JPH06214610A - Control method and its device - Google Patents

Abstract

PURPOSE:To improve the set-up control accuracy for deciding an action point for the satisfactory control of a controlled system having high nonlinearity and to ensure the satisfactory control of the controlled system. CONSTITUTION:A command is issued to an actuator 2 for controlling a controlled system 1 so that the system 1 has a desired operation. Meanwhile the relation between each parameter value of a model 5 of the system 1 and the input/output value of the model 5 is secured by a learning mechanism 11. This learning result is stored in a memory 20. Then the parameter value of the model 5 is decided based on the input/output measurement value of the system 1 and by making reference to the learning result stored in the memory 20. Then a command is produced based on the parameter value and given to the actuator 2.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a control method and apparatus for a controlled object, and more particularly to a setup control method and apparatus for determining an operating point of a non-linear control system such as a rolling mill.

[0002]

2. Description of the Related Art A linear control theory developed on the premise of a linear system enables highly accurate control. However, even if this linear control theory is applied to a control object with a strong non-linearity, for example, a rolling mill that rolls a material to be rolled, the desired control performance cannot be expected, and the product (steel material in the case of a rolling mill) cannot be expected. Quality cannot be improved. A control system of a rolling mill has a setup control system that obtains an operating point in a steady state where a controlled object does not change with time, and a method for improving the accuracy of the setup control is desired.
In thermal power generators and plant control, setup control is called setpoint control.

Usually, in a set-up control system, as in the case of obtaining a steady solution when solving a differential equation, the change in time is ignored and an optimum solution such as dynamic programming or linear programming is obtained to determine the operating point. .
Once the operating point is determined, the linear control theory follows that the controlled object is linearly approximated around the operating point, and an optimal solution is obtained for the transient solution of the differential equation as a regulator problem that makes the deviation from the operating point zero. Fordback control is performed.

The set-up control is specifically described in "Theory and practice of strip rolling" published by Japan Iron and Steel Institute, September 1984, pages 289 to 292.

[0005]

In the above-mentioned prior art, the control model for determining the operating point for setup control (setpoint control) is always changing. For example, in the case of a rolling mill, there are many unknown parameters, and even if the value of each parameter can be determined,
The heat generated during operation often changes the value of each parameter once determined, such as the value of frictional resistance, which is one of the parameters, or the roll expanding. However, since it is difficult to fix the changed parameter values accurately, the operating point determined by the setup control will be significantly increased, and the feedback control parameters will also be increased, making it possible to obtain good products (steel materials in the case of rolling mills). There is a problem that it cannot be obtained.

An object of the present invention is to improve the accuracy of set-up control (set point control) for determining an operating point for satisfactorily controlling a controlled object having a strong non-linearity, and to perform a good control method. And to provide the device.

[0007]

Means for Solving the Problems The above-described object is to provide each value of a parameter of a model of the controlled object and the model when giving an instruction to an actuator for controlling the controlled object so as to cause the controlled object to perform a desired operation. The relationship between the input / output state and the input / output state is obtained by learning and the learning result is stored in the storage means, and the parameter value of the model is determined based on the input / output measurement value of the control target with reference to the learning result of the storage means. It is achieved by obtaining a command based on the determined parameter value and giving it to the actuator.

Preferably, when the model of the controlled object is used as a setup model, and the command is given to the actuator, the input / output target of the controlled object is determined by the setup control based on the determined parameter value. Is obtained and given to the actuator, and feedback control is performed so that the deviation of the output state obtained from the controlled object as a result of the command from the target value of the output is made zero.

More preferably, the learning step is performed using a neural network.

More preferably, the step of determining the parameter is performed using a neural network.

[0011]

In the present invention, attention is paid to the fact that the drawback of the conventional setup control or set point control is that it cannot cope with the change in the parameter value of the control model, and the relationship between the control model and the parameter is learned and stored in advance. However, the parameter value is estimated from the actual operating state of the controlled object based on the learning result, and this value is used as the parameter value for the control setup control and the feedback control.

That is, in non-interference control such as a rolling mill whose operating point changes greatly, it is difficult to specify the parameters of the model of the controlled object instantaneously. In the conventional method, various parameters cause misalignment. Therefore, it was not possible to construct an accurate control model, and even if the modern control theory, which can improve performance if the control model is accurate, was applied, the ability could not be exhibited. However, according to the present invention, an accurate control model can be generated even when applied to a controlled object having strong nonlinearity, and control performance can be improved.

Therefore, it is possible to improve the accuracy of the setup control (set point control) for determining the operating point for satisfactorily controlling the controlled object having a strong non-linearity, and to perform good control.

[0014]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. As a preferred embodiment of the present invention, a non-linear controlled object, for example, strip thickness shape control of a rolling mill will be described.

The rolling mill, as shown in FIG. 3, for example, comprises a plurality of pairs of rolled materials 100 (only two rolls are shown in the illustrated example).
01, 202, 211, 212 and the like, and in each pair of rolls, the rotation speed of one roll 202, 212 is controlled by motors 203, 213, and the rolling position of each roll 202, 212 with respect to the rolled material 100. Are controlled by the hydraulic pressure reduction devices 204 and 214, respectively, so that the strip thickness of the rolling mill is set to a desired value. For example, a pair of rolls 211 and 212 which are actuators to be controlled
The control for the motor 213 and the hydraulic pressure reduction device 214 corresponding to the above is performed by giving the velocity command value V _i and the pressure reduction position command value P _i from the speed command control device 222 and the hydraulic pressure control device 220, respectively. The control of this rolling mill is performed as follows.

Here, only the control for the pair of rolls 211 and 212 will be described, but the same control is performed for the other roll pairs.

First, the set-up control system 4 outputs a target load p ₀ , a target speed v _R0 , a target plate thickness h ₀ , and a target tension τ _f0 by using the control model 5 of the rolling mill to be controlled. Subtractors 230, 232, 2 in the controller 2
34,236. On the other hand, the load value p from the load meter 217 for the roller 217, the speed value v _{R of} the motor 213 from the motor 213 that drives the roller 212, the plate thickness value h from the plate thickness meter 216 that measures the plate thickness of the rolled material, From the tensiometer 215 for measuring the tension of the rolled material 100, the tension value τ
_f is given to the subtractors 230, 232, 234 and 236, respectively, and subtraction with the above-mentioned p ₀ , v _R0 , h ₀ , τ _f0 is performed, and deviations Δp, Δv _R , Δh, Δτ _f are obtained.

On the other hand, the control system design system 3 obtains feedback parameters (coefficients) f _{11 to} f ₂₄ for the optimum control based on the control model 5 by the method described later and gives them to the control device 2. The control unit 2 uses these feedback coefficients f ₁₄ ,
The deviations Δp, Δv _R , Δ corresponding to f ₁₃ , f ₁₁ , and f ₁₂
h, Δτ _f and multipliers 240, 242, 244, 246
And the speed command deviation Δv _p = f ₁₄ Δ
p + f ₁₃ Δv _R + f ₁₁ Δh + f ₁₂ Δτ _f is obtained, and the adder 25
Give to 2.

Further, the feedback coefficients f ₂₄ , f ₂₃ , f
₂₁ and f ₂₂ and Δp, Δv, Δh and Δτ of the above deviations corresponding to
_f and are respectively multiplied by multipliers 241, 243, 245, 247, and the rolling position command deviation ΔS _p = f ₂₄ Δp + f
₂₃ Δv _R + f ₂₁ Δh + f ₂₂ Δτ _f is obtained and given to the adder 250.

The setup control system 21 further adds the speed set value (target speed command) S _p and the rolling position setting value (target rolling position command) v _p based on the control model to the adders 250, 2
52, and adders 250 and 252 add the command deviations ΔS _p and Δv _p and the commands S _p and v _p , respectively, and apply them to the hydraulic control device 220 and the speed control device 222.

FIG. 1A is a block diagram of a control device for controlling an object to be controlled according to an embodiment of the present invention. Control target 1
Receives a command from the control device 2 and operates with desired accuracy. The control device 2 has a feedback control system that determines the command deviation by receiving the state of the controlled object 1 and the control structure or parameters that determine the control design system 3, and the command deviation and the nonlinearity of the controlled object 1 A command is generated for the controlled object 1 in consideration of a motion point command (hereinafter referred to as a target value) from the setup control system 4 that determines the motion point from the nature. Here, the feedback control system is a DDC system (Dir
ect Digital Control),
Its function is to receive the state or controlled variable of the controlled object 1,
Configure a regular system that makes the deviation from the target value zero.

The setup control system 4 uses the model 5 of the controlled object 1 to determine the operating point. In the case of a rolling mill, this control model 5 is composed of a plastic curve (rolling load type), influence of tension, oil film thickness compensation, and the like. For example, the following "Hill's approximation formula" (Equation 1), which is an approximation of the "Brand &Ford" formula, is used as the rolling load formula.

[0023]

[Equation 1]

Here, b: rolled material plate width k _i : average deformation resistance κ _i : influence term of tension of i-th stand D _pi : friction correction term (the following 2) H _i : entrance side plate thickness of i-th stand h _i : Outgoing plate thickness of the i-th stand R _i : Work roll diameter of the i-th stand R _{i ′} : Roll flatness (“Hitchcock” formula).

[0025]

[Equation 2]

Here, a ₁ , a ₂ , a ₃ : constant u _i : friction coefficient (the following equation 3) r _i : rolling reduction of the i-th stand

[0027]

[Equation 3]

Here, u _{oi is a} constant (depending on the lubricating oil) u _{ci is a} constant (determined by an _actual value) v _{oi is a} roll speed A constant of a physical formula constituting the control model 5, for example, a friction coefficient u _i is It is affected by the operating temperature of the rolling mill, rolling oil (lubricating oil), etc., and changes considerably in actual rolling. That is, although the structure of the model 5 of the controlled object 1 is relatively clear so that it can be expressed by a mathematical expression, its parameters often change due to changes in operating conditions such as changes in roll speed. Here, constants a ₁ , a ₂ , a ₃ and friction coefficients u _i , k
_If the parameters _i , u _ci , u _ci, etc. cannot be determined, an accurate control model 5 cannot be created.

The reason why the parameters of the control model 5 change will be described below. If you consider the rolling mill as an elastic body,

[0030]

[Equation 4]

Here, H: entrance side plate thickness τ _b : rearward tension S: roll interval τ _f : front side tension h: output side plate thickness K: elastic modulus of rolled material P: load gauge meter formula (FIG. 5) is obtained. To be As shown in FIG. 4, S is the plate thickness immediately below the roll, and the plate thickness is reduced by P / K due to the spring effect to make the plate thickness thinner than the actual plate thickness. Load) is released and returns.

FIG. 5 is a diagram showing the equation (4), and FIG. 6 is a diagram showing the equation (1). In the plasticity curve of FIG. 6, the point of intersection with the x-axis is the entry side plate thickness when the load is zero.

The operating point of the rolling mill is the intersection of the characteristic line of the gauge meter type of FIG. 5 and the plasticity curve of FIG. 6 (FIG. 7).

Next, the plate thickness of the rolled material (entrance side plate thickness) is H → H.
_When it changes to ₀ , the plastic curve moves in parallel from the dotted line to the solid line as shown in FIG. 8 (in this case, the intersection with the x axis is from H to H ₀ ), the target plate thickness is from h to h ₀ , and the target load is P to P ₀
Change to.

At this time, if the automatic control works and the rolling position is changed from S to S ₁ , the gauge meter type moves from the dotted line to the solid line to obtain the relationship shown in FIG. 9, and the target load is P _{0 in} FIG. a P ₁ in FIG. 9, but the target thickness is h from h ₀
Return to.

That is, the automatic control is Δh = (h−h ₀ ), Δ
P = (P−P ₀ ), etc. is detected, and ΔS _p = f ₁ Δh + f ₂ ΔP
To obtain the reduction command deviation and obtain the target value S
In addition to the above, a reduction command S _p = S + ΔS _p is issued. here,
f ₁ and f ₂ are feedback coefficients.

[0037] However, the operating conditions (e _x, velocity v _p) parameter value of the plastic curve is changed by a change in, the plastic curve is changed as a solid line from the broken line in FIG. 10, the same output as in FIG. 9 In order to obtain the side plate thickness h, it is necessary to move the reduction position to S ₂ . The load at this time is P ₂ . That is, in this case, the plate thickness needs to be fed back by Δh = h′−h, and the load needs to be fed back by Δp = P′−P _0. However, the solid plastic curve in FIG. Since it is not known, only h ₀ −h and P ₁ −P ₀ are actually fed back. That is, with the feedback coefficient obtained from the plasticity curve in FIG. 9, it is not possible to generate a command to bring the reduction position to the required S ₂ . Therefore, it is necessary to obtain the correct parameter value according to the change of the operating condition and change the coefficients of f ₁ and f ₂ based on it.

Therefore, in this embodiment, the input and output of the controlled object 1 are input to the learning control device 6, and the learning control device 6 outputs the parameters of the control model 5. In this case, the parameters of the control model 5 can be determined based on the instantaneous operation of the controlled object 1, so that it is possible to identify the parameters that change with time, such as secular change and temperature change.

In the learning control device 6, the parameters and states of the controlled object model 5 (hereinafter,
Input to controlled object 1 and model 5 and controlled object 1 and model 5
The output from and the internal states of the controlled object 1 and the model 5 are collectively referred to as “state”. ) Is remembered.
Then, when the state of the controlled object 1 is input to the learning control device 6, the learning control device 6 refers to the relationship between the stored parameter and state and estimates the value of the parameter from the state at the time of operation. If the structure of the model 5 of the controlled object 1 is relatively accurate, the value of the parameter at that time can be accurately obtained.

FIG. 1B is a configuration diagram when the learning control mechanism 6 learns and stores the relationship between parameters and states. The input state changing unit 7 determines the input state of the control model 5, and the parameter changing unit 8 changes the parameter value and executes the simulation. As a result, the control model 5
As a result of the simulation, the output state 9 that changes according to the change in the parameter value is obtained. The learning control device 6 stores the relationship between these parameter values and the state of the control model 5. As the learning control device 6, a hierarchical neurocomputer having a function interpolation ability (generalization ability) or the like can be used.

FIG. 2 is a block diagram of the entire control device when a hierarchical neurocomputer is used as the learning control device 6. The neurocomputer is described in detail in "Introduction and Training Neuron Computer Computer", published by Technical Review, published by Kaoru Nakano, published on September 15, 1989, so only the parts related to this embodiment will be described here. Describe.

The neurocomputer is composed of a neural network 10, a learning mechanism 11 and a memory 20, and the learning mechanism 11 controls "learning" and "remembering".

At the time of learning, as shown in FIG. 11, a certain parameter value (e _x , friction rod number u _i ) from the parameter changing unit 8
When There given to the control model 5, the control model 5 on the input state of the input state changing unit 7 (e _x, the speed command value V _i, pressing position command value P _i) to the basis output state (e _x, load P, The velocity V _R , the plate thickness h, and the tension τ _f ) are obtained and given to the input layer of the neural network 10. An input state is further given to the input layer, and parameters are given to the output layer as a teacher signal, and the learning mechanism 11 controls these data to perform learning. When the neural network 10 is not learned, a random number or an appropriate numerical value is given to the weighting factor (weight) of the neural network 10, and the mortgage output 13 is input to the learning mechanism 11. On the other hand, at this time, the parameter output from the parameter changing mechanism 8 is input to the learning mechanism 11 as a teacher signal. The learning mechanism 11 corrects the weighting factor (weight) of the neural network 10 so as to minimize the squared deviation between the output of the parameter changing mechanism 8 and the output of the neural network 10. When the set of the input 12 and the output 9 of the control model 5 is input to the neural network 10, the output 13 becomes equal to the output of the parameter changing mechanism 8. These are called neural network learning.

Therefore, the parameter value u as shown in FIG.
The input / output state of each _i1 , u _i2 , ..., U _in , and the weight matrix ω of the neural network are obtained. These data are shown in FIG.
It is stored in the memory 20 as shown in FIG.

Although only the parameters u _i are used here for simplification of description, a plurality of parameters may be learned. In that case, input / output as described above is performed for each combination of parameter values. Get state and weight matrix.

In the learning mechanism 11 after the learning is completed, as shown in FIG.
Input status _{(e x, V i, P} i) of, also, the control model 5
Output state of the controlled object 1 as the output of _{(e x, P, V R} ,
h, τ _i ) is given to the input layer as an input of the neural network 10 which is a component of the learning mechanism 11. Then, the learning mechanism reads the weight matrix ω corresponding to the input state that has been input from the memory 20 and gives it to the neural network 10. Based on this weight matrix ω, the neural network 10 obtains an output, which is used as the parameter value of the controlled object 1. It is output to the control model 5 as (e _x , v _i ). When the control model 5 on the basis of the parameter value is changed, the target value of the up control system 4 is changed, the changed target value _{(e x, v p ',} s 9') is input to the controller 2 .

On the other hand, the control system design system based on the changed control model 5 has feedback coefficients (f _{11 to} f ₂₄ )
Is given to the control device 2. The control device 2 obtains a command deviation based on the feedback coefficient and the target value and performs feedback. In this way, the controller 2 receives the feedback of the target value and the output of the controlled object 1 and
Issue a command to. Control device 2 in the input state of controlled object 1
Command 14 and the output state 14 which is the output of the controlled object 1 are input to the learning mechanism 11.

The flow of the above operation will be described with reference to FIGS. FIG. 15 is a flowchart showing a learning procedure of the learning control device 6. At the time of learning, first, the parameter value changed by the parameter changing unit 8 is input to the control model 5 (step 101). The control model 5 carries out a simulation using the above-mentioned parameter values to obtain the input / output relationship of the control model 5 (step 102). Then, it is judged whether or not the required number of learnings have been completed (step 103), and the above steps 101, 10 are executed until the learning is completed.
It is configured to repeat 2. After learning,
The control is executed using the result. FIG. 16 is a flowchart showing the processing procedure.

In FIG. 16, during control execution, the states of the input and output of the controlled object 1 are input to the learning control device 6 (step 110), and the learning control device obtains the operating parameter values of the controlled object. This is the setup control system 4
Is output to the control model 5 (step 111). next,
Using this control model 5, the control system design system 3 determines the feedback parameter value and the feedforward parameter value of the control device 2 (step 112), and the setup control system determines the operating point in response to the parameter value from step 111. Then, the target value is output to the control device 2 (step 113).

The order of step 112 and step 113 can be exchanged. Also, step 11
Various modifications such as sometimes not performing the process may be considered for No. 2.

17 (a) to 17 (c) show operation examples when the above-described processing configuration is adopted. FIG. 17A shows an example at the time of learning, in which the parameter values of the control model 5 are changed to obtain the states of input and output of the control model 5. FIG. 17B shows the relationship between the parameter and the state obtained in FIG. 17A, that is, the parameter-state function of the control model 5. The neural network 10 is made to learn the state in the input layer so as to output the parameter.

FIGS. 18A to 18D show how to use the neural network 10 at the time of execution after learning is completed. The state of the controlled object 1 in FIG. 18A is input to the neural network 10. Then, the parameter value which is the learning result in the control model 5 at that time is output. This is called reverse setup. The output of the neural network 10 indicates the optimum parameter value of the control model 5 during the operation of the controlled object 1. Therefore,
When this parameter value is input to the control model 5, the control model 5 operates almost the same as the controlled object 1. As a result, the setup system and the control system design system based on the control model 5 output the optimum target value and the structure of the control system or the parameters of the control system, and can reduce the control deviation.

In the above embodiment, when the parameter value obtained by this setup by the neural network is incorrect, a new parameter value may be directly set by the parameter changing section 8 and given to the control model 5.

Next, feedback coefficients f ₁₁ , f ₁₂ , f ₁₃ , f ₁₉ based on modern control theory (for example, Takeshi Tsuchiya, Masashiba Egami, "Modern Control Engineering", P141 to P152, published in Sangyo Tosho, April 1991). , F ₂₁ , f ₂₂ , f ₂₃ , f ₂₄ will be described.

The rolling load formula (plasticity curve) includes the term of velocity tension as shown in the above-mentioned equations 1 and 3. As shown in FIG. 19, between the two pairs of rolls ((i-1) -th stand and i-th stand), the relationship between the speed and the tension of the rolled material is given by the following expression 5.

[0056]

[Equation 5]

Here, L: distance between stands E: Young's modulus b: width of rolled material v _ei : speed of i-th stand entry side rolled material, v _0i −1: (i-1) -th stand exit side rolling The material velocity τ is the tension between the (i-1) -th stand and the i-th stand, and is τ _{b when} viewed from the i stand.

By substituting such a relation into the above-mentioned (Equation 1), etc.

[0059]

[Equation 6]

[0060]

[Equation 7]

[0061]

[Equation 8]

[0062]

[Equation 9]

The relational expression of is obtained.

Here, the operating point is found (setup) by load balance or the like, and the operating point is set to τ _f0 , h ₀ , v _R0 , P _0.
And

Mathematical expressions 6 to 9 are tailed around the operating point.
or deployment. Here, only about Equation 6
r-expansion and omit about the equations 7-9.

[0066]

[Equation 10]

By substituting the equation 6 into the equation 10, the difference equation is obtained. Here, ∂g ₁ / ∂h etc. are influence coefficients.

[0068]

[Equation 11]

The symbols used in the equations 6 to 11 are Δh: thickness deviation Δτ _f : forward tension deviation Δv _R : speed deviation ΔP: load deviation Δv _P : speed command deviation ΔS _P : rolling position command deviation. .

The partial differentiation is h = h ₀ , τ _f = τ _f0 , v _R
= V _R0 , P = P ₀ , v _p = v _p0 , S _p = S _p0 .

Next,

[0072]

[Equation 12]

[0073]

[Equation 13]

The vector variable represents the state variable X and the command variable U, and

[0075]

[Equation 14]

[Equation 11]

[0077]

[Equation 15]

Similarly,

[0079]

[Equation 16]

[0080]

[Equation 17]

[0081]

[Equation 18]

It can be expressed as

The following Equation 19 is obtained by summarizing these Equations 15 to 18.
Becomes

[0084]

[Formula 19]

This is the equation to be obtained.

With such a configuration, the feedback coefficient is obtained by modern control theory.

Next, another embodiment of the present invention will be described with reference to FIGS.
2 is used for the explanation. In the figure, those having the same functions as those of the above-mentioned embodiment are designated by the same reference numerals, and their explanations are omitted.

In this embodiment, when the parameter value is determined based on the input / output state of the controlled object, the parameter value is determined based on the short-term data and the long-term data of the input / output state.

As described above, when the parameter value of the plasticity curve (rolling load formula) changes, it is necessary to change the optimum feedback coefficient. At this time, there are many factors that change the plasticity curve, and even if the rolled material has the same composition and the same plate thickness, plate width, etc., the curve becomes different immediately after the roll is renewed and when the roll is worn. However, in this case, the characteristics gradually change. That is, after the rolls are exchanged, a neuro operation is used for each operation, and the weighted average and moving average are operated on the parameter values at the time of operation and stored in the memory (304 in FIGS. 20 and 21) as long-term performance data Ps. To do.

In this case, even if the rolled material has the same steel type and the same specification, the plastic curve changes depending on the rolling speed, so that the parameters are stored for each operating state (for example, during acceleration / deceleration and during steady speed). .

As shown in FIG. 22, long-term performance data is obtained by rolling materials with the same specifications and steel types (for example, rolling materials 1, 2, ...
n), for each state value, for example, different speed points (eg,
These are data of the input / output states of the controlled object for each v ₁ , v ₂ ), and these are obtained, and these are subjected to neuro operation (reverse setup) to obtain the parameter value for each state value. But,
In rolling mills, the wear factors such as roll roughness (degree of wear) and the concentration of rolling oil change with time.These input / output states change over time, but the long-term actual data shows the change over time. Are averaged, and the change value (instantaneous value) such as the change coefficient is hard to be reflected.

On the other hand, since the output data of the rolling mill contains a lot of noise, it is necessary to perform noise elimination in the reverse setup.

In the processing of long-term performance data, noise is eliminated by using data between many rolled materials.

On the other hand, the short-term actual data is input / output data of one rolling material at different speeds v _{1 to} v _n for a short time as shown in FIG. Therefore, data reflecting the change value (instantaneous value) of the coefficient as described above can be obtained.

Similarly, in the short-term result data, noise removal is performed before reverse setup (block 308 in FIG. 21).
However, it is necessary to average data over a long period in order to reliably remove noise. However, if the data collection interval for averaging is lengthened, the speed state (acceleration / deceleration state, constant speed state) will be different, and therefore the length cannot be increased so much. Therefore, the parameter P _a obtained by the inverse setup using the neuro operation is determined. Can be noisy.

That is, the parameter value P _a obtained from the short-term actual data reflects the instantaneous operating state of the rolling mill,
It is easily affected by noise. On the other hand, the parameter P _s obtained from the long-term performance data is not easily affected by noise, but does not reflect the instantaneous state change of the rolling mill.

Therefore, the parameter calculation P _a = αP _a + (1
-Α) P _s is performed (block 32 in FIG. 21).
0), use long-term data and short-term data. in this case,
α takes a value of 0 to 1, and when α is close to 1, short-term data is reflected and long-term data is ignored, and when α = 0, long-term data is reflected and short-term data is ignored. When the value of α is small, it is used when the number of data used to obtain long-term data is large and the rolling is stable. When the value of α is large, it means that there is no long-term data, there is a large difference from the long-term data immediately after exchanging the rolls, or the case where a steel material that has not been rolled so far is used.

FIG. 2 shows an operation example when such a configuration is adopted.
An explanation will be given using 0. When n-1 rolling material is rolled, data corresponding to various speed states such as acceleration, steady running, deceleration low speed operation, etc. are collected and combined with the data collected from 1st to n-2nd rolling. , Neuro calculation (reverse setup) is performed to obtain the parameter value Ps of the long-term performance data (block 302), and the result is stored in the parameter table (304) for each operating state and specification.

At the time when the rolling of the nth rolled material is started, the neural network uses the short-term result data and sets the parameter value P.
_Determine a (block 310). For the setup, using the parameter values P _s and P _a , the above calculation P = αP _a
+ (1-α) P _s, and the parameter value P at that point
Then, the control model 5 is changed using this value P to perform the setup, and at the same time, the control system design tool 3 uses the control model to feedback the feedback parameters f ₁₁ , f ₁₂ , f ₁₃ , f ₁₄ , f ₂₁ , f ₂₂ , f ₂₃ ,
Obtain f ₂₄ and pass it to the feedback control system 2.

The feedback control system 2 obtains a command deviation from the state of the rolling mill and the feedback parameters, adds the target value for setup and the command deviation, and issues a command to the rolling mill. As a result, the feedback control system can generate a command suitable for the operation of the rolling mill, and thus makes a good response.

In the embodiment of the present invention, since the parameter value of the control model can be accurately determined, even if the plastic curve changes due to the change of the parameter value as shown in FIG. 24, the influence as shown in FIG. The coefficient can be accurately obtained, and therefore, the correct feedback coefficient can be obtained and the control performance can be improved.

In non-interference control such as in a rolling mill whose operating point changes greatly, it is difficult to specify the parameter value of the model of the control target instantaneously, and the conventional method causes various factors. The control model could not be constructed accurately, and even if the modern control theory, which can improve the performance if the control model is accurate, was applied, the ability could not be exhibited. However, by adopting the embodiment described above, an accurate control model can be generated even when applied to a control object with strong nonlinearity,
It is possible to improve control performance.

[0103]

According to the present invention, it is possible to improve the accuracy of setup control (set point control) for determining an operating point for favorably controlling a controlled object having a strong non-linearity and perform favorable control. Become.

[Brief description of drawings]

FIG. 1 is a block diagram of a control system (a) and a learning system (b) according to an embodiment of a control target control device of the present invention.

FIG. 2 is a block diagram showing an entire control device in the embodiment shown in FIG.

FIG. 3 is a block diagram of a control device when the embodiment shown in FIG. 1 is applied to a rolling mill.

FIG. 4 is a diagram showing a relationship between a rolled material and rolls.

FIG. 5 is a diagram showing characteristics of a gauge meter system.

FIG. 6 is a diagram showing a plasticity curve.

FIG. 7 is a diagram showing operating points of the rolling mill.

FIG. 8 is a diagram showing changes in operating points of the rolling mill due to changes in operating conditions.

FIG. 9 is a diagram showing changes in operating points of the rolling mill due to changes in operating conditions.

FIG. 10 is a diagram showing changes in operating points of the rolling mill due to changes in operating conditions.

FIG. 11 is a configuration diagram illustrating an operation of a learning control device during learning.

FIG. 12 is a diagram showing data obtained by learning.

FIG. 13 is a diagram showing a memory that stores data obtained by learning.

FIG. 14 is a configuration diagram illustrating an operation of a learning control device when the learning control device is executed using a learning result.

15 is a flowchart showing a learning processing procedure in the embodiment shown in FIG.

16 is a flowchart showing a processing procedure at the time of executing control in the embodiment shown in FIG.

FIG. 17 is a diagram illustrating a procedure at the time of learning in the embodiment shown in FIG.

18 is an explanatory diagram of how to use the neural network in the embodiment shown in FIG. 1. FIG.

FIG. 19 is a diagram showing a relationship between a rolled material and rolls.

FIG. 20 is a diagram showing an operation of another embodiment of the present invention.

FIG. 21 is a diagram showing a control procedure of the another embodiment.

FIG. 22 is a diagram showing long-term performance data in the above-mentioned another embodiment.

FIG. 23 is a diagram showing short-term performance data in the above-mentioned another embodiment.

FIG. 24 is a diagram showing a change in a model due to a change in a rolling mill parameter.

FIG. 25 is a diagram showing influence coefficients.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Kenichi Yoshioka 5-2-1 Omika-cho, Hitachi-shi, Ibaraki Hitachi Information Control System Co., Ltd. (72) Inventor Yu Saito 5-chome, Omika-cho, Hitachi-shi, Ibaraki No. 1 Incorporated company Hitachi Ltd. Omika Plant (72) Inventor Satoshi Hattori 5-21 Omika-cho, Hitachi City, Ibaraki Prefecture Incorporated Hitachi Ltd. Omika Plant (72) Inventor Masaaki Nakajima Omika Oita, Ibaraki Prefecture 5-2-1, Machi, Hitachi Ltd. Omika Factory (72) Inventor Takashi Okada 7-1, 1-1, Omika-cho, Hitachi City, Ibaraki Hitachi Ltd. Hitachi Research Laboratory (72) Inventor Yasuo Morooka Ibaraki 7-1-1, Omika-cho, Hitachi-shi, Hitachi, Ltd. Inside Hitachi Research Laboratory, Hitachi, Ltd. (72) Inventor, Executive professor Masahisa Omika, Hitachi-shi, Ibaraki 5-2-1, Machi Incorporated company Hitachi, Ltd. Omika factory

Claims (20)

[Claims] 1. A control target is caused to perform a desired operation,
In a control method for giving a command to an actuator for controlling the controlled object, a relationship between each value of the parameter of the model of the controlled object and the input / output state value of the model is obtained by learning, and the learning result is stored in a storage means, Based on the input / output measurement value of the controlled object, the parameter value of the model is determined with reference to the learning result of the storage means, and the command is obtained based on the determined parameter value and given to the actuator. Control method. 2. The control method according to claim 1, wherein the learning is performed using a neural network. 3. The control method according to claim 1, wherein the parameter value is determined by using a neural network. 4. The control method according to claim 1, wherein the controlled object is a non-linear control system. 5. The control method according to claim 4, wherein the controlled object is a rolling mill. 6. The model according to claim 1, wherein the model is a setup model, and target values of input and output of the controlled object are determined by setup control based on the determined parameter values, and the command is obtained and given to the actuator. At the same time, feedback control is performed so that the deviation of the output state obtained from the control target as a result of the command from the target value of the output is zero. 7. The control method according to claim 6, wherein the feedback control determines a feedback coefficient based on the determined parameter value. 8. The control method according to claim 7, wherein an influence coefficient is obtained based on the determined parameter value, and the feedback coefficient is obtained based on the influence coefficient. 9. The control method according to claim 1, wherein the parameter value is determined based on long-term and short-term input / output measurement values of the control target. 10. A control device for giving a command to an actuator for controlling a controlled object so as to cause the controlled object to perform a desired operation, and each value of a parameter of the model of the controlled object and an input / output state value of the model. A means for obtaining the relationship by learning, a storage means for storing the learning result, and a means for determining the parameter value of the model by referring to the learning result of the storage means based on the input / output measurement value of the controlled object. , A means for obtaining a command based on the determined parameter value and giving the command to the actuator. 11. The control device according to claim 10, wherein the learning means includes a neural network. 12. The control device according to claim 10, wherein the means for determining the parameter comprises a neural network. 13. The control device according to claim 10, wherein the controlled object is a non-linear control system. 14. The control device according to claim 13, wherein the control target is a rolling mill. 15. The model according to claim 10, wherein the model is a setup model, and means for giving the command to the actuator determines a target value of input / output of the controlled object by setup control based on the determined parameter value. A means for obtaining the command and giving it to the actuator; and a means for performing feedback control so as to make a deviation from a target value of the output of the output state obtained from the controlled object as a result of the command zero. Control device. 16. The control device according to claim 15, wherein the feedback control means determines a feedback coefficient based on the determined parameter value. 17. The control according to claim 16, wherein the means for determining the feedback coefficient is to obtain an influence coefficient based on the determined parameter value and obtain the feedback coefficient based on the influence coefficient. apparatus. 18. The control device according to claim 10, wherein the means for determining the parameter value is for determining the parameter value based on long-term and short-term input / output measurement values of the controlled object. 19. In a control method for giving a command to an actuator for controlling a controlled object so as to cause the controlled object to perform a desired operation, each value of a parameter of the model of the controlled object and an input / output state value of the model. Relationship is obtained by learning and is stored in the storage means as a learning result, and the parameter value of the model is determined based on the input / output measurement value of the controlled object with reference to the learning result of the storage means. A control method characterized in that the above model is tuned based on a parameter value, and a command is obtained based on the tuned model and given to the actuator. 20. In a control device for giving a command to an actuator for controlling a controlled object so as to cause the controlled object to perform a desired operation, each value of a parameter of the model of the controlled object and an input / output state value of the model. Means for obtaining the relationship by learning, means for storing the result of the learning, means for determining the parameter value of the model with reference to the learning result of the storage means based on the input / output measurement value of the controlled object, A controller that tunes the model based on the determined parameter value, obtains a command based on the tuned model, and gives the command to the actuator.