JPH02249004A - Process control method using neural circuit network model - Google Patents

Abstract

PURPOSE:To attain optimizing control in a shorter time and more effectively by identifying a neural circuit network model for identification by means of the input/output of a process being the object of control, inputting a test signal to the identified neural circuit network model so as to execute linearization and forming the model of the process by means of the neural circuit network model. CONSTITUTION:The neural circuit network model for identification is identified by using the input/output of the process 1, and the test signal is impressed on the identified model 3 so as to execute linearization. The linear relation is substituted for an evaluation function, and a manipulated variable setting the evaluation function to be minimum is obtained in an optimizing controller 5. Thus, the process model for a multi input/output system can simply be obtained, and the optimum manipulated variable can be obtained. Then, an optimum solution can be obtained in a short time compared to optimizing control by a present control theory and conventional solution of optimizing problem.

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 linking the controlled object, that is, creating a system that produces the same output when the same input as the controlled object is given. Normally, when designing a control system for a certain process, the objects to be controlled are first linked, but the amount of work required for this is enormous. Furthermore, even if a huge amount of work is performed, the desired model may not be obtained in many cases. On the other hand, 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 the HOpfeld model, which is a type of neural network model, can be used to solve certain optimization problems using the "principle of energy minimization." Using this, it is possible to find the optimal operation amount in a process control system.6 <Problem to be solved by the invention> However, since the neural network model is not a linear model, the optimal Control or station equipment methods cannot be used. Therefore, there was a problem in that it was not possible to perform control 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> An object of the present invention is to provide a process control method that performs optimal control using the learning function and self-organizing function that are characteristics of the neural network model.

Means for Solving the Problems> In order to solve the above problems, the present invention identifies a neural network model for identification at the input/output of a process to be controlled, and applies a test signal to the identified neural network model. is input and linearized, this linear relationship is substituted into the evaluation function of the optimization control device, and the manipulated variable that minimizes this evaluation function is determined.

By creating a process model using a neural network model,
Executing conventional optimization control in a shorter time and more effectively (Example) This invention uses a learning method such as pack propagation to identify a neural network model for identification based on the input and output of a process, and The linear relationship between the manipulated variable and the process variable is determined from the model, and this linear relationship is used to determine the optimal manipulated variable using an optimization control method. FIG. 1 shows the configuration of a process control method using a neural network according to the present invention. In FIG. 1, numeral 1 is a process to be controlled, into which a manipulated variable is input and a process variable is output. It is assumed that this process 1 has a sufficiently short time constant compared to the control period, and is a static process in which the influence of the previous operation settles down during the control period. Also, consider a process with multiple inputs and multiple outputs. Therefore, when expressed in discrete time format, yk+1=F'(IIJk) ・・・・・・・・・(
1) k: sample time + uk: N-dimensional input (operated amount) vector (ul,...
..., uN) yb: M-dimensional output (process amount) vector (y,
..., yH) F=can be expressed as a function vector representing the input/output characteristics of the process. 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. Pack propagation learning laws are well known, for example, E. Ruelhart, G.E.
Former nton and R, J, Will+als, “L
earn+ng Representations by
"Back-Propagating Errors"
Nature Vol, 323, 533-536 (19
86). 3 is a neural network model for identification of process 1, which is modeled using a multilayer perceptron type neural network model. This identification neural network model 3 is expressed by the following equation (2), and its parameters are adjusted according to the learning law 2.

Yl(+1=NN (1'uk)...= (2
) NN represents Neural Network (y), and lu = +1 as in equation (1). A test signal is applied to this identification neural network model 3, and its input/output sensitivity is calculated as shown in 4. A linear model is created from Reference numeral 5 denotes a known optimization control device, into which a target value and a process amount are input, and calculates a manipulated variable that minimizes a quadratic evaluation function and outputs it to the process 1.

The input neural network model 3 for identification is input with the manipulated variable as the input of process 1 and the process variable as the output, and these inputs change the connections within model 3 through self-organization, creating the model of process 1. be done. That is, it is possible to create a model that outputs the same output (process amount) when the same input (operation amount) as in process 1 is input.

This model can be expressed by equation (2) above. However, since it is inconvenient to handle as it is, by applying a test signal and checking the sensitivity of Model 3, the above (2)
) to linearize the equation. That is, "k+1 = (c5
NN/,)+u, )-n1H2・lu k...
Put it as (3). Z is a matrix of MXN and represents the sensitivity from Iuk to y/kt1. That is,
The element Zh.

is the sensitivity from U. to yh. Letting the vector of the j column of this matrix Z (sensitivity from old to J) be IN, Ill = (NN (luk+α1i−−NN (+
uk))/α φ・2N-dimensional i-th unit vector α・ :U It can be found by the amount of variation.

Next, the operation of the optimization control device 5 will be explained. Since we are considering a static model, the optimization control device 5 minimizes the next evaluation function J considering the magnitude of deviation from the magnitude of the manipulated variable and the target value of the controlled variable converged at the next sample time. Calculate the manipulated variable vector uh.

J=+uk-P ・+uk ten (Yk+I Y)”
・Q■ (yl, +1-y) ・・・・・・・・・(4) P:
Weight matrix of manipulated variables (diagonal matrix) Q: Weight matrix of controlled variables (diagonal matrix) 1u: Manipulated variable vector at time Vk: Controlled variable vector at time The evaluation function cannot be evaluated unless the relationship between IIJH and yk+1 is clear. Therefore, by substituting the relational expression (formula (3) above) of the linearized neural network model 3 for identification into this equation, we get J=+uH"-P-ru +(Z・+uk-y)"Q
・It becomes (2・+uk y). Applying a procedure similar to the weighted least squares method to this equation to find Ill k that minimizes J, we get T
-I T+u = (P0Z
-Q-Z) ・Z -Q-yh, and the optimum operation amount can be found.

FIG. 2 shows a flowchart of the entire algorithm.

First, a conventional control method such as PID control is used to output the manipulated variable and process input/output data'
Collect +1-1uk on /. Next, this input/output data
Neural network model 3 for identification by applying learning laws such as backpropagation using data Vk+1 and LLIk.
Determine. That is, NN of the above equation (2) is determined. This process continues until the model is identified. 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. In other words, the matrix Z in equation (3) above
Find each element of . Next, this linear relationship is substituted into the quadratic evaluation function of equation (4), and the operation 1h that minimizes this evaluation function J is determined and output.

Next, we will show an example in which the process control method using this neural network is applied to control the thickness profile in the width direction of a paper machine. In a paper machine, the profile of absolute dry basis weight in the width direction of the paper product is measured, and the opening profile of the slicing lip, which is the outlet for pulp, is controlled so that this profile matches a target value. In other words, if the opening vector of the slice lip is '7' + 1 - the absolute dry basis weight vector is River, then y/ is - [yt , yz , ......
・,,yt,1]".Here, A is an interference matrix, and if you change the opening degree of the slice lip in a certain part, the absolute dry basis weight of the surrounding area will also change, so it is usually as follows. It becomes a banded diagonal matrix.

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

Δlug = Q (s) ・M ・Δ'pkGc
(s): Transfer function M of PI controller: Distribution coefficient matrix for non-interference (=A-1) Δy/ = deviation between target value and measured value However, in such a method, (1) interference matrix A may have a different shape depending on the location, or the positional correspondence may deviate, so that it is not necessarily as shown in equation (5) above.

(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 percetron neural network model is used as the identification neural network model 3. 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 neurons near the input point and negative synaptic connections to neurons on both sides, and are said to be effective in explaining optical illusion phenomena and the Matsuha effect. It is something that The control method itself is the same as the method described above, so the explanation will be omitted. In this type of control, control can be achieved without idealizing the process characteristics to be linear and the interference characteristics to be symmetrical, diagonal, and uniform, so that higher quality control can be achieved.

In addition, the learning law uses CP in addition to backpropagation.
N (Counter Propagation Ne
twork), ART'(8 adaptive Re
5pance Theory) etc. can be applied.

Further, in some paper machines, there may be restrictions on the opening degree of the slicing lip, so a constraint condition may be attached to the amount of operation.

<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. This linear relationship is then substituted into the evaluation function to find the amount of operation that minimizes the evaluation function. Therefore, a process model can be easily obtained even for a multi-input multi-output system, and the optimum operation amount can be determined.

Furthermore, compared to optimization control based on current control theory or conventional methods for solving optimization problems, it is possible to obtain an optimal solution in a shorter calculation time.

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

[Brief explanation of drawings]

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... Optimization control device.

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, linearized, and this linear relationship is used as the evaluation function of the optimization control device. 1. A process control method using a neural network, characterized in that the amount of operation is determined such that the evaluation function is minimized.