Continuous casting slab longitudinal crack real-time prediction method based on deep learning

Abstract

The invention relates to the technical field of steelmaking continuous casting, in particular to a continuous casting slab longitudinal crack real-time prediction method based on deep learning, which comprises the following steps: data acquisition and preprocessing: acquiring continuous temperature time sequence data on a plurality of rows of thermocouples of a crystallizer copper plate; a hybrid neural network model is constructed, an input layer is used for receiving L * N-dimensional temperature time sequence data, L is a time window length, N is a crystallizer thermocouple number, a feature extraction module comprises a multi-scale feature extraction module and a feature interaction bidirectional LSTM layer, and a classification module constructs a network structure of a full connection layer, a Dropout layer and an output layer; according to the method, online prediction of slab longitudinal cracks is realized, deep features of temperature time sequence data are automatically learned through an end-to-end model, and prediction precision and real-time performance are remarkably improved in combination with a hybrid neural network architecture.

Description

Technical Field

[0001] This invention relates to the field of steelmaking continuous casting technology, and in particular to a method for real-time prediction of longitudinal cracks in continuously cast slabs based on deep learning. Background Technology

[0002] Longitudinal cracks in continuously cast slabs are a quality defect in steelmaking. These cracks directly deteriorate the quality of hot-rolled coils and can even disrupt production schedules. Improving the surface quality of hot-rolled coils is a crucial issue that needs to be addressed to achieve hot-rolling instead of cold-rolling in the continuous casting and rolling process. Surface defects are a significant factor affecting the overall quality of steel billets. Due to the high temperature and production process, traditional billet surface inspection relies primarily on visual inspection. This application utilizes cutting-edge digital technology combined with the formation mechanism of longitudinal cracks in continuously cast slabs to develop a longitudinal crack prediction model for slabs.

[0003] Research on the formation mechanism and prediction model of longitudinal cracks in slabs is still in its early stages both domestically and internationally, with significant room for further research. Correctly applying various statistical models, mechanistic models, and intelligent models, along with mathematical tools including data preprocessing, quality prediction, and parameter optimization algorithms, can establish reliable and accurate prediction methods. With the continuous development of steel production technology, research on longitudinal cracks in slabs will become increasingly in-depth, and corresponding longitudinal crack detection technologies need further improvement and development. Existing technologies largely rely on traditional machine learning methods such as Support Vector Machines (SVM), requiring manual extraction of typical features of temperature time series (such as temperature rise / fall amplitude, time intervals, etc.), which suffers from strong subjectivity in feature extraction and limited model generalization ability. Furthermore, traditional methods struggle to capture nonlinear relationships in complex time series, leading to insufficient prediction accuracy and real-time performance. Therefore, this invention proposes a real-time prediction method for longitudinal cracks in continuously cast slabs based on deep learning, which can also provide targeted optimization solutions for production processes. Summary of the Invention

[0004] This invention provides a real-time prediction method for longitudinal cracks in continuously cast slabs based on deep learning. It realizes online prediction of longitudinal cracks in slabs, automatically learns the deep features of temperature time series data through an end-to-end model, and combines a hybrid neural network architecture to significantly improve prediction accuracy and real-time performance.

[0005] To achieve the above objectives, the present invention employs the following technical solution:

[0006] A method for real-time prediction of longitudinal cracks in continuously cast slabs based on deep learning includes the following steps:

[0007] S1. Data Acquisition and Preprocessing: Continuous temperature time series data are acquired on the multi-row thermocouples on the copper plate of the crystallizer. Data in a time window of 100-120s, including longitudinal cracks and normal operating conditions, are extracted and the raw temperature data is standardized.

[0008] S2. Construction of Hybrid Neural Network Model:

[0009] (1) Input layer: Receives L×N dimensional temperature time series data, where L is the time window length and N is the number of thermocouples in the crystallizer;

[0010] (2) Feature extraction module: including a multi-scale feature extraction module and a feature interaction bidirectional LSTM layer;

[0011] The multi-scale feature extraction module uses a one-dimensional convolutional kernel to capture local temperature change patterns. It uses three different one-dimensional CNNs with 3×1, 5×1, and 7×1 convolutional kernels in parallel to extract local temperature change features. Then, the SE attention module weights the multi-channel thermocouple features, calculates the channel weights, encodes the casting speed and casting time into a temporal position vector, and concatenates it with the temperature features.

[0012] The feature interaction bidirectional LSTM layer extracts long-term dependencies in time series data. Before the CNN output is fed into the bidirectional LSTM, a timestamp embedding layer is added, and residual connections are used to bridge the CNN and LSTM to alleviate gradient vanishing.

[0013] (3) Classification module: Construct a network structure of fully connected layer, Dropout layer and output layer. The fully connected layer maps the features to a high-dimensional space. Set the parameters of the Dropout layer. The output layer uses the Sigmoid function to output the probability of longitudinal cracks.

[0014] S3, Model Training and Optimization;

[0015] S4. Online Real-time Prediction: Deploy the model to an industrial computer to receive thermocouple data in real time; use a sliding window to capture the latest L-second data and output the model's predicted probability of longitudinal cracks. If the probability exceeds the threshold of 0.9, trigger a longitudinal crack warning and locate the abnormal thermocouple position.

[0016] Furthermore, for the model training, a binary cross-entropy loss function is selected as the training objective, taking into account the characteristics of the binary classification task. The model convergence is driven by minimizing the difference between the predicted probability and the true label.

[0017] Furthermore, in step S3, the optimization adopts the Adam algorithm for adaptive moment estimation, with the initial learning rate set to 0.001, and the learning rate is dynamically adjusted in combination with a cosine decay strategy.

[0018] Furthermore, it also includes further optimization of model parameters, introduction of five-fold cross-validation technology, dividing the training set into 5 subsets, and iterating through 4 subsets for training and 1 subset for validation. Through multiple rounds of iteration, a configuration scheme of 128 is determined. An early stopping mechanism is implemented simultaneously during training. If the validation loss does not decrease after 10 consecutive training rounds, the training is terminated early and the current optimal model weights are saved.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] By adopting an end-to-end feature learning mechanism, manual intervention is avoided, the comprehensiveness of feature representation is improved, the hybrid model reduces the occurrence of steel leakage accidents, significantly improves product quality, and reduces production costs. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the thermocouple deployment according to an embodiment of the present invention;

[0022] Figure 2 This is an overall flowchart of the CNN-BiLSTM prediction model in an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram of the CNN-BiLSTM hybrid network architecture according to an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram illustrating the prediction process in an embodiment of the present invention. Detailed Implementation

[0025] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0026] This invention provides a real-time prediction method for longitudinal cracks in continuously cast slabs based on deep learning, comprising the following steps:

[0027] S1. Data acquisition and preprocessing;

[0028] Continuous temperature time-series data were collected from multiple rows of thermocouples on the copper plate of the crystallizer. A 120-second time window containing data from longitudinal cracks and normal operating conditions was extracted. The original temperature data was standardized to eliminate the influence of dimensions. Median filtering was added to eliminate instantaneous abnormal values ​​of thermocouples and avoid noise interference. Longitudinal crack samples were marked as 1, and normal samples as 0.

[0029] S2. Construction of hybrid neural network model;

[0030] Input layer: Receives L×N dimensional temperature time series data, where L is the time window length and N is the number of thermocouples in the crystallizer;

[0031] Feature extraction module: (1) Multi-scale feature extraction module: a one-dimensional convolutional kernel is used to capture local temperature change patterns. In parallel, three different convolutional kernels of 3×1, 5×1 and 7×1 are used to extract local temperature change features respectively. Then, the SE attention module is used to weight the multi-channel thermocouple features and calculate the channel weights. The process parameters such as casting speed and casting time are encoded into time-series position vectors and concatenated with the temperature features.

[0032] (2) Feature interaction bidirectional LSTM layer to extract long-term dependencies in time series data;

[0033] Before feeding the CNN output into the bidirectional LSTM, a timestamp embedding layer is added to encode the effect of speed changes on temperature; a residual connection is used to bridge the CNN and LSTM to alleviate gradient vanishing.

[0034] This architecture breaks through the dependence of traditional single models such as SVM on linearly separable features, and significantly improves the prediction accuracy under complex working conditions.

[0035] Classification module: Constructs a network structure of fully connected layer, Dropout layer, and output layer; the fully connected layer maps features to a high-dimensional space, the parameters of the Dropout layer are set to prevent overfitting, and the output layer uses the Sigmoid function to output the probability of longitudinal cracks.

[0036] S3, Model Training and Optimization;

[0037] A multi-dimensional strategy is employed to improve performance and generalization ability. First, considering the characteristics of binary classification tasks, a binary cross-entropy loss function is selected as the training objective, driving model convergence by minimizing the difference between the predicted probability and the true label. The optimization process uses the Adam algorithm of adaptive moment estimation, with an initial learning rate of 0.001, and dynamically adjusts the learning rate using a cosine decay strategy to ensure rapid convergence in the early stages of training while avoiding oscillations in the later stages. To further optimize the parameter model, a five-fold cross-validation technique is introduced, dividing the training set into five subsets, and iterating through four subsets for training and one subset for validation, ultimately determining a configuration of 128 through multiple iterations. To prevent overfitting, an early stopping mechanism is implemented simultaneously during training; if the validation loss does not decrease after 10 consecutive training iterations, training is terminated early and the current optimal model weights are saved.

[0038] S4. Online real-time forecasting;

[0039] The model is deployed to an industrial computer to receive thermocouple data in real time. A sliding window is used to capture the latest L-second data, and the model's predicted probability of longitudinal cracks is output. If the probability exceeds a threshold of 0.9, a longitudinal crack warning is triggered, and the location of the abnormal thermocouple is identified.

[0040] The following embodiments are implemented based on the technical solution of the present invention, providing detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments. Unless otherwise specified, the methods used in the following embodiments are conventional methods.

[0041] Example:

[0042] A method for real-time prediction of longitudinal cracks in continuously cast slabs based on deep learning, see [link to article]. Figure 2 The specific implementation steps are as follows:

[0043] S1. Data acquisition and preprocessing;

[0044] See the deployment of copper plate thermocouples in the crystallizer of a steel plant. Figure 1 The data is arranged in 6 rows and 11 columns, with a sampling frequency of 2Hz. The data is stored in a real-time database, and a 120s time window of data is extracted. The longitudinal crack sample is labeled as 1, and the normal working condition sample is labeled as 0.

[0045] Z-score standardization is used to eliminate dimensional differences:

[0046]

[0047] Where μ is the mean and σ is the standard deviation;

[0048] Data augmentation techniques are employed, which involve shifting the time series data randomly by 10% to the left or right to simulate fluctuations in production rhythm. Gaussian noise with a mean of 0 and a variance of 0.1 is injected to enhance the anti-interference capability. New samples are generated through linear interpolation to solve the problem of data imbalance.

[0049] The original dataset contained 400 cases, including 200 longitudinal crack samples and 200 normal samples. After data augmentation, the total number of samples was 1200, including 600 longitudinal crack samples and 600 normal samples. The training set and test set were divided in a 7:3 ratio, with 840 cases in the training set, 180 cases in the validation set, and 180 cases in the test set.

[0050] S2. Construction of hybrid neural network model, see Figure 3 ;

[0051] Input layer: Receives 120×11 dimensional data, i.e., 120 seconds × 11 columns of thermocouples;

[0052] CNN module: Conv1D layer and Pooling layer; Conv1D, 64 filters, kernel size 5, stride 1, ReLU activation function, extracts local temperature change features; MaxPooling1D layer, pooling window 2, stride 2, reduces dimensionality and preserves key features;

[0053] BiLSTM module: Bidirectional LSTM layer 1, with 128 hidden units, returns the complete sequence and captures bidirectional temporal dependencies; Bidirectional LSTM layer 2, with 64 hidden units, outputs the final temporal features;

[0054] Classification module: Fully connected layer, 32 neurons, ReLU activation; Dropout layer, ratio 0.3, to prevent overfitting; Output layer, 1 neuron, Sigmoid activation, outputs the probability of longitudinal cracks P∈[0,1].

[0055] S3, Model Training and Optimization;

[0056] A multi-dimensional strategy is employed to ensure the model's efficient learning and generalization capabilities. First, considering the characteristics of binary classification tasks, binary cross-entropy is used as the loss function to accurately measure the deviation between the predicted probability and the true label. The optimizer is Adam, with an initial learning rate set to 0.001. The learning rate is dynamically adjusted using a cosine decay strategy, which accelerates initial convergence and avoids oscillations in the later stages. The training batch size is set to 32, balancing computational efficiency and memory resource consumption. The maximum number of training epochs is set to 100, and an early stopping mechanism is introduced. If the validation set loss does not decrease for 10 consecutive epochs, training is terminated early, and the optimal weights are saved to effectively prevent overfitting.

[0057] To further optimize the model parameters, a 5-fold cross-validation method was adopted, dividing the training set into 5 subsets. The training was performed using 4 subsets and the validation was performed using 1 subset. Finally, a hyperparameter combination with a kernel size of 5 and 128 bidirectional LSTM units was selected. Experimental results show that the model achieves 99.2%, 98.5%, and 98.7% accuracy on the training, validation, and test sets, respectively, which is 6.3 percentage points higher than the traditional SVM method's 92.4%. This fully verifies the significant advantages of the proposed method in feature learning and prediction accuracy.

[0058] S4, Real-time Prediction Process;

[0059] See Figure 4 First, a sliding window is used to capture the latest 120 seconds of thermocouple data every second. Then, the data is standardized and input into the model to output the probability of longitudinal cracks, P. Finally, a threshold alarm is triggered. If P ≥ 0.9, an alarm is triggered and the abnormal thermocouple column is located.

[0060] Performance metrics:

[0061] accuracy Recall rate F1 score ROC-AUC 98％ 98.5％ 98.1％ 0.993

[0062] The longitudinal crack prediction model was deployed on the slab continuous casting line of a steel plant. The average time from data acquisition to alarm triggering was 45ms. Within one month, it successfully issued warnings for 6 longitudinal crack accidents, avoiding economic losses of about one million yuan.

Claims

1. A method for real-time prediction of longitudinal cracks in continuously cast slabs based on deep learning, characterized in that, Includes the following steps: S1. Data Acquisition and Preprocessing: Continuous temperature time series data are acquired on the multi-row thermocouples on the copper plate of the crystallizer. Data in a time window of 100-120s, including longitudinal cracks and normal operating conditions, are extracted and the raw temperature data is standardized. S2. Construction of Hybrid Neural Network Model: (1) Input layer: Receives L×N dimensional temperature time series data, where L is the time window length and N is the number of thermocouples in the crystallizer; (2) Feature extraction module: including a multi-scale feature extraction module and a feature interaction bidirectional LSTM layer; The multi-scale feature extraction module uses a one-dimensional convolutional kernel to capture local temperature change patterns. It uses three different one-dimensional CNNs with 3×1, 5×1, and 7×1 convolutional kernels in parallel to extract local temperature change features. Then, the SE attention module weights the multi-channel thermocouple features, calculates the channel weights, encodes the casting speed and casting time into a temporal position vector, and concatenates it with the temperature features. The feature interaction bidirectional LSTM layer extracts long-term dependencies in time series data. Before the CNN output is fed into the bidirectional LSTM, a timestamp embedding layer is added, and residual connections are used to bridge the CNN and LSTM to alleviate gradient vanishing. (3) Classification module: Construct a network structure of fully connected layer, Dropout layer and output layer. The fully connected layer maps the features to a high-dimensional space. Set the parameters of the Dropout layer. The output layer uses the Sigmoid function to output the probability of longitudinal cracks. S3, Model Training and Optimization; S4. Online Real-time Prediction: Deploy the model to an industrial computer to receive thermocouple data in real time; use a sliding window to capture the latest L-second data and output the model's predicted probability of longitudinal cracks. If the probability exceeds the threshold of 0.9, trigger a longitudinal crack warning and locate the abnormal thermocouple position.

2. The method for real-time prediction of longitudinal cracks in continuously cast slabs based on deep learning according to claim 1, characterized in that, The model training, tailored to the characteristics of binary classification tasks, uses the binary cross-entropy loss function as the training objective, driving model convergence by minimizing the difference between the predicted probability and the true label.

3. The method for real-time prediction of longitudinal cracks in continuously cast slabs based on deep learning according to claim 2, characterized in that, In step S3, the optimization adopts the Adam algorithm for adaptive moment estimation, with the initial learning rate set to 0.001, and the learning rate is dynamically adjusted in combination with a cosine decay strategy.

4. The method for real-time prediction of longitudinal cracks in continuously cast slabs based on deep learning according to claim 3, characterized in that, It also includes further optimizing model parameters, introducing five-fold cross-validation technology, dividing the training set into 5 subsets, and iterating through 4 subsets for training and 1 subset for validation. The configuration scheme of 128 is determined through multiple rounds of iteration. An early stopping mechanism is implemented simultaneously during training. If the validation loss does not decrease after 10 consecutive training rounds, the training is terminated early and the current optimal model weights are saved.

Images