with compensatory operations Table 2 Performance comparison of various controllers SAE = 120 k=1 155 |yref (k) − y(k)| CWNFS controller 352. The SAE values of the CWNFS controller. The corresponding errors of the CWNFS and SCNNF controllers are shown in Fig.95 273. The behaviors of the CWNFS controller when there is a change in the plant dynamics are shown in Fig. .41 280.2 Regulation performance Inﬂuence of impulse noise Effect of change in plant dynamics manually designed fuzzy controller 401.8 273. the PID controller.C. (b) The error curves of CWNFS controller and SCFNN controller. Lin et al.5 311. (a) Behavior of the CWNFS controller under the impulse noise for water bath system. and SCFNN are 262.2. 322. the GA-NFS controller. training data and learning parameters are used as those used in the ﬁrst set of simulations. 7(a).51 PID controller [16] 418. 6.50 270. the fuzzy controller.21 (a) (b) Fig.

5. and 270. Table 2 has shown that the proposed CWNFS controller has better performance than that of other methods. / Temperature control using neuro-fuzzy controllers with compensatory operations (a) (b) Fig.21. As for the PID controller. (a) Behavior of the CWNFS controller when a change occurs in the water bath system.156 C. For the fuzzy and PID controllers. and KD also have to be decided properly. Lin et al. they usually require a long time in design for achieving good performance. 7. The results show the good control and disturbance rejection capabilities of the trained CWNFS controller in the water bath system. no controller parameters have to be decided in advance. in general. the numbers of rules and membership functions have to be decided and tuned by hand. For the aforementioned simulation results. Although the structure of CWNFS controller is more complicated than the fuzzy and PID controllers. (b) The error curves of CWNFS controller and SCFNN controller. In the CWNFS controller. We only need to choose propose training patterns of the CWNFS controller. This study attempts to emphasize . 273. which are shown in the third row of Table 2. the CWNFS controller usually spends a relatively short time in design for achieving good performance. the parameters K p . therefore. KI . For the fuzzy controller.-J. however.

Finally.Y. Temperature control with a neural fuzzy inference network. A Wavelet-Based Neuro-Fuzzy System and Its Applications. J. and Cybern. Lin and C.Q. C. . 2309–2316. C.C. Process Control by On-Line Trained Neural Controllers. Acknowledgement This research is supported by the National Science Council of R. Man.-S. of International Conf.C. C. T.J. V. on Syst. Proc. Kandel. NJ: Prentice-Hall. Man. C. Neuro-Fuzzy and Soft Computing.T. Tanomaru and S.L. Orthonormal Bases of Compactly Supported Wavelets. Lin.J. Xu. Proc. 1995. Adaptive Control. Prentice Hall. Nagle. Reading. Lee.H. Genetic reinforcement learning through symbiotic evolution for fuzzy controller design. on Fuzzy Systems 9 (2001).F. Lin. 751–759. IEEE Trans. Electron.. Zhang and A. Nonlinear System Control Using Compensatory Neuro-Fuzzy Networks. Omatu.J. 5. 1997. 41 (1998).J. C. Wittenmark.H. Wavelet Neural Networks Employing Over-Complete Number of Compactly Supported NonOrthogonal Wavelets and Their Applications. Communications and Computer Sciences E86-A (2003). Pur. IEEE Trans. Yamakawa. Lin and C. Lee. 83–105. on Systems. on Systems. Juang. 1448–1453. Conclusion In this paper.H. 1 (1993).G.R. 1989. Samatsu. computer simulation results have shown that the proposed CWNFS controller has better performance than that of other methods. Pao. Phillips and H. In the future. C. ANFIS: Adaptive-Network-Based Fuzzy Inference System. on Fuzzy Syst.H. 200–211. on Ind. Li. Learning Control with Neural Networks. is proposed and applied to the water bath temperature system. Fuzzy Logic in Control Systems: Fuzzy Logic Controllers – Parts I.F.S. Fuzzy Control of pH Using Genetic Algorithms. Juang. Jang. Chen. Oregon. IEEE Int. Y.T. and Cybernetics C29 (1999). 511–521. 2003. MA: Addison-Wesley. we will apply the proposed CWNFS controller on a real water bath temperature control system. 20–24. IEEE Trans. 1391–1396. Ho. Karr and E.A. A TSK-type recurrent fuzzy network for dynamic systems processing by neural network and genetic algorithms.L. IEEE Trans. Appl. D. E. and Cybernetics B30 (2000). Sun and E. which combines TSK-type fuzzy model and wavelet neural networks.T. I. C. 1996. Lin et al. of IEEE Conf. Lin.F. 46–53. Gentry. Lin and C.T. Neural Fuzzy Systems: A Neuro-Fuzzy Synergism to Intelligent System. Self-Constructing Fuzzy Neural Network Speed Controller for Permanent-Magnet Synchronous Motor Drive. Man. Math. Digital Control System Analysis and Design. IEEE Trans.T.J. on Syst. J. Zhang and J. Man. Jang. II. Joint Conference on Neural Networks. under grant NSC 93-2213-E-324-008. 440–451.C. 404–435. C.C. The CWNFS controller can automatically construct and adjust free parameters itself by performing online supervised structure/parameter learning schemes concurrently. 290–302. Prentice-Hall. Compensatory Neurofuzzy Systems with Fast Learning Algorithms. IEICE Transactions on Fundamentals of Electronics. / Temperature control using neuro-fuzzy controllers with compensatory operations 157 the methodology and control abilities of the proposed CWNFS model. 20 (1990). Shen. 23 (1993). J. Lin and C. P. 9 (2001). J. Daubechies.. IEEE Transactions on Neural Networks 9 (1998). Juang and C. Portland. C.R. IEEE Trans.W. on Robotics and Automation. on Fuzzy Systems 10 (2002). 1989. on Neural Networks 3 (1994). Åstrom and B. Fuzzy Syst. References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] K. 17. IEEE Trans. C. 665–685. IEEE Trans. 155–170. Chen and Y. F. Cybern.C. 39 (1992). IEEE Trans. Mizutani. Fuzzy Wavelet Networks for Function Learning. Comm. the new CWNFS controller. Chin. C.C.O. Uchino and T. Lin and P.P.-S. C. Proc. Ch.-J.

- Fuzzy Logic (CT430)
- 6
- Diagonal Based Feature Extraction for Handwritten Alphabets Recognition System Using Neural Network
- Long Term Behavior Monitoring of an Arch Dam Using Adaptive Neuro-Fuzzy Model Technique
- Artificial Neural Networks
- On the diculty of training recurrent neural networks
- Neural Net 3rdClass
- Produccion de Urea
- Fuzzy numbers & Fuzzy Equation.pptx
- IJETTCS-2013-08-20-103
- K0266975.pdf
- APLICACION NEURONAS ARTIFICIALES
- Session 01_Paper 07
- Neural Networks
- A Comprehensive Analysis of Neural Solution to the Multitarget Tracking Data Association Problem
- Ensemble Neural Networks Using Interval Neutrosophic Sets and Bagging, by Pawalai Kraipeerapun, Chun Che Fung and Kok Wai Wong
- Grey Verhulst Model Based on BP Neural Network Optimization for Oil Production Forecasting
- 5o-239-benali
- Report
- nfs12
- CS4001 Neuro-Fuzzy Systems Lect6
- NFQ Reference guide
- Sensor based artificial neural network
- 05568614
- Watermarking
- UiTM Infoday Presentation Demo
- Research
- Wavelet transform
- 10.1.1.11.852
- 14790

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd