A method and system for predicting the width of a strip after continuous annealing and skin pass

Abstract

The present application relates to the technical field of continuous annealing and leveling width prediction, in particular to a continuous annealing and leveling after strip width prediction method and system, the method comprises the following steps: collecting continuous annealing and leveling site production data, identifying characteristic parameters that affect the strip width, and constructing a strip width prediction characteristic parameter dataset; preprocessing based on the strip width prediction characteristic parameter dataset, and dividing the preprocessing results to generate a training set and a test set; constructing a strip width prediction model based on a convolutional neural network and a bidirectional gated recurrent unit for predicting the strip width after continuous annealing and leveling; training the strip width prediction model based on the training set to generate a trained strip width prediction model. The present application uses data-driven machine learning technology to learn and capture complex nonlinear relationships from massive real-time data, thereby achieving more accurate strip width prediction.

Description

Technical Field

[0001] This invention relates to the field of continuous annealing and leveling width prediction technology, specifically to a method and system for predicting the width of strip steel after continuous annealing and leveling. Background Technology

[0002] As a key metal product widely used in manufacturing industries such as automobiles, home appliances, and aerospace, the width accuracy of strip steel is a core quality indicator that determines product yield, production efficiency, and production costs. Continuous annealing and leveling processes, as the final two critical steps in finished product production, directly affect the final width accuracy, thus impacting the economic benefits and market competitiveness of the entire production process. During continuous annealing, the combined effects of high temperature and tension significantly reduce the yield strength of the strip steel. When the tension exceeds the yield limit, the strip steel undergoes plastic deformation along its length, and according to the principle of constant volume, its transverse width narrows accordingly. In the leveling process, the synergistic effect of rolling force and tension induces transverse metal flow in the width direction, further leading to uncontrollable changes in width. However, existing prediction models based on traditional mechanistic analysis or single intelligent algorithms either fail to fully capture the complex characteristics of multi-factor coupling in the continuous annealing and leveling process or cannot adapt to the actual characteristics of scarce data and time-varying parameters in this process. This results in insufficient prediction accuracy and poor generalization ability, making it difficult to meet the urgent need for accurate strip steel width prediction in industrial production. Summary of the Invention

[0003] The purpose of this invention is to address the problems existing in the background technology by proposing a method and system for predicting the width of strip steel after continuous annealing and leveling.

[0004] The technical solution of this invention: A method for predicting the width of strip steel after continuous annealing and leveling, comprising: S1. Collect on-site production data of continuous annealing and leveling, identify the characteristic parameters that affect the strip width, and construct a dataset of characteristic parameters for strip width prediction. S2. Based on the strip width prediction feature parameter dataset, preprocess the data and divide the preprocessing results to generate training and test sets; S3. A strip width prediction model is constructed based on convolutional neural networks and bidirectional gated recurrent units for predicting the strip width after continuous annealing and leveling. S4. Train the strip width prediction model based on the training set to generate the trained strip width prediction model. S5. Input the test set into the trained strip width prediction model and output the strip width prediction result after continuous annealing and leveling.

[0005] As a further improvement to this technical solution, in step S1, the process of collecting on-site production data for continuous annealing and leveling, identifying characteristic parameters that affect the strip width, and constructing a dataset of predictive characteristic parameters for strip width includes the following steps: Collect production data from the continuous annealing and leveling site, including production data of strip steel from different batches and different steel grades; extract feature parameters that affect strip steel width based on the production data; summarize and organize each feature parameter according to the corresponding strip steel sample to generate a dataset of feature parameters for strip steel width prediction.

[0006] As a further improvement to this technical solution, in step S2, preprocessing is performed based on the strip width prediction feature parameter dataset, and the preprocessing results are divided to generate a training set and a test set, including the following steps: S2.1. Based on the strip width prediction feature parameter dataset, preprocessing is performed, and the feature parameters are screened using the Pearson correlation coefficient analysis method; S2.2 After completing feature filtering, the graph-based fence criterion is used to filter out outliers in the data to obtain the preprocessed feature parameter dataset. S2.3. Randomly partition the preprocessed dataset and generate training and test sets using stratified sampling.

[0007] As a further improvement to this technical solution, in step S2.2, the graph-based fence criterion is used to filter out outliers in the data, including the following steps: The first quartile, third quartile, and interquartile range (IQM) for each characteristic parameter are calculated separately. The IQM is the difference between the third quartile and the first quartile. The lower limit is the first quartile minus 1.5 times the IQM, and the upper limit is the third quartile plus 1.5 times the IQM. Each data point is evaluated item by item, and data with values ​​within the above upper and lower limits are retained as valid data for outlier screening.

[0008] As a further improvement to this technical solution, in step S3, a strip width prediction model is constructed based on a convolutional neural network and a bidirectional gated recurrent unit, including the following steps: S3.1 Receive the preprocessed feature parameter dataset as the input layer of the strip width prediction model, and transmit the input data to the convolutional neural network module; S3.2 Utilize two three-dimensional convolutional layers as convolutional neural network modules to extract local features from strip steel feature data; S3.3 Utilize a two-layer bidirectional gated loop unit structure as a bidirectional gated loop unit to extract the temporal characteristics in the strip steel production process; S3.4. The local features extracted by the convolutional neural network module and the temporal features extracted by the bidirectional gated recurrent unit are spliced ​​and fused using the fusion module. The fused features are then input into the fully connected layer for output. The mean square error is used as the loss function to output the strip width prediction value, thereby generating a strip width prediction model for predicting the strip width after continuous annealing and leveling.

[0009] As a further improvement to this technical solution, in step S4, the strip width prediction model is trained based on the training set to generate the trained strip width prediction model, including the following steps: S4.1. Based on the training set data, the Min-Max normalization method is used to process the data, and the processed two-dimensional data is converted into a 17×1×1 image format and then input into the strip width prediction model. S4.2 After receiving and processing the training set data, the strip width prediction model uses a convolutional neural network module to extract local features and a bidirectional gated recurrent unit to extract temporal features. The extracted local features and temporal features are then fused together, and the corresponding strip width prediction value is output through a fully connected layer. S4.3. Based on the prediction results and the actual strip width, calculate the loss function, and use the Adam gradient descent method to optimize the parameters of the strip width prediction model until the model training is completed and the trained strip width prediction model is generated.

[0010] As a further improvement to this technical solution, step S4.2 involves extracting local features using a convolutional neural network module, including the following steps: The first convolution operation is performed on the input 17×1×1 image format data to obtain output data with a dimension of 16×1×16, and the output data is non-linearly activated using a ReLU activation layer; the second convolution operation is performed using the same size convolution kernel, and the ReLU activation layer is used again for non-linear activation to obtain local feature data with a dimension of 15×1×32.

[0011] As a further improvement to this technical solution, step S4.2 involves extracting timing features using a bidirectional gated loop unit, including the following steps: The 15×1×32 local feature data output from the convolutional neural network module is flattened into a one-dimensional data with 480 columns and input into a forward gated recurrent unit. The gated recurrent unit is used to calculate the temporal dependencies of the input sequence, sequentially calculating the reset gate, update gate, candidate hidden state, and final hidden state to obtain the corresponding hidden state output. Based on the final hidden state, a 4-dimensional feature is output. At the same time, the input sequence is reversed and input into another gated recurrent unit with 4 hidden units for processing, outputting a 4-dimensional feature. The outputs of the forward and reverse gated recurrent units are concatenated to generate an 8-dimensional temporal feature.

[0012] As a further improvement to this technical solution, in step S5, the test set is input into the trained strip width prediction model, and the strip width prediction result after continuous annealing and leveling is output, including the following steps: The test set is input into the trained strip width prediction model. The convolutional neural network module is used to extract local features from the test set data, and the bidirectional gated recurrent unit is used to extract temporal features from the test set data. The extracted local features and temporal features are fused together, and the strip width prediction result after continuous unwinding and leveling is output through a fully connected layer. The predicted strip width is then compared and analyzed with the actual exit width to calculate the prediction error.

[0013] On the other hand, the present invention provides a strip width prediction system after continuous annealing and leveling, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the above-mentioned strip width prediction method after continuous annealing and leveling.

[0014] Compared with existing technologies, the above-mentioned technical solution of the present invention has the following beneficial technical effects: by utilizing data-driven machine learning technology, it can learn and capture complex nonlinear relationships from massive real-time data, thereby achieving more accurate strip width prediction. This high-precision prediction is of key significance for optimizing production efficiency and ensuring product quality, effectively overcoming the limitations of traditional models in complex and ever-changing production environments. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the overall method of the present invention; Figure 2 This is a structural diagram of the strip width prediction model built based on convolutional neural network (CNN) and bidirectional gated recurrent unit (BiGRU) of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1: Please refer to Figure 1-2 As shown, this embodiment provides a method for predicting the width of strip steel after continuous annealing and leveling, including the following steps: S1. Collect on-site production data of continuous annealing and leveling, identify the characteristic parameters that affect the strip width, and construct a dataset of characteristic parameters for strip width prediction. In this embodiment, the characteristic parameters that affect the strip width are identified, and a strip width prediction characteristic parameter dataset is constructed, including the following steps: Collect production data from the continuous annealing and leveling site, including production data for strip steel of different batches and grades. Based on the production data, extract characteristic parameters that affect strip width, including strip inlet width, furnace speed, heating section temperature, heating section tension, soaking section temperature, soaking section tension, slow cooling section temperature, slow cooling section tension, rapid cooling section temperature, rapid cooling section tension, aging section temperature, aging section tension, final cooling section temperature, final cooling section tension, leveling speed, leveling inlet thickness, leveling outlet thickness, tension before leveling, tension after leveling, leveling rolling force, and leveling extension. Summarize and organize each characteristic parameter according to the corresponding strip steel sample to generate a strip width prediction characteristic parameter dataset, providing input data for subsequent characteristic parameter data preprocessing.

[0018] S2. Based on the strip width prediction feature parameter dataset, preprocess the data and divide the preprocessing results to generate training and test sets; In this embodiment, the strip width prediction feature parameter dataset is preprocessed, and the preprocessing results are divided to generate training and testing sets, including the following steps: S2.1. Based on the strip width prediction feature parameter dataset, preprocessing is performed, and the feature parameters are screened using the Pearson correlation coefficient analysis method to reduce the feature dimension of the input data; S2.2 After completing feature selection, the Tukey's fences criterion is used to filter out outliers in the data, and data cleaning, missing value handling and Min-Max normalization are performed in sequence. For missing data, the median of the corresponding feature is used to fill in the missing data, and the normalized two-dimensional data is converted into a 17×1×1 image format to obtain the preprocessed feature parameter dataset. The Tukey's fences criterion is used to filter out outliers in the data, including the following steps: The first quartile Q1, third quartile Q3, and interquartile range (IQR) corresponding to each characteristic parameter are calculated separately. The IQR is the difference between the third quartile Q3 and the first quartile Q1. The lower limit is the first quartile minus 1.5 times the IQR, and the upper limit is the third quartile plus 1.5 times the IQR. Each data point is judged item by item. Data with values ​​within the above upper and lower limits are retained as valid data, and data exceeding the range are removed for outlier screening. S2.3. Based on the preprocessed dataset, randomly divide it into 80% as the training set and the remaining 20% ​​as the test set. Use stratified sampling to ensure that the distribution ratio of each category of data in the training set and the test set is basically the same, and generate the training set and the test set.

[0019] S3. A strip width prediction model is constructed based on convolutional neural networks (CNN) and bidirectional gated recurrent units (BiGRU) to predict the strip width after continuous annealing and leveling. (The convolutional neural network can extract local features from the multidimensional feature data in the strip production process and mine the correlation features between various process parameters; the bidirectional gated recurrent unit can make full use of the time sequence information before and after the production process to extract the time sequence change features in the strip production process; by fusing local features and time sequence features, the model can achieve comprehensive learning of the change law of various process parameters in the continuous annealing and leveling process, improve the fitting degree between the strip width prediction result and the actual exit width, and thus improve the prediction accuracy and prediction stability of the strip width after continuous annealing and leveling.) In this embodiment, a strip width prediction model is constructed based on a convolutional neural network (CNN) and a bidirectional gated recurrent unit (BiGRU). Figure 2 (As shown), including the following steps: S3.1 Receive the preprocessed feature parameter dataset as the input layer of the strip width prediction model. The input dimension is determined according to the number of features, and the input data is transmitted to the convolutional neural network (CNN) module. S3.2 Utilize two three-dimensional convolutional layers as convolutional neural network (CNN) modules. The kernel size is 2×1×1. After each convolution, the ReLU activation function is used to perform non-linear activation on the convolution result in order to extract local features from the strip steel feature data. S3.3. A two-layer bidirectional gated cyclic unit structure is used as a bidirectional gated cyclic unit (BiGRU) with 4 hidden layer nodes to extract the temporal features of the strip steel production process. S3.4. The local features extracted by the convolutional neural network (CNN) module and the temporal features extracted by the bidirectional gated recurrent unit (BiGRU) are spliced ​​and fused using the fusion module. The fused features are then input into the fully connected layer for output, and the mean square error is used as the loss function to output the strip width prediction value, thereby generating a strip width prediction model for predicting the strip width after continuous annealing and leveling.

[0020] S4. Train the strip width prediction model based on the training set to generate the trained strip width prediction model. In this embodiment, the strip width prediction model is trained based on the training set to generate the trained strip width prediction model, including the following steps: S4.1. Based on the training set data, the Min-Max normalization method is used to process the data, and the processed two-dimensional data is converted into a 17×1×1 image format and then input into the strip width prediction model. S4.2 After receiving and processing the training set data, the strip width prediction model uses a convolutional neural network module to extract local features and a bidirectional gated recurrent unit to extract temporal features. The extracted local features and temporal features are then fused together, and the corresponding strip width prediction value is output through a fully connected layer. The process of extracting local features using a convolutional neural network module includes the following steps: The first convolution operation is performed on the input 17×1×1 image data with a kernel size of 2×1×1, 16 output channels, and a stride of 1, resulting in output data with a dimension of 16×1×16. A ReLU activation layer is then used to perform non-linear activation on the output data. A second convolution operation is performed using the same kernel size, with the number of output channels set to 32. A ReLU activation layer is then used again for non-linear activation, resulting in local feature data with a dimension of 15×1×32. Furthermore, timing features are extracted using a bidirectional gated cyclic unit, including the following steps: The 15×1×32 local feature data output from the convolutional neural network module is flattened into a one-dimensional dataset containing 480 columns and input into a forward gated recurrent unit. The gated recurrent unit is used to calculate the temporal dependencies of the input sequence, sequentially calculating the reset gate, update gate, candidate hidden state, and final hidden state to obtain the corresponding hidden state output. Based on the final hidden state, a 4-dimensional feature is output. Simultaneously, the input sequence is reversed and input into another gated recurrent unit containing 4 hidden units for processing, outputting a 4-dimensional feature. The outputs of the forward and reverse gated recurrent units are concatenated to generate an 8-dimensional temporal feature. The formula for calculating the reset door is as follows: ; The formula for updating the gate is: ; The formula for calculating the candidate hidden state is: ; The final hidden state calculation formula is: ; In the formula, For the first The reset gate vector at each time step is used to control the degree to which the hidden state from the previous time step participates in the calculation of the current state. The Sigmoid activation function maps the calculation result to the range of 0 to 1. The weight matrix for input data to the reset gate, For the first The input vector at time t, This is the weight matrix from the previous hidden state to the reset gate. For the first The hidden state vector at time n, To reset the bias vector corresponding to the gate, For the first The update gate vector at each time step is used to control the ratio of information retained between the hidden state at the previous time step and the current state. To input data into the weight matrix of the update gate, This is the weight matrix from the hidden state of the previous time step to the update gate. To update the bias vector corresponding to the gate, For the first The candidate hidden state vector at time n. The hyperbolic tangent activation function maps the output to the range of -1 to 1. The weight matrix is ​​the input data to the candidate hidden state. For the first The reset gate vector at each time step. This is Hadamard's element-wise multiplication. This is the weight matrix from the hidden state in the previous time step to the candidate hidden state. This is the bias vector corresponding to the candidate hidden state; S4.3. Based on the prediction results and the actual strip width, calculate the loss function, and use the Adam gradient descent method to optimize the parameters of the strip width prediction model until the model training is completed and the trained strip width prediction model is generated.

[0021] S5. Input the test set into the trained strip width prediction model and output the strip width prediction result after continuous annealing and leveling. In this embodiment, the test set is input into the trained strip width prediction model, and the strip width prediction result after continuous annealing and leveling is output, including the following steps: The test set is input into the trained strip width prediction model. A convolutional neural network module is used to extract local features from the test set data, and a bidirectional gated recurrent unit is used to extract temporal features. The extracted local and temporal features are then fused, and the predicted strip width after continuous annealing and leveling is output through a fully connected layer. The predicted strip width is then compared with the actual exit width to calculate the prediction error. (The predicted strip width after continuous annealing and leveling is matched one-to-one with the corresponding actual exit width of the strip. The prediction error for each test sample is calculated, and the root mean square error between the prediction results and the actual exit width of all test samples is statistically analyzed. The degree of fit between the prediction results and the actual exit width is analyzed based on evaluation indicators, completing the comparative analysis of the predicted strip width and the actual exit width.) This verifies the accuracy of the trained strip width prediction model in predicting the strip width after continuous annealing and leveling.

[0022] This invention uses on-site production data from a continuous annealing leveling unit as an example: First, considering the comprehensive impact of high temperature during continuous annealing on the mechanical properties of the strip and the reduction effect of leveling rolling, 21 characteristic parameters were initially selected, including strip inlet width, furnace speed, heating section temperature, heating section tension, soaking section temperature, soaking section tension, slow cooling section temperature, slow cooling section tension, rapid cooling section temperature, rapid cooling section tension, aging section temperature, aging section tension, final cooling section temperature, final cooling section tension, leveling speed, leveling inlet thickness, leveling outlet thickness, pre-leveling tension, post-leveling tension, leveling rolling force, and leveling elongation. To reduce input dimensions, Pearson correlation coefficient analysis was used for feature selection, ultimately identifying 17 characteristic parameters that significantly affect strip width, such as inlet width, heating section temperature, soaking section temperature, heating section tension, soaking section tension, and leveling rolling force. The dataset of these characteristic parameters was then cleaned, missing values ​​were handled, and the data was standardized. For missing data points, the median of the feature was used for imputation. Min-Max normalization was applied to normalize the data, converting the normalized two-dimensional data into a 17×1×1 image format for easier processing by the convolutional neural network. The preprocessed dataset was then divided using stratified sampling, with 80% used as the training set and 20% as the test set. A random number seed was used to ensure uniform data distribution, guaranteeing that each category had approximately the same proportion in both the training and test sets. The training set data was used to train a strip width prediction model based on a convolutional neural network (CNN) and a bidirectional gated recurrent unit (BiGRU). Specifically, the input 17×1×1 image format data underwent a first convolution operation with a kernel size of 2×1×1, 16 output channels, and a stride of 1, resulting in output data of dimension 16×1×16. After non-linear activation using the ReLU activation function, a second convolution operation was performed with the same kernel, this time with 32 output channels, followed by another ReLU activation, resulting in local feature data of dimension 15×1×32. The aforementioned local feature data was flattened into a one-dimensional dataset containing 480 columns, and input into forward and reverse gated recurrent units (each containing 4 hidden units). The outputs of the two units were concatenated to generate an 8-dimensional temporal feature. After concatenating and fusing the local and temporal features, the predicted strip width was output through a fully connected layer. During training, the Adam gradient descent method was used for parameter optimization, with mean squared error as the loss function. The test set was input into the trained strip width prediction model to obtain the predicted strip width for each sample. The predicted width was compared with the actual exit width to calculate the prediction error and verify the model's accuracy in predicting the strip width after continuous annealing and leveling. In the first batch of test samples, data covering multiple steel grades were selected for verification. The results showed that the prediction error ranged from -1.7 mm to 2.4 mm, with the prediction deviation of most samples within ±1.5 mm. The prediction accuracy met the actual requirements of industrial production for strip width control.For example, the actual width of the 1180D steel sample at the outlet was 1208.9 mm, and the predicted width was 1207.5 mm, with an error of only 1.4 mm; the actual width of the 1470D steel sample was 1222.4 mm, and the predicted width was 1224.1 mm, with an error of -1.7 mm; another 980DP steel sample had an actual width of 1335.1 mm and a predicted width of 1333.6 mm, with an error of 1.5 mm. These results verify the feasibility and effectiveness of the method proposed in this invention for predicting the width of strip steel after continuous annealing and leveling. To further verify the model's generalization ability, strip steel data from different production batches were selected for prediction verification. In the second batch of test samples, multiple steel grades were also covered, and the prediction error range was further reduced to between -1.2mm and 1.1mm. For example, the actual width of the 1180D steel sample was 1278.2mm, and the predicted width was 1278.4mm, with an error of only -0.2mm; the actual width of the 590DP steel sample was 1278.4mm, and the predicted width was 1277.3mm, with an error of 1.1mm; the actual width of the DQ-IF steel sample was 1276.0mm, and the predicted width was 1277.2mm, with an error of -1.2mm. Both batches of data from different sources achieved good prediction results, indicating that the strip steel width prediction model constructed in this invention has good generalization ability and prediction stability, can adapt to the strip steel width prediction needs under different production conditions, and has high engineering practical value.

[0023] Example 2: This example provides a strip width prediction system after continuous annealing and leveling, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the strip width prediction method after continuous annealing and leveling described in Example 1 above.

[0024] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for predicting the width of strip steel after continuous annealing and leveling, characterized in that, include: S1. Collect on-site production data of continuous annealing and leveling, identify the characteristic parameters that affect the strip width, and construct a dataset of characteristic parameters for strip width prediction. S2. Based on the strip width prediction feature parameter dataset, preprocess the data and divide the preprocessing results to generate training and test sets; S3. A strip width prediction model is constructed based on convolutional neural networks and bidirectional gated recurrent units for predicting the strip width after continuous annealing and leveling. S4. Train the strip width prediction model based on the training set to generate the trained strip width prediction model. S5. Input the test set into the trained strip width prediction model and output the strip width prediction result after continuous annealing and leveling.

2. The method for predicting the width of strip steel after continuous annealing and leveling according to claim 1, characterized in that, In step S1, the following steps are included: collecting on-site production data for continuous annealing and leveling, identifying characteristic parameters that affect strip width, and constructing a dataset of predictive characteristic parameters for strip width. Collect production data from the continuous annealing and leveling site, including production data of strip steel from different batches and different steel grades; extract feature parameters that affect strip steel width based on the production data; summarize and organize each feature parameter according to the corresponding strip steel sample to generate a dataset of feature parameters for strip steel width prediction.

3. The method for predicting the width of strip steel after continuous annealing and leveling according to claim 1, characterized in that, In step S2, the strip width prediction feature parameter dataset is preprocessed, and the preprocessing results are divided to generate a training set and a test set, including the following steps: S2.

1. Based on the strip width prediction feature parameter dataset, preprocessing is performed, and the feature parameters are screened using the Pearson correlation coefficient analysis method; S2.2 After completing feature filtering, the graph-based fence criterion is used to filter out outliers in the data to obtain the preprocessed feature parameter dataset. S2.

3. Randomly partition the preprocessed dataset and generate training and test sets using stratified sampling.

4. The method for predicting the width of strip steel after continuous annealing and leveling according to claim 3, characterized in that, In step S2.2, the graph-based fence criterion is used to filter out outliers in the data, including the following steps: The first quartile, third quartile, and interquartile range (IQM) for each characteristic parameter are calculated separately. The IQM is the difference between the third quartile and the first quartile. The lower limit is the first quartile minus 1.5 times the IQM, and the upper limit is the third quartile plus 1.5 times the IQM. Each data point is evaluated item by item, and data with values ​​within the above upper and lower limits are retained as valid data for outlier screening.

5. The method for predicting the width of strip steel after continuous annealing and leveling according to claim 1, characterized in that, In step S3, a strip width prediction model is constructed based on a convolutional neural network and a bidirectional gated recurrent unit, including the following steps: S3.1 Receive the preprocessed feature parameter dataset as the input layer of the strip width prediction model, and transmit the input data to the convolutional neural network module; S3.2 Utilize two three-dimensional convolutional layers as convolutional neural network modules to extract local features from strip steel feature data; S3.3 Utilize a two-layer bidirectional gated loop unit structure as a bidirectional gated loop unit to extract the temporal characteristics in the strip steel production process; S3.

4. The local features extracted by the convolutional neural network module and the temporal features extracted by the bidirectional gated recurrent unit are spliced ​​and fused using the fusion module. The fused features are then input into the fully connected layer for output. The mean square error is used as the loss function to output the strip width prediction value, thereby generating a strip width prediction model for predicting the strip width after continuous annealing and leveling.

6. The method for predicting the width of strip steel after continuous annealing and leveling according to claim 1, characterized in that, In step S4, the strip width prediction model is trained based on the training set to generate the trained strip width prediction model, including the following steps: S4.

1. Based on the training set data, the Min-Max normalization method is used to process the data, and the processed two-dimensional data is converted into a 17×1×1 image format and then input into the strip width prediction model. S4.2 After receiving and processing the training set data, the strip width prediction model uses a convolutional neural network module to extract local features and a bidirectional gated recurrent unit to extract temporal features. The extracted local features and temporal features are then fused together, and the corresponding strip width prediction value is output through a fully connected layer. S4.

3. Based on the prediction results and the actual strip width, calculate the loss function, and use the Adam gradient descent method to optimize the parameters of the strip width prediction model until the model training is completed and the trained strip width prediction model is generated.

7. The method for predicting the width of strip steel after continuous annealing and leveling according to claim 6, characterized in that, In step S4.2, local features are extracted using a convolutional neural network module, including the following steps: The first convolution operation is performed on the input 17×1×1 image format data to obtain output data with a dimension of 16×1×16, and the output data is non-linearly activated using a ReLU activation layer; the second convolution operation is performed using the same size convolution kernel, and the ReLU activation layer is used again for non-linear activation to obtain local feature data with a dimension of 15×1×32.

8. The method for predicting the width of strip steel after continuous annealing and leveling according to claim 6, characterized in that, In step S4.2, the timing features are extracted using a bidirectional gated loop unit, including the following steps: The 15×1×32 local feature data output from the convolutional neural network module is flattened into a one-dimensional data with 480 columns and input into a forward gated recurrent unit. The gated recurrent unit is used to calculate the temporal dependencies of the input sequence, sequentially calculating the reset gate, update gate, candidate hidden state, and final hidden state to obtain the corresponding hidden state output. Based on the final hidden state, a 4-dimensional feature is output. At the same time, the input sequence is reversed and input into another gated recurrent unit with 4 hidden units for processing, outputting a 4-dimensional feature. The outputs of the forward and reverse gated recurrent units are concatenated to generate an 8-dimensional temporal feature.

9. The method for predicting the width of strip steel after continuous annealing and leveling according to claim 1, characterized in that, In step S5, the test set is input into the trained strip width prediction model, and the strip width prediction result after continuous annealing and leveling is output, including the following steps: The test set is input into the trained strip width prediction model. The convolutional neural network module is used to extract local features from the test set data, and the bidirectional gated recurrent unit is used to extract temporal features from the test set data. The extracted local features and temporal features are fused together, and the strip width prediction result after continuous unwinding and leveling is output through a fully connected layer. The predicted strip width is then compared and analyzed with the actual exit width to calculate the prediction error.

10. A strip width prediction system after continuous annealing and leveling, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes a computer program to implement the strip width prediction method after continuous annealing and leveling as described in any one of claims 1-9.

Images