LSTM-based rolling plate shape change trend prediction method

By using the CNN-Attention-LSTM model during the rolling process, processing time series data and capturing long-term dependencies, the problem that traditional methods cannot capture dynamic changes in plate shape is solved, and the efficiency and accuracy of plate shape prediction is achieved, reducing costs .

CN119047649APending Publication Date: 2024-11-29NORTHEASTERN UNIV CHINA
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411473615.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-22
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

Traditional plate shape prediction methods cannot effectively capture the dynamic changes and trends of plate shapes during rolling, resulting in inaccurate plate shape prediction.

Method used

The CNN-Attention-LSTM model based on LSTM is used, combined with 1D convolutional neural network and attention mechanism, time sequence data is processed, local features and long-term dependencies are captured, and plate-shaped values ​​are predicted.

Benefits of technology

It realizes efficient and accurate modeling of the time sequence data of the rolling process, improves the accuracy and computing speed of plate shape prediction, and reduces costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119047649A_ABST
    Figure CN119047649A_ABST
Patent Text Reader

Abstract

The invention discloses a rolling plate shape change trend prediction method based on LSTM, and relates to the technical field of strip steel rolling. Acquiring a continuous rolling production line industrial process time sequence data set; preprocessing data points in the continuous rolling production line industrial process time sequence data set; dividing into a training set and a test set according to a set proportion, and performing Min-Max normalization processing; determining an input step number; a CNN-Attention-LSTM (Long Short Term Memory) model is constructed; according to the determined number of input steps, training the constructed CNN-Attention-LSTM model by using a training set; and inputting the test set into the trained CNN-Attention-LSTM model, and obtaining a predicted plate shape value. The advantages of the 1D CNN in the aspect of local feature extraction are utilized, and the advantages that an attention mechanism distributes attention more flexibly and captures key features in a long-term dependency relationship and sequence more effectively are exerted, so that efficient and accurate modeling of time sequence data in the rolling process is realized. The operation speed is high, a large amount of production process data can be directly programmed on a computer to be put into use, and the cost is very low.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of strip steel rolling, and in particular to a method for predicting a changing trend of a rolled plate shape based on LSTM. Background Art

[0002] During the multi-stand continuous rolling process, the strip tension interacts with the strip thickness at the roll gap of each stand, forming a multi-input and multi-output system. The actual strip production process is complex, accompanied by fluctuations in various parameters, which can easily lead to plate shape defects. With the rapid development of industrial technology, the quality requirements for cold-rolled plate shape are increasing. High-quality plate shape is of great significance to improving product performance and reducing production costs.

[0003] From the perspective of control, the multivariable and nonlinear characteristics of the continuous rolling process are not conducive to the solution of traditional mathematical models, and machine learning is a powerful tool for solving such problems. At present, the data-driven plate shape regression prediction model has been widely used. In recent years, some scholars have proposed a plate shape prediction method based on neural networks. They use neural networks to extract features and perform regression analysis on the process parameters of the rolling process, and establish relevant prediction models to predict the deformation law of the plate. These methods have achieved remarkable results in the prediction process of rolling plate shape.

[0004] However, since the traditional regression prediction method assumes that the input variables are independent and have no time dependence, the model cannot capture the dynamic changes and trends of the plate shape over time. Therefore, studying and predicting the changing trend of the plate shape and realizing accurate prediction of the strip shape have become one of the current research hotspots. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a method for predicting the changing trend of rolled plate shape based on LSTM. The constructed CNN-Attention-LSTM (CAL) model can better handle the complexity of time series data in the rolling process and is more effective in capturing the local characteristics and long-term dependencies of time series data, thereby showing better performance in various plate shape prediction tasks.

[0006] A method for predicting the changing trend of rolling plate shape based on LSTM, comprising the following steps:

[0007] Step 1: Obtain the industrial process time series dataset of the continuous rolling production line;

[0008] The industrial process time series data set of the continuous rolling production line includes a plurality of data points, the data points are sampled at a set time interval and each data point corresponds to a moment, and each data point includes a plurality of features; the features include rolling process parameters and plate shape values ​​at different positions of the strip; the plate shape value is used to reflect the plate shape of the strip; the rolling process parameters include rolling force, roll gap difference, working roll bending force and intermediate roll bending force in the cold continuous rolling production process;

[0009] Step 2: preprocessing the data points in the continuous rolling production line industrial process time series data set to obtain a preprocessed continuous rolling production line industrial process time series data set; the preprocessing includes aligning the time series data, removing outliers, and filling in missing values;

[0010] Step 2.1: Perform time series data alignment operation on the plate shape values ​​of the data points;

[0011] Specifically: According to the rolling speed and the distance between the flatness detector and the rolling mill, the lag time of the flatness detector is calculated. t p , according to the sampling period of the data points T , move the flatness value in each data point forward t p / T time steps;

[0012] Step 2.2: Remove outliers from the data points after the time series data alignment operation;

[0013] Specifically, the mean and standard deviation of each feature in the time series data set of the industrial process of the continuous rolling production line are calculated, each data value of each feature in the time series data set of the industrial process of the continuous rolling production line is traversed, and the data value of the traversed feature is judged whether it meets the Pauta criterion according to the mean and standard deviation of each feature, and the data value of the feature that does not meet the Pauta criterion is marked as an outlier and deleted;

[0014] The Pauta criterion is:

[0015] In the formula, For the One data value for each feature, The industrial process time series data of continuous rolling production line is The average value of the features, The industrial process time series data of continuous rolling production line is The standard deviation of a feature.

[0016] Step 2.3: Filling the missing values ​​in the industrial process time series data set of the continuous rolling production line after removing the outliers; the missing value filling operation adopts a linear interpolation method;

[0017] Step 3: Divide the preprocessed continuous rolling production line industrial process time series data set into a training set and a test set according to a set ratio, and perform Min-Max normalization processing to obtain the processed training set and test set;

[0018] Step 4: Build a CNN-Attention-LSTM model;

[0019] The CNN-Attention-LSTM model includes a 1-dimensional convolutional neural network, an LSTM module, an attention module and a fully connected layer connected in sequence;

[0020] The 1-dimensional convolutional neural network is used to capture and extract local features in the input data;

[0021] The LSTM module includes a plurality of LSTM network layers connected in sequence, which are used to capture the long-term dependency in the input data from the local features output by the 1D convolutional neural network, learn the dynamic features in the input data, and obtain a high-level feature representation representing the input data; the number of LSTM network layer stacks is the number of hidden layers. num_layers_CAL , the number of hidden layer nodes in each LSTM network layer is hidden_size_cal;

[0022] The attention module uses the attention mechanism to calculate high-level feature representations, specifically by using softmax The function assigns weights to high-level feature representations and performs weighted summation to obtain the output of the attention module;

[0023] The fully connected layer receives the output result of the attention module and performs mapping to obtain the predicted board type value;

[0024] Step 5: Perform autocorrelation analysis on the flatness values ​​of the data points in the preprocessed continuous rolling production line industrial process time series data set to determine the number of input steps of the CNN-Attention-LSTM model;

[0025] Step 5.1: Calculate the autocorrelation coefficient of each flatness value at different positions of the strip;

[0026] In the formula, is the autocorrelation coefficient of the plate shape value, For time, is the length of the time series, i.e. the number of data points, for The flatness value at the moment, is the mean value of the plate shape, k is the lag order.

[0027] Step 5.2: Draw an autocorrelation diagram based on the autocorrelation coefficient of each flatness value at different positions of the strip, and determine the number of input steps based on the lag order in which the autocorrelation diagram falls into the confidence interval;

[0028] The confidence interval can be obtained as follows:

[0029] In the formula, represents the confidence interval.

[0030] Step 6: According to the determined number of input steps, the constructed CNN-Attention-LSTM model is trained using the training set to obtain a trained CNN-Attention-LSTM model;

[0031] Step 6.1: Under the determined number of input steps, set different learning rates and use the training set to train the constructed CNN-Attention-LSTM model. RMSE and MAE The smallest CNN-Attention-LSTM model;

[0032] Step 6.2: Optimize the number of hidden layers and the number of hidden layer nodes in the CNN-Attention-LSTM model to obtain the trained CNN-Attention-LSTM model.

[0033] Step 7: Input the test set into the trained CNN-Attention-LSTM model to obtain the predicted board type value;

[0034] The beneficial effects of the present invention are:

[0035] The present invention proposes a method for predicting the changing trend of rolling plate shape based on LSTM, which utilizes the advantages of 1D CNN in local feature extraction and the advantages of the attention mechanism in allocating attention more flexibly and capturing the key features in long-term dependencies and sequences more effectively, thereby realizing efficient and accurate modeling of rolling process time series data. The method has a fast operation speed and can be put into use by directly programming on a computer using a large amount of production process data, and the cost is very low. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A network structure diagram of a CNN-Attention-LSTM model (CAL) constructed in an embodiment of the present invention;

[0037] Figure 2 is an autocorrelation diagram of each plate-shaped channel in an embodiment of the present invention;

[0038] Figure 3 is a histogram showing the error comparison under different learning rates in an embodiment of the present invention;

[0039] Figure 4 It is a cloud diagram of error changes under different numbers of hidden layers and hidden layer nodes of different models in the embodiment of the present invention;

[0040] Figure 5 It is a scatter point fitting diagram of the predicted value and the actual value of the CAL model in the embodiment of the present invention;

[0041] Figure 6 This is a bar chart comparing evaluation indicators of different models in the embodiments of the present invention. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0043] Taking a five-stand cold rolling production line as an example, the network structure of the CNN-Attention-LSTM model (CAL) constructed by the present invention is as follows: Figure 1 As shown, a method for predicting the changing trend of rolling plate shape based on LSTM includes the following steps:

[0044] Step 1: Obtain the industrial process time series dataset of the continuous rolling production line;

[0045] The industrial process time series data set of the continuous rolling production line includes a number of data points, which are sampled at set time intervals and each data point corresponds to a moment, and each data point includes a number of features, which represent process parameters at a specific time point; the features include rolling process parameters and plate shape values ​​at different positions of the strip; the plate shape values ​​are used to reflect the plate shape of the strip; the rolling process parameters include rolling force, roll gap difference, working roll bending force and intermediate roll bending force in the cold rolling production process.

[0046] In this embodiment, the sampling time interval of the data points is 0.2 seconds, that is, 5 data points are collected per second, and the four rolling process parameters of the rolling force of the five stands, the roll gap difference of the five stands, the working roll bending force of the five stands and the intermediate roll bending force of the five stands are selected as the input features of the subsequent CNN-Attention-LSTM (CAL) model.

[0047] In this embodiment, the flatness values ​​of four key measurement points on the strip cross section are selected as the output values ​​of the subsequent CNN-Attention-LSTM (CAL) model. These measurement points are located at the left edge, left middle, right middle and right edge of the strip, respectively, and are required to directly reflect the general condition of the strip flatness.

[0048] After sorting the above features, several data points are obtained, each of which includes 8 features, where features 1-4 are input features and parameters 5-8 are output values, as shown in Table 1 below.

[0049] Table 1. Features displayed in data points serial number parameter unit 1 5-stand rolling force kN 2 5. Frame roller gap difference mm 3 5-stand working roll bending force kN 4 5. Bending force of the middle roll of the stand kN 5 Plate channel 2 IU 6 Plate channel 15 IU 7 Plate channel 16 IU 8 Plate channel 30 IU

[0050] Step 2: Preprocess the data points in the continuous rolling production line industrial process time series data set to obtain a preprocessed continuous rolling production line industrial process time series data set; the preprocessing includes time series data alignment, outlier removal and missing value filling, and the specific steps are as follows:

[0051] Step 2.1: Perform time series data alignment on the plate shape values ​​of the data points.

[0052] In order to avoid the misalignment of the flatness values ​​in the time series caused by the time difference, the time series data alignment operation of the flatness values ​​is first performed to obtain the industrial process time series data set of the continuous rolling production line with consistent input features and output flatness values ​​in the time series.

[0053] Specifically: According to the rolling speed and the distance between the flatness detector and the rolling mill, the lag time of the flatness detector is calculated. t p , according to the sampling period of the data points T , move the flatness value in each data point forward t p / T time steps;

[0054] In this embodiment, according to the rolling speed and the distance between the flatness detector and the rolling mill, the lag time of the flatness detector is calculated to be 1.4 seconds. According to the sampling period of 0.2 seconds, 7 flatness values ​​need to be manually deleted. Specifically, the flatness value in each data point is moved forward by 7 time steps to eliminate the time difference caused by the lag of the detector, so as to ensure the consistency of the input features and the output flatness values ​​in the time series;

[0055] Step 2.2: Remove outliers from the data points after the time series data alignment operation;

[0056] Specifically, the mean and standard deviation of each feature in the time series data set of the industrial process of the continuous rolling production line are calculated, each data value of each feature in the time series data set of the industrial process of the continuous rolling production line is traversed, and the data value of the traversed feature is judged whether it meets the Pauta criterion according to the mean and standard deviation of each feature, and the data value of the feature that does not meet the Pauta criterion is marked as an outlier and deleted;

[0057] The Pauta criterion is:

[0058] In the formula, For the One data value for each feature, The industrial process time series data of continuous rolling production line is The average value of the features, The industrial process time series data of continuous rolling production line is The standard deviation of a feature.

[0059] Step 2.3: Filling the missing values ​​in the industrial process time series data set of the continuous rolling production line after removing the outliers; the filling missing values ​​operation adopts a linear interpolation method.

[0060] In order to ensure the continuity of the time series and eliminate the missing data values ​​that may be caused by the removal of outliers, the linear interpolation method is used to interpolate the missing values.

[0061] Specifically, for a data point with missing data values, the linear interpolation method performs linear prediction based on its two adjacent non-missing data points to fill in the missing value.

[0062] In this example, missing values ​​are calculated according to the following formula:

[0063] In the formula, represents missing values, represents the time corresponding to the data point with missing data value, and Represents the values ​​of the two non-missing data points before and after the data point with missing data values. and represents the time instants corresponding to two non-missing data points.

[0064] In this embodiment, through the above preprocessing steps, 3741 data points are finally obtained for experiment, and some experimental data points are shown in Table 2.

[0065] Table 2 Display of some experimental data Serial number Rolling force / kN Roll gap difference / mm Working roll bending force / kN … Plate channel 2 / IU … 1 759.37 0.058012 86.006 -4.5368 2 757.34 0.058047 86.062 -4.5365 3 758.90 0.055828 85.069 -4.1765 4 758.87 0.054385 85.314 -3.9365 … … … … …

[0066] Step 3: Divide the preprocessed continuous rolling production line industrial process time series data set into a training set and a test set according to a set ratio, and perform Min-Max normalization processing to obtain the processed training set and test set.

[0067] In this example, in order to avoid data leakage, the above experimental data are divided into a training set and a test set, and then Min-Max normalization is performed on the above experimental data, and the data is linearly transformed so that the result value is mapped to the [0,1] interval.

[0068] Specifically, for the data value of each feature in the data point, the following formula is used in this embodiment to perform Min-Max normalization processing:

[0069] In the formula, For the One data value for each feature, for The data value after Min-Max normalization, is the minimum value of all data values ​​of this feature, It is the maximum value among all data values ​​of this feature.

[0070] Step 4: Based on the basic structure of LSTM, 1D CNN and attention mechanism are introduced to build the CNN-Attention-LSTM (CAL) model.

[0071] The CAL model includes a 1-dimensional convolutional neural network (1D CNN), an LSTM module, an attention module, and a fully connected layer, which are connected in sequence.

[0072] The 1D convolutional neural network (1D CNN) is used to perform convolution in the feature direction of the input to capture and extract local features in the input data.

[0073] The one-dimensional convolutional neural network includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer and a fully connected layer connected in sequence, wherein the activation function of all convolutional layers and the fully connected layers is ReLU.

[0074] The LSTM module includes a plurality of LSTM network layers connected in sequence, which are used to capture the long-term dependency in the input data from the local features output by the 1D convolutional neural network, learn the dynamic features in the input data, and obtain a high-level feature representation representing the input data; the number of LSTM network layer stacks is the number of hidden layers. num_layers_CAL , the number of hidden layer nodes in each LSTM network layer is hidden_size_cal.

[0075] The attention module uses the attention mechanism to calculate high-level feature representations, specifically by using softmax The function assigns weights to high-level feature representations and performs weighted summation to obtain the output of the attention module.

[0076] The fully connected layer receives the output result of the attention module for mapping to obtain the predicted board type value, thereby improving the model's ability to focus on key information in the time series, thereby improving the prediction performance.

[0077] Step 5: In order to determine whether the processed training set and test set obtained in step 3 meet the requirements of subsequent CNN-Attention-LSTM (CAL) model establishment and prediction, autocorrelation analysis is performed on the board shape values ​​of the four channels, that is, measuring the correlation of the board shape values ​​at different time points, which is helpful to determine the number of input steps.

[0078] Step 5.1: Calculate the autocorrelation coefficient of each flatness value at different positions of the strip.

[0079] In this example, the autocorrelation coefficient of each channel flatness value is calculated according to the following formula:

[0080] In the formula, is the autocorrelation coefficient of the plate shape value, and its value range is between -1 and 1. A value close to 1 indicates a strong positive correlation, a value close to -1 indicates a strong negative correlation, and a value close to 0 indicates no correlation. is time, since each data point corresponds to a moment, is the length of the time series, i.e. the number of data points. N is 3741, for The flatness value at the moment, is the mean value of the plate shape, k is the lag order.

[0081] Step 5.2: According to the autocorrelation coefficient of each flatness value at different positions of the strip, the autocorrelation diagrams of the four flatness channels are drawn, and the number of input steps is determined according to the lag order in which the autocorrelation diagrams of the four flatness channels fall into the confidence interval.

[0082] The confidence interval can be obtained as follows:

[0083] In the formula, represents the confidence interval.

[0084] In this embodiment, the autocorrelation diagrams of the four plate-shaped channels are drawn as follows: Figure 2As shown, plate-shaped channel 2, plate-shaped channel 15 and plate-shaped channel 30 fall into the confidence interval at a lag order of 150, and plate-shaped channel 16 also falls into the confidence interval at a lag order of about 200, which indicates that the experimental data has strong autocorrelation, so the subsequent CNN-Attention-LSTM (CAL) model can be used to predict it, and 150 steps are selected as the input steps of the subsequent CNN-Attention-LSTM (CAL) model.

[0085] Step 6: According to the determined number of input steps, the constructed CNN-Attention-LSTM (CAL) model is trained using the training set to obtain a trained CNN-Attention-LSTM (CAL) model.

[0086] Step 6.1: Under the determined number of input steps, set different learning rates and use the training set to train the constructed CNN-Attention-LSTM (CAL) model. RMSE and MAE The smallest CNN-Attention-LSTM (CAL) model.

[0087] The learning rates were set to 0.00001, 0.0001, 0.001, 0.005, 0.01, 0.05, and 0.1 for experiments. RMSE and MAE Minimum learning rate Lr。

[0088] In this embodiment, the RMSE and MAE values ​​of the model on the training set are calculated according to the following formulas.

[0089]

[0090]

[0091] In the formula, For the The predicted value of samples; For the The actual value of samples; L is the total number of samples in the training set.

[0092] In this embodiment, after testing, the errors under different learning rates are as follows: Figure 3 As shown in Figure 2, the overall performance of the model gradually increases with the increase of the learning rate, and the learning rate at which the optimal performance is achieved is 0.001. However, when the learning rate is further increased to 0.1, the model performance drops sharply. Therefore, the learning rate is set to Lr is 0.001.

[0093] Step 6.2: Tune the number of hidden layers and the number of hidden layer nodes in the CNN-Attention-LSTM (CAL) model to finally obtain the trained CNN-Attention-LSTM (CAL) model.

[0094] In this implementation, the number of hidden layers and the number of hidden layer nodes are taken according to the following formula and then combined in pairs to select the model that achieves the minimum RMSE and MAE The number of hidden layers when hid and the number of hidden layer nodes num.

[0095]

[0096] In the formula, hidden_size_CAL represents the number of hidden layer nodes, num_layers_CAL Represents the number of hidden layers.

[0097] In this example, after testing, the error under different numbers of hidden layers and hidden layer nodes of different models is as follows Figure 4 When the number of hidden layers is 1 and the number of hidden layer nodes is 112, the model achieves the best performance. RMSE and MAE They are 1.606 respectively IU and 1.229 IU Therefore, the number of hidden layers is set hidden_size_CAL The value is 1, the number of hidden layer nodes num_ layers_CAL The value is 112.

[0098] Step 7: Input the test set into the trained CNN-Attention-LSTM (CAL) model to obtain the predicted plate type value.

[0099] In this embodiment, the prediction effect of the CNN-Attention-LSTM (CAL) model is as follows: Figure 5 As shown in the figure, it can be seen that the error range between the predicted value and the actual value of the plate shape is ±2 IU , meeting the requirements of industrial production.

[0100] In particular, choose RMSE, MAE and R 2 As an evaluation index, the performance of the CNN-Attention-LSTM (CAL) model constructed in this embodiment is as follows: Figure 6 As shown in Figure 2, it can be seen that the CNN-Attention-LSTM (CAL) model outperforms other comparison models in all evaluation indicators, among which the CNN-Attention-LSTM (CAL) model RMSE The value is 1.598 IU , MAE The value is 1.226 IU , R 2 It is 0.950.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, it should be understood by those skilled in the art that the technical solutions described in the aforementioned embodiments can still be modified, or some or all of the technical features therein can be replaced by equivalents. Therefore, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A method for predicting the changing trend of rolling plate shape based on LSTM, characterized in that: The following steps are involved: Step 1: Obtain the industrial process time series dataset of the continuous rolling production line; Step 2: preprocessing the data points in the continuous rolling production line industrial process time series data set to obtain a preprocessed continuous rolling production line industrial process time series data set; the preprocessing includes aligning the time series data, removing outliers, and filling in missing values; Step 3: Divide the preprocessed continuous rolling production line industrial process time series data set into a training set and a test set according to a set ratio, and perform Min-Max normalization processing to obtain the processed training set and test set; Step 4: Build a CNN-Attention-LSTM model; Step 5: Perform autocorrelation analysis on the flatness values ​​of the data points in the preprocessed continuous rolling production line industrial process time series data set to determine the number of input steps of the CNN-Attention-LSTM model; Step 6: According to the determined number of input steps, the constructed CNN-Attention-LSTM model is trained using the training set to obtain a trained CNN-Attention-LSTM model; Step 7: Input the test set into the trained CNN-Attention-LSTM model to obtain the predicted board type value.

2. A method for predicting the changing trend of rolled plate shape based on LSTM according to claim 1, characterized in that: Step 1: The industrial process time series data set of the continuous rolling production line includes a number of data points, which are sampled at set time intervals and each data point corresponds to a moment, and each data point includes a number of features; the features include rolling process parameters and plate shape values ​​at different positions of the strip; the plate shape values ​​are used to reflect the plate shape of the strip; the rolling process parameters include rolling force, roll gap difference, working roll bending force and intermediate roll bending force in the cold rolling production process.

3. A method for predicting the changing trend of rolled plate shape based on LSTM according to claim 1, characterized in that: Step 2 specifically includes: Step 2.1: Perform time series data alignment on the plate shape values ​​of the data points; The lag time of the flatness detector is calculated based on the rolling speed and the distance between the flatness detector and the rolling mill. t p , according to the sampling period of the data points T , move the flatness value in each data point forward t p / T time steps; Step 2.2: Remove outliers from the data points after the time series data alignment operation; Calculate the mean and standard deviation of each feature in the time series data set of the industrial process of the continuous rolling production line, traverse each data value of each feature in the time series data set of the industrial process of the continuous rolling production line, and judge whether the data value of the traversed feature meets the Pauta criterion according to the mean and standard deviation of each feature, and mark the data value of the feature that does not meet the Pauta criterion as an outlier and delete it; Step 2.3: Filling the missing values ​​in the industrial process time series data set of the continuous rolling production line after removing the outliers; the filling missing values ​​operation adopts a linear interpolation method.

4. A method for predicting the changing trend of rolled plate shape based on LSTM according to claim 3, characterized in that: The Pauta criterion described in step 2.2 is: In the formula, For the One data value for each feature, The industrial process time series data of continuous rolling production line The average value of the features, The industrial process time series data of continuous rolling production line The standard deviation of a feature.

5. The method for predicting the changing trend of rolled plate shape based on LSTM according to claim 1, characterized in that: The CNN-Attention-LSTM model described in step 4 includes a 1D convolutional neural network, an LSTM module, an attention module, and a fully connected layer connected in sequence; The 1-dimensional convolutional neural network is used to capture and extract local features in the input data; The LSTM module includes a plurality of LSTM network layers connected in sequence, and is used to capture long-term dependencies in the input data from the local features output by the 1-dimensional convolutional neural network, learn the dynamic features in the input data, and obtain a high-level feature representation representing the input data; The number of LSTM network layers stacked is the number of hidden layers num_layers_CAL , the number of hidden layer nodes in each LSTM network layer is hidden_size_CAL ; The attention module uses the attention mechanism to calculate high-level feature representations, specifically by using softmax The function assigns weights to high-level feature representations and performs weighted summation to obtain the output of the attention module; The fully connected layer receives the output result of the attention module and maps it to obtain the predicted board type value.

6. The method for predicting the changing trend of rolled plate shape based on LSTM according to claim 1, characterized in that: Step 5 specifically includes: Step 5.1: Calculate the autocorrelation coefficient of each flatness value at different positions of the strip; Step 5.2: Draw an autocorrelation diagram based on the autocorrelation coefficient of each flatness value at different positions of the strip, and determine the number of input steps based on the lag order in which the autocorrelation diagram falls into the confidence interval.

7. The method for predicting the changing trend of rolled plate shape based on LSTM according to claim 6, characterized in that: The calculation method of the autocorrelation coefficient described in step 5.1 is: In the formula, is the autocorrelation coefficient of the plate shape value, For time, is the length of the time series, i.e. the number of data points, for The flatness value at the moment, is the mean value of the plate shape, k is the lag order.

8. The method for predicting the changing trend of rolled plate shape based on LSTM according to claim 7, characterized in that: The confidence interval described in step 5.2 can be obtained as follows: In the formula, represents the confidence interval.

9. The method for predicting the changing trend of rolled plate shape based on LSTM according to claim 5, characterized in that: Step 6 specifically includes: Step 6.1: Under the determined number of input steps, set different learning rates and use the training set to train the constructed CNN-Attention-LSTM model. RMSE and MAE The smallest CNN-Attention-LSTM model; Step 6.2: Adjust the number of hidden layers and the number of hidden layer nodes in the CNN-Attention-LSTM model to finally obtain the trained CNN-Attention-LSTM model.

Citation Information

Cited By

  • Cold-rolled sheet shape defect identification method based on multi-modal data

    CN120182724A