KR950003800B1 - Shape controller system by using artificial neural network - Google Patents

Abstract

No content.

Description

Shape control system using neural network

1 is a block diagram of a conventional shape control system.

2 is a configuration diagram of a shape control system of the present invention.

3 is a block diagram of a neural network applied to the present invention.

* Explanation of symbols for main parts of the drawings

1: neural network 2: process computer

3: cooling water amount control unit 4: middle roll position control unit

5: middle roll bending load control unit 6: working roll bending load control unit

7: work roll leveling control unit 8: rolling mill

9: strip 10: shape detection sensor

11 signal processing unit 12 Kalman filter

The present invention relates to a shape control system using an artificial neural network to improve the accuracy of the shape influence coefficient used as a control parameter in the shape control system of the cold rolling process.

The shape influence coefficient is an important parameter that represents the ratio of the actual shape change to the unit manipulated variable change of the shape control variable, and its value varies depending on the type of rolled material (determined by the thickness, width, and strength of the coil). Therefore, it is required to manage the unique optimal shape influence coefficient value for each of about 1400 material types.

The shape control system has an initial value table of shape influence coefficients for each type of material. Using this table value at the beginning of rolling of each material, once the closed loop control system is configured, the shape influence coefficient according to the variation of rolling conditions Since the optimal value of is also changed, the Kalman Filter's learning function is used to obtain a new shape influence coefficient box for each control cycle that accommodates these changes.

The conventional shape influence coefficient initial setting table values are obtained by experienced operators using correlation analysis or regression analysis. To reduce the complexity of the analysis process, the number of control variables is limited and the linear relationship between variables is used. Due to the constraints that need to be assumed, the general tendency is low accuracy and requires a lot of manpower and time to find each value for more than 1400 materials.

In the continuous rolling process, thickness control and forming control are simultaneously performed for one rolling mill. The control environment of the two control systems assumes a completely independent relationship. However, it is analyzed that the two control systems exhibit mutual interference during the actual rolling. Especially, the change of the rolling load or the rolling speed, which is a thickness control variable, has a great influence on the shape. Therefore, it is impossible to derive the exact shape influence coefficient reflecting the mutual interference effect of the thickness control system only by the correlation between the shape control variable and the shape considered when obtaining the shape influence coefficient in the existing Kalman filter.

This conventional shape control system, as shown in FIG. 1, has a process computer (not shown) for rolling strips in a rolling mill 8 including a backup roll 23, an intermediate roll 22, and a work roll 21. According to the initial set value table of 2), the coolant amount control part 3 limits the coolant supplied to the work roll 21, and the middle roll position control part 4 and the middle roll bending load control part 5 are the intermediate roll 22 Limiting the position and the load), the work roll bending load control unit 6 limits the bending load of the work roll 21, the work roll leveling control unit 7 provides a constant position control signal to the backup roll (23) Done. The shape of the strip 9 pressurized by this control system is detected as an analog waveform by the shape detection sensor 10 from the exit of the rolling mill 8, so that the quadratic approximation, symmetrical and asymmetrical component separation in the signal processing unit 11, The symmetrical and asymmetrical shape calculations and shape errors are provided to the Kalman filter 12 through the detection process, and the intermediate roll bending load control unit 5 and the work roll bending load control unit as feedback compensation values filtered and output from the Kalman filter 12. (6), the middle roll 22, the work roll 21, and the backup roll 23 by the work roll leveling control unit 7 are controlled.

Here, the analog signal for the performance shape for each control cycle detected by the shape detection sensor 10 is a fourth order approximation.

Separation of Symmetric and Asymmetric Components

Symmetric Geometry Computation

Asymmetric Geometry Calculation

Through the process, the performance shapes ∧1, ∧2, ∧4 are obtained, and this value is compared with the target shape to obtain the shape errors Δ∧1, Δ∧2, and Δ∧4. Each control unit 3, 4, 5, 6, 7 which obtains each control operation amount calculates a control operation amount necessary to correct this shape error, thereby forming a feedback control system in which a new rolling environment is set. At this time, the Kalman Filter 12 calculates the shape influence coefficient by comparing the change of the control operation amount and the shape error. The definition of each shape influence coefficient is as follows.

β11 ~ β22: Symmetrical shape influence coefficient

α: Asymmetric Shape Impact Factor

△ ∧i: Shape change amount

△ Fω: Work Roll Bending Force Change

△ Fi: Change in I.M.R Bending Force

△ SL: Work Roll Leveling Change

Since the shape influence coefficient is a time-varying parameter according to the change of rolling conditions, the Kalman filter adds or subtracts a certain amount proportional to the shape error value on the premise that the shape error during the previous control period is caused by the error of the shape influence coefficient. By doing so, self-learning is performed to find a new shape influence coefficient that can reduce the shape error. The process is as follows.

X (t): Shape influence factor (β11 ~ β22, α) in use at present control cycle

Y (t): Shape target value (∧i ') in current control cycle

Ψ (t-1): change of control operation amount of previous control period (△ Fw, △ Fi, △ SL)

K: Kalman filter gain

Y (t) -Ψ (t-1) X (t): Shape Error (△ ∧i)

As can be seen from the above equation [7], the function used for deriving the shape influence coefficient from Kalman filter is

The shape influence coefficient obtained from the Kalman filter is defined only as a linearized two-variable function expressed by, and it is not only able to accommodate the influence of the shape control variables except △ Fw on ∧2, but also the control system that changes the thickness control variables on the shape change. It is impossible to reflect the interference shape.

The use of such a low shape influence coefficient results in a significant reduction in the shape correction ability under normal rolling conditions, and especially in the beginning and the end of the rolling phase in which changes in the rolling conditions suddenly occur. Accurate shape influence coefficient derivation requires the definition of a new nonlinear multivariate function that can consider not only changes in shape control parameters but also thickness control variables other than ΔFw, △ Fi, and △ SL. Is difficult to solve.

An object of the present invention is to apply a neural network, which is evaluated to have excellent learning ability on nonlinear correlations between various variables. The thickness control variable is not considered in the Kalman filter as well as the shape control variable. The system can automatically calculate and obtain accurate shape influence coefficients to accommodate the influence of mutual control between two control systems, thereby reducing the manpower and time required for conventional analytical methods and reducing the shape steepness of cold rolled products. To provide a shape control system using a neural network that can be improved.

Hereinafter, the present invention will be described with reference to the accompanying drawings.

2 is a configuration diagram of the system of the present invention, in which the initial setpoint table data of the process computer 2 includes the coolant amount control section 3, the intermediate roll position control section 4, the intermediate roll bending load control section 5, and the work roll bending. The rolling mill is configured to be transmitted to the load control unit 6 and the work roll leveling control unit 7, respectively, and includes a work roll 21, an intermediate roll 22, and a backup roll 23 in accordance with these control units 3-7. (8) is configured to be controlled.

The strip 9 rolled through the rolling mill 8 is detected by the shape detection sensor 10 in the form of an analog waveform so that the shape error is detected by the signal processor 11 and then provided to the Kalman filter 12. And an error control signal that is self-learned and fed back by the Kalman filter 12 is provided to the intermediate roll bending load control unit 5, the work roll bending load control unit 6, and the work roll leveling control unit 7. do.

Meanwhile, the thickness control and shape control results detected by the thickness and shape sensors 24 and 25 are calculated by the process computer 2 to be provided to the neural network 1, and the shape influence calculated by the neural network 1. The coefficient output is configured to be provided to the initial setting table and the Kalman filter 12.

Referring to the operation of the present invention configured as described above is as follows.

The error backpropagation learning function of the multi-layer perceptron structure among the various structures of artificial neural network, a new field of artificial intelligence, enables not only shape control variables but also thickness control variables and shape influence coefficients. It is possible to derive the shape influence coefficient by the nonlinear multivariate function of

3 shows the structure of a multilayer perceptron neural network applied in the present invention to perform this function. The input of the neural network (1) is a shape control variable that affects the shape during the actual rolling process, in addition to the roll bending load (Fw, Fi) and the shape information ∧1, ∧2, ∧4, which are already considered by the Kalman filter. The symmetric shape influence coefficient (β11-) as the output to add and calculate the work side and drive side roll force, the entry and exit tension, and the rolling speed as roll position and thickness control parameters. β22) and asymmetric shape influence coefficient (α).

The learning goal of the neural network 1 is to reflect the influence of additional control variables in the calculation of the shape influence coefficient of the Kalman filter performed in the control cycle. When various input data 13 representing the current rolling conditions are applied to the neural network 1, the output 15 is obtained based on the current weight 14 and the shape influence coefficient value of the output and the Kalman filter ( 16) are compared to detect an error and then the weights 14 between each layer are corrected in the direction of reducing the error. This process is performed repeatedly until the error falls below a small enough value to learn the nonlinear correlation between all control variable inputs and the output of five shape impact coefficients.

When the shape influence coefficient is obtained from the neural network (1) where the learning is completed, the current rolling conditions are applied by inputting the values of each control variable in the current control period and the target value of each shape to be obtained by the change of these control variables. A satisfactory shape influence coefficient can be obtained. The reason why the neural network (1) is composed of 75 groups for deriving the shape influence coefficient in FIG. 3 is that the rolling conditions are greatly changed according to the material type (about 1400 kinds) of the cold rolling process, and the shape influence coefficient changes accordingly. This is because, if one neural network is configured, the learning ability is degraded due to the excessive amount of information to be learned. In other words, among the 1,400 kinds of materials, materials with similar rolling environment are grouped together and reclassified into 75 unit groups, and neural networks for each group are made to share the changes in rolling conditions to ensure sufficient learning ability. .

First, the cooling water amount control unit 3, the intermediate roll position control unit 4, the intermediate roll bending load control unit 5, the work roll bending load control unit 6, and the work roll leveling according to the initial setting table of the process computer 2 The control part 7 controls the work roll 21, the intermediate roll 22, and the backup roll 23 of the rolling mill 8, respectively, and performs rolling. Accordingly, when the shape detection sensor provided on the exit side of the rolling mill 8 detects the shape of the strip 9 as an analog waveform, the signal processing unit 11 generates the formation errors Δ∧1, Δ∧2, and Δ∧4 according to the above formulas (1)-(6). Is obtained and provided to the Kalman filter (12).

Accordingly, the Kalman filter 12 compares the change of the rolling mill control operation amount and the shape error, and calculates the shape coefficient according to the above formula ⑦-⑪.

On the other hand, based on the thickness and shape control results calculated by the process computer 2 according to the thickness and shape sensors 24 and 25, the neural network 1 considers the current rolling environment to obtain the desired shape. The coefficient of influence X '(t) is obtained and provided to the initial setting table of the process computer 2 and the Kalman filter 12.

That is, the neural network 1 continuously calculates the shape influence coefficient value for modifying the initialization table of the process computer 2 and the shape influence coefficient required for the self-learning process of the Kalman filter 12 continuously at every control cycle. You will perform two functions. Therefore, the equation 나타내는 representing the self-learning process of the Kalman filter has a new meaning as follows.

Therefore, in the Kalman filter 12, a value K (Y (T) -Ψ (t-1). Each intermediate roll bending load control unit controlling the work roll 21, the intermediate roll 22, and the backup roll 23 of the rolling mill 8 using a more accurate shape impact coefficient by performing a self-learning process of (5), the work roll bending load control section 6 and the work roll leveling control section 7 can feed back the more accurate rolling control operation amount.

As described above, according to the present invention, the neural network automatically derives the shape influence coefficient initial setting value applied to the shape control system and modifies the table of the process computer, thereby reducing the manpower, time, and control deadband at the initial stage of control. Shape steepness can be improved by shortening the band convergence time and improving the control ability in the normal rolling section.

Claims (1)

The controller 3-7 for controlling the angular rolls 21-23 of the rolling mill 8 according to the initial setting table of the process computer 2 and the shape of the strip 9 rolled by the rolling mill 8 are detected. The shape detection coefficients are calculated from the shape detection sensor 10, the signal processing unit 11 for detecting shape errors in the analog detection signal of the shape detection sensor 10, and the shape error coefficients of the signal processing unit 11. The Kalman filter 12, which feeds back the shape correction control signal to (5-7), receives various variables and shape information from the thickness and shape control results of the process computer 2 by the sensors 24 and 25. The output 15 is obtained based on the weight 14 of. The weight 14 of each layer is corrected by comparing this output with the shape influence coefficient value 16 of the Kalman filter. Recall shape influence coefficients to obtain target shape by learning with nonlinear correlation The initial setting value table and-control system using neural network comprises a neural network (1) provided to the Kalman filter 12 of the process computer (2).