Data-driven comprehensive evaluation method and device for performance of hot-rolled strip steel

Abstract

The invention provides a data-driven hot-rolled strip steel performance comprehensive evaluation method and device, and relates to the technical field of performance evaluation of industrial processes. The method comprises the steps that multi-dimensional data of the strip steel hot rolling production process are obtained and preprocessed; constructing a random forest prediction model to complete performance-related data missing values to obtain multi-dimensional complete data; a hot-rolled strip steel performance comprehensive rating scheme is designed according to the strip steel quality standard; a strip steel performance comprehensive evaluation model based on multi-task learning and a performance-related stacked auto-encoder is constructed, and multi-dimensional complete data is used for training; and according to the strip steel performance comprehensive evaluation model, a multi-level performance rating result of the hot-rolled strip steel is obtained. According to the method, scientific, accurate and multi-level real-time evaluation on the performance of the hot-rolled strip steel can be realized by mining the incidence relation between the process data and the performance related data.

Description

Technical Field

[0001] The present invention relates to the technical field of performance evaluation of industrial processes, and particularly to a data-driven comprehensive performance evaluation method and device for hot-rolled strip steel. Background Art

[0002] As the most widely used product with the highest technical content and industrial added value among steel products, hot-rolled strip steel is an important raw material in fields such as automobiles, military, aerospace, etc. The quality of hot-rolled strip steel will directly affect the performance of the final product. With the development of industrial intelligence and integration, modern hot-rolling production lines show new characteristics of diverse product varieties, mass production of multiple specifications, and customization according to demand, making the evaluation of product performance more complex. By scientifically and accurately evaluating the product performance, the overall distribution of product performance can be clearly and comprehensively grasped, providing a strong basis for enterprises to continuously improve product performance and make targeted process control.

[0003] The main evaluation indicators of strip steel performance include surface quality, dimensional accuracy, shape, mechanical properties, processability, etc. These indicators often rely on offline inspection and testing. The traditional strip steel performance evaluation method is to conduct sampling and uncoiling inspections, and issue a comprehensive quality inspection report based on probability statistics, manual experience, and enterprise standards. The evaluation results lack consistency, objectivity, and accuracy. At the same time, the standard only gives the requirements for different levels of a few indicators, the indicators are single and imperfect, the evaluation method is too rough, and the evaluation results often only distinguish between "qualified" and "unqualified" products, and cannot scientifically and accurately reflect the distribution of strip steel performance grades.

[0004] In the face of the urgent need for comprehensive evaluation, many scholars have introduced the fuzzy theory and established a fuzzy comprehensive evaluation model for strip steel performance using quality inspection data. The result of fuzzy comprehensive evaluation is not an absolute affirmation or negation, but is represented by a fuzzy set. Although the fuzzy method provides a performance evaluation result with high interpretability, the final performance of strip steel products is closely related to the specific implementation quality of each process in the production process. These traditional fuzzy evaluation methods have not broken away from the limitation of relying on offline inspection and testing for performance indicators, have not fundamentally solved the subjective factors brought by manual participation, and cannot reflect the influence of numerous process parameters on performance.

[0005] With the rapid development of sensing technology, big data, and industrial Internet technology, the data acquisition and information integration capabilities in the production process have been greatly improved. These technologies can not only monitor and record the detailed information of each process in real time but also provide new directions and possibilities for in-depth analysis of the impact of each production link on the performance of the final product. On this basis, data-driven modeling methods such as computer technology, machine learning, and deep learning have also begun to be applied in the field of strip steel performance evaluation. However, most of the existing methods only focus on a single evaluation index of strip steel performance, mine the potential relationship between the production process and this performance index from a large amount of historical data, and use this as the basis for defect detection and identification, still unable to give a comprehensive performance evaluation result.

[0006] Due to difficulties such as the difficulty of meeting the real-time requirement for obtaining performance index data, high-dimensional data coupling, poor interpretability of evaluation results, and lack of a unified evaluation standard, exploring the potential correlation between product performance and the production process and developing a comprehensive performance evaluation method suitable for the characteristics of modern hot rolling production lines have become urgent problems to be solved. Summary of the Invention

[0007] In order to solve the problems existing in the prior art, such as the inability to meet the real-time acquisition requirements of performance indicators, the time cost brought by waiting for sampling quality inspection data, the lack of consistency and accuracy of evaluation results due to the lack of a unified evaluation standard, and the inability to meet the requirement of mining the relevance of high-dimensional coupled data, the present invention provides a data-driven comprehensive performance evaluation method and device for hot-rolled strip steel. The technical solutions are as follows:

[0008] On the one hand, a data-driven comprehensive performance evaluation method for hot-rolled strip steel is provided, and the method includes the following steps:

[0009] S1. Obtain multi-dimensional data of the strip steel hot rolling production process and preprocess the multi-dimensional data; wherein, the multi-dimensional data includes process data and performance-related data;

[0010] S2. For the data missing in the performance-related data due to dependence on off-line sampling inspection, construct a random forest prediction model to complete the missing values and obtain multi-dimensional complete data;

[0011] S3. Integrate the strip steel quality standards, design a comprehensive performance rating scheme for hot-rolled strip steel, and divide the performance grades for the multi-dimensional complete data according to the comprehensive performance rating scheme for hot-rolled strip steel;

[0012] S4. Construct a comprehensive performance evaluation model for strip steel based on multi-task learning and performance-related stacked autoencoders, where each sub-aspect of strip steel performance is used as each sub-task in the multi-task learning framework, and performance-related deep features are extracted based on the performance-related stacked autoencoders in each sub-task;

[0013] S5. Train the comprehensive evaluation model of strip steel performance constructed with the multi-dimensional complete data, and use the Bayesian optimization strategy to optimize the hyperparameters of the model to obtain the optimal hyperparameter group that makes the model performance the best;

[0014] S6. According to the trained comprehensive evaluation model of strip steel performance, obtain the multi-level performance rating results of hot-rolled strip steel.

[0015] Optionally, step S1 specifically includes:

[0016] S11. Perform preliminary processing operations of normalization and outlier deletion on the collected multi-dimensional data;

[0017] S12. Divide the performance-related data in the multi-dimensional data into sub-blocks according to process knowledge. After division, the performance indicators in each sub-block jointly determine one aspect of the strip steel performance.

[0018] Optionally, step S2 specifically includes:

[0019] S21. Conduct a correlation analysis on the missing data in the process data and performance-related data to obtain feature variables that meet the correlation requirements as the input data for model training;

[0020] S22. Build a random forest prediction model, and input the input data that has been normalized and had outliers deleted into the initial random forest prediction model for training;

[0021] S23. Set the objective function of the random forest prediction model to minimize the prediction error, and use the Bayesian optimization strategy to optimize the model parameters to obtain the constructed random forest prediction model;

[0022] S24. Based on the constructed random forest prediction model, predict and fill in the missing performance-related data to obtain multi-dimensional complete data.

[0023] Optionally, step S3 specifically includes:

[0024] S31. Divide the strip steel performance indicators according to the strip steel quality standard;

[0025] S32. Use the entropy weight method to calculate the weights between the performance indicators of the same sub-aspect of the strip steel, that is, the secondary weights, and the weights of each performance sub-aspect for the comprehensive performance, that is, the primary weights;

[0026] S33. Use the method of multi-level performance indicator weighting to obtain the comprehensive performance grade of the strip steel;

[0027] S34. Divide the performance grades for the obtained multi-dimensional complete data according to the hot-rolled strip steel performance rating scheme.

[0028] Optionally, step S4 specifically includes:

[0029] S41. Based on the multi-task learning method, construct a comprehensive evaluation framework for strip steel performance, and regard each sub-aspect of strip steel performance as each sub-task in the multi-task learning framework;

[0030] S42. In each sub-task, use the performance-related stacked autoencoder to extract features from the multi-dimensional complete data to obtain deep features related to the performance of each sub-task;

[0031] S43. Integrate the deep features related to the performance of each sub-task, and use the performance-related stacked autoencoder to extract features again to obtain deep features related to the comprehensive performance.

[0032] Optionally, step S41 specifically includes:

[0033] Considering the multi-index characteristics of strip steel performance evaluation, introduce the multi-task learning strategy, select hard parameter sharing, enable multiple sub-tasks to learn in parallel in a unified model, and mine the coupling relationship between different performance indicators by sharing underlying features;

[0034] The loss function of multi-task learning is the weighted sum of the loss functions of multiple learning objectives, and its calculation formula is as follows:

[0035]

[0036] In the formula, λ k is the loss function weight coefficient of the kth sub-task, Loss k (x k ,y k ) is the loss function of the kth sub-task, x k ,y k are the input data and output data of the kth sub-task.

[0037] Optionally, step S42 specifically includes:

[0038] S421. Input the shared feature h 0 obtained from the multi-task learning shared layer into the first performance-related autoencoder network of each sub-task, and use the performance level of each sub-aspect as supervision during the training process to obtain the performance-related feature representation k represents the kth sub-task;

[0039] S422. In each sub-task, input the feature representation of the first performance-related autoencoder network into the second performance-related stacked autoencoder network, and use the performance level of each sub-aspect as supervision during the training process to obtain the performance-related feature representation

[0040] S423, and so on. Stack multiple performance-related stacked autoencoder networks hierarchically to obtain a performance-related deep feature representation. q represents the number of autoencoder networks.

[0041] Among them, the performance-related stacked autoencoder is constructed as follows:

[0042] Introduce performance level information into the autoencoder structure for supervised learning to construct a performance-related autoencoder, and stack multiple performance-related autoencoders hierarchically to construct the performance-related stacked autoencoder.

[0043] Optionally, step S5 specifically includes:

[0044] S51. Use the preprocessed multi-dimensional complete data to train the strip steel performance comprehensive evaluation model based on multi-task learning and performance-related stacked autoencoders.

[0045] S52. The loss function for training the strip steel performance comprehensive evaluation model is shown in the following formula:

[0046]

[0047] Among them, λ k is the loss function weight coefficient of the k-th sub-task, represents the loss function of the performance-related stacked autoencoder in the k-th sub-task;

[0048] S53. Use the Bayesian optimization strategy to tune the hyperparameters in the strip steel performance comprehensive evaluation model. The hyperparameters include the number of bottom-layer neurons shared by multiple tasks, the number of autoencoder networks, the number of neurons in the hidden layer of each autoencoder, and the learning rate, to obtain the optimal hyperparameter group that makes the model performance best.

[0049] S54. Save the trained strip steel performance comprehensive evaluation model and its hyperparameter set.

[0050] Optionally, step S6 specifically includes:

[0051] S61. According to the multi-dimensional complete data and the strip steel performance comprehensive evaluation model, use the deep features extracted in each sub-task to obtain the sub-item rating results representing each performance sub-aspect of the hot-rolled strip steel.

[0052] S62. According to the extracted deep features related to the comprehensive performance, obtain the comprehensive performance rating result of the hot-rolled strip steel.

[0053] On the other hand, a data-driven hot-rolled strip steel performance comprehensive evaluation device is provided for implementing the method described in any one of the above. The device includes:

[0054] A data processing module, configured to obtain multi-dimensional data in the hot strip rolling production process and preprocess the multi-dimensional data; wherein, the multi-dimensional data includes process data and performance-related data;

[0055] A data completion module, configured to build a random forest prediction model to complete the missing data in the performance-related data due to relying on off-line sampling inspection, and obtain multi-dimensional complete data;

[0056] A performance evaluation module, configured to design a comprehensive performance rating scheme for hot-rolled strip according to the comprehensive strip quality standards, and perform performance level division on the multi-dimensional complete data according to the comprehensive performance rating scheme for hot-rolled strip;

[0057] Build a comprehensive strip performance evaluation model based on multi-task learning and performance-related stacked autoencoders, where each sub-aspect of strip performance is used as each sub-task in the multi-task learning framework, and performance-related deep features are extracted based on the performance-related stacked autoencoder in each sub-task;

[0058] Train the constructed comprehensive strip performance evaluation model in combination with the multi-dimensional complete data, and use the Bayesian optimization strategy to optimize the model hyperparameters to obtain the optimal hyperparameter group that makes the model performance best;

[0059] A result output module, configured to obtain a multi-level performance rating result of hot-rolled strip according to the trained comprehensive strip performance evaluation model.

[0060] On the other hand, an electronic device is provided, and the electronic device includes:

[0061] A processor;

[0062] A memory, on which computer-readable instructions are stored, and when the computer-readable instructions are loaded and executed by the processor, the steps of the above data-driven comprehensive hot-rolled strip performance evaluation method are implemented.

[0063] On the other hand, a computer-readable storage medium is provided, and program codes are stored in the computer-readable storage medium, and the program codes can be called by a processor to execute the steps of the above data-driven comprehensive hot-rolled strip performance evaluation method.

[0064] The beneficial effects brought by the technical solution provided by the present invention at least include:

[0065] (1) The sub-item performance rating scheme and the comprehensive performance rating scheme constructed by the present invention integrate multiple strip quality standards, and realize the objectivity and consistency of the comprehensive strip performance rating.

[0066] (2) The performance index prediction model proposed by the present invention can capture the correlation features between process parameters and performance indexes, replace the sampling inspection values with the model prediction values, solve the problems of incomplete performance data and inability to meet the real-time acquisition requirements in the prior art, provide conditions for on-line real-time evaluation of strip steel performance, and reduce the economic cost and time cost brought by destructive inspection.

[0067] (3) The comprehensive strip steel performance evaluation model based on multi-task learning and performance-related stacked autoencoder (MTL-SPAE) constructed by the present invention can consider the correlation relationship between performance indexes, capture the deep features between multi-dimensional production data and different performance indexes of products, obtain multi-level rating results of strip steel performance, and at the same time, the multi-task parallel training mode also provides sub-item rating results of performance sub-aspects, provides information for comprehensive performance evaluation, and enhances the interpretability of the model. Thus, the overall distribution of product performance can be clearly and comprehensively grasped, providing a strong basis for enterprises to continuously improve product performance and make targeted process control. Description of the Drawings

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for description in the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0069] Figure 1 It is a flowchart of a data-driven comprehensive evaluation method for hot-rolled strip steel performance provided by an embodiment of the present invention;

[0070] Figure 2 It is a strip steel hot-rolling production flowchart provided by an embodiment of the present invention;

[0071] Figure 3 It is a schematic diagram of the SPAE model provided by an embodiment of the present invention;

[0072] Figure 4 It is a flowchart of comprehensive evaluation of hot-rolled strip steel performance based on MTL-SPAE provided by an embodiment of the present invention;

[0073] Figure 5 It is a schematic structural diagram of a data-driven comprehensive evaluation device for hot-rolled strip steel performance provided by an embodiment of the present invention;

[0074] Figure 6 It is a schematic structural diagram of an electronic device provided by an embodiment of the present invention. Detailed Embodiments

[0075] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the described embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0076] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to represent examples, illustrations, or explanations. Any embodiment or design solution described as an "example" in the present invention should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Rather, the use of the word "example" is intended to present concepts in a specific manner.

[0077] The embodiments of the present invention provide a comprehensive evaluation method for the performance of hot-rolled strip steel driven by data. This method can be implemented by an electronic device, which can be a terminal or a server. As Figure 1 shown, the processing flow of this method may include the following steps:

[0078] S1. Obtain multi-dimensional data of the strip hot-rolling production process and preprocess the multi-dimensional data; wherein, the multi-dimensional data includes process data and performance-related data.

[0079] In a feasible implementation manner, the strip hot continuous rolling process is a large-scale, long-process, multi-process, and complex-process steel production process. As Figure 2 shown, this production line consists of multiple processes such as a reheating furnace, rough rolling, finish rolling, laminar cooling, and coiling. Each process is accompanied by complex physical and chemical changes, with complex process mechanisms, numerous process variables, and diverse product steel grades. Each process cooperates with each other to jointly determine the comprehensive performance of the final product.

[0080] In the embodiments of the present invention, various sensing devices installed on the hot-rolling production line are used to collect production data from different processes and equipment as historical multi-dimensional data. The step S1 specifically includes:

[0081] S11. Perform preliminary processing operations of normalizing and removing outliers on the collected multi-dimensional data.

[0082] Due to the influence of industrial site environment interference, random noise, equipment failures, etc., the collected raw data often has non-ideal situations such as missing data values and anomalies. In order to obtain reliable analysis results, it is necessary to preprocess the raw data. This step performs normalization and outlier removal operations on the raw data.

[0083] Among them, the data normalization method can be based on the corresponding statistical characteristics of the data, and the calculation process is as follows:

[0084]

[0085] Among them, X*, μ, and σ are the normalized value, mean value, and variance value of the original data variable X, respectively.

[0086] Optionally, outlier elimination processing is performed on the variables approximately normally distributed in the original data through the [μ - 3σ, μ + 3σ] confidence interval rule. Before processing, first use the Q-Q plot to judge whether each variable conforms to the approximate normal distribution. In the Q-Q plot, when the data is approximately distributed along the diagonal straight line, it is considered that the variable satisfies the normal distribution, otherwise it does not.

[0087] S12. Sub-block divide the performance-related data in the multi-dimensional data according to process knowledge. After division, the performance indicators in each sub-block jointly determine one aspect of the strip performance.

[0088] In a feasible implementation, the performance of the hot-rolled strip is divided into multiple sub-aspects such as dimensional accuracy, shape, mechanical properties, surface quality, and process performance. Each performance sub-aspect contains multiple specific performance indicators. For example, dimensional accuracy includes thickness tolerance, width tolerance, diagonal tolerance, etc., mechanical properties include yield strength, tensile strength, elongation after fracture, etc., and process performance includes Erichsen value, plastic strain ratio, bend diameter, etc. Among them, some performance indicators can obtain performance-related production data through devices such as sensors and measuring instruments in the industrial field, such as thickness, width, etc.; some performance indicators need to wait for uncoiling sampling and destructive inspection and testing to obtain, such as yield strength, tensile strength, elongation after fracture, etc.

[0089] S2. For the data missing in the performance-related data due to dependence on offline sampling inspection, construct a random forest prediction model to complete the missing values and obtain multi-dimensional complete data.

[0090] In a feasible implementation, the step S2 specifically includes:

[0091] S21. Perform correlation analysis on the missing data in the process data and performance-related data to obtain characteristic variables that meet the correlation requirements as the input data for model training.

[0092] Optionally, the process variable correlation analysis method is measured by the mutual information value, and its calculation formula is:

[0093]

[0094] Wherein, X represents the first process variable, x represents the specific value of the first process variable X, Y represents the second process variable, y represents the specific value of the second process variable Y, I(X, Y) represents the mutual information value between the first process variable X and the second process variable Y, p(x, y) represents the joint distribution, p(x) represents the marginal distribution of the first process variable X, and p(y) represents the marginal distribution of the second process variable Y.

[0095] Further, according to the mutual information values between the process variables, characteristic variables with strong correlation with the missing performance indicators are obtained.

[0096] S22. Construct a random forest prediction model, and input the input data that has been normalized and with outliers removed into the initial random forest prediction model for training.

[0097] S23. Set the objective function of the random forest prediction model to minimize the prediction error, and use the Bayesian optimization strategy to optimize the model parameters to obtain the constructed random forest prediction model.

[0098] Among them, the hyperparameter optimization can be expressed as:

[0099]

[0100] Among them, f(x) represents minimizing the objective function, and the root mean square error (RMSE) is selected; x* is the hyperparameter combination that generates the minimum score, and x can take any value within the χ domain. The result of hyperparameter optimization is to find the model hyperparameters that generate the best score on the validation set metric.

[0101] S24. Based on the constructed random forest prediction model, predict and fill in the missing performance-related data to obtain multi-dimensional complete data.

[0102] S3. Integrate the strip quality standards, design a comprehensive performance rating scheme for hot-rolled strips, and perform performance grade division on the multi-dimensional complete data according to the comprehensive performance rating scheme for hot-rolled strips.

[0103] In a feasible implementation manner, the step S3 specifically includes:

[0104] S31. According to the strip quality standards, perform grade division on the strip performance evaluation indicators.

[0105] Optionally, according to the regulations in the national standard "GB / T 709-2019 Dimensions, Shape, Weight and Tolerances of Hot-Rolled Steel Plates and Strips", the allowable deviations required for different thickness and width specifications of the same steel grade are divided into intervals, and designed into 5 grades of "excellent", "sub-excellent", "good", "medium" and "poor", and the corresponding performance grade labels are 1, 2, 3, 4, 5, among which grades 1-4 are qualified products and grade 5 is unqualified product. According to the regulations in the national standard "GB / T 1591-2018 Low-Alloy High-Strength Structural Steel", the feasible ranges of chemical composition and mechanical properties of strip steel are divided into intervals, also divided into 5 grades, among which grades 1-4 are qualified products and grade 5 is unqualified product. Accordingly, different performance indicators are labeled with grades.

[0106] S32. Calculate the weights (secondary weights) among the performance indicators of the same sub-aspect of the strip steel using the entropy weight method, and the weights of each performance sub-aspect for the comprehensive performance (primary weights).

[0107] In a feasible implementation, the entropy weight method (Entropy Weight Method, abbreviated as EWM) is selected to calculate the weights of multiple performance indicators. In multi-attribute decision analysis, different indicators have different degrees of influence on the overall evaluation result, so weights need to be set. The core idea of the entropy weight method is to measure the amount of information of each indicator according to the size of the information entropy, and thus determine the importance of the indicators. This method can well avoid the interference of subjective factors and make the determination of weights more scientific and objective.

[0108] Specifically, the information entropy of a certain indicator is defined as:

[0109]

[0110] where p ij represents the proportion of the value of the i-th performance indicator in the total sample among the j performance indicators.

[0111] The larger the information entropy, the greater the variation of the indicator, that is, the greater the discrimination of different samples under this indicator. In the entropy weight method, the indicator with a large information entropy will obtain a larger weight. The weight w i of the i-th performance indicator is expressed as:

[0112]

[0113] where m is the total number of performance indicators, and E i is the information entropy of the i-th performance indicator.

[0114] S33. Use the method of multi-level performance indicator weighting to obtain the comprehensive performance grade of the strip steel.

[0115] The comprehensive performance score can be expressed as:

[0116]

[0117] Among them, w i is the weight of the i-th performance index, and N i represents the normalized value of the i-th performance index.

[0118] The comprehensive performance score is divided into 5 grades according to intervals, among which grades 1-4 are qualified products and grade 5 is unqualified products. Optionally, it can be defined that if one of the performance sub-aspects is unqualified, the comprehensive grade is defined as unqualified.

[0119] S34. Perform performance grade division on the obtained multi-dimensional complete data according to the hot-rolled strip performance rating scheme.

[0120] S4. Construct a comprehensive evaluation model for strip steel performance based on multi-task learning and performance-related stacked autoencoders, where each sub-aspect of strip steel performance is used as each sub-task in the multi-task learning framework, and performance-related deep features are extracted based on performance-related stacked autoencoders in each sub-task.

[0121] In the embodiments of the present invention, considering that the hot continuous rolling process has the characteristics of multiple processes, multiple variables, and strong coupling, there are many influencing factors on strip steel performance, and the performance indicators are mutually coupled. Multi-task learning (MTL) and performance-related stacked autoencoders (SPAE) are combined, and using the preprocessed multi-dimensional complete data, a comprehensive evaluation model for strip steel performance based on multi-task learning and performance-related stacked autoencoders (MTL-SPAE) is established. By mining the potential correlation between production data and the quality inspection of the final product performance grade, the performance grade of strip steel products can be evaluated in real time, potential quality problems can be discovered in time, and a basis can be provided for taking corresponding measures for improvement and optimization.

[0122] In a feasible implementation manner, the step S4 specifically includes:

[0123] S41. Based on the multi-task learning method, construct a comprehensive evaluation framework for strip steel performance, and use each sub-aspect of strip steel performance as each sub-task in the multi-task learning framework.

[0124] Most traditional machine learning only builds models for single tasks. For the multi-index characteristics of strip steel performance evaluation, it is necessary to train models for each evaluation index separately, which is redundant and each task is independent, and the coupling characteristics between the indexes cannot be reflected.

[0125] Therefore, considering the multi-index characteristics of strip steel performance evaluation, a multi-task learning strategy is introduced, and hard parameter sharing is selected to enable multiple tasks to learn in parallel in a unified model. By sharing underlying features, the coupling relationship between different performance indicators is mined, which can effectively improve data utilization and model generalization ability, and obtain better performance.

[0126] The loss function of multi-task learning is the weighted sum of the loss functions of multiple learning objectives, and its calculation formula is as follows:

[0127]

[0128] In the formula, λ k is the loss function weight coefficient of the k-th sub-task, and Loss k (x k , y k ) is the loss function of the k-th sub-task, and x k , y k are the input data and output data of the k-th sub-task.

[0129] S42. In each sub-task, use a performance-related stacked autoencoder to extract features from the multi-dimensional complete data to obtain deep features related to the performance of each sub-task.

[0130] In a feasible implementation, the step S42 specifically includes:

[0131] S421. Input the shared feature h 0 obtained from the multi-task learning shared layer into the first performance-related autoencoder network of each sub-task, and use the performance levels of each sub-aspect as supervision during the training process to obtain a performance-related feature representation k represents the k-th sub-task;

[0132] S422. In each sub-task, input the feature representation of the first performance-related autoencoder network into the second performance-related stacked autoencoder network, and use the performance levels of each sub-aspect as supervision during the training process to obtain a performance-related feature representation

[0133] S423. And so on, stack multiple performance-related stacked autoencoder networks layer by layer to obtain a deep feature representation related to performance q represents the number of autoencoder networks.

[0134] Among them, the performance-related stacked autoencoder is constructed as follows:

[0135] Introduce performance level information into the autoencoder structure for supervised learning, construct a performance-related autoencoder, and stack multiple performance-related autoencoders hierarchically to construct the performance-related stacked autoencoder.

[0136] An autoencoder (AE) is an unsupervised neural network model consisting of an encoder and a decoder. In the encoder, the data on the input layer is mapped to the hidden layer to learn the hidden feature representation of the original data. In the decoder, the feature data in the hidden layer reconstructs the original data on the output layer to obtain a new feature representation. Although AE has the ability to extract deep features of the original input data, since it is trained in an unsupervised manner, the information extracted cannot guarantee the correlation with the performance evaluation index. In the present invention, performance level information is introduced into the AE structure for supervised learning to construct a performance-related autoencoder (PAE). While reconstructing the original input data on the output layer, it also classifies the level to which the performance evaluation index belongs based on the hidden layer features. In this way, the model can take into account the feature information related to the performance evaluation level while extracting the essential features of the original data, which is more conducive to accurately distinguishing different performance levels. The encoding process and decoding process of PAE are as follows:

[0137] h = f(Wx T + b)

[0138]

[0139] where x represents the original input data, and h is the feature vector extracted by the hidden layer. represents the reconstructed vector of the original data. f(·) and g(·) are the activation functions of the encoder and decoder respectively, W ∈ R m×n and b ∈ R m are the weight matrix and bias vector of the encoder respectively, and are the weight matrix and bias vector for the decoder to reconstruct the original input data respectively, and are the weight matrix and bias vector for the decoder to predict the performance level respectively, n is the number of variables, m is the number of nodes in the hidden layer, n c is the number of performance levels, is the performance level label predicted by the model. The Softmax function outputs the posterior probability that the sample belongs to each performance level.

[0140] Its loss function is defined as follows:

[0141]

[0142] where N is the number of training samples, x iand are the i-th sample and its reconstruction, y i and are the true label and the predicted label of the i-th sample, respectively. is the reconstruction error of the original input data, is the cross-entropy classification loss of the performance evaluation result, and α and β are weight coefficients.

[0143] To gradually reduce performance-irrelevant information and learn deep features related to performance, multiple PAEs are stacked hierarchically to construct a deep supervised learning network, namely the performance-related stacked autoencoder SPAE. The network structure is as Figure 3 shown.

[0144] S43. Integrate the deep features related to the performance of each subtask, and use the performance-related stacked autoencoder to perform feature extraction again to obtain the deep features related to the comprehensive performance.

[0145] S5. Combine the multi-dimensional complete data pairs to train the strip performance comprehensive evaluation model, and use the Bayesian optimization strategy to tune the hyperparameters of the model to obtain the optimal hyperparameter group that makes the model performance optimal.

[0146] In a feasible implementation manner, the step S5 specifically includes:

[0147] S51. Use the preprocessed multi-dimensional complete data to train the strip performance comprehensive evaluation model based on multi-task learning and performance-related stacked autoencoder (MTL-SPAE). The specific training process of the hot-rolled strip performance comprehensive evaluation model based on MTL-SPAE is as Figure 4 shown.

[0148] S52. The loss function of the training of the strip performance comprehensive evaluation model is shown as follows:

[0149]

[0150] where λ k is the loss function weight coefficient of the k-th subtask, represents the loss function of the performance-related stacked autoencoder in the k-th subtask.

[0151] S53. Use the Bayesian optimization strategy to tune the hyperparameters in the strip performance comprehensive evaluation model. The hyperparameters include the number of underlying neurons shared by multiple tasks, the number of autoencoder networks, the number of neurons in the hidden layer of each autoencoder, and the learning rate, to obtain the optimal hyperparameter group that makes the model performance optimal.

[0152] S54. Save the trained strip performance comprehensive evaluation model and its hyperparameter set.

[0153] S6. Obtain the multi-level performance rating results of the hot-rolled strip according to the trained comprehensive evaluation model of strip performance.

[0154] In a feasible implementation manner, the step S6 specifically includes:

[0155] S61. According to the multi-dimensional complete data and the comprehensive evaluation model of strip performance, use the deep features extracted in each subtask to obtain the sub-rating results representing each performance sub-aspect of the hot-rolled strip;

[0156] S62. According to the deep features related to the comprehensive performance extracted, obtain the comprehensive performance rating result of the hot-rolled strip.

[0157] When a new coil is rolled on the hot-rolling production line, the specific values of the variables required for the model constructed by the present invention can be collected through sensors. If the value of the performance index variable is missing, data filling is performed through the constructed random forest prediction model. The completed multi-dimensional complete data is input into the constructed comprehensive evaluation model of strip performance, and the comprehensive performance grade of this coil and the sub-grade results representing each performance sub-aspect are output. If the comprehensive performance grade is excellent, it is considered that the overall performance of the strip product is excellent; if the comprehensive performance grade is other qualified situations, further check the sub-grade of each performance sub-aspect, and allocate subsequent plans according to different production requirements. If the comprehensive performance grade is poor, it is considered that the product is unqualified and needs to be treated for abnormality.

[0158] In the embodiment of the present invention, the potential correlation between the performance of the final strip product and the production process can be explored through data-driven modeling, and real-time online product performance grading evaluation can be completed. Enterprises can clearly and comprehensively grasp the overall distribution of product performance, realize pricing according to quality, improve economic benefits, and enhance market competitiveness. At the same time, the performance grade fluctuation of a batch of strips can also reflect the fluctuation of the production process or the rationality of the process parameter setting to a certain extent, which can provide a strong basis for enterprises to continuously improve product performance and make targeted process control.

[0159] Correspondingly, the embodiment of the present invention also provides a data-driven comprehensive evaluation device for the performance of hot-rolled strips, as Figure 5 shown. The device 500 includes:

[0160] A data processing module 510, configured to obtain multi-dimensional data of the strip hot-rolling production process and preprocess the multi-dimensional data; wherein, the multi-dimensional data includes process data and performance-related data;

[0161] A data filling module 520, configured to construct a random forest prediction model to fill in the missing data in the performance-related data due to relying on off-line sampling inspection, and obtain multi-dimensional complete data;

[0162] A performance evaluation module 530, which is used to synthesize the strip quality standards, design a comprehensive performance rating scheme for hot-rolled strips, and perform performance level division on the multi-dimensional complete data according to the comprehensive performance rating scheme for hot-rolled strips;

[0163] Construct a comprehensive strip performance evaluation model based on multi-task learning and performance-related stacked autoencoders, where each sub-aspect of strip performance is used as each sub-task in the multi-task learning framework, and performance-related deep features are extracted based on the performance-related stacked autoencoders in each sub-task;

[0164] Train the constructed comprehensive strip performance evaluation model in combination with the multi-dimensional complete data, and use the Bayesian optimization strategy to optimize the model hyperparameters to obtain an optimal hyperparameter group that makes the model performance the best;

[0165] A result output module 540, which is used to obtain a multi-level performance rating result of the hot-rolled strip according to the trained comprehensive strip performance evaluation model.

[0166] For the sake of convenience of description, Figure 5 only the main components of the device are shown. The device of this embodiment can be used to execute Figure 1 the technical solutions of the method embodiment shown, and its implementation principle and technical effects are similar, so they will not be elaborated here.

[0167] In an exemplary embodiment, the present invention also provides an electronic device, and the electronic device includes:

[0168] A processor;

[0169] A memory, on which computer-readable instructions are stored, and when the computer-readable instructions are loaded and executed by the processor, the steps of the data-driven comprehensive hot-rolled strip performance evaluation method as described above are implemented.

[0170] Figure 6 is a schematic structural diagram of an electronic device provided by an embodiment of the present invention. As Figure 6 shown, the electronic device 600 mainly consists of a processor 601 and a memory 602. The processor 601 can execute the steps of the data-driven comprehensive hot-rolled strip performance evaluation method of the present invention by running or executing software programs stored in the memory 602 and calling data stored in the memory 602. The memory 602 is used to store software programs for implementing the solution of the present invention and is controlled by the processor 601 to execute. The specific implementation manner can refer to the above method embodiment and will not be elaborated here.

[0171] In a specific implementation, there may be relatively large differences due to different configurations or performances. As an example, the processor 601 may include one or more CPUs, and the memory 602 may include one or more storage structures. Optionally, the memory may be a read-only memory (ROM), a random access memory (RAM), an electrically erasable programmable read-only memory (EEPROM), an optical disc, a magnetic tape, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer.

[0172] In an exemplary embodiment, the present invention also provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the steps of the data-driven comprehensive performance evaluation method for hot-rolled strip steel as described above. For example, the computer-readable storage medium may be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0173] It should be noted that in this document, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article or terminal device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or terminal device. Without further limitation, an element defined by the statement "comprising one..." does not exclude the existence of additional identical elements in the process, method, article or terminal device comprising the element.

[0174] References in the specification to "an embodiment", "embodiment", "exemplary embodiment", "some embodiments", etc. indicate that the embodiment may include a particular feature, structure, or characteristic, but not every embodiment necessarily includes the particular feature, structure, or characteristic. Additionally, when combining embodiments to describe a particular feature, structure, or characteristic, implementing such feature, structure, or characteristic in combination with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the relevant art.

[0175] It should be understood that the term "and / or" in this document is merely a description of the association relationship between associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. Here, A and B can be singular or plural. Additionally, the character " / " in this document generally represents an "or" relationship between the associated objects before and after, but it may also represent an "and / or" relationship, which can be specifically understood with reference to the context before and after.

[0176] In the present invention, "at least one" means one or more, and "a plurality of" means two or more. "At least one of the following" or a similar expression means any combination of these items, including any combination of single item(s) or plural item(s). For example, at least one of a, b, or c may represent: a, b, c, a - b, a - c, b - c, or a - b - c, where a, b, and c may be single or multiple.

[0177] It should be understood that in various embodiments of the present invention, the magnitudes of the serial numbers of the above - mentioned processes do not mean the order of execution is prior or subsequent. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0178] In several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0179] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0180] In addition, in each embodiment of the present invention, the functional units can be integrated in one processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit.

[0181] When the above-mentioned functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0182] The present invention covers any substitutions, modifications, equivalent methods, and solutions made within the essence and scope of the present invention. To enable the public to have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention. However, those skilled in the art can fully understand the present invention without these detailed descriptions. Additionally, well-known methods, processes, procedures, components, and circuits, etc., are not described in detail to avoid unnecessary confusion to the essence of the present invention.

[0183] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc., made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A data-driven comprehensive evaluation method for hot-rolled strip performance, characterized in that: The following steps are involved: S1. Acquire multidimensional data of a strip hot rolling production process and preprocess the multidimensional data; wherein the multidimensional data includes process data and performance-related data; S2. For the missing data in the performance-related data due to reliance on offline sampling inspection, a random forest prediction model is constructed to fill in the missing values ​​to obtain multi-dimensional complete data; S3. Design a comprehensive rating scheme for hot-rolled strip performance based on the comprehensive rating scheme for hot-rolled strip performance, and classify the multi-dimensional complete data into performance grades according to the comprehensive rating scheme for hot-rolled strip performance; S4. Construct a comprehensive evaluation model of strip steel performance based on multi-task learning and performance-related stacked autoencoders, in which each sub-aspect of strip steel performance is used as each sub-task in the multi-task learning framework, and in each sub-task, the performance-related deep features are extracted based on the performance-related stacked autoencoders; S5. Training the constructed strip steel performance comprehensive evaluation model in combination with the multidimensional complete data, and optimizing the model hyperparameters using a Bayesian optimization strategy to obtain an optimal hyperparameter group that optimizes the model performance; S6. Based on the trained comprehensive evaluation model of strip performance, a multi-level performance rating result of the hot-rolled strip is obtained.

2. The data-driven comprehensive evaluation method for hot-rolled strip performance according to claim 1, characterized in that: The step S1 specifically includes: S11, performing preliminary processing operations of normalizing the collected multidimensional data and deleting outliers; S12. Divide the performance-related data in the multidimensional data into sub-blocks according to process knowledge. After the division, the performance indicators in each sub-block jointly determine one aspect of the strip performance.

3. The data-driven comprehensive evaluation method for hot-rolled strip performance according to claim 1, characterized in that: The step S2 specifically includes: S21, performing correlation analysis on the missing data in the process data and performance-related data to obtain characteristic variables that meet the correlation requirements as input data for model training; S22, constructing a random forest prediction model, inputting the normalized and outlier-removed input data into the initial random forest prediction model for training; S23, setting the objective function of the random forest prediction model to minimize the prediction error, using the Bayesian optimization strategy to tune the model parameters, and obtaining a constructed random forest prediction model; S24. Based on the constructed random forest prediction model, the missing performance-related data are predicted and filled to obtain multi-dimensional complete data.

4. The data-driven comprehensive evaluation method for hot-rolled strip performance according to claim 1, characterized in that: The step S3 specifically includes: S31. Classify the performance indicators of the strip steel according to the strip steel quality standards; S32, using the entropy weight method to calculate the weights between the performance indicators of the same sub-aspect of the strip steel, i.e., the secondary weights, and the weights of each performance sub-aspect for the comprehensive performance, i.e., the primary weights; S33, using a multi-level performance index weighting method to obtain a comprehensive performance grade of the strip; S34. Classify the obtained multi-dimensional complete data into performance grades according to the hot-rolled strip performance rating scheme.

5. The data-driven hot-rolled strip performance comprehensive evaluation method according to claim 1, characterized in that: The step S4 specifically includes: S41. Based on the multi-task learning method, a comprehensive evaluation framework for strip steel performance is constructed, and each sub-aspect of strip steel performance is used as each sub-task in the multi-task learning framework; S42, in each subtask, using a performance-related stacked autoencoder to extract features from the multidimensional complete data to obtain deep features related to the performance of each subtask; S43. Integrate the deep features related to the performance of each subtask, use the performance-related stacked autoencoder to extract features again, and obtain the deep features related to the comprehensive performance.

6. The data-driven comprehensive evaluation method for hot-rolled strip performance according to claim 5, characterized in that: The step S41 specifically includes: Considering the multi-index characteristics of strip performance evaluation, a multi-task learning strategy is introduced, and hard parameter sharing is selected to enable multiple subtasks to be learned in parallel in a unified model, and the coupling relationship between different performance indicators is explored by sharing the underlying features; The loss function of multi-task learning is the weighted sum of the loss functions of multiple learning objectives, and its calculation formula is as follows: In the formula, λk is the weight coefficient of the loss function of the kth subtask, Loss k (x k ,y k ) is the loss function of the kth subtask, x k ,y k are the input data and output data of the kth subtask.

7. The data-driven comprehensive evaluation method for hot-rolled strip performance according to claim 5, characterized in that: The step S42 specifically includes: S421, the shared features h obtained by multi-task learning shared layer 0 Input to the first performance-related autoencoder network of each subtask, the training process uses the performance level of each sub-aspect as supervision to obtain performance-related feature representation k represents the kth subtask; S422, in each subtask, the feature representation of the first performance-related autoencoder network Input to the second performance-related stacked autoencoder network, the training process uses the performance level of each sub-aspect as supervision to obtain performance-related feature representation S423, and so on, multiple performance-related stacked autoencoder networks are stacked in layers to obtain performance-related deep feature representations q represents the number of autoencoder networks; Among them, the performance-related stacked autoencoder is constructed as follows: Performance level information is introduced into the autoencoder structure for supervised learning, a performance-related autoencoder is constructed, and multiple performance-related autoencoders are stacked in layers to construct the performance-related stacked autoencoder.

8. The data-driven comprehensive evaluation method for hot-rolled strip performance according to claim 1, characterized in that: The step S5 specifically includes: S51. Using the preprocessed multidimensional complete data, a strip steel performance comprehensive evaluation model based on multi-task learning and performance-related stacked autoencoders is trained; S52, the loss function of the strip steel performance comprehensive evaluation model training is as follows: Among them, λ k is the weight coefficient of the loss function of the kth subtask, represents the performance-dependent stacked autoencoder loss function in the kth subtask; S53, using a Bayesian optimization strategy to tune the hyperparameters in the comprehensive evaluation model of strip performance, wherein the hyperparameters include the number of multi-task shared bottom-layer neurons, the number of autoencoder networks, the number of neurons in the hidden layer of each encoder, and the learning rate, to obtain an optimal hyperparameter group that optimizes the model performance; S54, saving the trained comprehensive evaluation model of strip steel performance and its hyperparameter set.

9. The data-driven comprehensive evaluation method for hot-rolled strip performance according to claim 1, characterized in that: The step S6 specifically includes: S61, according to the multi-dimensional complete data and the comprehensive evaluation model of strip steel performance, using the deep features extracted from each subtask, obtaining sub-item rating results representing each performance sub-aspect of the hot-rolled strip steel; S62. Obtain comprehensive performance rating results of the hot-rolled strip based on the extracted deep features related to the comprehensive performance.

10. A data-driven hot-rolled strip performance comprehensive evaluation device, the device being used to implement the method according to any one of claims 1 to 9, characterized in that: The device comprises: A data processing module, used for acquiring multi-dimensional data of the strip hot rolling production process and pre-processing the multi-dimensional data; wherein the multi-dimensional data includes process data and performance-related data; A data completion module is used to construct a random forest prediction model to complete missing values ​​of the performance-related data due to reliance on offline sampling inspection, so as to obtain multi-dimensional complete data; A performance evaluation module, which is used to integrate the strip quality standards, design a comprehensive rating scheme for the performance of hot-rolled strip steel, and classify the multi-dimensional complete data into performance grades according to the comprehensive rating scheme for the performance of hot-rolled strip steel; A comprehensive strip performance evaluation model based on multi-task learning and performance-related stacked autoencoders is constructed, in which each sub-aspect of strip performance is used as a sub-task in the multi-task learning framework, and performance-related deep features are extracted in each sub-task based on the performance-related stacked autoencoder; The constructed strip steel performance comprehensive evaluation model is trained in combination with the multi-dimensional complete data, and the model hyperparameters are tuned using a Bayesian optimization strategy to obtain an optimal hyperparameter group that optimizes the model performance; The result output module is used to obtain the multi-level performance rating results of the hot-rolled strip according to the trained comprehensive evaluation model of the strip performance.

Images