JPH02161501A - Process control method using neural circuit network model - Google Patents

Abstract

PURPOSE:To input an objective value and a process variable to a model and to find out an optimum manipulated variable by identifying an identifying neural circuit network model by a process I/O, inputting a test signal to the model, linearizing a factor, and operating a parameter of the model from the linearized factor. CONSTITUTION:A learning rule 2 self-organizes a model by using the I/O information of the process 1 as a teaching signal, and a back propagation learning rule e.g. is used. The neural circuit network model 3 for identifying he process 1 is modeled by a multi-layer perception type neural circuit network model. The parameter of the model 3 is adjusted by the learning rule 2. A test signal is applied to the identified model 3 and the parameter of a neural circuit network model 5 for resolving an optimization problem is found out from the factor of a linearizing model, an objective value and a process variable. After determining the parameter, the model 5 is started from an proper initial state to output an optimum manipulated variable.

Description

DETAILED DESCRIPTION OF THE INVENTION <Industrial Application Field> The present invention relates to a novel process control method using a neural network.

<Prior Art> The application of neural network models to control systems is currently being realized in two ways: identification and solving optimization problems.

Identification means modeling the controlled object, that is, finding a thread that produces the same output when given the same input as the controlled object (1
Normally, when designing a control system for a certain process, the object to be controlled is first modeled, but the amount of work required for this is enormous. Also, even after a huge amount of work, can you still find the model you want? ), but it has been shown that a model can be obtained relatively easily by using a neural network model and making full use of various learning methods.

In order to perform appropriate control, it is necessary to solve an optimization problem. It has been shown that by using the 1lopfeld model, which is a type of neural network model, certain l1jL optimization problems can be solved using the "principle of energy minimization." It is conceivable to use this to find the optimal operator in a process control system.

<Problem to be solved by the invention> However, the model based on neural networks is not a linear model, so it is established in control theory! To 3i! There was a problem that i-control and station equipment methods could not be used, and therefore, control could not be performed just by obtaining a model.

In addition, solving optimization problems using the Hopfield model is limited to cases in which the objective function can be expressed in a quadratic form of the state variables of neurons, so there is also the problem that there are limited cases in which it can be applied to control methods.

<Objective of the Invention> The object of the present invention is to provide a general method for controlling processes using the learning function, self-organizing function, and optimal solution search function that are characteristics of the neural network model.

In order to solve the above problems, the present invention identifies a neural network model for identification using the input and output of the process to be controlled, and tests the identified neural network model. Input the signal and linearize it, calculate the parameters of the l1opfield type neural network model for solving the optimization problem from the linearized coefficients, and input the target value and professional override to this 1lopfield type neural network model. The plan was to request that Uni be destroyed.

<Example> Fig. 1 shows the configuration of a process control method using a neural network according to the present invention. In Fig. 1, 1 is a process to be controlled, and a manipulated variable is input to control the process Output. It is assumed that the time constant of this process is 1 minute shorter than the control period, and that the influence of the previous operation settles down during the control period. Also, considering a process with multiple inputs and multiple outputs, and expressing it in 5,1ift distributed time form, W k,1 = IF + 11k)...
...(1) k: Sun-wable time utk: N-dimensional input (manipulated amount) petal (l) 1.
・・・・un) y, 2M-dimensional output (process amount) Peter Le(y ・
...yH) 1.゛: Does not represent the input/output characteristics of the process, and is expressed as a function vector. 2 is a learning law that self-organizes a model using the input/output data of process 1 as a teacher signal; for example, a pack propagation learning law is used. Back 10 baggage learning rules are, for example, D.
elhart, G., [,1linton and R.
Willians, J., “Learning Repr.
esentations by Back-Propag
"Ating Errors" Nature Vol.
323, 533-536 (1986). 3 is a neural network model for identification of process 1,
It is modeled using a multilayer percept type neural network model.

The parameters of this identification neural network model 3 are adjusted according to the learning law 2. The identification neural network model 3 is expressed by the following equation (2).

y/+1=NN(ILIk) ・・・・・・(2)N
N is +1 for y/, which represents the neural network, and Iuk is the same as equation (1).Next, as shown in 4, apply a test signal to the identified neural network model 3 and calculate its input/output. The sensitivity is linearized, and the parameters of the Hopfield neural network model 5 for solving the M optimization problem are calculated from the coefficients, target values, and process quantities of the linearized model. Once the parameters are determined, the neural network model 5 is activated from an appropriate initial state to output an appropriate manipulated variable.

11opfiel [F-shaped neural gyrus! i! State MW and output value vi of the first knee 1-17 in the 8-network model
is the following formula % ′1゛, ・, from the 1st neuron to the 1st uni IJ
'Strength of synaptic connection to one Ron 1;: Bias value g of i-th degree neuron']: For example, sigmoidal analog nonlinear monotone like (1/2) [1+tanh(x/xo)] It is an increasing function and represents a function that approaches a unit j-step function when xO→0.

In this model, the strength of the synaptic connection is ′]゛丙；1゛J
When there is symmetry as in the following equation (5)
It has the property of spontaneously changing its state so as to always minimize the energy function E expressed by .

(5) On the other hand, the purpose of control is to bring the output y closer to the target value, in other words, to minimize the evaluation function of the following equation (6), The purpose is to find the manipulated variable 11.

J ”” It yk+1-112 ・・・・・・・・・
...(6) Therefore, express the problem so that equation (6) matches equation (5), set the synaptic connection strength T, , 1, and calculate the above from an appropriate initial value ■. (3) and (4) It is sufficient to repeatedly calculate the equations and converge the evaluation function J to the minimum optimal state.

In order to perform concrete calculations, the amount of operation must be expressed numerically by neurons, and for that purpose, the operation Ji u ,
u 2...UH is expressed as
Expressed as Ron's output value.

■ ・: State variable 1, J which is the output value of the neuron, 0 or 1. The subscripts i and j are the neuron numbers indicating the j-th second and l-mouth representing the manipulated variable 13.

ul is −L≦u1≦1. can only accommodate integer values, but
By appropriately converting the scale, the problem can be reduced to a general problem. In order to make the equations (5) and (6) match, +1 to y/ must be expressed in the quadratic form of the neuron's state variable V.

1'' Therefore, linearize the above equation (2), y(,1= (aNN/c)Uk) ・1uk=Z ・
+u (・・・・・・・・・ (8) Place < , (lj L5.) N N / => I
II h (= Z ) is an MXN matrix representing the sensitivity from (Ilk to yk+1. That is, its l'h
, il element Zhi represents the collapse from Ill i to yIk+1, and if we set the sentencing vector of Z as ~i, then N −tNN (u +a−e,) −NN (uk
)) + k ++ / α ) ...... (9) Calculate. However, α・ is the variation amount of U when calculating the sensitivity, and e is the N-dimensional i-th unit vector I, which is actually obtained by adding the dest signal to the identified neural network model 3 for identification. .

Next, in order to incorporate the constraint conditions for optimal control into the evaluation function of equation (6), a penalty function is added to this evaluation function. In other words, replace with . A (M: number) is a weighting function for the penalty function.

After making the above preparations, the above-mentioned formulas (5) and (6) are compared. Since we changed the expression of neuron scent in equation (7) above, if we rewrite equations (3) to (5) above using this expression, Meeting-AN+ , / 2 ・・・ ・・・ ・・
・ (14) When the coefficients of equations (13) and (14) are ratio-pruned, we get "'i, j; i', J"""Σ(Zhizh
i"^・l +2Aδ(i, J) δ(1-, J") ・・・・・・・・・ (15) ・・・・・・・・・(13) The above By substituting equations (7) and (8) into equation (IO), the following can be obtained.

FIG. 2 shows a flowchart of the entire algorithm.

First, the manipulated variable is output using conventional PID control, etc., and the process input/output data y/ is +1
, jug 6 Next, this person output data “k+
A neural network model 3 for identification is determined by applying a learning law such as backpropagation using 1-cLlk. That is, NN of the above equation (2) is determined. Once the model is identified, a test signal is added to the identified model to check its sensitivity, and the coefficients of the linear model of the identification neural network model 3 are determined. That is, Z in the above equation (8) is determined. Next, from this coefficient Z, the parameters of the neural network model 5 for the optimization problem, that is, the synaptic connection strength 'I' and the bias value I', are determined by the equations (15) and (16). Next, from the appropriate initial value? Equations (3) and (4) are calculated an appropriate number of times to obtain the output value V of the new X7, and the optimum operation amount is obtained by applying equation (7). Since the equations (3) and (4) above automatically minimize the energy function equation (5), the evaluation interval number <10) equation also automatically becomes the minimum. In this way, the optimum manipulated variable can be determined.

Next, we will show an example in which the process control method using this neural network is applied to the widthwise thickness profile control of a paper machine. The profile of the opening degree of the slicing lip, which is the pulp discharge 11, is controlled so that this profile corresponds to the standard value. In other words, if the opening vector of the slice lip is +1 to +1, and the absolute dry basis weight vector is l1l (then, So1. ""ryl, 3'2,...
...,yH]".Here, A is the interference matrix, and if you change the opening degree of the switch knob in a certain part,
Since the absolute dry basis weight of the surrounding area also changes, it usually becomes a band diagonal matrix as shown below.

In order to make such a process non-interfering, the variation ΔU of the opening degree manipulation amount of the slice lip is determined as follows.

ΔIll = G (s) ・M・Δy7kc Δyk: Deviation between target value and measured value However, in such a method, (1) the shape of the interference matrix A differs depending on the location, or the positional correspondence shifts, so the (17) Formula % Formula % (2) Since the model is expressed deterministically, the dynamic characteristics of the process are not considered.

(3) M may not be a symmetric band diagonal matrix like A.

There are drawbacks such as. Therefore, by applying the control method using the neural network model described above, these drawbacks can be eliminated. In this case, a berceptron type neural network model is used as the neural network model 3 for identification. The reason for using this is that the interference characteristics of the paper machine closely resemble a neural network model with lateral inhibitory connections. In addition,
Lateral inhibitory connections have positive synaptic connections to the bifurcated-ron of the human tvA point near chamber and negative synaptic connections to both +1117 neurons. This is considered effective for explanation. Since the control method is the same as the method described above, the explanation will be omitted.

In this type of control, it is possible to perform control without idealizing the process characteristics such that they are linear and the −1-interval characteristics are symmetrical, diagonal, and uniform, so that higher quality control can be achieved.

In addition, the learning law uses CP in addition to backpropagation.
N (C01lnlQr PrOpagaf!On N
ejWOrk), A11T(Ha+1apHVe l
1eSDallCe rheory), etc. can be applied.

In addition, the lloρfield model automatically converges the energy function to a minimum, but there is no minimum value, so it searches from some initial values and evaluates the required operation h: quantity. Optimize the one that minimizes the function 1
1: It may be output as a quantity.Furthermore, it may be calculated using the previous manipulated quantity as an initial value.

<Effects of the Invention> As explained above in detail based on the embodiments, in this invention, a neural network model for identification is identified using the input/output of a process, and a test signal is applied to this identified model. Then, the parameters of the neural network model for solving the optimization problem were determined from the coefficients of this linearization. Therefore, a process model can be easily obtained even for a multi-input multi-output system, and f&appropriate manipulated variables can be determined.

Furthermore, compared to the fi'a control based on current control theory or the usual method for solving optimization problems, the optimal solution can be found in a shorter computation time.

Furthermore, since the identification model is linearized to obtain the parameters of the solution model, it can also be applied to nonlinear processes.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of a process control method using a neural network according to the present invention, and FIG. 2 is a flowchart showing a control procedure. 1... Process, 2... Learning law, 3... Neural network model for identification, 5... Neural port N for solving optimization problem
net model. No.! Figure most realistic mockery 1

Claims (1)

[Claims] A neural network model for identification is identified based on the input and output of the process to be controlled, a test signal is input to the identified neural network model and linearized, and the Hopfield form is calculated from the coefficients of this linearized model. A process control method using a neural network model, characterized in that parameters of the neural network model are determined, and a target value and a process amount are input into the Hopfield type neural network model to determine an optimal operation amount.