TWI792240B - Method for adjusting control parameters used in rolling mill process - Google Patents

Abstract

A method for adjusting control parameters used in rolling mill process, which includes: (a) providing a virtual rolling mill; (b) providing control parameters corresponding to an input shape as input parameters of the virtual rolling mill; (c) generating a predicted shape according to machine parameters, the input shape and the input parameters by the virtual rolling mill; (d) inputting the predicted shape and a target shape corresponding to the input shape to an evaluation function to generate a shape score; (E) inputting the shape score to an optimization model to generate adjusted control parameters; (f) taking the adjusted control parameters as the input parameters; and (g) repeating steps (c) to (f) until the optimal control parameters for the input shape are generated.

Description

Adjustment method for control parameters of rolling process

The present invention relates to the technical field of rolling process, in particular to a method for adjusting control parameters for rolling process.

General steel coil products need to go through rolling process, such as hot rolling and cold rolling to form the final product. In order to make the mechanical properties of the product comply with predetermined specifications, in the conventional technology, the mechanical properties are measured after the product is produced, and if the mechanical properties do not meet the specifications, at least one stage of process control parameters is adjusted. The current control method of a single rolling mill considers the concept of time series, and adjusts parameters according to the decision-making of the fuzzy control system built in the rolling mill. The single rolling mill will send back a piece of data every 20 milliseconds, including the exit shape, machine parameters and control parameters, etc. The built-in fuzzy system of the system compares the difference between the returned export shape and the target shape with several built-in shape difference templates, and uses the ratio as a reference for adjusting control parameters.

However, the rolling process will be affected by machine parameters, steel strip characteristics, control parameters, and other factors at the same time. At the same time, it contains many controllable and uncontrollable factors, which cannot be explained by existing physical, mathematical or chemical models. Due to the long-term, unstable state of the problem, and many nonlinear factors, it cannot be processed by differential equations. This often results in a final shape that does not match the target shape, thus increasing production costs.

Therefore, how to effectively improve the control parameters used in the rolling process so that the exit shape is consistent with the target shape is a topic of concern to those skilled in the art.

An object of the present invention is to provide a method for adjusting the control parameters of the rolling process, so as to reduce the error between the exit shape and the target shape, thereby reducing the cost and improving the efficiency of the process.

To achieve the above-mentioned purpose, the present invention provides a method for adjusting control parameters of a rolling process, said method comprising the following steps: (a) providing a virtual rolling mill, which corresponds to the rolling machine used for the rolling process ; (b) providing a plurality of control parameters corresponding to the input shape as a plurality of input parameters of the virtual rolling mill; (c) generating a predicted shape according to the plurality of machine parameters, the input shape and the input parameters through the virtual rolling mill ; (d) inputting the predicted flatness and the target flatness corresponding to the input flatness into the evaluation function to generate a flatness score; (e) inputting the flatness score into the optimization model to generate a plurality of adjusted control parameters; (f) using the adjusted control parameters as the input parameters; and (g) repeating steps (c) to (f) until the control parameters most suitable for the input shape are produced.

According to some embodiments of the present invention, after step (d), the adjustment method further includes: (h) comparing the shape score with a threshold value, wherein when the shape score is less than or equal to the threshold value, step (e) is not performed Go to step (g).

According to some embodiments of the present invention, after step (d), the adjustment method further includes: (h) comparing the shape score with a threshold value, wherein when the shape score is less than or equal to the threshold value, step (e) is not performed Go to step (g).

According to some embodiments of the present invention, the virtual rolling mill is a machine learning model established by ensemble learning and/or deep learning.

According to some embodiments of the present invention, the virtual rolling mill is established using an extreme gradient boosting (eXtreme Gradient Boosting, XGBoost) algorithm.

According to some embodiments of the present invention, the virtual rolling mill is established using a deep neural network (Deep Neural Networks, DNN) model.

According to some embodiments of the present invention, the evaluation function compares the predicted shape to the target shape to generate a shape score that reflects an error condition between the predicted shape and the target shape.

According to some embodiments of the present invention, the evaluation function is one of root mean square error, R squared, mean absolute percentage error, mean absolute error, and Hill inequality coefficient.

According to some embodiments of the present invention, the control parameters include temperature, rolling speed, rolling force, and tension of the rolled piece.

According to some embodiments of the present invention, the optimization model uses a Cross-Entropy algorithm to generate the adjusted control parameters according to the shape score.

The present invention utilizes the XGBoost model of integrated learning and/or the integrated neural network model of deep learning to create a virtual rolling mill, and continuously predicts the status of equipment or systems through data integration. In addition, the virtual rolling mill is further combined with an optimization model to adjust and calculate the most appropriate control parameters only by giving necessary parameters, thereby improving the overall quality and avoiding the high cost and high risk that may be caused by direct testing in real equipment. Therefore, it can be widely used in the prediction of unknown parts in the manufacturing process.

In order to make the above and other objects, features, and advantages of the present invention more comprehensible, preferred embodiments of the present invention will be exemplified below in detail together with the attached drawings.

FIG. 1 is a schematic diagram of an electronic device 100 for adjusting control parameters used in a rolling process according to some embodiments of the present invention. The electronic device 100 can be various forms of computers, computers or server systems. The electronic device 100 includes a storage unit 110 and a processing unit 120 . The storage unit 110 can be a volatile memory or a non-volatile memory. The storage unit 110 stores a plurality of historical data previously used in the rolling process, including a plurality of control parameters, a plurality of machine parameters, a plurality of entry shape data and a plurality of exit shape data. In addition, the storage unit 110 also stores a plurality of instructions, so that the processing unit 120 executes these instructions to implement a method for adjusting control parameters. The processing unit 120 may be a central processing unit, a microprocessor, a microcontroller, a digital signal processor, an application-specific integrated circuit, and the like.

Please also refer to FIG. 2 , which is a flowchart of a method 200 for adjusting control parameters used in a rolling process according to some embodiments of the present invention. First, in step S210, a virtual rolling mill is provided. The virtual rolling mill corresponds to the rolling machine used for the rolling process. The virtual rolling mill simulates a physical rolling machine to predict the exit shape of the steel coil after being rolled by the rolling machine under the condition of receiving the same parameters and the entry shape of the corresponding steel coil, which is called predicted shape. Specifically, the processing unit 120 can access a plurality of historical parameters and data used by the rolling machine in the rolling process from the storage unit 110, and establish a machine learning model based on these information, which is called a virtual rolling mill in the present invention. In some embodiments, an ensemble learning-based extreme gradient boosting (eXtreme Gradient Boosting, XGBoost) algorithm can be used to build a virtual rolling mill. In some embodiments, a deep neural network (Deep Neural Networks, DNN) model of deep learning can be used to build a virtual rolling mill. In some embodiments, the XGBoost algorithm and the DNN model can be used together, and the prediction performances of the two can be compared and combinations of different hyperparameters can be tested to obtain the most suitable prediction model.

Taking the XGBoost algorithm as an example, XGBoost is an improved method of Gradient Boosting Decision Trees (GBDT). GBDT is a boosting method that can be used for both regression and classification problems. In short, GBDT uses the negative gradient of the loss function as a residual approximation in the boosting tree algorithm for regression problems to fit a regression tree. The objective function of the GBDT algorithm is only the loss function, and XGBoost has improved on this basis, adding a regular term to prevent the model from being overly complex.

The basic concept of XGBoost is to continuously add trees and grow a tree by continuously performing feature splits. Every time a tree is added, a new function is learned to fit the last predicted residual. In each tree, it will fall to a corresponding leaf node, and each leaf node corresponds to a score. Finally, you only need to add up the scores corresponding to each tree to get the predicted value of the sample.

Specifically, the objective function (Obj) of XGBoost can be shown in the following formula (1), which consists of two parts. The first part is the loss function, which is used to measure the gap between the predicted score and the real score, and the other part is the regularization The transformation term (such as formula (2)), this part is also different from the gradient boosting algorithm. The regularization term includes the number T of leaf nodes and the score ω of leaf nodes. γ can control the number of leaf nodes, and λ can control the score of leaf nodes from being too large to prevent overfitting. In other words, XGBoost is to find an f that minimizes the objective function.

The present invention uses the XGBoost algorithm to train multiple parameters used in the rolling process in the past to establish a virtual rolling mill, which has multiple advantages. For example, it uses a number of strategies to prevent overfitting, such as regularization terms. In addition, the optimization of the objective function utilizes the second derivative of the loss function with respect to the function to be sought. Furthermore, it supports parallelization. Although there is a serial relationship between trees, nodes at the same level can be parallelized. Specifically, for a certain node, the best split point is selected in the node, and the candidate split point calculation gain is parallelized with multiple threads, so the training speed is fast.

It should be noted that although the virtual rolling mill in this embodiment is implemented by using the XGBoost algorithm and/or the DNN model as an example, the present invention is not limited thereto. As long as the integrated learning and/or deep learning is used to realize the virtual rolling mill to simulate the output of the physical rolling mill, it is within the scope of the present invention.

After the virtual rolling mill is established, the present invention further optimizes the established virtual rolling mill to try to obtain more suitable control parameters. Returning to Fig. 2, next, in step S220, a plurality of control parameters are provided as a plurality of input parameters of the virtual rolling mill. Then in step S230, a predicted shape is generated through the virtual rolling mill according to a plurality of machine parameters, an input shape and a plurality of input parameters. It should be understood that in these two steps, the input shape and machine parameters are fixed values, while the control parameters are variables. The input shape is the known shape of the steel strip before rolling process. The machine parameters are the machine characteristics of the virtual rolling mill corresponding to the physical rolling machine, such as roll diameter, hardness or material of the rolling mill, tensile strength, yield stress, and elongation, etc. . The control parameters may be, for example, temperature, rolling speed, rolling force, and tension of the workpiece to be rolled, and so on.

Next, in step S240, the predicted flat shape and the target flat shape corresponding to the input flat shape are input into the evaluation function to generate a flat shape score. The target strip shape represents the actual exit strip shape produced by the physical rolling machine after passing the entry strip shape through the rolling process, and the purpose of the present invention is to combine the preset strip shape produced (simulated) by the virtual rolling mill with the The exit plate shape produced by the solid rolling machine under the same parameters is consistent. Therefore, the evaluation function can be used to compare the predicted flatness with the target flatness to generate a flatness score. The flatness score can reflect the error between the predicted flatness simulated by the virtual rolling mill and the actual target flatness.

In some embodiments, the evaluation function used to evaluate the error may include root mean square error (Root-Mean-Square Error, RMSE), R square (R square), mean absolute percentage error (Mean Absolute Percentage Error, MAPE ), mean absolute error (Mean Absolute Error, MAE), and Hill inequality coefficient (Theil Inequality Coefficient, TIC) and so on. In this embodiment, the evaluation function is root mean square error, but the invention is not limited thereto.

Next, in step S250, it is determined whether the shape score generated by the evaluation function is less than or equal to a threshold value. As mentioned earlier, the flatness score represents the error between the predicted flatness of the virtual mill simulation and the actual target flatness. Therefore, if the shape score is less than or equal to the threshold value, it means that the error between the predicted shape and the target shape is very small, that is, the exit shape predicted by the virtual rolling mill is almost consistent with the actual exit shape, so the corresponding entry shape is completed. adjustment of the control parameters.

If the shape score is greater than the threshold value, it means that there is still a certain error between the predicted shape and the target shape. At this time, step S260 is performed to adjust the original control parameters. In step S260, the shape score is input into the optimization model to generate a plurality of adjusted control parameters. The optimization model can adjust the control parameters originally provided to the virtual rolling mill according to the flatness score (that is, the error between the predicted flatness and the target flatness), thereby improving the prediction result of the virtual rolling mill. In some embodiments, the optimization model uses a cross-entropy (Cross-Entropy) algorithm to generate adjusted control parameters according to the shape score, but the invention is not limited thereto.

Specifically, entropy represents the uncertainty of a random variable or the entire system, and the greater the entropy, the greater the uncertainty of the random variable or system. The cross entropy is used to measure the cost of eliminating the uncertainty of the system by using the strategy specified by the non-real distribution under the given real distribution. Crossover means the intersection of the true distribution with the non-true distribution. In the present invention, the optimization model uses cross-entropy as a loss function and uses the shape score as a reference, and its purpose is to minimize the cross-entropy.

Therefore, after the control parameters generated by the optimization model, proceed to step S270, using these adjusted control parameters as input parameters, and then return to step S230, through the virtual rolling mill according to these adjusted control parameters, the original machine The parameter predicts the shape of the entrance to generate a new predicted shape, and then proceeds to steps S240 to S270, repeating these steps until the error between the predicted shape and the shape of the exit is less than or equal to the threshold value.

The present invention builds a virtual rolling mill with an XGBoost model of integrated learning and/or an integrated neural network model of deep learning, and continuously predicts the status of equipment or systems through data integration. In addition, the virtual rolling mill is further combined with an optimization model to adjust and calculate the most appropriate control parameters only by giving necessary parameters, thereby improving the overall quality and avoiding the high cost and high risk that may be caused by direct testing in real equipment. Therefore, it can be widely used in the prediction of unknown parts in the manufacturing process.

Although the present invention has been disclosed with preferred embodiments, it is not intended to limit the present invention. Anyone skilled in this art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the present invention The scope of protection shall be determined by the scope of the attached patent application.

100: Electronic device 110: storage unit 120: processing unit 200: Adjustment method S210～S270: Steps

FIG. 1 is a schematic diagram of an electronic device for adjusting control parameters used in a rolling process according to some embodiments of the present invention. FIG. 2 is a flowchart illustrating a method for adjusting control parameters used in a rolling process according to some embodiments of the present invention.

200: Adjustment method

S210~S270: steps

Claims (7)

A method for adjusting control parameters of a rolling process, comprising: (a) providing a virtual rolling mill corresponding to a rolling machine used in the rolling process; (b) providing a rolling machine corresponding to an input shape A plurality of control parameters of the virtual rolling mill are used as a plurality of input parameters of the virtual rolling mill; (c) a predicted flat shape is generated through the virtual rolling mill according to a plurality of machine parameters, the input flat shape and these input parameters; (d) the predicted flat shape is input Shape and a target shape corresponding to the input shape to an evaluation function to generate a shape score; (h) compare the shape score with a threshold value, wherein when the shape score is greater than the threshold value, Perform steps (e) to (g); when the shape score is less than or equal to the threshold value, do not perform steps (e) to (g); (e) input the shape score into an optimization model to generate a plurality of adjusted control parameters; (f) using these adjusted control parameters as the input parameters; and (g) repeating steps (c) to (f) until the control parameters most suitable for the input shape are generated. parameters; wherein the optimization model uses a cross-entropy algorithm to generate the adjusted control parameters according to the shape score. The adjustment method according to claim 1, wherein the virtual rolling mill is a machine learning model established by ensemble learning and/or deep learning. The adjustment method as claimed in claim 2, wherein the virtual rolling mill is established using a limit gradient boosting algorithm. The adjustment method as claimed in item 2, wherein the virtual rolling mill is established using a deep neural network model. The adjustment method as claimed in claim 1, wherein the evaluation function compares the predicted flat shape with the target flat shape to generate the flat shape reflecting an error condition between the predicted flat shape and the target flat shape Fraction. The adjustment method according to claim 1, wherein the evaluation function is one of root mean square error, R square, mean absolute percentage error, mean absolute error, and Hill inequality coefficient. The adjustment method as claimed in item 1, wherein the control parameters include temperature, rolling speed, rolling force, and tension of the rolled piece.

Images