Design Of Fuzzy Neural Network Based Control System For Cement Rotary Kiln.pdf

201O 2nd International Asia Conference on Informatics in Control, Automation and Robotics

Design of Fuzzy Neural Network Based Control System for Cement Rotary Kiln Zheng LI School of Electrical Engineering and Information Science Hebei University of Science and Technology Shijiazhuang, China [email protected] Abstract-This paper presents a fuzzy neural network control system for the process of cement production with rotary cement kiln. Since the dynamic characteristics and reaction process parameters are with large inertia, pure hysteresis, nonlinearity and strong coupling, a fuzzy neural network controller combining both the advantages of neural network and fuzzy control is applied. This fuzzy neural network controller adjusts the parameters of membership functions in order to acquire the required control performance. The fuzzy neural network is an adaptive neural network whose parameters can be corrected by learning algorithms automatically. The main control system structure includes three control loops as the pressure control loop, the burning zone control loop and the back-end of kiln temperature control loop. The simulation results show the effectiveness of the control scheme with quick response time and lower overshoot, also with small temperature deviation.

the temperature higher; and when the feed volume of raw material increases, the reaction material in the kiln is added to make the temperature higher. But when the temperature increases to a certain value, since the material can not get a fully reaction, the temperature of inner kiln drops. So the input volume of inner kiln material should be in a certain proportional relationship with the feed volume of coal to make them in a fully reaction state. The feed volume of coal and raw material are controlled by the speed of coal feed motor and raw material motor respectively. The rotary kiln should be remaining a micro-negative pressure state, because in the positive pressure state, the ventilation is poor and the fuel can not be burned completely; in the large negative pressure state, the fast ventilation will take away the heat. The inner pressure of kiln is controlled by the speed of flue blower. The whole system can be shown as figure 1.

Index Terms-FNN, neural network, control system, cement rotary kiln

I.

Head-end of the kiln

INTRODUCTION

CoollnC

Zoot

BwrinC

Zoot

Reac1ion

Exothermic Zone

Decomposilt

Zone

Dryinc Ptt-htatlnc Raw Material Zoot Zone

Coal Powder

In the process of cement production, the rotary kiln

calcination is the most important technology link which includes complicated physical and chemical reaction process with large inertia, pure hysteresis, nonlinearity and strong coupling characteristics. It is hard to derive the exact mathematical model and can not reach satisfied results with conventional control algorithms [1-5]. Up to now, the cement rotary kilns are mainly controlled manually or semiautomatically, which is based on the experience of operators to attain acceptable performances with low production rate. This paper presents the application of fuzzy neural networks to implement the monitoring, analysis and optimization based on the field bus technology to the conventional cement production defects. II.

Figure I. Cement rotary kiln system

III.

A.

System Structure ofthe Cement Rotary Kiln Figure 2 denotes the control system of cement rotary kiln. The control links contain the burning zone temperature (feed volume of coal control), back-end of kiln temperature (feed volume of raw material control), and the inner pressure (blower speed control) three parts [6]. The control system contains AID, D/A converter and I/O modules together with a number of sensors or transformers. There are three control loops in the system, which are the pressure control loop, burning zone control loop and back-end of kiln temperature control loop. The

ANALYSIS OF CEMENT ROTARY KILN

Cement rotary kiln thermal system decides the production, quality and energy consumption. There are several factors impact the thermal system of rotary kiln, including the rotation speed, the feed volume of coal, the feed volume of raw material and inner pressure of rotary kiln. When the kiln rotation speed increases, the temperature drops slightly and usually the speed is kept constantly. When the feed volume of coal increases, the reaction of decomposition furnace can be exacerbated to make

ack-endorlhe kiln

_---t-~

FUZZY NEURAL NETWORK SYSTEM DESIGN

290 CAR 2010

PID neuron network control algorithm is implemented by the IPC of the highest level in the system. IPC PID Neuron Networks Controller

Modules

€tHau.

where Xi' is the command error between the desired command and the actual output value, N denotes the number of iterations,

y:

Field Bus

I

Layer 2: Membership Layer

Convert

In this layer, each node performs a membership function. The Gaussian function is adopted as the membership function. For the jth node

va AID

is the output of the FNN.

L....::Dor/A,----------,

2

net~ (N) = y~(N) = If(net~ (N)) where Figure 2. Cement rotary kiln control system

mij

and

(T ij

(x j

- mij )2

(3)

(Tij)2

= exp(net~(N)),j = 1,...,n

(4)

are, respectively, the mean and the

standard deviation of the Gaussian function in the jth term of 2

Fuzzy Neural Network Control System The fuzzy neural network control system can be shown as in figure 3. In the figure,yi(k) (i=1,2,3) denote the actual output of the burning zone temperature, back-end of kiln temperature and the inner pressure respectively. ydi denote the desired output of the system. The fuzzy neural network is tuned by online learning algorithm. The 3 sub-networks are connected in parallel. The controlled objects of this system have 3 inputs and 3 outputs.

the ith input linguistic variable x j to the node of layer 2, and n is the total number of the linguistic variables with respect to the input nodes.

B.

Layer 3: Rule Layer Each Each node k in this layer is denoted by II, which multiplies the input signals and outputs the result of product. For the kth rule node

net~(N) =

n

w~kx~(N),

(5)

j

y~ (N)

= I: (net~ (N)) = net~ (N), k = 1,...,1

(6)

where X~ represents the jth input to the node of layer 3; W~k' the weights between the membership layer and the rule layer,

iy

are assumed to be unity; 1 = (n / is the number of rules with complete rule connection of each input node has the same linguistic variables.

YoO

y'

Layer 4: Output Layer The single node a in this layer is labeled with L, which computes the overall output as the summation of all input signals Figure 3 Cement rotary kiln fuzzy neural network control system

net: (N) y: (N)

ROht')'

, ,2,3

0

= 1,2,3

wto is the output action strength

of the 0 th output associated with the kth rule;

xt represents

the kth input to the 0 th node of layer 4. The output of the FNN is the is the controlled signal.

(1) (2)

Kiln

y'

291 Tr"' ....notu.... Mrllsu........."f

= 104 (net: (N)) = net: (N)

where the connecting weight

• and the net

Cement

(7)

(8)

Layer 1: Input Layer

yo

wtoxt

k

A four-layer FNN as shown in Fig. 4, which comprises the input (the i layer), membership (the j layer), rule (the k layer) and output layer (the a layer), is adopted to implement the FNN controller in this study. The signal propagation and the basic function in each layer of the FNN are introduced as follows [7]:

Tr"' ....nofu.... ,\Ira'o...." ...",

=I

To described the online learning algorithm of the FNN using the supervised gradient decent method, the following equation is dermed as

The whole fuzzy set can be derived from every fuzzy input variable using the maximum belonging function. The intensity of each rule can be written as

OJk =PA(x\)APB(X 2) ,

~

So by inference, the result of each rule can be derived as

ak

~

y:

~

y;

~

= OJkApc; (u)

(15)

The fuzzy determination method is based on the weighted average method and the exact control variable u can be derived from

y;

(16)

~o

IV.

_':__ 0

Figure 4. Structure ofFNN

TABLE I.

The purpose of BP neural network algorithm is to fmd the learning parameters, as the last layer connection weights OJdk ' the center value of the belonging function and width value, making J the minimum value of complete process.

NO

The adjustment of weights adopts the first order of gradient method, so

dOJ dk

0.1 = -lJ-= lJY; (1- Y; )(Ydi OOJdk

d

Y; )Ok

= lJ8k O kd (10) (II)

8 k =Y;(1-Y;)(Ydi -Y;)

0.1

CONTROL SYSTEM SIMULATION

In the previous section, a fuzzy neural network controller for the rotary cement kiln has been developed. In this section, it is tested on the simulation model. Based on the above training data of real system, 1000 sets data points are used for training the neural network model as shown in table I. In MATLAB simulation programs, the learning effects are simulated as shown in figure 5.

- ) I

(12)

uning process.

SOMERARAMETERS DATA OF REAL CEMENT ROTARY KiLN

Wind Speed (m/s)

Rotatio n speed of coal motor

Bumning zone temperatur e

Raw material motor speed

(rim)

(0C)

(rim)

Backend of kiln temper ature

Kiln rotatio n speed(r

(0C)

1m)

I

24.13

643.15

1378.23

1234.61

670.12

464

2

22.69

638.26

1363.21

1219.86

659.36

471

3

23.88

651.76

1354.47

1310.37

671.23

479.43

4

24.35

649.38

1321.83

1287.69

668.32

473.24

5

25.23

658.36

1409.91

1269.71

672.43

469.56

6

24.76

652.82

1405.68

1295.62

668.85

472.11

...

...

...

...

...

'"

...

>e accelerated by >n can be written

'OJdk(t-l)] (13) t relations of one

)----'-'y;'---

y; C k

i=I,2, ... ,7;

o

The given input of the system is the step function T] = [t( (k),t z (k),t) (k)y = [1,O.8,O.6f , with training 100 times, and each time is accompanied with 300 sampling points. The controlled object deal with quantization (normalization), the back-end temperature of the kiln is 0 °C ~ 600 °c , corresponding to 0-1, with sampling period of T=2ms; the burning zone temperature is 0 0 C ~ 1400 °C , corresponding to 0-1, with sampling period of T=2ms; the inner pressure OP-20000P, corresponding to 0-1, with sampling period of

T=2ms.

~o

_x~_ _

(14)

}

292

Perfonnance is 0.00575525, Goal is 0

10' r-~-~-~~-~-~~-~~~---,

0.9 0.75 0.6

~ ~

0.45 0.3 0.15

10" '--------'-_----'-_'--------'-_----'-_'-------'-_---'-------''------J

o

If stop Training 1

10

20

30

40

50

60

70

80

90

100

0 0

100 Epochs

Figure 5. Error curves of network training process

20

40

60

80

100S

(c)

In the second condition, the initial step response is 1; = [I, (k),12(k),13(k)f = [0.65,0.76,0.85t ' and the given value

Figure 6. System step response curves after 200 times training (a) t1(k) , (b) t2(k), (c) t3(k)

changed to 1; = [I, (k),12(k),13(k)f = [0.65,0.7,0.68t for testing the control ability at the time of 35s, trained with 100 times and each time is accompanied with 300 sampling points. After 200 times of training, the response curves of the system can be shown as figure 6.

From the results, the presented control system operated steadily with a quick response time and lower overshoot, also with small temperature and pressure deviation, which proves the effectiveness of the control scheme. V. CONCLUSION

0.7 0.6

The paper presents the application of fuzzy neural networks as controllers to control the temperature and pressure of the cement rotary kiln, and simulation results based on MATLAB/SIMULINK were derived. The results show that the presented method can reach satisfied performance by network learning and self-adjusting without establishing accurate mathematical model and independent on the model identification. It is expected that the presented method can give a reference for further development of advanced control schemes of rotary cement kiln.

f"--./

0.5

~

;: 0.4 0.3 0.2 0.1

REFERENCES

0 0

20

40

80

60

100 S

[1]

[21

(a) 0.8

[3]

0.7

[4]

0.6 0.5

~

<;:J 0.4

[5]

0.3 0.2

[61

Luyben W L. Process Modeling, Simulation and Control for Chemical Engineers. New York: McGraw-Hili, 1990.

[71

F. 1. Lin, W. 1. Huang, and R. 1. Wai, "A supervisory fuzzy neural network control system for tracking periodic inputs," IEEE Trans. Fuzzy Systems, vol. 7, no. I, February 1999, pp. 41-52.

0.1 0 0

20

40

60

80

100S

(b)

293

Mintus F, Hamel S, Krumm W. "Wet Process Rotary Cement Kilns: Modeling and Simulation". Clean Techn. Environ. Policy, 2006,8(2): 112-122. Mujumdar K S,Arora A,Ranade V V. "Modeling of Rotary Cement Kilns: Applications to Reduction in Energy Consumption". Ind. Eng. Chern. Res., 2006,45(7): 2315-2330. Mujumdar K S,Ranade V V. "Simulation of Rotary Cement Kilns Using a One-Dimensional Model," Chemical Engineering Reasearch and Design, 2006, 84(A3): 165-177. Mujumdar K S,Ganesh K V,Kulkarni S B,Ranade V V. "Rotary Cement Kiln Simulator (Rocks): Integrated Modeling of Pre-Heater,Calciner, Kiln and Clinker Cooler," Chemical Engineering Science, 2007, 62(9): 2590-2607. Wang Z,Yuan M Z,Wang B,Wang H,Wang T R. "Dynamic Model of Cement Precalcination Process," Proceedings of the 27 th lASTED International Conference on Modeling, Identification, and Control. Innsbruck,Austria; lASTED, 2008: 160-165.