Multi-index prediction method for quality of continuously rejected products based on unbalanced index preference data

Abstract

The present invention provides a multi-index prediction method for the quality of strip products oriented to unbalanced index preference data, and relates to the technical field of data processing in the cold rolling production control process of steel enterprises. First, by defining the preference relationship of data samples for different performance indicators, this preference relationship is quantitatively represented, and then integrated into a multi-task deep learning prediction model for feature calculation and preference processing, while using an evolutionary algorithm to realize the autonomous optimization construction of a multi-task deep learning model, thereby improving the performance of the multi-task prediction model. The present invention proposes a quantitative calculation method for sample-index preference, and applies it to the construction of a multi-task deep learning model, thereby improving the accuracy of the multi-index online prediction model for the quality and performance of strip steel, thereby helping the cold rolling production process to improve the quality and stability of strip steel products, and improving the intelligent level of quality control.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing in a cold rolling production control process of a steel enterprise, and in particular to a multi-index prediction method for quality of a continuous rolling product facing index preference imbalance data. Background Art

[0002] Continuous annealing is an important process in the cold rolling process of the steel industry. It can improve the physical properties and processing performance of cold-rolled strip steel, making it more suitable for various industrial applications. In the process of continuous annealing, since it is impossible to realize online detection of quality indicators, accurate prediction of multiple mechanical performance indicators can effectively reduce product quality fluctuations in the production process and improve the product quality control level of steel enterprises. At present, most machine learning-based continuous annealing strip performance prediction models can usually only predict a single performance indicator. For example, the invention patent "A method for online detection of continuous annealing strip quality based on hybrid ensemble learning" with patent number ZL201710159565.8 and the invention patent "A method for online ensemble learning prediction of hardness of continuous annealing products" with patent number ZL201410843307.8 cannot use a model to predict multiple performance indicators at the same time. In terms of multi-indicator prediction, since the data samples in actual production generally have different prediction capabilities for different performance indicators, the machine learning model established based on the data set can accurately predict some quality indicators, but cannot accurately predict other indicators (called indicator preference imbalance). Different from the traditional unbalanced classification problem, the unbalanced characteristic of this indicator preference makes it difficult for the existing multi-indicator regression modeling methods to achieve good results in practical applications. Therefore, the current multi-indicator prediction methods (such as the invention patent with application number CN115392104A "A method for predicting the mechanical properties of cold-rolled continuous annealing strip based on annealing process") usually do not consider the imbalance problem of data samples for different indicator preferences. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a multi-indicator prediction method for the quality of continuously rejected products with unbalanced indicator preference data in view of the deficiencies in the above-mentioned prior art. The method first defines the preference relationship of data samples for different performance indicators, quantifies this preference relationship, and then integrates it into a multi-task deep learning prediction model for feature calculation and preference processing. At the same time, an evolutionary algorithm is used to realize the autonomous optimization construction of the multi-task deep learning model to improve the performance of the multi-task prediction model.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0005] A method for predicting the quality of continuously-degraded products with multiple indicators for indicator preference imbalance data includes the following steps:

[0006] Step 1: Build a sample set;

[0007] Collect n production process data samples of steel strips under normal production conditions from the historical data of the cold rolling annealing unit to establish a sample set S; each sample contains v features, corresponding to the v production process data of the steel strip, and performs m outputs, corresponding to m quality performance indicators of the steel strip;

[0008] Step 2: Data preprocessing;

[0009] Normalizing n sample data in the established sample data set S;

[0010] Step 3: Quantitative evaluation of sample-indicator preference;

[0011] According to the n sample data obtained in step 2, construct the sample-indicator preference matrix P;

[0012] Step 4: Dataset division;

[0013] The n sample data preprocessed in step 2 are randomly divided into 0.8n and 0.2n samples, respectively, to form the training set D train and the test set D test ; If 0.8n and 0.2n are not integers, they are rounded down;

[0014] Step 5: Build a hybrid deep learning network model CNN-Transformer network model;

[0015] Step 6: Obtain the production process data of each strip in real time through the OPC interface, and use the constructed CNN-Transformer network model to make online predictions on the three key performance indicators of the strip: yield strength, tensile strength and elongation. When the deviation between the predicted value of a performance indicator and the set target value exceeds a given threshold, a timely alarm is issued to ensure the stability of the quality of the continuous annealing strip products.

[0016] Furthermore, the specific method of step 3 is:

[0017] Step 3.1: Data clustering;

[0018] The n sample data obtained in step 2 are clustered using the density clustering algorithm DBSCAN to obtain c clusters;

[0019] Step 3.2: Establish a single indicator regression prediction model;

[0020] In each cluster, t different machine learning methods are used to establish t different single-indicator regression prediction models for each quality performance indicator;

[0021] Step 3.3: Calculate sample-indicator preference;

[0022] For each cluster, the t single index regression prediction models established for each quality performance index in step 3.2 are respectively K Fold cross validation is performed to obtain the respective mean square error MSE, and then the mean square error is averaged to construct the sample-indicator preference matrix ,in, , p ij Represents the preference of all samples in the i-th cluster for the j-th performance indicator, i=1,2,…,c; j=1,2,…,m.

[0023] Furthermore, in step 3.2, each single indicator regression prediction model is constructed using the machine learning library Scikit-learn.

[0024] Furthermore, the specific method of step 5 is:

[0025] Step 5.1: Use the training set D train , train a m A multi-task deep convolutional neural network (CNN) with 100 outputs, each of which is a quality performance indicator;

[0026] Step 5.2: A Transformer network is superimposed on the back of the CNN model obtained by training to construct a hybrid deep learning network model CNN-Transformer, in which a new representation layer is added to the encoder unit Encoder-block of the Transformer network to process the preference matrix P obtained in step 3;

[0027] Step 5.3: Use the training set D train The CNN-Transformer network was trained to obtain a multi-index prediction model for the quality of continuously rejected products that combines sample-index preference features.

[0028] Step 5.4: Use the test set D test Test and evaluate the CNN-Transformer network.

[0029] Furthermore, in step 5.1, the multi-task deep convolutional neural network CNN includes 3 convolutional layers, each convolutional layer uses 128 output channels and two sets of convolution kernels of different sizes [[3,5], [5,7]], and the convolutional layer introduces nonlinear transformation through the ReLU activation function.

[0030] Furthermore, the specific method of step 5.2 is:

[0031] Step 5.2.1: Transformer module design;

[0032] The output dimension of the multi-task deep convolutional neural network CNN is used as the input of the Transformer network, and a new representation layer is added to the encoder unit Encoder-block of the Transformer network to process the preference matrix P obtained in step 3;

[0033] Step 5.2.2: Fully connected layer design; the network includes 3 fully connected layers, each of which contains 256 hidden units, and there is a ReLU activation function behind each hidden unit;

[0034] Step 5.2.3: Set weight decay; the weights of the convolutional layer and the linear layer are initialized using the Xavier initialization method. During training, the entire network continuously optimizes parameters to minimize the loss function through back propagation and gradient descent. The learning rate is set to 0.001 and the weight decay is set to 0.001.

[0035] Furthermore, the step 5.2.1 specifically includes:

[0036] Step 5.2.1.1: Embedding of preference matrix;

[0037] Convert the sample-indicator preference matrix P into an embedding vector compatible with Transformer;

[0038] Step 5.2.1.2: Modify input;

[0039] Concatenate the preference matrix P with the Transformer input to obtain a new matrix representation P mix ;

[0040] Step 5.2.1.3: Modify the encoder structure;

[0041] The modified matrix P mix Added between adjacent layers of the Transformer encoder unit.

[0042] The beneficial effects of adopting the above technical scheme are as follows: the multi-indicator prediction method for the quality of annealed products for data with unbalanced indicator preferences provided by the present invention proposes a quantitative calculation method for sample-indicator preference, and applies it to the construction of a multi-task deep learning model, thereby improving the accuracy of the multi-indicator online prediction model for the quality performance of annealed strip steel, thereby helping the cold rolling production process to improve the quality and stability of strip steel products and enhance the intelligent level of quality control. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A data clustering distribution diagram provided by an embodiment of the present invention;

[0044] Figure 2 A data preference layer network design diagram of a Transformer module provided in an embodiment of the present invention;

[0045] Figure 3 A flow chart of embedding a sample-indicator preference matrix into an encoder unit according to an embodiment of the present invention;

[0046] Figure 4 A flow chart of a multi-index prediction model for the quality of continuously rejected products based on unbalanced index preference data provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0048] For the cold rolling annealing production process, this embodiment proposes a multi-index prediction method for annealing product quality based on indicator preference imbalance data, which is described in detail as follows.

[0049] Step 1: Build a sample set.

[0050] 15201 production process data samples of strip steel under normal production conditions were collected from the historical data of the cold rolling annealing unit to establish a sample set S. Each sample contains 56 features, corresponding to 56 production process data of the strip steel (including strip steel specifications, temperature setting values ​​of the annealing process, etc.), and 3 outputs, corresponding to the 3 main quality performance indicators of the strip steel (yield strength, tensile strength, and elongation).

[0051] The collection item information of the continuous retreat production process data in this embodiment is shown in Table 1.

[0052] Table 1 Collection item information of continuous retreat production process data

[0053] Collection item name unit Thickness of continuous withdrawal plan mm Width of continuous withdrawal plan mm Elongation % Upper surface roughness μm Lower surface roughness μm Furnace speed m / min … … Heating temperature ℃ Finishing rolling temperature ℃ Coiling temperature ℃ R2 outlet actual temperature ℃ C (carbon content) % Mn (Manganese content) % S(sulfur content) % P (Phosphorus content) %

[0054] Step 2: Data preprocessing: The 15201 sample data in the established sample data set S are normalized, and the data are scaled to a range to ensure the regularity of the data.

[0055] Step 3: Quantitative evaluation of sample-indicator preference. Based on the 15,201 sample data obtained in step 2, a sample-indicator preference matrix is ​​constructed, which includes the following steps:

[0056] Step 3.1: Data clustering. The 15201 sample data obtained in step 2 are clustered using the density clustering algorithm DBSCAN to obtain c = 17 clusters, as shown in Figure 1 shown.

[0057] Step 3.2: Establish a single indicator regression model.

[0058] In each cluster, machine learning methods were used to establish t = 8 different single-index regression models for each mechanical performance index, including linear regression LR ( Linear Regression, LR ), Ridge Regression RR ( Ridge Regression, RR ), Elastic Net Regression EN ( Elastic Net Regression, EN ), Decision Tree Regression DTR ( Decision Tree Regression, DTR ), Random Forest Regression RFR ( Random Forest Regression, RFR ), Gradient Boosting Regression GBR ( Gradient Boosting Regression, GBR ), Support Vector Regression SVR ( Support Vector Regression, SVR ), Lasso return( Least Absolute Shrinkage and Selection Operator Regression , Lasso ). Each single indicator regression model is constructed using the machine learning library.

[0059] Step 3.3: Calculate sample-index preference.

[0060] For each cluster, the t=8 models established for each quality performance indicator in the previous step are cross-validated with 5-fold to obtain their respective mean squared errors (MSE), and then the mean squared errors are averaged to construct the sample-indicator preference matrix. ,in, , p ij Represents the preference of all samples in the i-th cluster for the j-th performance indicator, i=1,2,…,c; j=1,2,…,m.

[0061] Step 4: Dataset division. The n = 15201 samples preprocessed in step 2 are randomly divided into 0.8n = 12160 and 0.2n = 3040 samples, forming the training set D train and the test set D test .

[0062] Step 5: Build a hybrid deep learning network model, including the following steps:

[0063] Step 5.1: Use the training set D train , train a multi-task deep convolutional neural network with 3 outputs CNN The network consists of three convolutional layers, each of which uses 128 output channels and two sets of convolution kernels of different sizes [[3,5],[5,7]]. The convolutional layer introduces nonlinear transformation through the ReLU activation function, and each output of the network is a quality performance indicator.

[0064] Step 5.2: In the obtained CNN A Transformer network is superimposed on the back of the model to construct a hybrid deep learning network model CNN-Transformer, in which a new representation layer is added to the encoder unit Encoder-block of the Transformer network to process the preference matrix P obtained in step 3. The establishment of this model includes the following steps:

[0065] Step 5.2.1: Transformer module design: Take the output dimension of the convolutional network as the input of the Transformer network, and add a new representation layer in the Encoder-block of the Transformer to process the preference matrix P obtained in step 3, such as Figure 2 As shown, the number of multi-head attention mechanisms is set to 4, the number of repeated encoder units in the Transformer network is 2, and the size of the intermediate layer of the feedforward neural network in the encoder is 256. The specific design steps include:

[0066] Step 5.2.1.1: Embedding of preference matrix: Convert the sample-indicator preference matrix P into an embedding vector compatible with Transformer, such as Figure 3 shown.

[0067] Step 5.2.1.2: Modify the input; concatenate the preference matrix P with the Transformer input to obtain a new matrix representation P mix .

[0068] Step 5.2.1.3: Modify the encoder structure; convert the modified matrix P mix Added between adjacent layers of the Transformer encoder unit.

[0069] Step 5.2.2: Fully connected layer design; the network includes 3 fully connected layers, each of which contains 256 hidden units, and there is a ReLU activation function after each hidden unit to improve the nonlinear ability of the model;

[0070] Step 5.2.3: Set weight decay; the weights of the convolutional layer and the linear layer are initialized using the Xavier initialization method. During training, the entire network continuously optimizes parameters to minimize the loss function through back propagation and gradient descent. The learning rate is set to 0.001 and the weight decay is set to 0.001.

[0071] Step 5.3: Use the training set D train The CNN-Transformer network is trained to obtain a multi-index prediction model for the quality of continuously rejected products that integrates sample-index preference characteristics. The model flow chart is shown in the figure below: Figure 4 shown.

[0072] Step 5.4: Use the test set D test Test and evaluate the CNN-Transformer network.

[0073] Step 6: Model application. In actual production, the production process data of each strip is obtained in real time through the OPC interface, and the three key performance indicators of the strip, namely, yield strength, tensile strength and elongation, are predicted online using the constructed CNN-Transformer network model. When the deviation between the predicted value of a certain performance indicator and the set target value exceeds a given threshold, a timely alarm is issued to ensure the stability of the quality of the strip product. Table 2 shows the results of the experimental design and verification of the multi-indicator prediction method for the quality of the strip product for the unbalanced indicator preference data in this embodiment on 12,160 training samples.

[0074] Table 2 Evaluation results of mechanical property prediction indicators

[0075] Mechanical properties MAE without considering sample-metric preference MSE without considering sample-index preference MAE considering sample-metric preference MSE considering sample-index preference Yield Strength 0.0319 0.0029 0.0290 0.0024 tensile strength 0.0222 0.0012 0.0185 0.0010 Elongation 0.0595 0.0061 0.0355 0.0027

[0076] Table 2 shows the experimental results of the test samples (samples that the model has not been trained on). In the experiment, the test set samples were independently run 30 times, and the mean square error (MSE) and mean absolute error (MAE) were calculated for analysis. From the results, it can be seen that before the data is processed with preference, the prediction accuracy of the constructed prediction model for the data does not show obvious differences, and the performance tends to be stable.

[0077] Next, through the multi-index prediction method for the quality of continuously shrunk products facing the unbalanced data of indicator preference described in this embodiment, the sample-indicator preference expression is calculated based on the data of the continuously shrunk production process, and this feature is integrated into the multi-task deep learning model for comprehensive consideration. As shown in Table 3, it is found that the model is most sensitive to the preference for elongation, followed by tensile strength and yield strength. On another level, it also verifies the concept of the present invention. The sample-indicator preference can be quantitatively represented by a regression model established with unbalanced data. At the same time, it also shows that the method of the present invention has certain advantages in comprehensively considering the sample-indicator preference to construct a multi-task deep learning model, and can provide a reference for the strategic design of indicator preference in the field of multi-task learning.

[0078] Table 3 Comparison of mechanical performance index improvement rates considering sample-index preference treatment

[0079] Mechanical properties MAE improvement rate MSE improvement rate Yield Strength 9% 17% tensile strength 17% 17% Elongation 40% 56%

[0080] 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 above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, 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 quality of continuously degraded products with multiple indicators for indicator preference imbalance data, characterized by: The steps include: Step 1: Build a sample set; Collect n production process data samples of steel strips under normal production conditions from the historical data of the cold rolling annealing unit to establish a sample set S; each sample contains v features, corresponding to the v production process data of the steel strip, and performs m outputs, corresponding to m quality performance indicators of the steel strip; Step 2: Data preprocessing; Normalizing n sample data in the established sample data set S; Step 3: Quantitative evaluation of sample-indicator preference; According to the n sample data obtained in step 2, construct the sample-indicator preference matrix P; the specific method is: Step 3.1: Data clustering; The n sample data obtained in step 2 are clustered using the density clustering algorithm DBSCAN to obtain c clusters; Step 3.2: Establish a single indicator regression prediction model; In each cluster, t different machine learning methods are used to establish t different single-indicator regression prediction models for each quality performance indicator; Step 3.3: Calculate sample-indicator preference; For each cluster, the t single index regression prediction models established for each quality performance index in step 3.2 are respectively K Fold cross validation is performed to obtain the respective mean square error MSE, and then the mean square error is averaged to construct the sample-indicator preference matrix ,in, , p ij represents the preference of all samples in the i-th cluster for the j-th performance indicator, i=1,2,…,c; j=1,2,…,m; Step 4: Dataset division; The n sample data preprocessed in step 2 are randomly divided into 0.8n and 0.2n samples, respectively, to form the training set D train and the test set D test ; If 0.8n and 0.2n are not integers, they are rounded down; Step 5: Build a hybrid deep learning network model CNN-Transformer network model; the specific method is: Step 5.1: Use the training set D train , train a m A multi-task deep convolutional neural network (CNN) with 100 outputs, each of which is a quality performance indicator; Step 5.2: A Transformer network is superimposed on the back of the CNN model obtained by training to construct a hybrid deep learning network model CNN-Transformer, in which a new representation layer is added to the encoder unit Encoder-block of the Transformer network to process the preference matrix P obtained in step 3; Step 5.3: Use the training set D train The CNN-Transformer network was trained to obtain a multi-index prediction model for the quality of continuously rejected products that combines sample-index preference features. Step 5.4: Use the test set D test Test and evaluate the CNN-Transformer network; Step 6: Obtain the production process data of each strip in real time through the OPC interface, and use the constructed CNN-Transformer network model to make online predictions on the three key performance indicators of the strip: yield strength, tensile strength and elongation. When the deviation between the predicted value of a performance indicator and the set target value exceeds a given threshold, a timely alarm is issued to ensure the stability of the quality of the continuous annealing strip products.

2. The method for predicting the quality of continuously degraded products based on unbalanced indicator preference data according to claim 1, characterized in that: In step 3.2, each single indicator regression prediction model is constructed using the machine learning library Scikit-learn.

3. The method for predicting the quality of continuously degraded products based on unbalanced indicator preference data according to claim 1, characterized in that: In step 5.1, the multi-task deep convolutional neural network CNN includes 3 convolutional layers, each of which uses 128 output channels and two sets of convolution kernels of different sizes [[3,5], [5,7]], and the convolutional layer introduces nonlinear transformation through the ReLU activation function.

4. The method for predicting the quality of continuously degraded products with multiple indicators based on indicator preference imbalance data according to claim 1, characterized in that: The specific method of step 5.2 is: Step 5.2.1: Transformer module design; The output dimension of the multi-task deep convolutional neural network CNN is used as the input of the Transformer network, and a new representation layer is added to the encoder unit Encoder-block of the Transformer network to process the preference matrix P obtained in step 3; Step 5.2.2: Fully connected layer design; the network includes 3 fully connected layers, each of which contains 256 hidden units, and there is a ReLU activation function behind each hidden unit; Step 5.2.3: Set weight decay; the weights of the convolutional layer and the linear layer are initialized using the Xavier initialization method. During training, the entire network continuously optimizes parameters to minimize the loss function through back propagation and gradient descent. The learning rate is set to 0.001 and the weight decay is set to 0.

001.

5. The method for predicting the quality of continuously degraded products based on unbalanced indicator preference data according to claim 4, characterized in that: The step 5.2.1 specifically includes: Step 5.2.1.1: Embedding of preference matrix; Convert the sample-indicator preference matrix P into an embedding vector compatible with Transformer; Step 5.2.1.2: Modify input; Concatenate the preference matrix P with the Transformer input to obtain a new matrix representation P mix ; Step 5.2.1.3: Modify the encoder structure; The modified matrix P mix Added between adjacent layers of the Transformer encoder unit.

Images