Method for realizing strip deviation prediction of cold rolling continuous annealing furnace based on SMOTE algorithm and deep neural network

Abstract

The present application relates to a kind of based on SMOTE algorithm and deep neural network realizes the method for the prediction processing of cold-rolled continuous annealing furnace strip deviation, comprising the following steps: preliminary processing and label annotation are carried out to data;Carry out combination and feature selection;Deep student framework is built by deep neural network, and loss function is optimized using optimizer;The performance of classification model is evaluated and the model is optimized in super parameter.This application also relates to a kind of for realizing based on SMOTE algorithm and deep neural network to carry out the device for the prediction processing of cold-rolled continuous annealing furnace strip deviation, processor and medium.The method for the prediction processing of cold-rolled continuous annealing furnace strip deviation based on SMOTE algorithm and deep neural network, device, processor and medium of the present application are not wasted in production process The historical data generated, make full use of data analysis, prevent strip deviation;Realize the reduction of cost, the optimization of resources, so that strip production quality is improved, production efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of cold rolling processes, and more particularly to the field of cold rolling continuous annealing furnace technology. Specifically, it refers to a method, apparatus, processor, and medium for predicting and processing strip deviation in cold rolling continuous annealing furnaces based on the SMOTE algorithm and deep neural networks. Background Technology

[0002] Cold-rolled continuous annealing furnace strip steel is an important metallic material that undergoes continuous annealing treatment during the cold rolling process. Continuous annealing refers to the continuous heating and cooling of strip steel under specific temperature conditions to improve its mechanical properties and physical characteristics. Cold-rolled continuous annealing furnace strip steel is widely used in automobile manufacturing, aerospace, home appliances, building materials, and other fields.

[0003] However, strip steel may face a series of technical problems and challenges during the continuous cold rolling annealing process. Among them, strip steel deviation is a common problem in the production process of continuous cold rolling annealing furnace. It refers to the phenomenon that the strip steel deviates from the predetermined position when it passes through the continuous annealing furnace, which may lead to problems such as uneven strip steel width, edge deformation, and reduced production efficiency. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, device, processor and medium for predicting and processing strip deviation in cold rolling continuous annealing furnaces based on the SMOTE algorithm and deep neural network, which is simple to operate, has high production efficiency and wide applicability.

[0005] To achieve the above objectives, the present invention provides a method, apparatus, processor, and medium for predicting and processing strip misalignment in cold-rolled continuous annealing furnaces based on the SMOTE algorithm and deep neural networks, as follows:

[0006] The method for predicting and handling strip misalignment in cold-rolled continuous annealing furnaces based on the SMOTE algorithm and deep neural networks is characterized by the following steps:

[0007] (1) The data were initially processed and labeled by the mode replacement method. After the data was analyzed, the SMOTE method was used to oversample the small sample data and undersample the large sample data to normalize the data.

[0008] (2) Feature selection is performed by combining expert experience method and XGBOOST algorithm;

[0009] (3) Using the data from (1) and (2) as input, a deep neural network is constructed using the deep student framework. BN layers and regularization are used to prevent overfitting, and an optimizer is used to optimize the loss function.

[0010] (4) The performance of the classification model was evaluated by T-SNE dimensionality reduction visualization and ten-fold cross-validation groups, and the hyperparameters of the model were tuned by observing the model results.

[0011] Preferably, step (1) specifically includes the following steps:

[0012] (1.1) The mode substitution method is used to correct outliers and fill in missing values;

[0013] (1.2) Label the tags;

[0014] (1.3) Analyze the sample balance and use the SMOTE method to oversample small sample data;

[0015] (1.4) Perform uniform random undersampling on large sample data and undersample the synthetic sample obtained from oversampling;

[0016] (1.5) The Min-Max standardization method is used to unify the units of measurement to achieve data standardization.

[0017] Preferably, step (1.3) specifically includes the following steps:

[0018] (1.3.1) Identify the minority class samples in the dataset;

[0019] (1.3.2) For each minority class sample, select a certain number of samples from its nearest neighbors as candidates for the synthetic sample;

[0020] (1.3.3) Interpolate based on the distance between samples to calculate new synthetic samples;

[0021] (1.3.4) Repeat step (1.3.3) until the preset oversampling ratio or the specified number of samples is reached.

[0022] Preferably, step (2) specifically includes the following steps:

[0023] (2.1) Perform initial screening for each feature and remove unimportant variables;

[0024] (2.2) The importance of features is sorted by the XGBOOST algorithm, a threshold is set, and features less than the threshold are removed.

[0025] Preferably, step (2.2) specifically includes the following steps:

[0026] (2.2.1) Use the XGBoost algorithm to train the model on the prepared dataset;

[0027] (2.2.2) Perform feature importance assessment by calculating the split contribution of each feature when constructing the decision tree to determine the importance of the feature;

[0028] (2.2.3) Based on the results of feature importance assessment, feature selection is performed.

[0029] Preferably, step (3) specifically includes the following steps:

[0030] (3.1) Construct a deep neural network model based on the Tensorflow framework;

[0031] (3.2) Add a regularization term to prevent overfitting;

[0032] (3.3) The Adam optimizer is used to optimize the cross-entropy loss function.

[0033] Preferably, step (4) specifically includes the following steps:

[0034] (4.1) The performance of the classification model is evaluated by ten-fold cross-validation;

[0035] (4.2) Use the T-SNE dimensionality reduction visualization method to evaluate the model classification performance;

[0036] (4.3) Based on the training process information output by the model, the hyperparameters are tuned and a suitable combination of hyperparameters is selected.

[0037] Preferably, step (4.1) specifically includes the following steps:

[0038] (4.1.1) Randomly divide the dataset into multiple subsets of equal size;

[0039] (4.1.2) Perform iterative training. For each iteration, select one fold as the validation set and the remaining folds as the training set. Use each fold as a validation set once.

[0040] The device for predicting strip deviation in cold-rolled continuous annealing furnaces based on the SMOTE algorithm and deep neural networks is characterized by comprising:

[0041] A processor is configured to execute computer-executable instructions;

[0042] The memory stores one or more computer-executable instructions, which, when executed by the processor, implement the various steps of the method described above for predicting and processing strip deviation in a cold-rolled continuous annealing furnace based on the SMOTE algorithm and deep neural networks.

[0043] The processor used to implement the prediction of strip deviation in cold-rolled continuous annealing furnace based on the SMOTE algorithm and deep neural network is characterized in that the processor is configured to execute computer-executable instructions, which, when executed by the processor, implement the various steps of the method for predicting strip deviation in cold-rolled continuous annealing furnace based on the SMOTE algorithm and deep neural network.

[0044] The computer-readable storage medium is characterized in that it stores a computer program that can be executed by a processor to implement the various steps of the above-described method for predicting and processing strip deviation in a cold-rolled continuous annealing furnace based on the SMOTE algorithm and deep neural networks.

[0045] The present invention employs a method, apparatus, processor, and medium based on the SMOTE algorithm and deep neural network to predict and process strip deviation in cold-rolled continuous annealing furnaces. This method avoids wasting historical data generated during the production process, fully utilizes the data for analysis, and prevents strip deviation. It achieves cost reduction, resource optimization, and improved strip production quality and efficiency. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the data balance analysis results of the method for predicting and processing strip deviation in cold rolling continuous annealing furnace based on the SMOTE algorithm and deep neural network of the present invention.

[0047] Figure 2a This is a schematic diagram of the classification results after random undersampling of label0 in the method for predicting and processing strip deviation in cold rolling continuous annealing furnace based on SMOTE algorithm and deep neural network of the present invention.

[0048] Figure 2b This is another schematic diagram showing the results of data balance analysis for the method of predicting and processing strip deviation in cold rolling continuous annealing furnace based on SMOTE algorithm and deep neural network of the present invention.

[0049] Figure 3 This is a schematic diagram of the importance ranking results of the XGBOOST variable in the method for predicting and processing strip deviation in cold rolling continuous annealing furnace based on the SMOTE algorithm and deep neural network of the present invention.

[0050] Figure 4 This is a schematic diagram showing the model training process of the method for predicting and processing strip deviation in a cold-rolled continuous annealing furnace based on the SMOTE algorithm and deep neural network according to the present invention.

[0051] Figure 5This is another schematic diagram of the model training process for the method of predicting and processing strip deviation in cold rolling continuous annealing furnace based on SMOTE algorithm and deep neural network of the present invention.

[0052] Figure 6 This is a schematic diagram of the deep neural network prediction results of the method for predicting and processing strip deviation in cold rolling continuous annealing furnace based on the SMOTE algorithm and deep neural network of the present invention.

[0053] Figure 7 This diagram illustrates the results of ten cross-validation tests of the method for predicting and processing strip deviation in a cold-rolled continuous annealing furnace based on the SMOTE algorithm and deep neural network, as presented in this invention.

[0054] Figure 8 This is a schematic diagram of the T-SNE dimensionality reduction visualization results of the method for predicting and processing strip deviation in cold-rolled continuous annealing furnaces based on the SMOTE algorithm and deep neural network according to the present invention.

[0055] Figure 9 This is a schematic diagram of the overall structure of the method for predicting and processing strip deviation in a cold-rolled continuous annealing furnace based on the SMOTE algorithm and deep neural network according to the present invention.

[0056] Figure 10 This is a flowchart illustrating the method for predicting and processing strip deviation in a cold-rolled continuous annealing furnace based on the SMOTE algorithm and deep neural network according to the present invention. Detailed Implementation

[0057] To more clearly describe the technical content of the present invention, the following description is provided in conjunction with specific embodiments.

[0058] The present invention provides a method for predicting and processing strip misalignment in cold-rolled continuous annealing furnaces based on the SMOTE algorithm and deep neural networks, comprising the following steps:

[0059] (1) The data were initially processed and labeled by the mode replacement method. After the data was analyzed, the SMOTE method was used to oversample the small sample data and undersample the large sample data to normalize the data.

[0060] (2) Feature selection is performed by combining expert experience method and XGBOOST algorithm;

[0061] (3) Using the data from (1) and (2) as input, a deep neural network is constructed using the deep student framework. BN layers and regularization are used to prevent overfitting, and an optimizer is used to optimize the loss function.

[0062] (4) The performance of the classification model was evaluated by T-SNE dimensionality reduction visualization and ten-fold cross-validation groups, and the hyperparameters of the model were tuned by observing the model results.

[0063] In a preferred embodiment of the present invention, step (1) specifically includes the following steps:

[0064] (1.1) The mode substitution method is used to correct outliers and fill in missing values;

[0065] (1.2) Label the tags;

[0066] (1.3) Analyze the sample balance and use the SMOTE method to oversample small sample data;

[0067] (1.4) Perform uniform random undersampling on large sample data and undersample the synthetic sample obtained from oversampling;

[0068] (1.5) The Min-Max standardization method is used to unify the units of measurement to achieve data standardization.

[0069] In a preferred embodiment of the present invention, step (1.3) specifically includes the following steps:

[0070] (1.3.1) Identify the minority class samples in the dataset;

[0071] (1.3.2) For each minority class sample, select a certain number of samples from its nearest neighbors as candidates for the synthetic sample;

[0072] (1.3.3) Interpolate based on the distance between samples to calculate new synthetic samples;

[0073] (1.3.4) Repeat step (1.3.3) until the preset oversampling ratio or the specified number of samples is reached.

[0074] In a preferred embodiment of the present invention, step (2) specifically includes the following steps:

[0075] (2.1) Perform initial screening for each feature and remove unimportant variables;

[0076] (2.2) The importance of features is sorted by the XGBOOST algorithm, a threshold is set, and features less than the threshold are removed.

[0077] In a preferred embodiment of the present invention, step (2.2) specifically includes the following steps:

[0078] (2.2.1) Use the XGBoost algorithm to train the model on the prepared dataset;

[0079] (2.2.2) Perform feature importance assessment by calculating the split contribution of each feature when constructing the decision tree to determine the importance of the feature;

[0080] (2.2.3) Based on the results of feature importance assessment, feature selection is performed.

[0081] In a preferred embodiment of the present invention, step (3) specifically includes the following steps:

[0082] (3.1) Construct a deep neural network model based on the Tensorflow framework;

[0083] (3.2) Add a regularization term to prevent overfitting;

[0084] (3.3) The Adam optimizer is used to optimize the cross-entropy loss function.

[0085] In a preferred embodiment of the present invention, step (4) specifically includes the following steps:

[0086] (4.1) The performance of the classification model is evaluated by ten-fold cross-validation;

[0087] (4.2) Use the T-SNE dimensionality reduction visualization method to evaluate the model classification performance;

[0088] (4.3) Based on the training process information output by the model, the hyperparameters are tuned and a suitable combination of hyperparameters is selected.

[0089] In a preferred embodiment of the present invention, step (4.1) specifically includes the following steps:

[0090] (4.1.1) Randomly divide the dataset into multiple subsets of equal size;

[0091] (4.1.2) Perform iterative training. For each iteration, select one fold as the validation set and the remaining folds as the training set. Use each fold as a validation set once.

[0092] The present invention relates to an apparatus for predicting strip misalignment in cold-rolled continuous annealing furnaces based on the SMOTE algorithm and deep neural networks, wherein the apparatus comprises:

[0093] A processor is configured to execute computer-executable instructions;

[0094] The memory stores one or more computer-executable instructions, which, when executed by the processor, implement the various steps of the method described above for predicting and processing strip deviation in a cold-rolled continuous annealing furnace based on the SMOTE algorithm and deep neural networks.

[0095] The present invention provides a processor for implementing strip deviation prediction processing in cold-rolled continuous annealing furnaces based on the SMOTE algorithm and deep neural networks. The processor is configured to execute computer-executable instructions, which, when executed by the processor, implement the various steps of the method for implementing strip deviation prediction processing in cold-rolled continuous annealing furnaces based on the SMOTE algorithm and deep neural networks.

[0096] The computer-readable storage medium of the present invention stores a computer program that can be executed by a processor to implement the various steps of the above-described method for predicting and processing strip deviation in a cold-rolled continuous annealing furnace based on the SMOTE algorithm and a deep neural network.

[0097] To address existing industry challenges, this invention innovates a method for predicting strip deviation in cold-rolled continuous annealing furnaces. It provides an analytical and predictive method applicable to strip steel in cold-rolled continuous annealing furnaces, which can solve current industry problems. This invention utilizes big data and machine learning to analyze strip steel data from cold-rolled continuous annealing furnaces, thereby diagnosing strip deviation issues.

[0098] To accomplish this task, we must first collect relevant data on processes such as furnace roll speed, furnace temperature in each section, and strip temperature, and establish a data analysis and preprocessing platform. Based on this, we can develop analysis and prediction models and model evaluation systems to achieve the analysis and prediction of strip deviation in cold rolling continuous annealing furnaces.

[0099] This invention provides a method for predicting strip misalignment in cold-rolled continuous annealing furnaces based on the SMOTE algorithm and deep neural networks. The method consists of four steps:

[0100] Step 1: Data Analysis and Preprocessing. The first part of this step involves data analysis, including but not limited to addressing sample balance issues. The second part is data preprocessing, which includes outlier detection and correction, missing value imputation, and data standardization. The third part is labeling. The fourth part uses the SMOTE algorithm to synthesize samples from categories with smaller sample sizes to achieve sample balance. The purpose of this step is to refine the industrial data, obtain labeled annotations, and make the model's training results more consistent with actual industrial conditions.

[0101] Step Two: Feature Selection. The first part of this step involves initial feature screening based on expert experience; the second part uses XGBoost to assess the importance of features, resulting in the final selection. The purpose of this step is to eliminate redundant features, improve the modeling process, enhance training efficiency and accuracy, and make the model more robust.

[0102] Step 3: Model Building and Training. Construct a deep neural network, adding Batch Normalization (BN) layers and Dropout to prevent overfitting. Use the data processed in Steps 1 and 2 as model input, dividing the dataset into training, validation, and test sets.

[0103] Step Four: Model Evaluation and Hyperparameter Tuning. The first part of this step evaluates the performance of the proposed classification model using ten-fold cross-validation and compares it with other methods to demonstrate the superiority of this patent's effect. The second part uses the T-SNE dimensionality reduction visualization method to evaluate the model's classification performance. The third part optimizes the hyperparameters based on the training process information output by the model, selecting appropriate hyperparameter combinations.

[0104] The present invention provides a method for predicting strip deviation in cold-rolled continuous annealing furnaces based on the SMOTE algorithm and deep neural networks. This system comprises the following four main stages:

[0105] (1) The data were initially processed and labeled by the mode replacement method. After the data was analyzed, the SMOTE method was used to oversample the small sample data and undersample the large sample data. Finally, the data was normalized.

[0106] (2) Feature selection is achieved by combining expert experience method and XGBOOST algorithm;

[0107] (3) Using the data from stage (1) data analysis and preprocessing and stage (2) feature selection as input, a deep neural network is constructed through the deep student framework, and BN layers and regularization are used to prevent overfitting, and an optimizer is used to optimize the loss function;

[0108] (4) The performance of the classification model was evaluated by T-SNE dimensionality reduction visualization and ten-fold cross-validation groups, and the hyperparameters of the model were tuned by observing the model results.

[0109] The data analysis and preprocessing process includes the following specific steps:

[0110] (1.1) The mode substitution method is used to correct outliers and fill in missing values;

[0111] (1.2) Label the variables based on the variable “CPCDev8”;

[0112] (1.3) Analyze the sample balance and use the SMOTE method to oversample small sample data;

[0113] (1.4) Perform uniform random undersampling on large sample data and undersample the synthetic samples obtained from oversampling. The purpose of this step is to reduce sample similarity and eliminate redundant samples.

[0114] (1.5) The Min-Max standardization method is used to unify the units of measurement to achieve data standardization.

[0115] The feature selection process consists of two steps: first, preliminary feature screening based on expert experience; and second, final feature screening based on the XGBOOST method. This step includes the following specific steps:

[0116] (2.1) Based on expert experience, each feature is analyzed and variables that are obviously unimportant are removed.

[0117] (2.2) Based on the XGBOOST algorithm, the importance of features is sorted, and finally a threshold is set to remove features that are less than the threshold.

[0118] Using a deep neural network model, this step involves the following specific steps:

[0119] (3.1) Construct a deep neural network model based on the Tensorflow framework;

[0120] (3.2) Use regularization methods such as BN layer and Dropout to prevent overfitting;

[0121] (3.3) The Adam optimizer is used to optimize the cross-entropy loss function.

[0122] The model evaluation module includes the following specific steps for determining hyperparameters and validating model performance:

[0123] (4.1) The combined validation method based on T-SNE dimensionality reduction visualization classification results and 10-fold cross-validation can effectively evaluate the model performance and select the best model. At the same time, this method can also reduce the overfitting and underfitting problems of the model and improve the generalization ability of the model.

[0124] (4.2) Based on empirical rules, select parameters for various hyperparameters of the neural network, such as the number of neurons in each layer, learning rate, and number of iterations, and optimize the model.

[0125] (4.3) The superiority of this method is verified by comparison with other methods.

[0126] In a specific embodiment of the present invention, a method for analyzing and predicting strip deviation in a cold-rolled continuous annealing furnace is provided, including the following:

[0127] Step 1: This invention provides a method for constructing a data analysis and preprocessing module. This module can complete the supplementation of important missing values, the correction of outliers, sample balance analysis, sample synthesis using the SMOTE method, and uniform random undersampling of redundant samples.

[0128] This module can perform the following functions:

[0129] (1) The data analysis and preprocessing module has data repair capabilities:

[0130] The data analysis and preprocessing module designed in this invention uses the mode substitution method to fill in missing values ​​and correct outliers. Due to the complex environment of strip steel production, issues such as measurement deviations and missing data records exist. Data repair has a profound impact on the model, preventing data anomalies caused by the complex industrial environment.

[0131] (2) The data analysis and preprocessing module can perform analysis of sample balance;

[0132] (3) The data analysis and preprocessing module has the functions of sample undersampling and oversampling.

[0133] The data analysis and preprocessing module designed in this invention can use SMOTE to oversample small sample data and use uniform random undersampling to remove some redundant samples.

[0134] The key computational steps of the SMOTE method are described below:

[0135] 1) Identify the minority class samples:

[0136] First, it is necessary to identify the minority class samples in the dataset, that is, the classes that need to be oversampled.

[0137] 2) Select nearest neighbor samples:

[0138] For each minority class sample, a certain number of samples are selected from its nearest neighbors as candidates for the synthetic sample.

[0139] 3) Calculate the synthetic sample:

[0140] For each minority class sample, after selecting several nearest neighbor samples, a new synthetic sample is obtained through interpolation. Typically, interpolation can be performed based on the distance between samples. Specifically, for a minority class sample, its k nearest neighbor samples are selected, and then a sample is randomly selected from these nearest neighbor samples, and a new synthetic sample is generated through interpolation.

[0141] 4) Repeat step 3):

[0142] For each minority class sample, the calculation of the synthetic sample is repeated until the preset oversampling ratio or the specified number of samples is reached.

[0143] Combine synthetic and original samples: The generated synthetic samples are merged with the original samples to form a new balanced dataset.

[0144] (4) The data analysis and preprocessing module has data standardization function.

[0145] The data analysis and preprocessing module designed in this invention adopts the Min-Max standardization method to unify the data dimensions.

[0146] Step Two: This invention provides a method for constructing a feature selection module, which can filter feature variables and key parameters related to strip deviation prediction, including variables such as furTempJPF, furTempRTF, RTFTempSetModel, and furHeatModel. The feature selection module designed in this invention can achieve the following functions:

[0147] (1) Based on expert experience, analyze each feature and delete obviously unimportant variables.

[0148] (2) The importance of features is sorted based on the XGBOOST algorithm, and finally a threshold is set to remove features that are less than the threshold.

[0149] The reasons for deciding to use the XBGOOST algorithm for feature selection are as follows:

[0150] 1. Improve the predictive accuracy of the model;

[0151] 2. Remove irrelevant variables;

[0152] 3. Improve learning efficiency to reduce computing and storage costs, etc.

[0153] The key calculation steps of XGBOOST are described below:

[0154] 1) Model Training

[0155] The XGBoost algorithm is used to train the model on the prepared dataset. During training, XGBoost iteratively optimizes the loss function to gradually build multiple decision tree models and combines them into the final ensemble model.

[0156] 2) Feature Importance Assessment

[0157] After training, XGBoost provides a method for evaluating feature importance. It determines feature importance by calculating the split contribution of each feature during decision tree construction. These contributions can be normalized based on their average or sum to derive an importance score for each feature.

[0158] XGBoost provides two commonly used metrics to determine contribution: weight and gain.

[0159] Weight: Represents the sum of the number of times a feature is used across all decision trees. The calculation formula is:

[0160]

[0161] Where f represents the feature, T represents the number of decision trees, and f t f represents the feature used for splitting in decision tree t. t =f is an indicator function, when f t =1 when f, otherwise 0.

[0162] Gain: Represents the sum of the contributions of features to splitting in a decision tree. The calculation formula is:

[0163]

[0164] Where f represents the feature, T represents the number of decision trees, λ is the regularization parameter, and G... L G is the sum of the first derivatives of the gradients of the left subtree before and after the split. R It is the sum of the first derivatives of the gradients of the right subtree before and after the split, and the first term inside the square brackets. The second term represents the information gain of the left subtree after splitting at the selected node in the model. The third term represents the information gain of the right subtree after splitting at the selected node in the model. The information gain of the representative model before splitting at this node.

[0165] 3) Feature Filtering

[0166] Based on the results of feature importance assessment, feature selection can be performed. A threshold is set, and only features with importance scores higher than this threshold are selected. This reduces model complexity and computational cost while retaining the most predictive features.

[0167] Step 3: This invention provides a method for constructing and training a prediction model for strip deviation in a cold-rolled continuous annealing furnace, using a deep neural network to analyze and predict strip deviation problems.

[0168] Industrial data is characterized by high dimensionality, large volume, and complex relationships between features. Deep neural network models can be used to fully learn the complex nonlinear relationships in industrial data.

[0169] The following explains the principles of deep neural network models:

[0170] The basic structure of a deep neural network consists of a multi-layered structure composed of multiple neurons. Each neuron receives input from neurons in the previous layer, processes it through a non-linear activation function, and then passes it to neurons in the next layer. Each neuron has a set of adjustable weight parameters and a bias term. By continuously adjusting these parameters and bias terms, the deep neural network can gradually learn the abstract representation and mapping rules of the input data.

[0171] The forward propagation process of a deep neural network can be represented by the following formula:

[0172] y (l) =f (l) (W (l) y (l-1) +b (l) )……(3)

[0173] Among them, y (l) W represents the output of the l-th layer. (l) b represents the weight matrix between layer l and layer (l-1). (l) f represents the bias term of the l-th layer. (l) (·) represents the nonlinear activation function of the l-th layer, y (0) This indicates the input data.

[0174] Deep neural networks are trained using the backpropagation algorithm. The basic idea is to calculate the partial derivatives of the loss function with respect to the weights and biases, and then update the network parameters based on these partial derivatives. In the backpropagation algorithm, we need to calculate the error of the output layer and propagate it forward layer by layer to calculate the error signal for each neuron, and then update the weights and biases based on the error signals.

[0175] The objective function of a deep neural network is the cross-entropy loss function, and its calculation formula is as follows:

[0176]

[0177] Where N represents the training samples, K represents the number of output categories, and y (i，k) This represents the true label of the i-th sample. θ represents the model output value of the i-th sample, and θ represents the model parameters.

[0178] Step 4: This invention provides a model evaluation and hyperparameter tuning module, which can evaluate deep neural network models by combining T-SNE dimensionality reduction visualization and 10-fold cross-validation; and perform hyperparameter tuning.

[0179] (1) Cross-validation with ten folds:

[0180] This validation method divides the dataset into 10 equal-sized subsets. Each time, 9 subsets are used as the training set, and the remaining subset is used as the validation set. By repeating this process 10 times, each time using a different validation set, a more stable evaluation of the model's performance can be obtained.

[0181] First, the dataset is partitioned. The dataset is randomly divided into 10 equal-sized subsets. These subsets are typically called folds and are numbered from 1 to 10.

[0182] Secondly, the training and validation processes are iterated. In each iteration, one fold is selected as the validation set, and the remaining nine folds are used as the training set. There are ten possible combinations, meaning each fold serves as a validation set once.

[0183] (2) T-SNE dimensionality reduction visualization:

[0184] Feature representations are extracted from the penultimate layer of the deep neural network model. These features carry a high-level abstraction of the original data by the model. The T-SNE algorithm is then used to perform dimensionality reduction and visualization of these high-level features.

[0185] By extracting the feature representations from the penultimate layer of a deep neural network and mapping them to a two-dimensional space using the T-SNE algorithm, the original data can be presented in a new, lower-dimensional space. This visualization helps us intuitively observe the distribution of data points of different categories in the reduced-dimensional space.

[0186] By observing the T-SNE dimensionality reduction visualization, one can identify the degree of separation between categories, cluster structure, and outliers.

[0187] This invention uses an implementation case based on strip steel data from a cold-rolled continuous annealing furnace in a large steelmaking enterprise as a specific example:

[0188] In the implementation case, the factory provided raw data of cold-rolled continuous annealing furnace strip steel recorded in batches over the past few years. Each batch of data, along with several indicators, was recorded in a CSV file. The CSV file contained outliers and anomalies. According to the data analysis and preprocessing steps designed in this invention, the aforementioned raw data needed to be analyzed and preprocessed.

[0189] Data analysis and preprocessing, including correcting discrete values, imputing missing values, and normalizing the data, are performed before data balance analysis. This work is based on Python, and the computer used for this operation needs to have Pandas, Scikit-learn, Matplotlib, NumPy, and other related libraries and environments configured beforehand.

[0190] After data preprocessing, the balance of the data is analyzed. Figure 1 It demonstrates the problem of data imbalance.

[0191] Depend on Figure 1 It is evident that the sample size difference between label0 (normal sample) and label1 (skewed sample) is very large.

[0192] First, we tried random undersampling of label0 (normal samples) to reduce the sample size of label0 to the same level as label1, but the results were not good. The results are as follows... Figure 2a As shown.

[0193] Analysis revealed that the main reason for inaccurate results is that simply undersampling normal samples leads to a significant loss of useful information. Using generative networks would increase computational burden and time costs, which is counterproductive for industrial production. Therefore, the SMOTE method is needed to retain a large amount of useful information from label0 while simultaneously expanding label1 with additional information from smaller samples.

[0194] First, the SMOTE method was used to oversample label1, and then a small amount of undersampling was performed on label0. The sample balance analysis results are as follows: Figure 2b As shown.

[0195] The first part of the feature selection step in the implementation case is the preliminary selection of features based on expert experience. In this embodiment, this step specifically includes analyzing the strip steel production process, consulting with frontline workers, and communicating with experts from relevant research institutes. This process effectively extracts feature variables related to the strip steel production process mechanism, on-site personnel operating experience, and the main control loops of the equipment.

[0196] The second part of the feature selection step in the implementation case is the final feature screening based on the XGBOOST method. Redundant feature variables are removed from the sample data, and XGBOOST is used to rank the features by importance. Thus, we screen the sample data to obtain the most important variables and remove redundancy, as shown in the following results. Figure 3 As shown.

[0197] In summary, the feature selection steps applied to this enterprise have been established. Feature selection was achieved through the data analysis and preprocessing in step one of this invention, which greatly improved the training speed and model accuracy of the model training module.

[0198] In the implementation case, the model training module uses a deep neural network model for training. The training process is as follows: Figure 4 and Figure 5 As shown. The final result is as follows. Figure 6 As shown in Table 1, the prediction results are as follows.

[0199] Table 1. Prediction results of deep neural networks

[0200]

[0201] In summary, the model training module applied to this plant has been established, realizing the analysis and prediction model for the strip steel data quality of this plant.

[0202] The model hyperparameter tuning and model evaluation sections in the embodiments.

[0203] 1. The results of ten cross-validation tests are as follows: Figure 7 As shown.

[0204] 2. Visualization results of T-SNE dimensionality reduction are as follows: Figure 8 As shown.

[0205] The comparison results with other methods are shown in Table 2.

[0206] Table 2 Performance of other algorithms

[0207] Model accuracy accuracy Recall rate F1 XGB 0.8500 0.8609 0.9075 0.8831 LR 0.8230 0.8542 0.8636 0.8583 SVM 0.7324 0.8118 0.7683 0.7771 RF 0.8081 0.8540 0.8354 0.8443

[0208] In summary, the hyperparameter tuning and model evaluation of the model applied to this plant have been completed, realizing the parameter selection and model evaluation of the analysis and prediction model for the strip steel quality of this plant.

[0209] For the specific implementation scheme of this embodiment, please refer to the relevant descriptions in the above embodiments, which will not be repeated here.

[0210] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0211] It should be noted that in the description of this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this invention, unless otherwise stated, "a plurality of" means at least two.

[0212] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0213] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution device. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0214] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The corresponding program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiments.

[0215] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0216] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0217] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0218] The present invention employs a method, apparatus, processor, and medium based on the SMOTE algorithm and deep neural network to predict and process strip deviation in cold-rolled continuous annealing furnaces. This method avoids wasting historical data generated during the production process, fully utilizes the data for analysis, and prevents strip deviation. It achieves cost reduction, resource optimization, and improved strip production quality and efficiency.

[0219] In this specification, the invention has been described with reference to specific embodiments thereof. However, it will be apparent that various modifications and variations can be made without departing from the spirit and scope of the invention. Therefore, the specification and drawings should be considered illustrative rather than restrictive.

Claims

1. A method for predicting and processing strip misalignment in cold-rolled continuous annealing furnaces based on the SMOTE algorithm and deep neural networks, characterized in that... The method includes the following steps: (1) The data were initially processed and labeled by the mode replacement method. After the data was analyzed, the SMOTE method was used to oversample the small sample data and undersample the large sample data to normalize the data. (2) Feature selection is performed by combining expert experience method and XGBOOST algorithm; (3) Using the data from (1) and (2) as input, a deep neural network is constructed using the deep student framework. BN layers and regularization are used to prevent overfitting, and an optimizer is used to optimize the loss function. (4) The performance of the classification model was evaluated by T-SNE dimensionality reduction visualization and ten-fold cross-validation groups, and the hyperparameters of the model were tuned by observing the model results; Step (1) specifically includes the following steps: (1.1) The mode substitution method is used to correct outliers and fill in missing values; (1.2) Label the items; (1.3) Analyze the sample balance and use the SMOTE method to oversample small sample data; (1.4) Perform uniform random undersampling on large sample data and undersample the synthetic sample obtained from oversampling; (1.5) The Min-Max standardization method is used to unify the units of measurement to achieve data standardization; Step (1.3) specifically includes the following steps: (1.3.1) Identify the minority class samples in the dataset; (1.3.2) For each minority class sample, select a certain number of samples from its nearest neighbors as candidates for the synthetic sample; (1.3.3) Interpolate based on the distance between samples to calculate new synthetic samples; (1.3.4) Repeat step (1.3.3) until the preset oversampling ratio or the specified number of samples is reached; Step (2) specifically includes the following steps: (2.1) Perform initial screening for each feature and remove unimportant variables; (2.2) The importance of features is ranked by the XGBOOST algorithm, a threshold is set, and features with importance less than the threshold are removed; Step (2.2) specifically includes the following steps: (2.2.1) Use the XGBoost algorithm to train the model on the prepared dataset; (2.2.2) Perform feature importance assessment by calculating the splitting contribution of each feature when constructing the decision tree to determine the importance of the feature; (2.2.3) Based on the results of feature importance assessment, feature selection is performed.

2. The method for predicting and processing strip misalignment in cold-rolled continuous annealing furnaces based on the SMOTE algorithm and deep neural network as described in claim 1, characterized in that, Step (3) specifically includes the following steps: (3.1) Construct a deep neural network model based on the Tensorflow framework; (3.2) Add a regularization term to prevent overfitting; (3.3) Use the Adam optimizer to optimize the cross-entropy loss function.

3. The method for predicting and processing strip misalignment in cold-rolled continuous annealing furnaces based on the SMOTE algorithm and deep neural network as described in claim 1, characterized in that, Step (4) specifically includes the following steps: (4.1) The performance of the classification model is evaluated by 10-fold cross-validation; (4.2) Use the T-SNE dimensionality reduction visualization method to evaluate the model classification performance; (4.3) Based on the training process information output by the model, the hyperparameters are tuned and a suitable combination of hyperparameters is selected.

4. The method for predicting and processing strip misalignment in cold-rolled continuous annealing furnaces based on the SMOTE algorithm and deep neural network as described in claim 3, characterized in that, Step (4.1) specifically includes the following steps: (4.1.1) Randomly divide the dataset into multiple subsets of equal size; (4.1.2) Perform iterative training. For each iteration, select one fold as the validation set and the remaining folds as the training set. Use each fold as a validation set once.

5. An apparatus for predicting and processing steel strip misalignment in a cold-rolled continuous annealing furnace based on the SMOTE algorithm and a deep neural network, characterized in that, The device includes: A processor is configured to execute computer-executable instructions; The memory stores one or more computer-executable instructions, which, when executed by the processor, implement the steps of the method for predicting and processing strip deviation in a cold-rolled continuous annealing furnace based on the SMOTE algorithm and a deep neural network, as described in any one of claims 1 to 4.

6. A processor for predicting steel strip misalignment in cold-rolled continuous annealing furnaces based on the SMOTE algorithm and deep neural networks, characterized in that, The processor is configured to execute computer-executable instructions, which, when executed by the processor, implement each step of the method for predicting and processing strip deviation in a cold-rolled continuous annealing furnace based on the SMOTE algorithm and a deep neural network, as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, It stores a computer program that can be executed by a processor to implement the various steps of the method for predicting and processing strip deviation in a cold-rolled continuous annealing furnace based on the SMOTE algorithm and deep neural network as described in any one of claims 1 to 4.

Images