JP2968631B2 - Hot rolling process control method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to control of inter-stand tension and looper angle of a material in a hot rolling process.

[0002]

2. Description of the Related Art In recent years, various automatic control systems have been provided in a manufacturing process such as a steel process, so that quality has been advanced, energy saving and labor saving have been promoted. For example, in many cases, a looper is arranged between stands of a hot finishing mill, and a system for controlling the steel strip tension and the plate thickness between stands by controlling the looper is often used.

Conventionally, the looper control generally uses a PI control device and controls the tension and the looper angle so as not to interfere with each other. Further, a method of performing optimal control of a looper by performing multivariable control has also been implemented.

[0004]

However, due to a change in process characteristics or a change in material, the control performance,
That is, the response of the process output to the control input often deteriorates. This kind of responsiveness can be improved by re-adjusting the control gain, but in order to always keep the responsiveness of the control system in an appropriate state, usually,
It requires manual work and is a considerable burden.
In addition, methods and apparatuses for adaptively modifying the control gain online, such as model reference adaptive control and self-tuning control, have been proposed.However, these methods can be applied only because of the relationship between the process input and output. Only when there is a linear relationship.

An object of the present invention is to improve, maintain, and prevent deterioration of control characteristics in looper control.

[0006]

According to the present invention, there is provided a method for controlling a hot rolling process in a hot rolling process in which a looper is provided between rolling stands. Detecting at least one of the inter-stand tensions of the material as a process output value, and using a time series data of a control target value and a process output value corresponding thereto,
Process output value from the current time to n + k time before and k +
The control target value from one time ago to k + m time ago is input to the multilayer neural network, and the weight coefficient of the neural network is learned so that the output of the circuit becomes the control target value from the current time to k time ago. Then, in the learned neural network, desired response values from the current k time point to the current time point, process output values from the current time point to the current time point n, and m from the current time point to the current time point.
A control target value before the time is input to generate a current final target value, and the process is controlled based on the final target value.

[0007]

According to the hot rolling process control method of the present invention, in order to make the response of the control system as close to the desired response as possible, a neural network, that is, a neural network, generates a virtual control target value from the desired control response. This is performed by using a minus mark. Generally, the control object in the hot rolling process control includes the angle of the looper and the tension applied to the looper.
In the present invention, at least one of them is controlled. In the neural network used in the hot rolling process control method, learning is performed using a learning rule (for example, back propagation) of a neural network using a target value of the actual hot rolling process control and its response output.

For example, in the three-layer neural network shown in FIG. 2, when data xi (i = 1 to n1) is input to the input layer, the state of the nerve cells (see FIG. 3) from the input layer to the intermediate layer Is represented by the following equation.

[0009]

(Equation 1)

Here, Wij (2) is a weight coefficient from the input layer to the intermediate layer. n2 indicates the number of nodes of the intermediate layer. The output of the intermediate layer is expressed by the following equation.

[0011]

Yi = fi (ui) (2) Here, fi is a firing function, and more specifically, a monotonically increasing function is used. Further, the state change of the nerve cell from the intermediate layer to the output layer is expressed by the following equation.

[0012]

(Equation 3)

Here, Wij (3) is a weight coefficient from the hidden layer to the output layer. n3 indicates the number of nodes of the output layer. The output of the output layer is represented by the following equation.

[0014]

Yi = fi (ui) (4) In this case, learning of the weight coefficient Wij is performed as follows.

The weight coefficient wij (3) from the output layer to the hidden layer
Learning:

[0016]

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.

[0017]

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

[0018]

(Equation 7)

[0019]

## EQU00008 ## Wij (2) =. Eta..delta.j.xi + Wij (2) (8) where: i = 1 to n1, j = 1 to n2 In the hot rolling process control method of the present invention, the control target value Using the time-series data of the process output values, the process output values from the current time to the time n + k and the control target values from the time k + 1 to the time k + m are input to the multilayer neural circuit. The control target value from the time to the time before the time k is given as a teacher signal to learn the weight coefficient of the neural circuit.

Next, a desired response (target value before compensation), a past process output value, and a previously output target value (target value after compensation) to the neural network (pre-compensator) having the learned weighting coefficient. When a value is input, a target value to be set first (target value after compensation) is output. By inputting the compensated target value (final target value) to the control system, an actual response close to a desired response, that is, a high response control performance can be obtained.

That is, the desired response (target value before compensation) from the current k time point to the next time point and the control system output value (the process output value) from the current time to the current time n points before the current neural network Value) and the current control target value (target value before compensation) from one time ago to m time ago to generate the current final target value.

In the example shown in FIG. 2, a three-layer neural circuit is used. However, in a system without nonlinearity, a two-layer neural circuit, or in a system with nonlinearity, a higher-order fourth- or fifth-order neural circuit is used. A similar pre-compensator may be configured using a neural circuit having a hierarchy of.

[0023]

FIG. 1 shows the configuration of a control system according to one embodiment. FIG.
1 shows a part of the hot rolling process, and two control systems are configured using a looper 4 installed at an intermediate position between the rolling stands 2 and 3 adjacent to each other. One of the two control systems includes an actual looper angle α detected by the angle detector 9 installed on the looper 4 and a looper angle target value αr input from a process control computer (not shown). The other is a looper angle control system that drives the looper motor 7 based on the above, the other is the actual tension β of the material detected by the tension detector 11 installed on the looper 4, and the computer for process control. And a looper tension control system for driving the mill motor 5 based on the tension target value βr input from the motor.

The compensation units 20 and 30 are provided to improve a response delay caused based on the characteristics of the process of each control system. The looper angle target value αr is input to the compensation unit 20 and converted into a compensated looper angle target value αr ′, and the tension target value βr is calculated by the compensation unit 3.
0, and is converted to a compensated tension target value βr ′. The actual looper angle α is calculated from the compensated looper angle target value αr ′.
Is input to the angle control device (PI control device) 10, and the value obtained by subtracting the actual tension β from the compensated tension target value βr ′ is input to the tension control device (PI control device) 12. In order to eliminate the interference between the looper angle control system and the tension control system, the output Uαo of the angle control device 10 and the tension control device 1
2 output Uβo is input to the cross controller 14,
These are converted into control amounts Uα and Uβ, respectively. Control amount Uα
Is input to the looper motor speed controller (ACR) 8,
The control amount Uβ is input to a mill motor speed controller (ASR) 6. The looper motor speed controller 8 controls the looper motor 7 to adjust the angle of the looper 4, and the mill motor speed controller 6 controls the speed of the mill motor 5 to reduce the tension. adjust.

The compensating unit 20 comprises a pre-compensator 21, an integrator 22, a subtractor and an adder. Learning device 2
3 is a characteristic (weight coefficient) of the pre-compensator 21 by learning.
Is adjusted to the optimal state. The integrator 22 is provided to compensate for a control offset error described later.
Similarly, the compensation unit 30 includes a pre-compensator 31, an integrator 3
2, a pre-compensator 3 comprising a subtractor and an adder
The weighting coefficient of 1 is learned by the learning device 33. The integrator 32 is provided for compensating for a control offset error.

FIG. 2 shows an example of a neural network used for the pre-compensators 21 and 31. In this case, a neural network of three layers is constructed. 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 hierarchically combined neural network.

FIG. 4 shows the configuration of the neural network provided in the learning devices 23 and 33. In the learning device 23, the nodes of y (t) to y (tnk) in the input layer of the neural circuit are obtained.
Time series data of the actual looper angle α obtained at the output of the angle detector 9 (n + k from the current time to n + k times before).
+1 data) is input, and u (tk-1) of the input layer is input.
The time series data of the compensated looper angle target value αr ′ (excluding the output of the integrator) (k + m time before k + 1 time) is added to the nodes u to (t−k−m). The previous m data) are input, and the nodes of u (t) to u (tk) in the output layer are input.
The time series data of the compensated looper target angle value αr ′ (k + 1 data from the current time to the time before k time) is given as the teacher data to the output layer, and the output layer u (t ) To u (tk)
The weighting factor of the neural circuit is modified so that the output of each node approaches the teacher data. As a learning formula for correction,
In this example, a back propagation algorithm is used.

In the learning unit 33, the tension detector 11 is connected to the nodes y (t) to y (tnk) of the input layer of the neural circuit.
, The obtained time-series data of the actual tension β (n + k + 1 data from the current time to the time before n + k) is input, and u (tk−1) to u (t) of the input layer are input. -Km), the past time series data of the compensated tension target value βr ′ (excluding the output of the integrator) (k + m from the time before k + 1)
M data up to the time before) are input, and u in the output layer
The compensated tension target value βr ′ is added to the nodes (t) to u (tk).
Of time series (k + 1 data from the current time to k time before) are given as teacher data, and u in the output layer
The weight coefficient of the neural circuit is corrected so that the output of each node of (t) to u (tk) approaches the teacher data. In this example, a back-propagation algorithm is used as a learning equation for correction.

In this embodiment, the learning devices 23 and 33 both have k = 10, n = 10, and m = 10, the number of input layers is 31, the number of intermediate layers is 21, and the number of output layers is 10 It is configured as an individual.

In this example, the following function is used as the function f (x).

[0031]

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

[0032]

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 the configuration of the neural network included in the pre-compensators 21 and 31. The configuration of this neural network is the same as the neural network of the learning device shown in FIG. However, the contents of the input data in the input layer are different. Also,
Only one node of the output layer is actually used.
The weighting coefficient of each node of this neural network is
And 33 are given as they are in the neural network.

In the pre-compensator 21, yd of the input layer
The time series data of the k pre-compensation target values αr from the current k time point to the one time point are added to the nodes (t + k) to yd (t + 1).
(I.e., desired response), and y (t) to y (t) in the input layer.
−n), time series data of n + 1 control system output values (that is, angle α) from the current time to the current time n is input to the node u (t−1) of the input layer. Time series data of m past compensated target values αr ′ from one time before to m time before are input to the nodes 〜 u (tm). As a result, the post-compensation target value αr '
(However, the output of the integrator is not included) is obtained.

In the pre-compensator 31, yd of the input layer
The time series data of the k pre-compensation target values βr from the current k time point to the one time point are added to the nodes (t + k) to yd (t + 1).
(I.e., desired response), and y (t) to y (t) in the input layer.
-N), time series data of (n + 1) control system output values (that is, angle β) from the current time to the current time before n are input to the node u (t−1) of the input layer. The time series data of m past compensated target values βr ′ from the current one time point to the current time point m are input to nodes u to (t−m). As a result, the compensated target value βr ′ is added to the u (t) node of the output layer.
(However, the output of the integrator is not included) is obtained.

In this embodiment, the time series data when the control target value is changed stepwise for each control system is learned by the learning devices 22 and 32. In this case, the network of the learning device is used. The teacher data to be set in the output layer has a very complicated time-series pattern. Therefore, in this embodiment, formulas for calculating the inverse models of the looper angle control and the tension control process (calculation formulas for converting the process output value into the input value) are created in advance, and the process is performed when learning is performed. The difference between the actual output value (α, β) and the desired response value is converted into a process input value, that is, a compensated target value (αr ′, βr ′) by each inverse model calculation formula. Has been created. Based on the teacher data, each weighting factor of the network is modified.

FIG. 7 shows an example of the response characteristic of the looper angle control system after some learning has been performed, and FIG. 8 shows an example of the response characteristic of the tension control system (when the output of the integrator is not added). . Referring to FIG. 7, the target value .alpha.r changes stepwise. If the target value .alpha.r is applied to the control system without compensation, the looper angle changes gradually, and the target value .alpha. Approaches the value. However, the post-compensation target value αr ′ output from the compensation unit 20 changes in a very complicated time-series pattern, and as a result, the looper angle α quickly changes to the level of the post-change target value αr. Is controlled so as to converge. As for the response characteristics of the tension control system shown in FIG. 8, it can be seen that the response characteristics of the tension β are greatly improved by the compensation of the target value βr. Although not shown, according to this embodiment, even when various disturbances are applied to the control system or when the target value changes in a sine wave shape,
It was confirmed that the process output could be controlled to the target value in a very short time.

Referring to FIG. 8, it can be seen from the response characteristics in the case where the pre-compensator 31 is provided that the level at which the tension β converges slightly deviates from the input target value βr. According to experiments performed by the present inventor, it has been confirmed that this kind of offset does not decrease even when the number of times of learning is increased. Therefore, in this embodiment, as shown in FIG.
Compensation units 20 and 3
0. That is, a value obtained by subtracting the output u of the pre-compensator (21, 31) from the target value d of the input of the pre-compensator (21, 31) is input to the integrators (22, 32), and the output of the integrator is output. Is added to the output u of the pre-compensator (21, 31), and the value u ′ is output as the compensated target value (αr ′, βr ′). This prevented the occurrence of offset.

[0039]

As described above, according to the present invention, the responsiveness of looper angle control and tension control is greatly improved. Further, in the present invention, since the characteristics of the control system can be sequentially corrected by learning, it is possible to automatically prevent the response characteristics from deteriorating due to a change with time. Therefore, the maintenance work of the control device becomes unnecessary, and the operation can be made unmanned.
In addition, improvement in control performance leads to product cost reduction and yield improvement.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration example of a part of a control system of a hot rolling process for implementing 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 a configuration of a neural circuit of the learning device shown in FIG.

FIG. 5 is a block diagram showing a configuration of a neural circuit of the pre-compensator shown in FIG. 1;

FIG. 6 is a block diagram showing a transfer function of the control system of FIG. 1;

FIG. 7 is a waveform chart showing an example of response characteristics of a looper angle control system.

FIG. 8 is a waveform chart showing an example of response characteristics of the tension control system.

[Explanation of symbols]

1: Steel strip 2, 3: Rolling stand 4:
Looper 5: Mill motor 6: Mill motor speed controller 7: Looper motor 8: Looper motor speed controller 9: Angle detector 10: Angle controller 1
1: tension detector 12: tension controller 14: cross controller 20, 30: compensation unit 2
1: Pre-compensator 22: Integrator 23: Learning device 3
1: Pre-compensator 32: Integrator 33: Learning device α:
Actual looper angle αr: Looper angle target value αr ': Looper after compensation
Power angle target value β: Actual tension βr: Tension target value βr ': Tension target value after compensation

Claims (1)

(57) [Claims] In a hot rolling process control method in a hot rolling process in which loopers are provided between rolling stands, at least one of an angle of a looper and a tension between stands of a material is detected as a process output value, and control is performed. Using the time series data of the target value and the process output value corresponding to the target value, the process output value from the current time to the time n + k and the control target value from the time k + 1 to the time k + m are input to the multilayer neural network, The weighting factor of the neural network is learned so that the output of the circuit becomes the control target value from the current time to the time k before the current time.
A desired final response value is generated by inputting a desired response value up to the current time, a process output value from the current time to the current time n, and a control target value from the current time 1 to the current time m,
A method for controlling a hot rolling process, comprising controlling a process based on the final target value.