Method and device for predicting quality performance of strip steel in hot continuous rolling process

Abstract

The invention provides a strip steel hot continuous rolling process quality performance prediction method and device, and belongs to the technical field of industrial process performance prediction. The method comprises the following steps: acquiring process data of a strip steel hot continuous rolling process; preprocessing the data set, and taking the preprocessed data as a to-be-screened data set; performing variable screening on the to-be-screened data set in combination with an XGBoost algorithm and strip steel process mechanism knowledge to obtain a model input data set; a strip steel quality performance prediction model is constructed, the strip steel quality performance prediction model is a CNN-LSTM model constructed in combination with an attention mechanism module, and the CNN-LSTM model is constructed based on a CNN feature extraction layer and an LSTM module; taking the data set as the input of a prediction model, outputting a long-time-sequence prediction value of the model on the strip steel quality performance, and obtaining an optimal model parameter through repeated iterative training; and online real-time data in the strip steel hot continuous rolling process are input into the trained strip steel quality performance prediction model, and the strip steel quality performance can be predicted in real time.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial process performance prediction, and particularly to a method and device for predicting the quality performance of a hot strip continuous rolling process. Background Art

[0002] The hot strip continuous rolling process is a steel production process with complex mechanisms, large scale, high efficiency, and multiple processes, which is a typical complex process industry. This production line mainly consists of a heating furnace, a rough rolling mill, a hot output roller table and a flying shear, a finishing mill, a laminar cooling, and a coiling unit, etc. In modern industrial manufacturing, especially for a complex and continuous production process like hot strip continuous rolling, ensuring the high-quality standards of the final product is a multi-dimensional challenge. With the transformation of manufacturing towards intelligent manufacturing, quality performance prediction has become a key research area, aiming to predict the quality status of products during the production process in advance through advanced data analysis techniques and model construction, so as to achieve optimized control of the production process, reduce the scrap rate, and improve production efficiency and product quality.

[0003] The data-driven modeling method is a main method in the current field of quality performance prediction. Thanks to the progress of sensor technology and the application of the Internet of Things, a large amount of production parameters can now be collected in real time, including but not limited to temperature, pressure, speed, etc. These rich data provide a basis for deeply understanding the relationship between various factors in the production process and product quality. However, due to the characteristics of strong coupling, dynamics, and multi-dimension of data in the industrial environment, how to effectively extract and utilize this information has become the focus of research. Summary of the Invention

[0004] The present invention provides a method and device for predicting the quality performance of a hot strip continuous rolling process. The technical solution is as follows:

[0005] On the one hand, a method for predicting the quality performance of a hot strip continuous rolling process is provided. This method is applied to an electronic device and includes:

[0006] S1. Obtain the process data of the hot strip continuous rolling process;

[0007] S2. Preprocess the data set, and use the preprocessed data as the data set to be screened;

[0008] S3. Combine the XGBoost algorithm and the knowledge of the strip process mechanism to screen variables from the data set to be screened, and obtain the model input data set;

[0009] S4. Build a strip quality performance prediction model, use the said data set as the input of the strip quality performance prediction model, and the output is the long-time series prediction value of the strip quality performance by the model. After multiple iterative trainings, obtain the optimal model parameters. Among them, the strip quality performance prediction model is a CNN-LSTM model constructed by combining an attention mechanism module. The CNN-LSTM model includes a CNN feature extraction layer, an LSTM module, an attention mechanism module, a Softmax layer, and an output layer;

[0010] The CNN feature extraction layer includes a convolutional layer, a max pooling layer, a fully connected layer, and a Dropout layer. The convolutional layer is used to extract feature maps from the input sequence, while maintaining the position sensitivity and translational invariance of these features. After the convolutional operation, a ReLU activation function is applied to introduce non-linearity into the model. The max pooling layer is used to reduce the spatial size of the feature map and select the maximum value within each window as the output. The fully connected layer is used for dimensional transformation, and then a ReLU activation function is applied to introduce non-linearity into the model. The Dropout layer is used to prevent overfitting;

[0011] The LSTM module consists of four parts: a cell state, a forget gate, an input gate, and an output gate, and can capture long-term dependencies in the time series using the hidden state;

[0012] The attention mechanism module determines the weights of each part in the input sequence by calculating the similarity between the query and the key, and then performs weighted summation on the values according to these weights to generate a context vector reflecting the most important information in the input, thereby enhancing the model's attention to key information;

[0013] The Softmax layer is used to convert the output of the previous layer into a probability distribution;

[0014] The output layer is used to output the prediction result;

[0015] S5. Input the online real-time data of the hot strip rolling process into the trained strip quality performance prediction model to obtain the real-time quality performance prediction result.

[0016] Optionally, the preprocessing of the data set in S2, and using the preprocessed data as the data set to be screened, includes:

[0017] S21. Centralize the data set to obtain the centralized data;

[0018] S22. Normalize the centralized data, and the normalized data is used as the data set to be screened.

[0019] Optionally, combining the XGBoost algorithm with strip steel process mechanism knowledge in S3 to perform variable screening on the dataset to be screened, and obtaining a model input dataset, including:

[0020] S31. Construct a decision tree using the XGBoost algorithm, use the number of times of feature splitting as a variable importance measurement index to perform feature scoring on the dataset to be screened, and determine the set of variables to be deleted according to the feature scores;

[0021] S32. Combine strip steel process mechanism knowledge to perform deletion processing on the set of variables to be deleted, and obtain a model input dataset.

[0022] Optionally, the construction and training process of the strip steel quality performance prediction model in S4 includes:

[0023] S41. In the feature extraction stage of the strip steel quality performance prediction model, use the training data that has undergone data preprocessing and variable screening as the input of the CNN feature extraction layer;

[0024] S42. Extract the time series correlation of the output variables of the CNN through the LSTM module, calculate the hidden layer state of the input feature at the current moment from the hidden layer state of the previous moment of the LSTM and the input variables at the current moment, and obtain the variables input to the attention mechanism module;

[0025] S43. Convert the scores in the attention mechanism module into weights through the Softmax layer, and output the long-time prediction value of the strip steel quality performance by weighted summation;

[0026] S44. Select an appropriate loss function to compare the difference between the predicted value and the actual value of the model. The backpropagation algorithm is used to calculate the gradient of the loss function with respect to each weight. According to the calculated gradient, an optimization algorithm is used to update the weights in the network;

[0027] S45. After multiple iterative trainings, the strip steel quality performance prediction model gradually improves its internal weights and gradually approaches the optimal solution that can minimize the loss function.

[0028] Optionally, inputting the online real-time data of the strip steel hot continuous rolling process into the trained strip steel quality performance prediction model to obtain real-time quality performance prediction results, including:

[0029] S51. Obtain the online real-time data of the strip steel hot continuous rolling process;

[0030] S52. Perform preprocessing and variable screening operations on the real-time dataset to obtain a model real-time input dataset;

[0031] S53. Use the real-time input data set as the input of the trained strip quality performance prediction model to predict the strip quality performance in real time.

[0032] On the other hand, a strip hot continuous rolling process quality performance prediction device is provided. The device includes:

[0033] A data acquisition module for acquiring process data of the strip hot continuous rolling process;

[0034] A preprocessing module for preprocessing the data set and using the preprocessed data as the data set to be screened;

[0035] A variable screening module for combining the XGBoost algorithm and strip process mechanism knowledge to screen variables of the data set to be screened to obtain a model input data set;

[0036] A model construction module for implementing the construction and training process of the strip quality performance prediction model. Among them, the strip quality performance prediction model is a CNN-LSTM model constructed by combining an attention mechanism module. The CNN-LSTM model includes a CNN feature extraction layer, an LSTM module, an attention mechanism module, a Softmax layer, and an output layer;

[0037] The CNN feature extraction layer includes a convolutional layer, a max pooling layer, a fully connected layer, and a Dropout layer. The convolutional layer is used to extract feature maps from the input sequence while maintaining the position sensitivity and translational invariance of these features. After the convolutional operation, a ReLU activation function is applied to introduce nonlinearity into the model. The max pooling layer is used to reduce the spatial size of the feature map and select the maximum value in each window as the output. The fully connected layer is used for dimensional transformation, and then a ReLU activation function is applied to introduce nonlinearity into the model. The Dropout layer is used to prevent overfitting;

[0038] The LSTM module is composed of four parts: a cell state, a forget gate, an input gate, and an output gate, and can capture long-term dependencies in the time series using the hidden state;

[0039] The attention mechanism module determines the weights of each part in the input sequence by calculating the similarity between the query and the key, and then weights and sums the values according to these weights to generate a context vector reflecting the most important information in the input, so as to enhance the model's attention to key information;

[0040] The Softmax layer is used to convert the output of the previous layer into a probability distribution;

[0041] The output layer is used to output the prediction result;

[0042] A real-time prediction module, configured to input online real-time data of the hot strip continuous rolling process into the trained strip quality performance prediction model to obtain real-time quality performance prediction results.

[0043] Optionally, the preprocessing module is further configured to:

[0044] S21. Centralize the data set to obtain centralized data;

[0045] S22. Normalize the centralized data, and the normalized data is used as the data set to be screened.

[0046] Optionally, the variable screening module is further configured to:

[0047] S31. Use the XGBoost algorithm to construct a decision tree, adopt the number of times of feature splitting as a variable importance measurement index, score the features of the data set to be screened, and determine the variable set to be deleted according to the feature scores;

[0048] S32. Combine the preset strip mechanism knowledge to delete the variable set to be deleted to obtain the model input data set;

[0049] Optionally, the model construction module is further configured to:

[0050] S41. In the feature extraction stage of the strip quality performance prediction model, use the training data after data preprocessing and variable screening as the input of the CNN feature extraction layer;

[0051] S42. Extract the time series correlation of the output variables of the CNN through the LSTM module, calculate the hidden layer state of the input feature at the current moment from the hidden layer state of the previous moment of the LSTM and the input variables at the current moment, and obtain the variables input to the attention mechanism module;

[0052] S43. Convert the scores in the attention mechanism module into weights through the Softmax layer, and output the long-time prediction value of the model for the strip quality performance through weighted summation;

[0053] S44. Select an appropriate loss function to compare the difference between the predicted value and the actual value of the model. The backpropagation algorithm is used to calculate the gradient of the loss function with respect to each weight. According to the calculated gradient, an optimization algorithm is used to update the weights in the network;

[0054] S45. After multiple iterative trainings, the strip quality performance prediction model gradually improves its internal weights and gradually approaches the optimal solution that can minimize the loss function.

[0055] Optionally, the real-time prediction module is further configured to:

[0056] S51. Obtain the on-line real-time data in the strip hot continuous rolling process;

[0057] S52. Perform operations such as preprocessing and variable screening on the real-time data set to obtain the real-time input data set of the model;

[0058] S53. Use the real-time input data set as the input of the trained prediction model to predict the strip quality performance in real time.

[0059] On the other hand, an electronic device is provided. The electronic device includes a processor and a memory. At least one instruction is stored in the memory, and the at least one instruction is loaded and executed by the processor to implement the above-mentioned strip hot continuous rolling process quality performance prediction method based on process data.

[0060] On the other hand, a computer-readable storage medium is provided. At least one instruction is stored in the storage medium, and the at least one instruction is loaded and executed by a processor to implement the above-mentioned strip hot continuous rolling process quality performance prediction method based on process data.

[0061] The beneficial effects brought by the technical solution provided by the present invention at least include:

[0062] In the present invention, a machine learning method is introduced into the strip data preprocessing and variable screening process to improve the accuracy and interpretability of the model.

[0063] A CNN-LSTM model is constructed to improve the model accuracy. Among them, a CNN feature extraction layer is constructed to capture the spatial features between input variables. The internal gate operation mode of LSTM is used to transfer the time series relationship between industrial process samples through the cell state at different times, which can alleviate the gradient disappearance phenomenon. LSTM can effectively learn long-term dependence relationships. While considering spatial information, it can also capture the time dynamic changes in the data. An attention mechanism module is introduced to improve the model's ability to capture the long-term time series dependence of the sequence. At the same time, the attention mechanism module allows the model to pay more attention to the parts that are more important for the current task when processing sequence data, and ignore the irrelevant or noisy parts. This helps to enhance the model's attention to relevant features, helps the model better understand and utilize key information, and improves the accuracy and efficiency of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] 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 the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0065] Figure 1 It is a schematic flow chart of the method for predicting the quality performance in the strip hot continuous rolling process provided by the embodiment of the present invention;

[0066] Figure 2 It is a schematic flow chart of the strip hot continuous rolling process provided by the embodiment of the present invention;

[0067] Figure 3 It is a schematic flow chart of the finish rolling process provided by the embodiment of the present invention;

[0068] Figure 4 It is a schematic diagram of the CNN feature extraction layer in the CNN-LSTM model provided by the embodiment of the present invention;

[0069] Figure 5 It is a schematic diagram of the overall framework of the strip quality performance prediction model provided by the embodiment of the present invention;

[0070] Figure 6 It is a schematic structural diagram of the device for predicting the quality performance in the strip hot continuous rolling process provided by the embodiment of the present invention;

[0071] Figure 7 It is a schematic structural diagram of an electronic device provided by the embodiment of the present invention. Detailed implementation manners

[0072] To make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments.

[0073] The embodiment of the present invention provides a method for predicting the quality performance in the strip hot continuous rolling process. This method can be implemented by an electronic device, and the electronic device can be a terminal or a server. As Figure 1 shown in the flow chart of the method for predicting the quality performance in the strip hot continuous rolling process, the processing flow of this method can include the following steps:

[0074] S1. Obtain the process data of the strip hot continuous rolling process;

[0075] In a feasible implementation manner, the strip hot continuous rolling process is a steel production process with complex mechanisms, large scale, high efficiency and multiple processes, and is a typical complex process industry. As Figure 2 shown, this production line mainly consists of a heating furnace, a rough rolling mill, a hot delivery roller table and a flying shear, a finish rolling mill, a laminar cooling and a coiling unit, etc. The embodiment of the present application uses the process data of the finish rolling mill as the training data and the test data to illustrate the present application. As Figure 3As shown in the figure, the finishing mill is usually composed of seven stands. Each stand is connected through a control loop. There are two backup rolls above and below each stand, and two work rolls in the middle. A hydraulic system is equipped to provide the required rolling force and bending roll force, which can control the thickness of the steel and the smooth forward movement of the strip steel. Therefore, the finishing process data collected on-site (including process variables, composition variables, operation variables, etc.) will be used as training data and test data.

[0076] S2. Preprocess the data set, and use the preprocessed data as the data set to be screened.

[0077] In this embodiment, preferably, the data set can be preprocessed based on the corresponding statistical characteristics of the data set, which may specifically include the following steps:

[0078] S21. Centralize the data set to obtain the centralized data.

[0079] As an alternative embodiment, the mean value of the whole sample can be subtracted from each variable in the data to obtain the change value of each sample relative to the overall average value.

[0080] S22. Normalize the centralized data, and the normalized data is used as the data set to be screened.

[0081] As an alternative embodiment, the centralized data can be divided by its standard deviation to calibrate each variable to unit variance, so that individual special variables will not be in a dominant position.

[0082] S3. Combine the XGBoost algorithm and the strip steel process mechanism knowledge to screen the variables of the data set to be screened, and obtain the model input data set.

[0083] Among them, the strip steel mechanism knowledge refers to the operating rules and principles that affect the quality of the final product inside the strip steel and in the strip steel hot rolling system, which may include the influence of strip steel rolling process parameters (such as rolling force, rolling speed, heating temperature, etc.) on the quality performance of the strip steel.

[0084] Optionally, combining the XGBoost algorithm and the strip steel mechanism knowledge in S3 to screen the variables of the data set to be screened and obtain the model input data set includes:

[0085] S31. Use the XGBooost algorithm to construct a decision tree, use the number of times of feature splitting as the variable importance measurement index, score the features of the data set to be screened, and determine the variable set to be deleted according to the feature scores.

[0086] S32. Combine the preset strip steel mechanism knowledge to delete the variable set to be deleted, and obtain the model input data set.

[0087] S4. Build a prediction model for strip quality performance. Use the above dataset as the input of the prediction model, and the output is the long-term prediction value of the strip quality performance by the model. After multiple iterative trainings, obtain the optimal model parameters.

[0088] Among them, the prediction model for strip quality performance is a CNN-LSTM model constructed by combining an attention mechanism module. The CNN-LSTM model includes a Convolutional Neural Network (CNN) feature extraction layer, a Long Short-Term Memory (LSTM) module, an attention mechanism module, a Softmax layer, and an output layer.

[0089] Optionally, as Figure 4 shown, the CNN feature extraction layer includes a convolutional layer, a max pooling layer, a fully connected layer, and a Dropout layer. The convolutional layer is used to extract feature maps from the input sequence while maintaining the position sensitivity and translational invariance of these features. After the convolutional operation, a ReLU activation function is applied to introduce non-linearity into the model. The pooling layer is used to reduce the spatial size of the feature map and selects the maximum value within each window as the output. The fully connected layer is used for dimensional transformation, and then a ReLU activation function is applied to introduce non-linearity into the model. The Dropout layer is used to prevent overfitting.

[0090] Optionally, the LSTM module consists of four parts: a cell state, a forget gate, an input gate, and an output gate, and can capture long-term dependencies in the time series using the hidden state.

[0091] The LSTM model mainly uses a gate structure. Each LSTM cell receives the hidden layer output h t-1 from the previous moment, the cell state C t-1 and the input x t at the current moment, and calculates the hidden layer output h t and the cell state C t at the current moment. Compared with the RNN, the LSTM can simultaneously remember and learn the long-term and short-term time series relationships between different moments through the transmission of the cell state and the design of the gate structure, is suitable for the processing of time series data, and the internal calculation method of the LSTM can effectively avoid the gradient disappearance problem originally existing in the RNN.

[0092] The specific calculation formulas are shown as formulas (1)-(6) below, where W f , W i , W c , W o and b f , b i , bc , b o respectively represent the weight and bias parameters in the model, σ(·) and tanh(·) represent the Sigmoid and Tanh activation functions respectively, and * represents the Hadamard product.

[0093] f t = σ(W f · [h t-1 , x t + b f ) (1)

[0094] i t = σ(W t · [h t-1 , x t + b t ) (2)

[0095]

[0096] o t = σ(W o · [h t-1 , x t + b o ) (5)

[0097] h t = o t * tanh(C t ) (6)

[0098] Through the cyclic input model structure, the LSTM can consider the influence of past moments when extracting the sample features at the current moment, so as to effectively mine and extract the temporal characteristics between samples in the industrial process.

[0099] Optionally, the attention mechanism module mainly determines the weights of each part in the input sequence by calculating the similarity between the query and the key, and then performs weighted summation on the values according to these weights to generate a context vector reflecting the most important information in the input, so as to enhance the model's attention to key information. The working process of the attention mechanism module is as follows:

[0100] 1. The input sequence generates query (Query, Q), key (Key, K), and value (Value, V) vectors through linear transformation. Specifically, assume the input sequence is X = [x1, x2,..., x n , where n represents the sequence length, and each x i is a vector. Through the weight matrices W Q , W K , and W V , the input is respectively converted into the query W Q , the key K, and the value V, that is:

[0101] Q = XW Q (7)

[0102] K = XW K (8)

[0103] V = XW V (9)

[0104] 2. Calculate the similarity score between the query and all keys. In this embodiment, the preferred similarity function is the scaled dot product, which is defined as follows:

[0105]

[0106] where d k represents the dimension of the key vector, and the square root in the denominator is used for scaling to prevent the inner product from being too large and causing the Softmax function to saturate.

[0107] 3. Use the Softmax function to convert the above scores into a probability distribution, and the formula is as follows:

[0108]

[0109] Here, α ij represents the attention of the i-th query to the j-th key.

[0110] 4. According to the obtained probability distribution as weights, perform a weighted sum on all value vectors to generate the context vector C:

[0111] C = αV (12)

[0112] More specifically, for each query, the corresponding context vector can be expressed as:

[0113]

[0114] 5. The context vector is used as the input of the subsequent layer (output layer) to help generate the final output.

[0115] Optionally, the construction and training process of the strip quality performance prediction model in S4 may include the following steps S41 - S45:

[0116] S41. In the feature extraction stage of the strip quality performance prediction model, use the training data that has undergone data preprocessing and variable screening as the input of the CNN feature extraction layer;

[0117] S42. Extract the time series correlation of the output variables of the CNN through the LSTM module, and calculate the hidden layer state of the input features at the current moment from the hidden layer state of the previous moment of the LSTM and the input variables at the current moment to obtain the variables input to the input attention mechanism module;

[0118] Because the LSTM can effectively mine and represent the long-term and short-term time series relationships existing in the data, and industrial objects often have obvious time series characteristics, and since different variables are at different time lag positions in industrial production, long-term and short-term time series relationships may exist simultaneously. Therefore, in this embodiment, the LSTM is used to learn the time relationship of the industrial process.

[0119] S43. Convert the scores in the attention mechanism module into weights through the Softmax layer, and output the long-time series prediction value of the strip steel quality performance after weighted summation.

[0120] S44. Select an appropriate loss function to compare the difference between the predicted value and the actual value of the model, and the backpropagation algorithm is used to calculate the gradient of the loss function with respect to each weight. According to the calculated gradient, an optimization algorithm is used to update the weights in the network;

[0121] S45. After multiple iterative trainings, the strip steel quality performance prediction model gradually improves its internal weights to obtain the optimal model parameters.

[0122] The overall framework of the strip steel quality performance prediction model is as Figure 5 shown.

[0123] S5. Input the online real-time data of the strip steel hot continuous rolling process into the trained strip steel quality performance prediction model to obtain the real-time quality performance prediction result.

[0124] Optionally, inputting the online real-time data of the strip steel hot continuous rolling process into the trained strip steel quality performance prediction model in S5 to obtain the real-time quality performance prediction result may specifically include the following steps:

[0125] S51. Obtain the online real-time data of the strip steel hot continuous rolling process.

[0126] S52. Perform preprocessing and variable screening operations on the real-time data set to obtain the real-time input data set of the model;

[0127] S53. Use the real-time input data set as the input of the trained prediction model to predict the strip steel quality performance in real time.

[0128] Optionally, use the coefficient of determination (R 2) and two commonly used performance metrics, the coefficient of determination (R - squared) and the root mean square error (RMSE), are used as evaluation metrics for prediction performance. The R 2 - squared metric represents the proportion of the total variance explained by the regression model. This metric is less than 1, and the closer it is to 1, the better the model performance. The RMSE metric represents the average level of the deviation between the predicted values and the true values. This metric is greater than 0, and the closer it is to 0, the better the model performance. Formulas (14) and (15) are their calculation formulas. Where m is the number of test samples, y i and represent the true value and the predicted value of sample i respectively, is the average value of the true values of all test samples.

[0129]

[0130] In the embodiments of the present invention, machine learning methods are introduced into the strip data pre - processing and variable screening processes to improve the accuracy and interpretability of the model.

[0131] A CNN - LSTM model is constructed to improve the model accuracy; the spatial features between input variables are captured by constructing the CNN input layer.

[0132] The internal gate operation mode of LSTM is used to transfer the temporal relationship between industrial process samples through the cell state at different times, which can alleviate the problem of gradient disappearance. LSTM can effectively learn long - term dependence relationships and can capture the temporal dynamic changes in the data while considering the spatial information.

[0133] An attention mechanism module is introduced to improve the model's ability to capture the long - term temporal dependence of sequences. At the same time, the attention mechanism module allows the model to pay more attention to the parts that are more important for the current task when processing sequence data, and ignore those irrelevant or noisy parts. This helps to enhance the model's attention to relevant features, helps the model better understand and utilize key information, and improves the accuracy and efficiency of the model.

[0134] As Figure 6 shown, the embodiments of the present invention also provide a device 600 for predicting the quality performance of the hot strip rolling process. This device 600 is applied to the method for predicting the quality performance of the hot strip rolling process. Referring to Figure 6 , this device 600 includes:

[0135] A data acquisition module 610, which is used to acquire the process data of the hot strip rolling process;

[0136] A pre - processing module 620, which is used to pre - process the data set and use the pre - processed data as the data set to be screened;

[0137] The variable screening module 630 is used to combine the XGBoost algorithm and the strip process mechanism knowledge to screen variables for the dataset to be screened, and obtain the model input dataset;

[0138] The model construction module 640 is used to implement the construction and training process of the strip quality performance prediction model. Among them, the strip quality performance prediction model is a CNN-LSTM model constructed by combining an attention mechanism module. The CNN-LSTM model includes a CNN feature extraction layer, an LSTM module, an attention mechanism module, a Softmax layer, and an output layer;

[0139] The CNN feature extraction layer includes a convolutional layer, a max pooling layer, a fully connected layer, and a Dropout layer. The convolutional layer is used to extract feature maps from the input sequence, while maintaining the position sensitivity and translational invariance of these features. After the convolutional operation, a ReLU activation function is applied to introduce non-linearity into the model. The max pooling layer is used to reduce the spatial size of the feature map and select the maximum value within each window as the output. The fully connected layer is used for dimensional transformation, and then a ReLU activation function is applied to introduce non-linearity into the model. The Dropout layer is used to prevent overfitting;

[0140] The LSTM module consists of four parts: a cell state, a forget gate, an input gate, and an output gate, and can capture long-term dependencies in the time series using the hidden state;

[0141] The attention mechanism module determines the weights of each part in the input sequence by calculating the similarity between the query and the key, and then performs weighted summation on the values according to these weights to generate a context vector reflecting the most important information in the input, thereby enhancing the model's attention to key information;

[0142] The Softmax layer is used to convert the output of the previous layer into a probability distribution;

[0143] The output layer is used to output the prediction result;

[0144] The real-time prediction module 650 is used to input the online real-time data of the hot strip rolling process into the trained strip quality performance prediction model to obtain the real-time quality performance prediction result.

[0145] Optionally, the preprocessing module 620 is used for:

[0146] S21. Centralize the dataset to obtain the centralized data;

[0147] S22. Normalize the centralized data, and the normalized data is used as the dataset to be screened.

[0148] Optionally, the preprocessing module 620 is further configured to:

[0149] S21. Centralize the data set to obtain centralized data;

[0150] S22. Normalize the centralized data, and the normalized data is used as the data set to be screened.

[0151] Optionally, the variable screening module 630 is configured to:

[0152] S31. Use the XGBoost algorithm to construct a decision tree, use the number of times of feature splitting as a variable importance measurement index, perform feature scoring on the data set to be screened, and determine the variable set to be deleted according to the feature scores;

[0153] S32. Combine the preset strip steel mechanism knowledge to perform deletion processing on the variable set to be deleted to obtain the model input data set.

[0154] Optionally, the model construction module 640 is further configured to:

[0155] S41. In the feature extraction stage of the strip steel quality performance prediction model, use the training data after data preprocessing and variable screening as the input of the CNN feature extraction layer;

[0156] S42. Extract the time series correlation of the output variables of the CNN through the LSTM module, calculate the hidden layer state of the input feature at the current moment from the hidden layer state of the previous moment of the LSTM and the input variable at the current moment, and obtain the variable for the input attention mechanism module;

[0157] S43. Convert the scores in the attention mechanism module into weights through the Softmax layer, and output the long-time prediction value of the model for the strip steel quality performance through weighted summation;

[0158] S44. Select an appropriate loss function to compare the difference between the predicted value and the actual value of the model. The backpropagation algorithm is used to calculate the gradient of the loss function with respect to each weight. According to the calculated gradient, an optimization algorithm is used to update the weights in the network;

[0159] S45. After multiple iterative trainings, the strip steel quality performance prediction model gradually improves its internal weights and gradually approaches the optimal solution that can minimize the loss function.

[0160] Optionally, the real-time prediction module 650 is further configured to:

[0161] S51. Obtain the online real-time data of the strip steel hot continuous rolling process;

[0162] S52. Preprocess the real-time data set, perform operations such as variable screening, etc., to obtain the real-time input data set for the model;

[0163] S53. Use the real-time input data set as the input of the trained prediction model to predict the strip quality performance in real time.

[0164] Figure 7 It is a schematic structural diagram of an electronic device 700 provided by an embodiment of the present invention. The electronic device 700 may vary greatly due to different configurations or performances, and may include one or more processors (central processing units, CPUs) 701 and one or more memories 702. Among them, at least one instruction is stored in the memory 702, and the at least one instruction is loaded and executed by the processor 701 to implement the steps of the above-mentioned strip hot continuous rolling process quality performance prediction method based on process data.

[0165] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions. The above instructions can be executed by a processor in a terminal to complete the above-mentioned strip hot continuous rolling process quality performance prediction method based on process data. For example, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0166] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above embodiments can be completed by hardware, or can be completed by a program instructing relevant hardware. The program can be stored in a computer-readable storage medium, and the above-mentioned storage medium can be a read-only memory, a magnetic disk, or an optical disc, etc.

[0167] 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 in the protection scope of the present invention.

Claims

1. A method for predicting the quality performance in the hot continuous rolling process of strip steel, characterized in that, The method includes: S1. Obtain the process data of the hot strip continuous rolling process; S2. Preprocess the data set, and use the preprocessed data as the data set to be screened; S3. Combine the XGBoost algorithm and the strip process mechanism knowledge to screen variables from the data set to be screened, and obtain the model input data set; S4. Build a strip quality performance prediction model, use the data set as the input of the strip quality performance prediction model, and the output is the long-time series prediction value of the strip quality performance by the model. After multiple iterative trainings, obtain the optimal model parameters. Among them, the strip quality performance prediction model is a CNN-LSTM model constructed by combining an attention mechanism module. The CNN-LSTM model includes a CNN feature extraction layer, an LSTM module, an attention mechanism module, a Softmax layer, and an output layer; The CNN feature extraction layer includes a convolutional layer, a max pooling layer, a fully connected layer, and a Dropout layer. The convolutional layer is used to extract feature maps from the input sequence while maintaining the position sensitivity and translational invariance of these features. After the convolutional operation, a ReLU activation function is applied to introduce non-linearity into the model. The max pooling layer is used to reduce the spatial size of the feature map and select the maximum value in each window as the output. The fully connected layer is used for dimensional transformation, and then a ReLU activation function is applied to introduce non-linearity into the model. The Dropout layer is used to prevent overfitting; The LSTM module consists of four parts: cell state, forget gate, input gate, and output gate, and can capture long-term dependencies in the time series using the hidden state; The attention mechanism module determines the weights of each part in the input sequence by calculating the similarity between the query and the key, and then weights and sums the values according to these weights to generate a context vector reflecting the most important information in the input, so as to enhance the model's attention to key information; The Softmax layer is used to convert the output of the previous layer into a probability distribution; The output layer is used to output the prediction result; S5. Input the online real-time data of the hot strip continuous rolling process into the trained strip quality performance prediction model to obtain the real-time quality performance prediction result.

2. The method according to claim 1, wherein The preprocessing of the data set in S2, and using the preprocessed data as the data set to be screened, includes: S21. Centralize the data set to obtain the centralized data; S22. Normalize the centralized data, and the normalized data is used as the data set to be screened.

3. The method according to claim 1, wherein The combination of the XGBoost algorithm and the strip process mechanism knowledge in S3 to screen variables from the data set to be screened and obtain the model input data set includes: S31. Use the XGBoost algorithm to build a decision tree, use the number of times of feature splitting as the variable importance measurement index to score the features of the data set to be screened, and determine the variable set to be deleted according to the feature scores; S32. Combine the strip process mechanism knowledge to delete the variable set to be deleted to obtain the model input data set.

4. The method according to claim 1, wherein The construction and training process of the strip quality performance prediction model in S4 includes: S41. In the feature extraction stage of the strip quality performance prediction model, the training data after data preprocessing and variable screening is used as the input of the CNN feature extraction layer; S42. The LSTM module is used to extract the time series correlation of the output variables of the CNN. The hidden layer state of the previous moment of the LSTM and the input variables of the current moment are used to calculate the hidden layer state of the input features at the current moment, and the variables for the input attention mechanism module are obtained; S43. The Softmax layer is used to convert the scores in the attention mechanism module into weights, and the long-term prediction value of the strip quality performance is output through weighted summation; S44. An appropriate loss function is selected to compare the difference between the predicted value and the actual value of the model. The backpropagation algorithm is used to calculate the gradient of the loss function with respect to each weight. According to the calculated gradient, an optimization algorithm is used to update the weights in the network; S45. After multiple iterations of training, the strip quality performance prediction model gradually improves its internal weights and gradually approaches the optimal solution that can minimize the loss function.

5. The method according to claim 1, wherein The input of the online real-time data of the hot strip rolling process in S5 into the trained strip quality performance prediction model to obtain the real-time quality performance prediction result includes: S51. Obtain the online real-time data of the hot strip rolling process; S52. Perform preprocessing and variable screening operations on the real-time data set to obtain the real-time input data set for the model; S53. Use the real-time input data set as the input of the trained strip quality performance prediction model to predict the strip quality performance in real time.

6. A quality performance prediction device for the hot continuous rolling process of strip steel, characterized in that, The device includes: A data acquisition module for acquiring the process data of the hot strip rolling process; A preprocessing module for preprocessing the data set and using the preprocessed data as the data set to be screened; A variable screening module for combining the XGBoost algorithm and the strip process mechanism knowledge to screen the variables of the data set to be screened to obtain the model input data set; A model construction module for implementing the construction and training process of the strip quality performance prediction model. Among them, the strip quality performance prediction model is a CNN-LSTM model constructed by combining an attention mechanism module. The CNN-LSTM model includes a CNN feature extraction layer, an LSTM module, an attention mechanism module, a Softmax layer, and an output layer; The CNN feature extraction layer includes a convolutional layer, a max pooling layer, a fully connected layer, and a Dropout layer. The convolutional layer is used to extract feature maps from the input sequence while maintaining the position sensitivity and translational invariance of these features. After the convolutional operation, a ReLU activation function is applied to introduce non-linearity into the model. The max pooling layer is used to reduce the spatial size of the feature map and select the maximum value in each window as the output. The fully connected layer is used for dimensional transformation, and then a ReLU activation function is applied to introduce non-linearity into the model. The Dropout layer is used to prevent overfitting; The LSTM module consists of four parts: cell state, forget gate, input gate, and output gate, and can capture long-term dependencies in time series using hidden states; The attention mechanism module determines the weights of each part in the input sequence by calculating the similarity between the query and the key, and then performs weighted summation on the values according to these weights to generate a context vector reflecting the most important information in the input, thereby enhancing the model's attention to key information; The Softmax layer is used to convert the output of the previous layer into a probability distribution; The output layer is used to output the prediction result; The real-time prediction module is used to input the online real-time data of the hot strip continuous rolling process into the trained strip quality performance prediction model to obtain the real-time quality performance prediction result.

7. The device according to claim 6, characterized in that, The preprocessing module is further used for: S21. Centralize the data set to obtain the centralized data; S22. Normalize the centralized data, and the normalized data is used as the data set to be screened.

8. The device according to claim 6, characterized in that, The variable screening module is further used for: S31. Use the XGBoost algorithm to construct a decision tree, use the number of feature splits as a variable importance metric, score the features of the data set to be screened, and determine the set of variables to be deleted according to the feature scores; S32. Combine the preset strip mechanism knowledge to delete the set of variables to be deleted to obtain the model input data set.

9. The device according to claim 6, characterized in that, The model construction module is further used for: S41. In the feature extraction stage of the strip quality performance prediction model, use the training data after data preprocessing and variable screening as the input of the CNN feature extraction layer; S42. Extract the time series correlation of the output variables of the CNN through the LSTM module, and calculate the hidden layer state of the input feature at the current moment from the hidden layer state of the previous moment of the LSTM and the input variable at the current moment to obtain the variable input to the attention mechanism module; S43. Convert the scores in the attention mechanism module into weights through the Softmax layer, and output the long-time sequence prediction value of the strip quality performance by weighted summation; S44. Select an appropriate loss function to compare the difference between the predicted value and the actual value of the model. The backpropagation algorithm is used to calculate the gradient of the loss function with respect to each weight. According to the calculated gradient, an optimization algorithm is used to update the weights in the network; S45. After multiple iterative trainings, the strip quality performance prediction model gradually improves its internal weights and gradually approaches the optimal solution that can minimize the loss function.

10. The device according to claim 6, characterized in that, The real-time prediction module is further used for: S51. Obtain the online real-time data of the hot strip continuous rolling process; S52. Perform operations such as preprocessing and variable screening on the real-time data set to obtain the real-time input data set of the model; S53. Use the real-time input data set as the input of the trained prediction model to predict the strip quality performance in real time.

Images