Strip plate shape intelligent regulation and control method based on adaptive fusion model

Abstract

The application discloses a strip plate shape intelligent regulation and control method based on an adaptive fusion model, which comprises the following steps: step 1, collecting historical rolling production process data and performing pretreatment; step 2, constructing a rolling production process data set based on the pretreated data; step 3, classifying plate shape quality according to the ratio of strip crown and target thickness and setting a category label; step 4, establishing an adaptive strip outlet plate shape diagnosis model containing multiple classifiers based on a DS theory, and training the diagnosis model through the rolling production process data set; step 5, inputting rolling production process data under a new rolling schedule into the classifier of the trained adaptive strip outlet plate shape diagnosis model to obtain a real-time classification prediction result of the strip, and if the prediction classification is under-crown or over-crown, step 6 is executed; otherwise, the process parameters are not adjusted; and step 6, dynamically adjusting and optimizing the process parameters based on an inverse linear quadratic control (ILQ) according to the prediction result of step 5.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of rolling control, and relates to a strip plate shape intelligent regulation and control method based on an adaptive fusion model. BACKGROUND

[0002] In the steel rolling process, the strip crown is a key indicator of product quality, and its stability and uniformity directly affect the subsequent cold rolling and downstream processing. Accurate prediction of the strip crown is crucial for improving rolling efficiency and product quality. Machine learning methods have been widely used in rolling production due to their advantages in handling complex nonlinear problems. However, due to the limitations of a single machine learning algorithm, it is difficult to effectively integrate the advantages of different models, thereby limiting the model accuracy and directly affecting the model decision results. DS evidence theory has the ability to handle conflicts and uncertainties, and can integrate prediction results between different algorithms, thereby improving the credibility and consistency of the prediction.

[0003] To improve the control accuracy of strip plate shape, domestic scholars have conducted in-depth research. Chinese patent "CN116689506A strip plate shape control method" obtains the real-time plate shape value and target plate shape value of the strip, adjusts the inclination of the rolling rod and the intermediate roll to eliminate the deviation between them, improves the common rolling defects of the eighteen-rod production line, and improves the plate shape quality of the strip rolling. Chinese patent "CN117655118B multi-model fusion strip plate shape control method and device" constructs an integrated prediction model that integrates multiple intelligent algorithms, uses error compensation to improve the accuracy of the model, and realizes dynamic regulation and control of the plate shape quality in the rolling process through crown feedback and flatness feedback to ensure the plate shape quality of the full length. Chinese patent "CN101758084B model adaptive plate shape prediction control method" decomposes the plate shape mode, establishes a plate shape control model using rolling force and incoming crown, dynamically corrects according to historical data and real-time rolling parameters, eliminates transmission time lag, adjusts the feedback controller in real time, and realizes fast dynamic control of the plate shape.

[0004] The above methods have improved the strip plate shape quality control to some extent, but more rely on parameter setting and model fusion to optimize the control strategy, lacking more detailed feedback mechanisms and working condition adaptability. In actual production, the strip plate shape is affected by multiple dynamic factors, and traditional methods often control based on fixed parameters or simple models, making it difficult to respond to working condition fluctuations in real time. In addition, existing plate shape control methods do not fully address the conflict and uncertainty problem in data fusion, resulting in limited control accuracy and difficulty in fully adapting to complex rolling environments. Especially in the actual dynamic regulation and control process, the decision interpretability and feedback mechanism are not effectively integrated, affecting the controllability and stability of the system, and ultimately leading to limitations in plate shape quality control effect. SUMMARY

[0005] To solve the above technical problems, the present application provides a strip plate shape intelligent regulation and control method based on an adaptive fusion model, which uses DS evidence theory to formulate multiple fusion strategies, obtains high-credibility diagnostic results and decision logic, and generates plate shape regulation and control feedback information, thereby realizing intelligent regulation and control of strip plate shape.

[0006] The present application provides a strip plate shape intelligent regulation and control method based on an adaptive fusion model, which includes:

[0007] Step 1: Collect historical rolling production process data and perform preprocessing;

[0008] Step 2: Construct a rolling production process data set based on the preprocessed data;

[0009] Step 3: Classify the plate shape quality based on the ratio of strip crown to target thickness, including under-crown, qualified crown and over-crown, and set the class label;

[0010] Step 4: Establish an adaptive strip outlet plate shape diagnostic model containing multiple classifiers based on DS theory, and train the diagnostic model through the rolling production process data set;

[0011] Step 5: Input the rolling process parameter data under the new rolling schedule into the classifier of the trained adaptive strip outlet plate shape diagnostic model to obtain the real-time classification prediction result of the strip, and if the prediction classification is under-crown or over-crown, execute Step 6; otherwise, do not adjust the process parameters;

[0012] Step 6: Based on the prediction result of Step 5, dynamically adjust and optimize the process parameters based on inverse linear quadratic control ILQ.

[0013] The present application provides a strip plate shape intelligent regulation and control method based on an adaptive fusion model, which has the following beneficial effects:

[0014] The regulation method can effectively improve the intelligent regulation level of the strip rolling process through data processing, prediction diagnosis and dynamic process regulation. Compared with the traditional method, the present application overcomes the problem of high precision requirement for parameter adjustment in traditional plate shape control by introducing multiple intelligent classifiers, DS evidence theory and ILQ control. Specifically, first, collect and process multi-dimensional historical rolling process parameters, set plate shape classification criteria and set category labels, thereby providing high-precision plate shape prediction information. Then, the multi-source data (from multiple intelligent classifiers) is fused by DS evidence theory, effectively dealing with the uncertainty and conflict problems in the prediction process. Further, the SHAP model is used to clarify the decision logic, ensuring the explainability and transparency of the regulation process, and providing plate shape feedback information. Finally, the ILQ controller adjusts and optimizes the rolling force, bending force and rolling speed key parameters in real time through feedback information, realizing the dynamic and accurate regulation of the strip. Compared with the traditional control method, the present application can adjust the control strategy in real time during the dynamic rolling process, significantly improve the stability and precision of the plate shape quality, optimize the production efficiency and reduce the energy consumption, thereby significantly improving the intelligent and automatic level of the strip production process. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a flowchart of a strip plate shape intelligent regulation method based on an adaptive fusion model of the present application;

[0016] Figure 2 is a SHAP value waterfall chart in an embodiment of the present application;

[0017] Figure 3 is a flowchart of dynamic adjustment and optimization of process parameters based on inverse linear quadratic control (ILQ) in an embodiment of the present application. DETAILED DESCRIPTION

[0018] The specific embodiments of the present application will be further described in detail below in conjunction with the drawings and examples. The following examples are used to illustrate the present application, but not to limit the scope of the present application.

[0019] As shown in Figure 1 , a strip plate shape intelligent regulation method based on an adaptive fusion model of the present application comprises:

[0020] Step 1: Collect historical rolling production process data and perform preprocessing;

[0021] The rolling production process data includes rolling force, bending force, roll shifting amount, rolling speed, strip inlet temperature, strip inlet thickness, strip outlet temperature, strip outlet thickness, strip outlet width and crown.

[0022] The preprocessing includes data cleaning, missing value processing, outlier detection and processing, to ensure the integrity and accuracy of the data.

[0023] Step 2: Based on the pre-processed data, a rolling production process dataset is constructed, and is divided into a training set and a test set in a ratio of 7:3.

[0024] In this embodiment, a large amount of representative data is collected from a certain hot rolling 2160mm production line, as shown in Table 1, and the dataset includes 2796 groups of samples.

[0025] Table 1. Rolling production process data

[0026]

[0027] Step 3: The strip crown to target thickness ratio is used for shape quality classification, including under-crown, qualified crown and over-crown, and the category label is set, specifically:

[0028] When the strip crown to target thickness ratio is less than 0.8, it is defined as under-crown, and the category label is 0. When the strip crown to target thickness ratio is 0.8-1.8, it is defined as qualified crown, and the category label is 1. When the strip crown to target thickness ratio is greater than 1.8, it is defined as over-crown, and the category label is 2. In this embodiment, the target thickness is 3mm.

[0029] Step 4: Based on the DS theory, an adaptive strip exit shape diagnosis model containing multiple classifiers is established, and the diagnosis model is trained through the rolling production process dataset, specifically:

[0030] Step 4.1: Select multiple intelligent classifiers, and divide the rolling production process dataset into a training set and a test set in a ratio of 7:3, train each classifier, and calculate the posterior probability and reliability factor of each classifier based on the shape quality classification result of each group of samples, specifically:

[0031] Step 4.11: Select multiple intelligent classifiers, including random forest RF, light gradient boosting machine LightGBM, extreme gradient boosting XGBoost, K-neighbor KNN, decision tree DT and gradient boosting decision tree GBDT.

[0032] In specific implementation, 2796 groups of samples are divided into a training set and a test set in a ratio of 7:3, and 29-dimensional data is finally determined as the input of the classifier, as shown in Table 1, and the strip crown is the output of the classifier.

[0033] Step 4.12: Taking the rolling force, bending force, roll shifting amount, rolling speed, strip inlet temperature, strip inlet thickness, strip outlet temperature, strip outlet thickness, strip outlet width as the input of the classifier, and the crown as the output of the classifier; each time training uses 80% of the training set data randomly extracted, the test set remains unchanged, and each classifier is trained and predicted independently for 10 times.

[0034] Step 4.13: For each group of samples, the number of times each category appears in 10 predictions is counted to calculate the posterior probability P(A i |X), and the calculation formula is:

[0035]

[0036] Where i = 0, 1, 2, A i represents the category i; represents the number of times the category i appears in 10 predictions. For example, the category 0 appears 3 times, then P(A0|X) = 0.3.

[0037] Step 4.14: The reliability factor evaluates the performance of the classifier through the statistical information of the confusion matrix , which is an HxH matrix, where H represents the number of categories, and the form of the confusion matrix is as follows:

[0038]

[0039] Where C ij represents the number of samples that are actually of category j but are predicted to be of category i, and the correct recognition ability of the classifier in predicting a certain category is calculated by the following formula:

[0040]

[0041] Where C ii represents the number of samples correctly recognized by the classifier as category i; represents the number of samples that are actually of category i; R i is recorded as Rf(A i ), which is the reliability factor of the classifier.

[0042] Step 4.2: Calculate the quality function of each classifier according to the posterior probability and the reliability factor, fuse the quality functions of multiple classifiers using DS evidence theory, and select the category with the largest combined quality function as the final plate shape diagnosis result; if the plate shape diagnosis result deviates from the actual category, add a fuzzy membership function to adjust the weight of each classifier in the combination process, update the quality function of each classifier, and re-fuse the quality functions until the diagnosis result and the actual category are consistent, and select the classifier in the combination as the final classifier of the self-adaptive strip outlet plate shape diagnosis model, which is specifically:

[0043] Step 4.21: Define the identification framework Θ = {under convexity, qualified convexity, over convexity}, corresponding to the class labels 0, 1, 2 respectively.

[0044] Step 4.22: Calculate the quality function of each classifier according to the posterior probability and the reliability factor:

[0045] m r (A i )=P(A i |X)×Rf(A i )

[0046] Where m r (A i ) represents the quality function of the sample belonging to the category i predicted by the rth classifier.

[0047] Step 4.23: Synthesize the quality functions of multiple classifiers to obtain the final confidence. For the quality functions of two classifiers, the Dempster synthesis rule is:

[0048]

[0049] Where A represents the synthesized category, B ∈ Θ, C ∈ Θ.

[0050] Step 4.24: Combine multiple classifiers and calculate the synthesized quality function of each combination. Select the category with the maximum synthesized quality function as the final plate shape diagnosis result, and select the classifier in the combination as the final classifier of the adaptive strip exit plate shape diagnosis model.

[0051] Step 4.25: If the plate shape diagnosis result deviates from the actual category, add a fuzzy membership function to adjust the weight of each classifier in the synthesis process, and update the quality function of each classifier in real time as m r (A i )':

[0052] m r (A i )'=μ(x)·m r (A i )

[0053] Where μ(x) represents the membership degree of x:

[0054]

[0055] Where x = P(A i|X), a, b and c represent fuzzy parameters, a, b, c ∈ [0, 1] (a < b < c); the values of a, b and c are adjusted according to the deviation between the prediction result and the actual category until the diagnosis result is consistent with the actual category.

[0056] Repeat step 4.23, use the updated quality function to perform DS synthesis, and obtain a new diagnosis result.

[0057] Step 5: input the rolling production process data under the new rolling schedule into the trained adaptive strip exit shape diagnosis model classifier to obtain the real-time classification prediction result of the strip, if the prediction classification is under-convexity or over-convexity, execute step 6; otherwise, do not adjust the process parameters.

[0058] Step 6: based on the prediction result of step 5, dynamically adjust and optimize the process parameters based on inverse linear quadratic control ILQ, as shown in Figure 3 , specifically:

[0059] Step 6.1: input the rolling production process data under the new rolling schedule and the predicted convexity output by the classifier of the diagnosis model into the SHAP model to obtain the SHAP value of each rolling production process data. As shown in Figure 2 , the horizontal axis represents the target value, and the vertical axis represents the name and value of different input features input into the SHAP model. The first 9 data output by the SHAP model are shown in the figure, including: F6 bending force, F6 rolling force, F5 rolling speed, strip width, F6 roll shifting amount, F5 bending force, F4 rolling speed, F4 rolling force and F5 roll shifting amount. In this embodiment, a SHAP waterfall chart at a certain time is obtained, for example, the F6 bending force SHAP value is +2.59, indicating that the F6 bending force is the main influencing factor of the current shape, and the increase of the value will cause the shape deviation to increase.

[0060] Step 6.2: establish the ILQ system state equation to clearly define the relationship between the state variables and the control input, and provide a mathematical model for ILQ controller design:

[0061]

[0062] Wherein, X is a state variable, that is, rolling production process data under a new rolling schedule; U is a control input, that is, an adjustment amount of a key rolling parameter, and the key rolling parameter is rolling production process data except the roll shifting amount and the strip width in the first 9 data output by the SHAP model. In the embodiment, the control input is an adjustment amount of the F6 bending force, the F6 rolling force, the F5 rolling speed, the F5 bending force, the F4 rolling speed and the F4 rolling force. E is a control matrix, reflecting the influence of the control input on the system state, in order to facilitate engineering implementation, it is considered that each control input only acts on the corresponding state variable, and is set as a unit matrix. D is a system state matrix, describing the dynamic relationship between the state variables, and the autocorrelation coefficient and the cross-correlation coefficient of the state variables are obtained by regression analysis on the historical data, as the elements on the diagonal and the remaining adjacent elements, respectively.

[0063] Step 6.3: The target of the ILQ controller is to minimize the deviation of the target variable, that is, to minimize the state deviation X T QXand control cost U T RU, to ensure that the system reaches the optimal performance under the dynamic changing working condition, and the objective function J is defined as:

[0064]

[0065] Wherein, Q is a state weight matrix, which is a diagonal matrix, and the elements on the diagonal are SHAP values of the key rolling parameters, for adjusting the importance of each state variable; R is a control input weight matrix, for adjusting the control strength of each state variable, and in order to prevent drastic adjustment, it is set as a unit matrix.

[0066] The minimization problem of the objective function is converted into the solution of the Riccati equation:

[0067] P=D T PD-D T PE(R+E T PE) -1 E T PD+Q

[0068] The ILQ feedback gain matrix K is calculated through the solution P of the Riccati equation:

[0069] K=(R+E T PE) -1 E T PD

[0070] The rolling process parameter adjustment amount U is calculated to realize real-time adjustment of the system state:

[0071] U=-KX

[0072] In this embodiment, the current rolling process parameters are known, F6 bending force is 550 kN, F6 rolling force is 9549.411 kN, F5 rolling speed is 7.835 m / s, strip width is 1276.348 mm, F5 bending force is 480 kN, F4 rolling speed is 5.621 m / s, F4 rolling force is 12069.873 kN, then:

[0073]

[0074] According to the feedback information generated in step 4.1, Q and R are constructed as follows:

[0075]

[0076] The adjusted rolling parameters are as follows: F6 bending force is 495 kN, F6 rolling force is 8594.47 kN, F5 rolling speed is 7.0515 m / s, strip width is 1276.348 mm, F5 bending force is 432 kN, F4 rolling speed is 5.0589 m / s, and F4 rolling force is 10862.89 kN.

[0077] The ILQ controller is connected with each actuator through a PLC and is based on the action of the U adjustment mechanism. The actuator includes a hydraulic cylinder and a main motor. The hydraulic cylinder is used to adjust the size of the rolling force and the bending force, and the main motor is used to adjust the size of the rolling speed.

[0078] The above description is only the preferred embodiment of the present application, and is not intended to limit the idea of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for intelligent control of strip shape based on an adaptive fusion model, characterized in that, include: Step 1: Collect historical rolling production process data and preprocess it; Step 2: Construct a rolling production process dataset based on the preprocessed data, and divide it into a training set and a test set in a 7:3 ratio; Step 3: Classify the strip shape quality based on the ratio of strip crown to target thickness, including under-crown, qualified crown, and over-crown, and set category labels; Step 4: Based on DS theory, establish an adaptive strip exit shape diagnostic model containing multiple classifiers, and train the diagnostic model using a rolling production process dataset. Specifically: Step 4.1: Select multiple intelligent classifiers and divide the rolling production process dataset into training and test sets in a 7:3 ratio. Train each classifier and calculate the posterior probability and reliability factor of each sample based on each classifier according to the plate shape quality classification results. Step 4.2: Calculate the quality function of each classifier based on the posterior probability and reliability factor, use DS evidence theory to fuse the quality functions of multiple classifiers, and select the class with the largest quality function after synthesis as the final plate shape diagnosis result; If the plate shape diagnosis result deviates from the actual category, a fuzzy membership function is added to adjust the weights of each classifier in the synthesis process, the quality function of each classifier is updated, and the quality function is fused again until the diagnosis result is consistent with the actual classification. Then, the classifier in the combination is selected as the final classifier of the adaptive strip steel export plate shape diagnosis model. Step 5: Input the rolling production process data under the new rolling schedule into the classifier of the trained adaptive strip exit shape diagnosis model to obtain the real-time classification prediction result of the strip. If the prediction classification is under-convexity or over-convexity, proceed to step 6; otherwise, do not adjust the process parameters. Step 6: Based on the real-time classification and prediction results from Step 5, dynamically adjust and optimize the process parameters using inverse linear quadratic control ILQ, specifically as follows: Step 6.1: Input the rolling production process data under the new rolling schedule and the predicted crown output by the classifier of the diagnostic model into the SHAP model to obtain the SHAP value of each rolling production process data. Step 6.2: Establish the state equations of the ILQ system, clarify the relationship between state variables and control inputs, and provide a mathematical model for the design of the ILQ controller: Where X is the state variable, i.e. the rolling production process data under the new rolling schedule; U is the control input, i.e. the adjustment amount of the key rolling parameters, which are the rolling production process data other than roll shifting amount and strip width in the first 9 data outputs of the SHAP model; D is the system state matrix, describing the dynamic relationship between state variables; E is the control matrix, describing the influence of control variables on the system. Step 6.3: The goal of the ILQ controller is to minimize the deviation of the target variable, i.e., to minimize the state deviation X. T QX and control costs U T RU ensures that the system achieves optimal performance under dynamically changing operating conditions. The objective function J is defined as: Where Q is the state weight matrix, which is a diagonal matrix, and the elements on the diagonal are the SHAP values ​​of key rolling parameters, used to adjust the importance of each state variable; R is the control input weight matrix, used to balance control energy consumption and system performance. The problem of minimizing the objective function is transformed into solving the Riccati equation: P=D T PD-D T PE(R+E T ON) -1 It is T PD+Q Calculate the ILQ feedback gain matrix K using the solution P of the Riccati equation: K=(R+E T EITHER) -1 IN T PD Calculate the adjustment amount U of the rolling process parameters to achieve real-time adjustment of the system state: U = -KX The ILQ controller is connected to each actuator via a PLC and adjusts the action of the actuator based on the U-shaped adjustment mechanism. The actuator includes a hydraulic cylinder and a main motor. The hydraulic cylinder is used to adjust the rolling force and bending force, and the main motor is used to adjust the rolling speed.

2. The intelligent control method for strip shape based on an adaptive fusion model as described in claim 1, characterized in that, The rolling production process data includes: rolling force, bending force, roll shifting amount, rolling speed, strip inlet temperature, strip inlet thickness, strip outlet temperature, strip outlet thickness, strip outlet width, and crown.

3. The intelligent control method for strip shape based on an adaptive fusion model as described in claim 1, characterized in that, The preprocessing in step 1 includes: data cleaning, missing value handling, and outlier detection and handling, to ensure the integrity and accuracy of the data.

4. The intelligent control method for strip shape based on an adaptive fusion model as described in claim 1, characterized in that, Step 3 specifically involves: When the ratio of strip crown to target thickness is less than 0.8, it is defined as under-crown, and the category label is 0; When the ratio of strip crown to target thickness is 0.8 to 1.8, it is defined as qualified crown, and the category label is 1; When the ratio of strip crown to target thickness is greater than 1.8, it is defined as excessive crown, and the category label is 2.

5. The intelligent control method for strip shape based on an adaptive fusion model as described in claim 1, characterized in that, Step 4.1 specifically involves: Step 4.11: Select multiple smart classifiers, including Random Forest (RF), Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting (XGBoost), K-Nearest Neighbors (KNN), Decision Tree (DT), and Gradient Boosting Decision Tree (GBDT); Step 4.12: Use rolling force, bending roll force, roll shifting amount, rolling speed, strip inlet temperature, strip inlet thickness, strip outlet temperature, strip outlet thickness, and strip outlet width as inputs to the classifier, and convexity as the output of the classifier; use 80% of the training set data randomly selected for each training session, while keeping the test set unchanged, and perform 10 independent training and prediction sessions for each classifier; Step 4.13: For each sample group, calculate the posterior probability P(A) by counting the number of times each category appears in 10 predictions. i |X), the calculation formula is: Among them, i=0,1,2,A i Indicates category i; This represents the number of times category i appears in 10 predictions; Step 4.14: Reliability factor through confusion matrix The performance of the classifier is evaluated using statistical information. The confusion matrix is ​​an H×H matrix, where H represents the number of classes. The confusion matrix has the following form: Among them, C ij This represents the number of samples that are actually of class j but are predicted as class i. The classifier's ability to correctly identify a class is calculated using the following formula: Among them, C ii This represents the number of samples that the classifier correctly identifies as class i; This represents the actual number of samples of category i; R i denoted as Rf(A) i ), which serves as a reliability factor for the classifier.

6. The intelligent control method for strip shape based on an adaptive fusion model as described in claim 5, characterized in that, Step 4.2 specifically involves: Step 4.21: Define the recognition framework Θ = {under-convexity, adequate convexity, over-convexity}, corresponding to the category labels 0, 1, and 2 respectively; Step 4.22: Calculate the quality function of each classifier based on the posterior probability and reliability factor: m r (A i )=P(A i |X)×Rf(A i ) Where, m r (A i ) represents the quality function that the sample belongs to class i as predicted by the r-th classifier; Step 4.23: Combine the quality functions of multiple classifiers to obtain the final confidence level. For the quality functions of two classifiers, the Dempster combination rule is as follows: Where A represents the synthesized category, B∈Θ, C∈Θ; Step 4.24: Combine multiple classifiers and calculate the composite quality function of each combination. Select the class with the largest composite quality function as the final strip shape diagnosis result, and select the classifier in the combination as the final classifier of the adaptive strip steel export strip shape diagnosis model. Step 4.25: If the plate shape diagnosis result deviates from the actual category, a fuzzy membership function is added to adjust the weights of each classifier in the synthesis process, and the quality function of each classifier is updated in real time to m. r (A i )′: m r (A i )'=μ(x)·m r (A i ) Where μ(x) represents the membership degree of x: Where, x = P(A) i |X), where a, b, and c represent fuzzy parameters, a, b, c ∈ [0, 1] (a < b < c); the values ​​of a, b, and c are adjusted according to the deviation between the prediction result and the actual category until the diagnosis result is consistent with the actual category; Repeat step 4.23, using the updated quality function to perform DS synthesis, and obtain new diagnostic results.

Images