KR19990035067A - Shape Control Method Using Fuzzy-Neural Network - Google Patents

Abstract

The present invention relates to a shape control method for a cold rolled steel sheet using an iterative learning network control system. In consideration of the nonlinear characteristics of a shape control system, a fuzzy rule is generated in the design of a nonlinear system to generate a fuzzy rule to generate a shape parameter calculation model. Since the number of rules is too exponentially proportional to the number of variables in the system, the fuzzy-neural network controller is constructed by applying to the shape parameter calculation using neural networks to increase the accuracy of the model. The main purpose is the shape control method applied to the shape parameter calculation, and it is possible to control the shape with high accuracy not only for the thick material but also for the complex shape of the strip below the heavy material. There is an effect that can be effectively applied to the field.

Description

Shape Control Method Using Fuzzy-Neural Network

The present invention relates to a method for controlling the shape of a cold rolled steel sheet by a control system using a repetitive learning network, and more particularly, a complex shape for a strip below a heavy material as well as a thick material using a repetitive learning network control system. The present invention also relates to a shape control method of a cold rolled steel sheet capable of controlling the shape with high accuracy.

In general, in the production of steel, the cold rolling process is a process of finishing rolling a strip-shaped steel sheet rolled from 1 mm to several mm by hot rolling into a thin steel sheet used as an exterior plate material for automobiles or households by cold rolling. Since this cold rolling process is the final process of making a steel plate into the size and material normally required by a user, high precision is required for the quality of plate | board thickness, a shape, etc.

In the cold rolling process, rolling mills are initially set and controlled by a process computer and a sequence. Among them, a setup model for initial setting control is controlled by a technique based on the conventional rolling theory. The set-up model is based on the operating conditions such as the final product calculation method, the interval between the rolls of the rolling mill, the rolling speed, and the tension on the rolled material. As a calculation, it is the most important control step affecting the productivity, stability and quality of the cold rolling process.

Since such conventional rolling theory requires variables that are difficult to measure, such as the coefficient of friction between the roll and the rolled material and the deformation resistance of the rolling control, they must be calculated based on their answers, which may cause a large error in the calculation result. It is a big obstacle.

Therefore, in order to suppress the occurrence of quality aging or poor operation, the experienced operator always monitors the calculation result of the setup model manually, and in case of abnormality, the operation value is stabilized by correcting the calculated values such as the interval between rolls and rolling speed. .

Operation can be stabilized by manual adjustment of the operator, but this leads to an extension of the setting change time, and the setting change in the operation is usually performed at a low rolling speed, thereby sacrificing productivity.

In addition, the control system is designed to mathematically design the characteristics of the system to be controlled and to configure the controller using this, most of the shape control plant (plant) is inherently non-linear characteristics and the system (parameter) of the system (parameter) is constantly changing, so Mathematical design using conventional system theory is very difficult. In addition, as the size of the system increases, the number of parameters increases geometrically, which complicates the system and has uncertainty characteristics such as noise and time-varying system. Therefore, the classical control theory cannot achieve satisfactory performance.

In order to solve this problem, the repetitive learning network control system with excellent control function and applied learning function has been developed and applied based on the past operation experience which the operator performs manually.

Unlike the conventional approach, the control method of the shape control equipment using the iterative learning network is to learn the dynamic characteristics of the shape control equipment by learning the transfer function of the system with the weight, and then configure the controller using the information. This solves the complexity of increasing parameters and has a strong adaptability to real-time processing and system changes.

Korean Patent Publication No. 94-13639, "Shape Control System Using Neural Networks," introduces a function of deriving the adjustment of the initial setting list using neural networks, but uses a Kalman filter having many nonlinear factors. ) Falls.

Korean Patent Application No. 97-11944, "Shape Control Method of Cold Rolled Steel Sheet Using Neural Network," performs a control operation without using a shape influence coefficient of a control target.

Shape detectors, which are the most important equipment for automatic shape control of cold rolled steel sheets, are generally classified into contact type and non-contact type. A representative type of contact type is a stressometer, and a representative type of non-contact type is magnetic suction type. A measuring instrument is mentioned.

In the contact type, the shape detector has the form of a deflector roll, and a method of measuring contact with a plate by arranging pressure sensors at regular intervals along the width direction of the roll.

The non-contact magnetic suction type uses the fact that the distance between the detector and the steel sheet is inversely related to the sheet tension. The conventional method of controlling the shape of the cold rolled steel sheet by the non-contact type is as follows.

As shown in FIG. 1, the measuring object strip (cold rolled steel sheet) 13 that exits the work roll 12 of the finishing mill is a shape detector 1 installed at a lower distance from the bottom of the strip 1. The shape data is measured by the shape detector 1 while passing through.

The principle of measuring the shape data by the shape detector 1 is that when the strip 13 passes through the magnetization state by applying a constant current to the excitation coil of the shape detector 1, The degree of attraction due to the magnetic force varies depending on the gap between (13) and the shape detector 1, and the degree of attraction is measured by the voltage value.

The detected signal of the shape detector 1 is input to the program logic controller 14, and the input signal undergoes a process of detecting a bad signal, deleting and correcting the bad signal.

The shape signal Cl received through the shape detector 1 is preprocessed to detect defects in the shape data.If there are three or more outliers in a channel other than the two channels, or two outliers are continuously If there is, the report data is deleted as shape defect data. Otherwise, the shape data is corrected by preprocessing which approximates the shape data using both neighboring data.

The corrected shape data U (i) is converted into an elongation El (i) in the width direction represented by the following equation.

ΔU (i): U ₀ -U (i),

E: Young's modulus (2 × 10 ⁴ kg / mm ² ),

U ₀ : shape good unit tension,

U (i): Shape failure tension.

U ₀ and U (i) of Equation 1 are expressed as follows.

G: gain rate,

B: plate width,

P ₀ : shape defect unit displacement,

F: pressure applied to the coil width direction by the magnetic attraction force of the shape detector,

P (i): Shape defect displacement (mm) (P (i) = B · P0).

Substituting Equations 2 and 3 into Equation 1, the elongation El (i) is converted into Equation 4 below.

In the next process, we perform preprocessing to approximate the function to the shape data, smooth them exponentially, smooth them in the plate width direction, and smooth each individual sensor.

Equation 4 is obtained by approximating the transformed amount of shape stretching as shown in Equation 4 by the fourth order function. In Equation 5 below, X represents a distance from the center of the plate and if the value of X is 1 and -1, it means the end of the plate.

y = λ ₀ + λ ₁ X + λ ₂ X ² + λ ₃ X ³ + λ ₄ X ⁴

The process of deriving the shape parameter from the fourth order function of Equation 5 is as follows.

First, the function is symmetric about the plate center. y _s = λ ₂ X ² + λ ₄ X ⁴ And asymmetrical components y _a = λ ₁ X ¹ + λ ₃ X ³ After separating the coefficients λ _2, λ ₄ and λ ₁ , λ ₃ by linear transformation of Equations 6 and 7 Λ ₂ , Λ ₄ Wow Λ ₁ , Λ ₃ Find the shape pattern with.

In Equations 6 and 7 Λ ₂ = y _s (1), Λ ₄ = y _s (1 / )ego Λ ₁ = y _a (1), Λ ₃ = y _a (1 / )to be.

About symmetrical components Λ ₂ Denotes the amount of stretching in the judgment unit. Λ ₄ Is 1 / for the normalized X axis The stretching amount of the part is shown. About asymmetric components Λ ₁ Also represents the amount of stretching in the judgment unit. Λ ₃ Is 1 / for the normalized X axis The stretching amount of the part is shown.

The symmetrical component model and the asymmetrical component model are represented as follows.

Symmetric Component Model

Asymmetric Component Model

β _{11 to} β ₂₂ : symmetric component influence coefficient

ΔF _W : Work roll bending force change

ΔF _i : Mid-roll bender force change

α _{11 to} α ₂₂ : asymmetric component impact coefficient

ΔDF _W : Work roll bender level change

ΔD _s : Roll gap change amount

β _{11 to} β ₂₂ and α _{11 to} α ₂₂ are coefficients indicating the effect of setting change amount of the operating stage on the change of the shape parameter.Asymmetric control factors include work roll bender force level change and roll gap change amount. Asymmetric control is performed by changing only the roll interval.

Therefore, the asymmetric component equation model of Equation 9 is as follows.

Shape parameters calculated from the above equations ( Λ ₁ To Λ ₄ ) Is approached to the target shape using the method of Mountain climbing search method.

The ridge determination method first determines the search width based on the current input value, and then determines 17 search points according to the search width.When the search point is determined, the shape change at this search point is predicted through the shape equation model. It is a method to select the point where the evaluation function set among the three search points is the minimum as a control input.

The performance of the shape control system depends on how well the shape change at each search point is predicted by the mountain climbing method, that is, how close the shape influence coefficient is to the actual value, and the shape influence coefficient β ₁₁ in the mathematical model for symmetry control. ? ₂₂ has the following meaning.

₁₁ g of β in the formula for example, β ₁₁ is the shape parameter is changed only when the work rolls vendor force without changing the intermediate roll force vendor Λ ₂ It is a value that quantitatively indicates how much will change. This value is not fixed but changed because it is affected by disturbances that appear in various forms, such as noise, characteristics of shape control equipment, plate material, and rolling speed.

The coefficient of influence currently used depends on theoretical calculation formulas for plate width and thickness, using averaged values after several repeated experiments, and when the thickness, width and range of steel grades are changed by making a table by plate width, thickness, and steel grade. Each time it is loaded into the SCC (Supervise Computer Control) 15 and can be used.

The conventional shape control method described above uses a quasi-optimal control technique that approximates the relationship between the bender force and the shape coefficient to a linear model. The bender force and the shape coefficient are represented by the following nonlinear model. Can be represented.

In the above formula, G ₂ (F _W , F _I ) and G ₄ (F _W , F _I ) represent nonlinear components, and the ridge method ignores these nonlinear components and uses the following linear model.

Λ ₂ = β ₁₁ F _W + β ₁₂ F _I

Λ ₂ = β ₁₁ F _W + β ₁₂ F _I

However, this linear model has a problem that the neglected nonlinear components can have a big influence on the shape control when used in the complex shape control of the strip below the middle thin film. have.

The actual cold rolling control system is a complex nonlinear time-varying multi-input multi-output system and it is difficult to find a mathematical model of dynamic characteristics.

Λ ₂ = β ₁₁ F _W + β ₁₂ F _I

Λ ₄ = β ₂₁ F _W + β ₂₂ F _I

However, non-linear components neglected during the linearization process can be a non-negligible factor in strips below the heavy material, which can cause many problems when controlling the shape of thin strips below the heavy material by the ridge system based on the linearized model. .

The present invention is to solve the above problems, cold rolling using a repetitive learning network control system suitable for designing and controlling a nonlinear system, which can control the shape with high precision even for the complex shape for the thick material or the strip below the heavy material. It is an object of the present invention to provide a method for controlling the shape of a steel sheet.

1 is a state diagram showing the structure of a shape control system;

2 is a state diagram showing a neural network structure of the present invention;

3 is a state diagram showing a control system structure of the present invention;

4 is a state diagram showing the structure of a fuzzy-network controller of the present invention;

5 is a state diagram showing the program structure of the present invention;

Figure 6 is a graph showing the experimental results of the middle roll buckle form,

7 is a graph showing the experimental results of the edge (edge) shape,

8 is a graph showing the results of the M-shaped experiment,

9 is a graph showing the experimental results of the W form

<Description of Major Symbols in Drawing>

1: Shape detector 10: Backup roll

11: Medium roll 12: Work roll

13: strip 14: program logic controller

15: SCC 16: Hydraulic pressure

17: Medium roll bender 18: Work roll bender

19: Middle roll shift 20: Initial set value

21: Geometry Data Performance

In order to achieve the above object, the present invention applies a nonlinear characteristic, an adaptive learning function, a large parallel processing function, and a characteristic of a repetitive learning network having many degrees of freedom to a cold rolled shape control system.

The most commonly used method of preliminary information of the shape control equipment is the configuration of an iterative learning network emulator. As shown in FIG. 2, the dynamic characteristics of the shape control facility are learned using the iteration learning network emulator in advance, and then the iteration learning network controller is learned using the iteration learning network emulator for the shape control facility.

In order to examine the correlation between the shape data of time (t + 1) and the previous control data, the data of the emulator were analyzed using the correlation coefficient. This analysis method is used to analyze the linear causal relationship between given variables in the statistical approach for time series prediction.

The correlation coefficient r of the random variables X and Y is defined as follows.

The correlation coefficient ranges from -1 to 1, indicating that the larger the absolute value, the greater the linear correlation.

The emulator's inputs and outputs are determined from the correlation of shape data.

FIG. 2 shows a neural network structure. The iterative learning network has 10 input layers, 2 output layers, 2 hidden layers, and each hidden layer has 30 nodes.

The sigmoid function of the output node is Λ ₂ , Λ ₄ The linear function is used to learn random values without knowing the maximum value of. The sigmoid function of the remaining nodes to be.

Λ ₂ Wow Λ ₄ Acts as a bias for nonlinear control equipment, Λ ₂ Wow Λ ₄ To minimize the influence of Λ ₂ Wow Λ ₄ Of the output through another iterative learning network ΔΛ ₂ Wow ΔΛ ₄ Connect to

The control system of the present invention has a structure in which the shape control data input from the shape control equipment by the input / output data recognition device is input to the neural network learner through the fuzzy rule generation system and the shape control equipment interface device as shown in FIG.

Deadzone in front of the controller Λ ₂ Wow Λ ₄ Were controlled until all of them were in 3 I unit and maximum 5 control was performed.

The structure of the fuzzy neural network controller is as shown in FIG. The controller generates a fuzzy rule by using an adaptive purge controller in consideration of the nonlinear characteristics of the shape control system and uses it in the shape parameter calculation model.

The fuzzy rule constructed in the present invention is expressed as follows.

R _i : If X ₁ is A ₁ ⁱ , X ₂ is A ₂ ⁱ ,… , X _m is A _m ⁱ

y ⁱ = a ₁ ⁱ X ₁ + a ₂ ⁱ X ₂ +… + a _m ⁱ X _m

In the above formula R _i Is the i th fuzzy rule at each stand, A _j ⁱ Is a fuzzy set, X _j Is an input variable, y ⁱ Is the output of the i th fuzzy rule, and a _j ⁱ Is a constant that must be determined for each rule.

Input variable value X ₁ , X ₂ ,… , X _m Each fuzzy rule is applied and the fuzzy result is inferred using Mandami's inference method, and the deduced fuzzy result is calculated as one result using the area center method .

Where n is the number of fuzzy rules applied, g ⁱ Represents the minimum value of membership belonging to the fuzzy set for each rule applied as follows.

A _j ⁱ (X _j ) Is X _j An input variable called A _j ⁱ Denotes the degree of belonging included in the fuzzy set. However, since the number of rules becomes too exponentially proportional to the number of variables in the system, the accuracy of the model is improved by applying to the shape parameter calculation using neural networks.

By applying the fuzzy-neural network controller configured as described above to the shape parameter calculation, various problems resulting from the conventional nonlinearity can be solved.

The overall program structure of the present invention is shown in FIG.

The present invention will be described in more detail with reference to the following Examples.

(Example)

In the present invention by simulating the steel sheet having a shape defect Λ ₂ Wow Λ ₄ The shape was controlled until all were within 3 I units , and a maximum of 5 controls were performed.

The results of the shape control on the shape of the middle roll buckles are shown in the following table and the results are shown in FIG. 6.

Control frequency ΔF _I ΔF _W Λ ₂ Λ ₄ 0 - - -20 -7.5 One -1.666 -1.390 -5.272 -3.142 2 -0.927 -0.738 -1.783 -2.141

The shape control results for the edge wave form are shown in the following table and the results are shown in FIG.

Control frequency ΔF _I ΔF _W Λ ₂ Λ ₄ 0 - - 20 7.5 One 1.666 1.390 7.221 1.649 2 1.016 0.816 2.570 -0.126

The shape control results for the M-wave shape are shown in the following table and the results are shown in FIG. 8.

Control frequency ΔF _I ΔF _W Λ ₂ Λ ₄ 0 - - 7.5 15 One 1.398 1.134 2.978 6.960 2 0.898 0.707 1.538 2.945

The shape control results for the W-wave shape are shown in the following table and the results are shown in FIG. 9.

Control frequency ΔF _I ΔF _W Λ ₂ Λ ₄ 0 - - -7.5 -15 One -1.398 -1.134 -2.790 -8.628 2 -0.950 -0.749 -1.460 -5.283 3 -0.661 -0.513 -0.901 -3.519 4 -0.476 -0.366 -0.600 -2.582

In the graphs of FIGS. 6 to 9, the numbers on the horizontal axis represent shapes. When the values on the horizontal axis are 0, the center portion of the strip is -1 and 1, and the edges of the strip, that is, the work side and the drive side. ) And the vertical axis indicates the degree of shape defect.

The shape control method of the cold rolled steel sheet using the iterative learning network control system of the present invention as described above takes into account the nonlinear components ignored by the conventional cold rolling controller, so that the nonlinear components neglected when used in the complex shape control of the strip below the medium thickness This problem can be greatly influenced by the shape control to solve the problem that a lot of defects can be generated in the shape of the strip below the heavy thin.

Therefore, it is possible to control the shape with high accuracy not only in the thick material but also in the complex shape for the strip below the heavy material, and it is suitable for the design and control of the nonlinear system, and it can be effectively applied to the important applications of the industry.

Claims (1)

The process of learning the dynamic characteristics of the shape control equipment by using the iterative learning network emulator, and learning the iterative learning network controller using the iterative learning network emulator for the shape control equipment. when ; The process of determining the input and output of the emulator from the correlation analysis of the shape data so that the iterative learning network has a structure of 10 input layers, 2 output layers, 2 hidden layers, and 30 hidden nodes. and ; In consideration of the nonlinear characteristics of the shape control system, the controller arranges an adaptive purge controller in the nonlinear system to generate a fuzzy rule for use in the shape parameter calculation model. R _i : If X ₁ is A ₁ ⁱ , X ₂ is A ₂ ⁱ ,… , X _m is A _m ⁱ y ⁱ = a ₁ ⁱ X ₁ + a ₂ ⁱ X ₂ +… + a _m ⁱ X _m ; Since the number of rules increases exponentially in proportion to the number of variables in the system, the configuration includes a process of controlling the shape by applying neural networks to the shape parameter calculation to increase the accuracy of the model. Shape Control Method Using Fuzzy-Neural Network