Hot rolling and rough rolling head warping and buckling prediction and optimization method and system based on deep learning
Through deep learning-based methods, KA-CNN and Transformer modules are used to extract the characteristics of rolling process parameters, and combined with multi-agent optimization, the problem of difficulty in accurately predicting rolling process parameters with traditional methods is solved, and high-quality products and adaptability are achieved. Strong production conditions.
Patent Information
- Application Number
- CN202411134394.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-19
- Publication Date
- 2025-01-21
AI Technical Summary
Traditional methods are difficult to achieve accurate prediction of rolling process parameters, cannot guarantee the high quality of rolled products, and cannot adapt to rapidly changing production conditions.
A method based on deep learning is used to extract features of rolling process parameters through KA-CNN adaptive convolution module and Transformer module, and multi-agent is used to optimize rolling process parameters and deep learning module to predict product quality data.
It achieves accurate prediction of rolling process parameters, improves product quality, can adapt to rapidly changing production conditions, and improves the intelligence, efficiency and predictability of the system.
Smart Images
Figure CN119338037A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of hot rolling prediction and optimization technology, and in particular to a method and system for hot rolling rough rolling buckle head prediction and optimization based on deep learning. Background Art
[0002] With the rapid development of technology in the steel industry, especially the continuous advancement of artificial intelligence technology, higher requirements are placed on the control and optimization of the rolling process. As the core link in the metal rolling process, the problem of bending the head of the hot rolling rough rolling - the buckle phenomenon has become a technical problem that needs to be solved urgently in the rolling field. In the hot rolling production process, after the slab is rough rolled, the head often undergoes asymmetric deformation, resulting in the bending of the rolled piece, that is, upturning or buckling. This phenomenon directly affects the product quality data and limits production efficiency.
[0003] The traditional prediction methods currently used in production rely on the operator's experience and trial and error, which makes it difficult to achieve accurate prediction of rolling process parameters, fail to guarantee the high quality of rolled products, and fail to adapt to rapidly changing production conditions. Summary of the invention
[0004] In view of the above analysis, the embodiments of the present invention aim to provide a hot rolling and rough rolling buckle head prediction and optimization method and system based on deep learning, so as to solve the problems that traditional methods are difficult to achieve accurate prediction of rolling process parameters, cannot guarantee the high quality of rolled products, and cannot adapt to rapidly changing production conditions.
[0005] On the one hand, an embodiment of the present invention provides a method for predicting and optimizing the buckle head of hot rolling rough rolling based on deep learning, comprising:
[0006] Collecting and preprocessing the rolling process parameters of the process to be predicted to obtain the preprocessed rolling process parameters;
[0007] The pre-processed rolling process parameters are input into the pre-trained deep learning module to obtain the predicted product quality data;
[0008] The product quality data includes: product dimensional accuracy, whether buckle heads occur and the degree of buckle heads;
[0009] The deep learning module is used to sequentially use the KA-CNN adaptive convolution module and the Transformer module to extract features of the pre-processed rolling process parameters; and to predict product quality data Q using the extracted features.
[0010] Based on the further improvement of the above method, the KA-CNN adaptive convolution module includes:
[0011] The convolution layer with actively changing convolution kernel size is used to select the convolution kernel based on the size of the input data sequence to obtain the feature map matrix of the data sequence;
[0012] Activation function layer;
[0013] Max pooling layer;
[0014] M*M convolutional layer;
[0015] Attention pooling layer, including weight matrices that change as the module prediction progresses.
[0016] Based on the further improvement of the above method, the convolution layer with actively changing convolution kernel size includes a preset larger convolution kernel 1 and a smaller convolution kernel 2, and KNN;
[0017] Use KNN to classify the input data sequence (Pnorm, Qnorm), including:
[0018] Calculate the variance of all data in the data sequence, treat data with values greater than the variance as scattered data, and treat data with values less than or equal to the variance as clustered data;
[0019] If there are more discrete data, use a larger convolution kernel 1, otherwise use a smaller convolution kernel 2;
[0020] The selected convolution kernel is used to perform a convolution operation on the input data sequence to obtain the feature map matrix of the data sequence.
[0021] Based on the further improvement of the above method, the activation function layer includes a Swish activation function and an XTan activation function;
[0022] Calculate the loss value of the feature map matrix input Swish activation function. When the loss value is greater than the switching threshold, continue to use the Swish activation function in the training of the module. When the loss value is less than or equal to the switching threshold, input the data output by the Swish activation function into the XTan activation function, and use the XTan activation function in the training of the module. The XTan activation function is obtained using the following formula:
[0023] XTan(y)=sin(π·tanh(y))
[0024] Among them, y represents the data output by the Swish activation function.
[0025] Based on the further improvement of the above method, in the M*M convolution layer, the size of M is the sum of convolution kernel 1 and convolution kernel 2.
[0026] Based on the further improvement of the above method, the attention pooling layer includes:
[0027] A fully random matrix with a value range of 0-1 is used as the initial weight matrix corresponding to the data matrix, and the initial weight matrix and the data matrix corresponding to the weight matrix have the same size;
[0028] The data matrix and the weight matrix are calculated by dot product to obtain the updated weight matrix;
[0029] Use softmax to normalize the updated weight matrix to obtain the normalized weight matrix;
[0030] The normalized weight matrix and the data matrix are dot-producted, and a second normalization process is performed to obtain a normalized data matrix with weights.
[0031] Based on the further improvement of the above method, the collecting and preprocessing of the rolling process parameters of the process to be predicted includes:
[0032] Vector P = [T, N, F, G] represents the collected rolling process parameters in the hot rolling rough rolling process, where: T represents the rolling temperature, N represents the rolling mill motor speed, F represents the rolling force, and G represents the roll gap;
[0033] The collected P is preprocessed using the following formula:
[0034] Pnorm=(P-μP) / σP
[0035] Among them, Pnorm represents the normalized rolling process parameter vector, μP represents the mean of the rolling process parameter P, and σP represents the standard deviation of the rolling process parameter P.
[0036] A further improvement based on the above method also includes using multi-agents to optimize rolling process parameters and deep learning modules; the multi-agents include a strategy function Agent;
[0037] The strategy function Agent includes:
[0038] The proportional coefficient K for updating the historical rolling process parameter P is calculated using the following formula:
[0039]
[0040] Wherein, Q^ represents the desired product quality data, and Q represents the product quality data corresponding to the historical rolling process parameter P;
[0041] The updated rolling process parameter Pnew is calculated using the following formula:
[0042] Pnew=[K·T P ,-K·N P ,K.FP ,±K·G P ]
[0043] Among them, T P 、N P 、F P and G P Indicates the corresponding value of the historical rolling process parameter P, G P The weight coefficient is selected according to the deviation of slab thickness in product quality data ±K;
[0044] Use the trained deep learning model to calculate the prediction result Qnew of the product quality data of Pnew;
[0045] Determine whether Qnew meets expectations. If so, use Pnew. Otherwise, continue to update the rolling process parameters using the following formula:
[0046]
[0047] Pnew=[K′·T Pnew ,-K′·N Pnew ,K′·F Pnew ,±K′·G Pnew ]
[0048] Among them, T Pnew 、N Pnew 、F Pnew and G Pnew Indicates the corresponding value of the rolling process parameter Pnew before updating, G Pnew The weight coefficient is selected according to the deviation of slab thickness in product quality data ±K;
[0049] Determine whether the product quality data Qne obtained by actual production using Pnew meets expectations. If so, update the deep learning module; otherwise, continue to adjust and optimize the deep learning module.
[0050] Based on the further improvement of the above method, the optimization of rolling process parameters and deep learning modules by using multi-agents also includes online updating, and the online updating includes:
[0051] After collecting the rolling process parameters Pnew and the corresponding product quality data Qne that meet the production needs, Pnew and Qne are normalized and added to the existing training data set to complete the update of the training data set;
[0052] The deep learning module is trained using the updated training data set to obtain an online updated deep learning module.
[0053] On the other hand, an embodiment of the present invention provides a hot rolling rough rolling buckle head prediction and optimization system based on deep learning, including: a data acquisition module, a data preprocessing module, a deep learning module and a multi-agent module; wherein,
[0054] A data acquisition module, used to collect rolling process parameters of the process to be predicted;
[0055] A data preprocessing module is used to preprocess the collected rolling process parameters;
[0056] The deep learning module includes a KA-CNN adaptive convolution module and a Transformer module, which are used to extract features of the pre-processed process parameters; and use the extracted features to predict product quality data Q; wherein the product quality data Q includes: the dimensional accuracy of the product, whether warping occurs, and the degree of warping.
[0057] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:
[0058] 1. The present invention improves the feature extraction capability of rolling process parameters, enhances the generalization capability, realizes the comprehensive processing of large-scale and high-dimensional data, can provide accurate buckle head prediction results, and effectively overcomes the limitations of traditional methods;
[0059] 2. This solution combines multi-agents to optimize rolling process parameters and deep learning modules, improves systematization and predictability, solves the optimization problem of complex systems, and realizes intelligent, efficient and global optimization of rolling process parameters;
[0060] 3. The present invention has good scalability and flexibility, and can be adjusted and optimized according to actual needs to meet the requirements of different rolling processes.
[0061] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;
[0063] Figure 1 It is a flowchart of a method for predicting and optimizing buckle heads of hot-rolled rough-rolled slabs based on deep learning according to one embodiment of the present invention;
[0064] Figure 2 A block diagram of a hot-rolled rough-rolled slab buckle prediction and optimization system based on deep learning according to an embodiment of the present invention;
[0065] Figure 3 This is a structural diagram of a KA-CNN adaptive convolution module according to an embodiment of the present invention. DETAILED DESCRIPTION
[0066] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.
[0067] A specific embodiment of the present invention discloses a method for predicting and optimizing the buckle head of a hot-rolled rough-rolled slab based on deep learning, such as Figure 1 shown.
[0068] S11: Collect and preprocess the rolling process parameters of the process to be predicted to obtain the preprocessed rolling process parameters.
[0069] Specifically, the process to be predicted refers to a process in which the degree of buckle head of the product obtained by the process is unknown, and the rolling process parameters include rolling temperature, rolling mill motor speed, rolling force and roll gap;
[0070] Specifically, the collected vector P = [T, N, F, G] represents the collection of rolling process parameters during the hot rolling and rough rolling process, where: T represents the rolling temperature, N represents the rolling mill motor speed, F represents the rolling force, and G represents the roller gap.
[0071] Furthermore, the collected data is preprocessed, including normalizing the data using the following formula:
[0072] Pnorm=(P-μP) / σP
[0073] Among them, Pnorm represents the rolling process parameter vector after normalization, μP represents the mean of the rolling process parameter P, and σP represents the standard deviation of the rolling process parameter P. It should be noted that normalization helps to improve the training efficiency and prediction accuracy of the deep learning module.
[0074] S12: Input the preprocessed rolling process parameters into a pre-trained deep learning module to obtain predicted product quality data.
[0075] Product quality data includes the product's dimensional accuracy, whether the product has warping and the degree of warping.
[0076] Specifically, the deep learning module includes the independently designed KA-CNN adaptive convolution module and Transformer module in sequence; among them, the KA-CNN adaptive convolution module is used to extract the local features Flocal of the rolling process parameters, and the Transformer module is used to process long sequence data and capture the long-term dependency Flong, thereby improving the module's feature extraction capability for the rolling process parameters and enhancing the module's generalization capability.
[0077] Specifically, the KA-CNN adaptive convolution module includes:
[0078] The convolution layer with actively changing convolution kernel size is used to select the convolution kernel based on the size of the input data sequence to obtain the feature map matrix of the data sequence;
[0079] Activation function layer;
[0080] Max pooling layer;
[0081] M*M convolutional layer;
[0082] Attention pooling layer, including weight matrices that change as the module prediction progresses.
[0083] Among them, the convolution layer with active change of convolution kernel size includes: a preset larger convolution kernel 1 and a smaller convolution kernel 2, and KNN (K nearest neighbor, k-Nearest Neighbor); using KNN to classify the input data sequence (Pnorm, Qnorm), including: calculating the variance of all data in the data sequence, treating data with values greater than the variance as scattered data, and treating data with values less than or equal to the variance as aggregated data; if there are more discrete data, use a larger convolution kernel 1, otherwise use a smaller convolution kernel 2; realize active change of convolution kernel size, so as to better extract the features of the input data sequence; illustratively, the parameter K value in KNN is 4, and the Euclidean distance is used as the distance metric. The selected convolution kernel is used to perform convolution operation on the input data sequence to obtain a feature map matrix; it should be noted that the size of the convolution kernel in the traditional convolution network is usually fixed, and the size of the convolution kernel in this embodiment changes with the update of the module during the training process, which can better capture the features of the input data sequence at different scales.
[0084] Furthermore, the obtained feature map matrix is input into a custom activation function layer, including: Swish activation function and self-designed XTan activation function; wherein, the output of the Swish activation function is obtained using the following formula:
[0085] Swish(x)=x·σ(x)
[0086] Among them, x represents the input feature map matrix, σ(x) represents the sigmoid function;
[0087] The self-designed XTan activation function is obtained using the following formula:
[0088] XTan(y)=sin(π·tanh(y))
[0089] Among them, y represents the data output by the Swish activation function;
[0090] It should be noted that the XTan activation function uses the periodicity of the sine function to capture the periodic changes of long input time series; in addition, the non-monotonicity of the XTan activation function helps the module learn more complex function mappings; the tanh function inside the XTan activation function provides adaptive normalization of the input, so that the XTan function has different response strengths to inputs of different sizes;
[0091] Specifically, the Swish and XTan activation functions have different mathematical properties and are suitable for different types of data matrices and network structures. The initial activation function of this embodiment adopts Swish, and the activation degree is adaptively adjusted according to the input value. The activation degree is low for small input values and high for large input values; the loss value of the feature map matrix input Swish activation function is calculated, and when the loss value is greater than the switching threshold, the Swish activation function is continued to be used in the training of the module; when the loss value is less than or equal to the switching threshold, the data output by the Swish activation function is input into the XTan activation function, and the XTan activation function is used in the training of the module to achieve the switching of the activation function from faster to more accurate; exemplarily, the switching threshold is 0.2.
[0092] The data matrix output by the custom activation function layer is input into the maximum pooling layer for pooling operation, and then input into the M*M convolution layer to further extract the features of the data matrix; where the size of M is the sum of convolution kernel 1 and convolution kernel 2.
[0093] Furthermore, the data matrix output by the M*M convolutional layer is input into the attention pooling layer. The attention pooling layer includes a weight matrix that changes with the module prediction process, which is used to perform attention pooling on the data matrix to obtain a normalized data matrix with weights; including:
[0094] A fully random matrix with a value range of 0-1 is used as the initial weight matrix, and the initial weight matrix and the data matrix corresponding to the weight matrix are of the same size; the data matrix and the weight matrix are calculated by dot product to obtain an updated weight matrix; the updated weight matrix is normalized using softmax to obtain a normalized weight matrix, so that the numerical differences of each point in the weight matrix are further reduced; the normalized weight matrix and the data matrix are dot-producted to assign weights to each point in the data matrix, and a second normalization process is performed to obtain a normalized data matrix with weights. It is worth noting that the attention pooling layer includes a weight vector matrix that changes with the module prediction process, which can further extract the relationship between different features in the data set and improve the generalization and accuracy of the module.
[0095] It is worth noting that the KA-CNN adaptive convolution module uses KNN to adjust the size of the convolution kernel, which can more flexibly process features of different sizes, thereby improving the module's ability to recognize complex data structures. The attention mechanism enables the module to focus on key areas in the data matrix, especially when there are significant features, improving the expressiveness of the features.
[0096] Furthermore, the data matrix obtained after attention pooling is node-expanded to obtain one-dimensional long data Flocal, and then Flocal is input into the Transformer module for subsequent processing.
[0097] Among them, the Transformer module calculates the attention weight sequence A through the self-attention mechanism, performs dot product on the obtained attention weight sequence A and the corresponding position data of Flocal, and sums them to obtain Flong; wherein A represents the attention weight corresponding to different position data in Flocal.
[0098] Furthermore, the prediction result d of the product quality data is obtained by weighted summing of Flocal and Flong and calculating through a fully connected layer. For example, the weight coefficients of Flocal and Flong are both 0.5.
[0099] Specifically, the training data set of the deep learning module is obtained by the following method, including:
[0100] Collect historical rolling process parameters and corresponding product quality data, where: vector P = [T, N, F, G] represents the rolling process parameters in the hot rolling and rough rolling process, where: T represents the rolling temperature, N represents the rolling mill motor speed, F represents the rolling force, and G represents the roller gap; the corresponding product quality data vector Q = [W, S, E], where W represents the dimensional accuracy of the product, S represents whether the product has warping, and E represents the degree of warping. Using W, S and E, it can be obtained whether the slab has warping and the degree of warping.
[0101] Furthermore, the collected data is preprocessed, including normalizing the data using the following formula:
[0102] Pnorm=(P-μP) / σP
[0103] Qnorm=(Q-μQ) / σQ
[0104] Among them, Pnorm represents the normalized rolling process parameter vector, Qnorm represents the normalized product quality data vector, μP represents the mean of the rolling process parameter P, σP represents the standard deviation of the rolling process parameter P, μQ represents the mean of the product quality data Q, and σQ represents the standard deviation of the product quality data Q. It should be noted that normalization helps to improve the training efficiency and prediction accuracy of the deep learning module. After normalization, Pnorm and Qnorm are concatenated to form a long time series data (Pnorm, Qnorm) as the input for deep learning module training.
[0105] Furthermore, in order to ensure that the module can accurately predict product quality data, the mean square error is selected as the loss function and the module parameters are optimized through the back propagation algorithm and the gradient descent method. By updating the module parameters, the module's prediction results are closer and closer to the true value. Finally, the training is completed when the module's prediction accuracy meets the engineering requirements.
[0106] After the module training is completed, this embodiment uses multi-agent to optimize the rolling process parameters and deep learning modules; wherein, the multi-agent includes a strategy function Agent, which is used to: use the desired product quality data Q^ to estimate the update proportional coefficient K of the historical rolling process parameter P; use K and P to calculate the updated rolling process parameter Pnew; use the trained deep learning model to calculate the predicted result Qnew of the product quality data of Pnew; judge whether Qnew meets expectations, if so, use Pnew, otherwise continue to update the rolling process parameters; judge whether the product quality data Qne actually produced using Pnew meets expectations, if so, update the deep learning module, otherwise continue to adjust and optimize the deep learning module.
[0107] Specifically, for the current rolling process parameter P, the product quality data Q corresponding to P, and the desired product quality data Q^, the proportional coefficient K for P update is calculated using the following formula:
[0108]
[0109] Furthermore, the updated rolling process parameter Pnew is calculated using the following formula:
[0110] Pnew=[K·T P ,-K·N P ,K.F P ,±K·G P ]
[0111] Among them, T P 、N P 、F P and G P Indicates the corresponding value of the historical rolling process parameter P, G P Select ±K based on the deviation of slab thickness in product quality data;
[0112] It should be noted that the ideal state in industrial manufacturing is that the slab is flat, so the dimensional accuracy in Q^, namely the width, thickness and elongation, is the main evaluation index. The updated rolling process parameter Pnew is not just a theoretical value, it represents the direction and degree of adjustment in the actual production process. Next, the updated rolling process parameter Pnew is input into the trained deep learning module to calculate the corresponding prediction result Qnew of the product quality data. The difference between Qnew and Q^ is calculated. If the difference is less than the update threshold, it means that Pnew meets expectations; illustratively, the update threshold is 0.1Q^; if the difference is greater than or equal to the update threshold, K is continued to be updated using the following formula to obtain the updated K':
[0113]
[0114] Pnew=[K′·T Pnew ,-K′·N Pnew ,K′·F Pmew ,±K′·G Pnew ]
[0115] Among them, T Pnew 、N Pnew 、F Pnew and G Pnew Indicates the corresponding value of the rolling process parameter Pnew before updating, G Pnew The weight coefficient of is selected according to the deviation of the slab thickness in the product quality data ±K; this step is the key to verifying the optimization results of the multi-agent system and an important step to ensure that the module can continue to adapt to changes in the production environment. By continuously updating K' and Pnew, Qnew is close to Q^, and Pnew that meets production needs is obtained.
[0116] Next, the predicted result Qnew of the product quality data is compared with the actual product quality data Qne obtained by using the updated rolling process parameters Pnew to evaluate the prediction accuracy and optimization effect of the module; specifically, if the difference between Q^ and Qne is less than the update threshold, it is considered that Qne meets expectations, and the optimization strategy adopted by the deep learning module is successful and can continue to be used; if Q^ and Qne are greater than or equal to the update threshold, it means that the deep learning module needs further adjustment.
[0117] Through such cyclic iteration and optimization, we can continuously improve and optimize the rolling process parameters of hot rolling and rough rolling, so that product quality can be continuously improved and production efficiency can be improved. At the same time, the deep learning module is adjusted in a timely manner so that it can adapt to changes in the production environment and maintain the accuracy and effectiveness of the module.
[0118] In addition, the Agent also includes online updates, which include: after collecting the rolling process parameters Pnew and the corresponding product quality data Qne that meet the production needs, normalize Pnew and Qne and add them to the existing training data set to complete the update of the training data set; use the updated training data set to train the deep learning module to obtain the deep learning module after online update; then, use the updated training data set to train the deep learning module, and the training loss function is still the mean square error to obtain the deep learning module after online update.
[0119] like Figure 2 As shown, a specific embodiment of the present invention discloses a hot rolling rough rolling buckle head prediction and optimization system based on deep learning, including: a data acquisition module, a data preprocessing module, a deep learning module and a multi-agent module; wherein,
[0120] Data acquisition module 1, used for collecting rolling process parameters of the process to be predicted;
[0121] Data preprocessing module 2, used for preprocessing the collected rolling process parameters;
[0122] Deep learning module 3, including KA-CNN adaptive convolution module 31, see the structure diagram Figure 3 , and a Transformer module 32, which is used to extract features of the pre-processed process parameters; and, includes a prediction module 33, which is used to predict product quality data Q using the extracted features; wherein the product quality data Q includes: the dimensional accuracy of the product, whether the buckle head occurs and the degree of the buckle head.
[0123] It should be noted that the system and method of the present invention are based on the same inventive concept, and the specific contents can be referenced from each other and will not be described repeatedly here; including data acquisition process, data preprocessing process, etc.
[0124] Compared with the prior art, the solution provided in this embodiment improves the feature extraction capability of rolling process parameters, enhances the generalization capability, realizes the comprehensive processing of large-scale, high-dimensional data, can provide accurate buckle prediction results, and effectively overcomes the limitations of traditional methods; this solution combines multi-agents to optimize the rolling process parameters and deep learning modules, improves the systematicness and predictability, solves the optimization problems of complex systems, and realizes the intelligent, efficient and global optimization of rolling process parameters; this solution has good scalability and flexibility, and can be adjusted and optimized according to actual needs to meet the needs of different rolling processes.
[0125] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.
[0126] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.
Claims
1. A method for predicting and optimizing the buckle head of hot rolling rough rolling based on deep learning, characterized in that: include: Collecting and preprocessing the rolling process parameters of the process to be predicted to obtain the preprocessed rolling process parameters; The pre-processed rolling process parameters are input into the pre-trained deep learning module to obtain the predicted product quality data; The product quality data includes: product dimensional accuracy, whether buckle heads occur and the degree of buckle heads; The deep learning module is used to sequentially use the KA-CNN adaptive convolution module and the Transformer module to extract features of the pre-processed rolling process parameters; and to predict product quality data Q using the extracted features.
2. The method for predicting and optimizing the hot rolling and rough rolling buckle head based on deep learning according to claim 1, characterized in that: The KA-CNN adaptive convolution module includes: The convolution layer with actively changing convolution kernel size is used to select the convolution kernel based on the size of the input data sequence to obtain the feature map matrix of the data sequence; Activation function layer; Max pooling layer; M*M convolutional layer; Attention pooling layer, including weight matrices that change as the module prediction progresses.
3. The method for predicting and optimizing the hot rolling and rough rolling buckle head based on deep learning according to claim 2 is characterized in that: The convolution layer with actively changing convolution kernel size includes a preset larger convolution kernel 1 and a smaller convolution kernel 2, and KNN; Use KNN to classify the input data sequence (Pnorm, Qnorm), including: Calculate the variance of all data in the data sequence, treat data with values greater than the variance as scattered data, and treat data with values less than or equal to the variance as clustered data; If there are more discrete data, use a larger convolution kernel 1, otherwise use a smaller convolution kernel 2; The selected convolution kernel is used to perform a convolution operation on the input data sequence to obtain the feature map matrix of the data sequence.
4. The method for predicting and optimizing the hot rolling and rough rolling buckle head based on deep learning according to claim 2, characterized in that: The activation function layer includes a Swish activation function and an XTan activation function; Calculate the loss value of the feature map matrix input Swish activation function. When the loss value is greater than the switching threshold, continue to use the Swish activation function in the training of the module. When the loss value is less than or equal to the switching threshold, input the data output by the Swish activation function into the XTan activation function, and use the XTan activation function in the training of the module. The XTan activation function is obtained using the following formula: XTan(y)=sin(π·tanh(y)) Among them, y represents the data output by the Swish activation function.
5. The method for predicting and optimizing the hot rolling and rough rolling buckle head based on deep learning according to claim 2, characterized in that: In the M*M convolutional layer, the size of M is the sum of convolution kernel 1 and convolution kernel 2.
6. The method for predicting and optimizing the hot rolling and rough rolling buckle head based on deep learning according to claim 2, characterized in that: The attention pooling layer comprises: A fully random matrix with a value range of 0-1 is used as the initial weight matrix corresponding to the data matrix, and the initial weight matrix and the data matrix corresponding to the weight matrix have the same size; The data matrix and the weight matrix are calculated by dot product to obtain the updated weight matrix; Use softmax to normalize the updated weight matrix to obtain the normalized weight matrix; The normalized weight matrix and the data matrix are dot-producted, and a second normalization process is performed to obtain a normalized data matrix with weights.
7. The method for predicting and optimizing the hot rolling and rough rolling buckle head based on deep learning according to claim 1, characterized in that: The collecting and preprocessing of rolling process parameters of the process to be predicted includes: Vector P = [T, N, F, G] represents the collected rolling process parameters in the hot rolling rough rolling process, where: T represents the rolling temperature, N represents the rolling mill motor speed, F represents the rolling force, and G represents the roll gap; The collected P is preprocessed using the following formula: Pnorm=(P-μP) / σP Among them, Pnorm represents the normalized rolling process parameter vector, μP represents the mean of the rolling process parameter P, and σP represents the standard deviation of the rolling process parameter P.
8. The method for predicting and optimizing the hot rolling and rough rolling buckle head based on deep learning according to claim 1, characterized in that: It also includes the optimization of rolling process parameters and deep learning modules using multi-agents; The multi-agent includes a strategy function Agent; The strategy function Agent includes: The proportional coefficient K for updating the historical rolling process parameter P is calculated using the following formula: Wherein, Q^ represents the desired product quality data, and Q represents the product quality data corresponding to the historical rolling process parameter P; The updated rolling process parameter Pnew is calculated using the following formula: Pnew=[K·T P ,-K·Np,K·F P ,±K·G P ] Among them, T P 、N P 、F P and G P Indicates the corresponding value of the historical rolling process parameter P, G P The weight coefficient is selected according to the deviation of slab thickness in product quality data ±K; Use the trained deep learning model to calculate the prediction result Qnew of the product quality data of Pnew; Determine whether Qnew meets expectations. If so, use Pnew. Otherwise, continue to update the rolling process parameters using the following formula: Pnew=[K′·T Pnew ,-K′·N Pnew ,K′·F Pnew ,±K′·G Pnew ] Among them, T Pnew 、N Pnew 、F Pnew and G Pnew Indicates the corresponding value of the rolling process parameter Pnew before updating, G Pnew The weight coefficient is selected according to the deviation of slab thickness in product quality data ±K; Determine whether the product quality data Qne obtained by actual production using Pnew meets expectations. If so, update the deep learning module; otherwise, continue to adjust and optimize the deep learning module.
9. The method for predicting and optimizing the hot rolling and rough rolling buckle head based on deep learning according to claim 8, characterized in that: The optimization of rolling process parameters and deep learning modules by using multi-agents also includes online updating, and the online updating includes: After collecting the rolling process parameters Pnew and the corresponding product quality data Qne that meet the production needs, Pnew and Qne are normalized and added to the existing training data set to complete the update of the training data set; The deep learning module is trained using the updated training data set to obtain an online updated deep learning module.
10. A hot rolling rough rolling buckle head prediction and optimization system based on deep learning, characterized in that: include: Data acquisition module, data preprocessing module, deep learning module and multi-agent module; among them, A data acquisition module, used to collect rolling process parameters of the process to be predicted; A data preprocessing module is used to preprocess the collected rolling process parameters; A deep learning module includes a KA-CNN adaptive convolution module and a Transformer module, which are used to extract features of pre-processed process parameters; and a prediction module, which is used to predict product quality data Q using the extracted features; wherein the product quality data Q includes: the dimensional accuracy of the product, whether warping occurs, and the degree of warping.