JP3083172B2 - Pre-compensator and pre-compensation method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a control device and a control method for improving the response of a control system such as a steel process.

[0002]

2. Description of the Related Art In recent years, a manufacturing process such as a steel process has been equipped with a control system to improve quality and save energy.
Labor saving has been promoted.

However, the control performance often deteriorates due to a change in the process or the like. Therefore, the operation technician needs to readjust the control device.
Also, a method and apparatus for adaptively modifying the control gain online have been proposed. Model reference type adaptive control and self-tuning control are specific methods. However, these methods can be applied only when the input / output relationship of the process is linear.

[0004]

SUMMARY OF THE INVENTION An object of the present invention is to improve, maintain, and prevent deterioration of process control characteristics.

[0005]

In order to solve the above-mentioned problems, a precompensator according to the present invention uses time series data of a response output of a closed loop control system with respect to a target value, and uses the time series data from the current time to n + k times before. Control system output value and k before time k + 1
The target value up to + m time before is input to the multilayer neural circuit,
Means for learning the weighting factor of the neural circuit so that the output of the circuit becomes the target value from the current time to the time before the time k, and a desired response from the current time k to the time 1 from the learned neural circuit. Inputting the control system output values from the current time to the current time n and the control target values from the current time one to the current time m, and outputting the control target values from the current time to the time k ahead; Means for weighting the set control target value to generate a current final target value.

Further, the control method of the present invention uses time series data of a response output of a closed loop control system with respect to a target value,
The control system output value from the current time to the time before n + k time and k + 1
Learning the weighting factor of the neural circuit so that the target value from the time before to the time before k + m time is the input of the multilayer neural circuit, and the output of the circuit becomes the target value from the current time to the time before k time; Input the desired response from the current k time point to the one time point ahead, the control system output value from the current time to the current n time point and the control target value from the current one time point to the m time point before the current neural network. And outputting a control target value from the present to the point in time k later, and further weighting the output control target value to generate a current final target value.

[0007]

According to the precompensator and the compensation method of the present invention, a neural network is used to generate a virtual control target value from a desired control response so as to make the response of the control system to the control target of the process as close as possible to the control target. That is, this is performed using a neural network. In the neural network used in the pre-compensator and the compensation method, learning is performed using a learning rule of a neural network (for example, back propagation) using a target value of an actual process control and its response output.

In the three-layer neural circuit shown in FIG. 2, when data xi (i = 1 to n1) is input to the input layer, the change of the state of the nerve cell (FIG. 3) from the input layer to the intermediate layer is as follows.

[0009]

(Equation 1)

Here, Wij (2) is a weight coefficient from the input layer to the intermediate layer. n2 indicates the number of intermediate layers.

The output of the hidden layer is

[0012]

Yi = fi (ui) (2) Here, fi is a firing function, and more specifically, a monotonically increasing function is used.

Next, the state change of the nerve cell from the intermediate layer to the output layer is as follows:

[0014]

(Equation 3)

Here, Wij (3) is a weight coefficient from the hidden layer to the output layer. n3 indicates the number of output layers.

The output of the output layer is:

[0017]

Yi = fi (ui) (4)

At this time, learning of Wij is performed as follows.

Learning of wij (3) from the output layer to the hidden layer is as follows:

[0020]

Δj ⁰ = (dj−yj) fj ′ (uj) (j = 1 to n3) (5) where dj is called a teacher signal, and in the present invention, Means the actual output. fj 'means the derivative of fj.

[0021]

[6] Wij (3) = η · δj 0 · yi + Wij (3) ··· (6) However: i = 1~n2, learning of Wij (2) from j = 1~n3 intermediate layer to the input layer Is

[0022]

(Equation 7)

[0023]

## EQU00008 ## Wij (2) =. Eta..delta.j.xi + Wij (2) (8) where: i = 1 to n1, j = 1 to n2.

In the apparatus and method of the present invention, the control system output value from the current time to the time before n + k time and k are used by using the time series data of the response output of the closed loop control system to the target value.
The target value from +1 time to k + m time is input to the three-layer type neural circuit, and the output of this circuit gives the target value from the current time to k time as a teacher signal to learn the weighting factor of the neural circuit.

Next, when a desired response, a past output value, and a target value at that time are input to a pre-compensator having learned neural circuit information, a target value to be set first is output. By setting the target value, an actual response close to a desired response, that is, a high response control performance can be obtained.

That is, in the learned neural circuit, the desired response from the current k time point to the current time point, the control system output value from the current time to the current time point n, and the control system output value from the current time point to the current time n. The previous control target value is input, the control target value from the present to the point in time k is output, and the output control target value is weighted to generate the current final target value. In the present invention, a three-layer god system circuit is used. However, a system without nonlinearity has a two-layer neural circuit, or a system with nonlinearity has higher-order fourth and fifth-order layers. A similar pre-compensator may be configured using a neural circuit.

[0027]

Next, an embodiment of the present invention will be described with reference to the drawings.

FIG. 1 shows the positioning of the precompensator and the process control system of the present invention. 1 is a target process, and 2 is a control device of the process, which constitutes a feedback system. Reference numeral 3 denotes a pre-compensator that generates a target value such that the process control system has a high response. A learning device 4 learns the characteristics of the pre-compensator.

FIG. 2 shows an example of a neural circuit used in the pre-compensator 3, which in this case constitutes a three-layer neural circuit. Each node of the circuit performs weighted addition of its input as shown in FIG. 3 and calculates an output with a monotonically increasing nonlinear function. These nodes constitute a neural circuit that is hierarchically combined.

FIG. 4 shows an example of a neural circuit used for learning characteristics to be possessed by the pre-compensator 3. A time series of a response output of a closed loop control system to a target value of process control is provided in an input layer of the neural circuit. Using the data, the control system output values y (t) to y (tnk) from the current time to n + k times before, and k
The target value u (t−k−) from the time before +1 time to the time before k + m time
1) to u (tkm) are input to a three-layer neural circuit, and the output of the circuit is a target value u (t) from the current time to the time before k time.
The weighting factor of the neural circuit is learned so as to become u (tk). As the learning expression, the above-described back propagation algorithm or the like is used.

In this example, k = 10, n = 10, m = 10
The number of input layers is 31, the number of intermediate layers is 21, and the number of output layers is 10.

The following function is used as f (x).

[0033]

F (x) = 1−0.5γ ² / (x−θ + γ) ² x ≧ θ (9)

[0034]

F (x) = 0.5γ ² / (x−θ−γ) ² x <θ (10) The time step of the time series is 0.02 seconds. Learning is 1
It is repeated several thousand times using time-series data for seconds.

FIG. 5 shows an example of using the learned precompensator.
3 shows a neural circuit for generating a target value. The neural circuit and the contents of FIG. 4 are exactly the same. The difference is the contents of the input data of the input layer. In the learned neural circuit, the current k
Desired response y (t + k) to y (t) from time point to one time point
+1) and the control system output values y (t) to y (t−n) from the current time to the current time n and the control target values u (t−1) to u from the current time one time to the current time m (tm), and outputs the control target values u (t + k) to u (t) from the present to the point in time k. This output u (t) or the output control target value u (t) to u (t + k) U (t) is generated and set as the output of the pre-compensator.

[0036]

[Equation 11]

Here, w0 = 1.0, w1
.., W10 = 0.0.

FIG. 6 shows the response when the response is degraded, and FIG. 7 shows the improved response when using the precompensator and method according to the invention. Here, the response y is a dimensionless numerical value, 0.
3 is the target value. In FIG. 6, the response slowly rises slowly, but in FIG. 7, a high-speed response is realized by generation of the target value U (t).

[0039]

As described above, according to the present invention, the deterioration of the response of the process control system can be prevented by the learning type pre-compensator. This eliminates the need for maintenance work on the control device, and also makes it possible to make the control device unmanned. Improvements in control performance have led to cost reductions and improved yields of products, leading to significant effects.

[Brief description of the drawings]

FIG. 1 is a block diagram of a process control system using a precompensator according to the present invention.

FIG. 2 is a block diagram of a neural circuit used in the embodiment.

FIG. 3 is a block diagram showing a nerve cell at a node in FIG. 2;

FIG. 4 is a block diagram showing inputs and outputs of a neural circuit at the time of learning of the predistorter of the embodiment.

FIG. 5 is a block diagram showing input / output of a neural circuit when generating a target value using the pre-compensator of the embodiment.

FIG. 6 is a graph showing a response example of a deteriorated process control system.

FIG. 7 is a graph showing an example of an improved response when the predistorter of the embodiment is used.

[Explanation of symbols]

 1: Process 2: Control device 3: Pre-compensator 4: Learning device

Continuation of the front page (56) References JP-A-2-15718 (JP, A) JP-A-2-64788 (JP, A) JP-A-4-271486 (JP, A) JP-A-4-220758 (JP) (A) JP-A-4-205195 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06G 7/60 G06C 15/00 G06F 15/18

Claims (2)

(57) [Claims] In a process control apparatus, a control system output value from a current time to a time before an (n + k) th time and k are used by using time series data of a response output of a closed loop control system to a target value.
Means for taking a target value from +1 time ago to k + m time ago as an input to the multilayer neural circuit, and learning a weight coefficient of the neural circuit such that an output of the circuit becomes a target value from the current time to k time ago; And the desired response from the current k time point to the current time point, the control system output value from the current time to the current time point n, and the control from the current time point to the current time point m using the learned neural network information. A front end having a means for inputting a target value, outputting a control target value from the present to a point in time k later, and further weighting the output control target value to generate a current final target value; Compensator. 2. A process control method, comprising: using a time series data of a response output of a closed loop control system with respect to a target value to obtain a control system output value from a current time to an (n + k) th time and k
Learning the weighting factor of the neural circuit so that the target value from time +1 to time k + m is input to the multilayer neural circuit, and the output of the circuit becomes the target value from the current time to time k before the time; In the learned neural circuit, a desired response from the current k time point to the current time point, a control system output value from the current time point to the current time point n, and a control target value from the current time point to the current time point m are described. A pre-compensation method comprising: inputting and outputting a control target value from the present to a point in time k later; and weighting the output control target value to generate a current final target value.