Multi-task Slab Quality Defect Prediction Method and Device Based on Adaptive Loss

Abstract

The present invention provides a multi-task continuous casting billet quality defect prediction method and device based on adaptive loss, which relates to the field of artificial intelligence technology. The method includes: obtaining a continuous casting production sample data set; preprocessing the continuous casting production sample data set to obtain a preprocessed sample data set; constructing an initial multi-task learning model based on TabNet; training the initial multi-task learning model based on TabNet according to the preprocessed sample data set to obtain a trained multi-task learning model based on TabNet; obtaining continuous casting production data to be predicted; preprocessing the continuous casting production data to be predicted to obtain preprocessed continuous casting production data to be predicted; inputting the preprocessed continuous casting production data to be predicted into the trained multi-task learning model based on TabNet to obtain a prediction result of the internal quality of the continuous casting billet. The use of the invention can improve the performance of predicting internal quality defects of continuous casting billets.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly to a multi-task continuous casting billet quality defect prediction method and device based on adaptive loss. Background Art

[0002] The development of the steel industry is closely related to the industrial level of a country. China's steel industry has entered the era of intelligent manufacturing and urgently needs to carry out intelligent transformation and upgrading. In the whole process of steel production, the prediction of tissue properties is a crucial link, covering the quality prediction of continuous casting billets and the mechanical property prediction of rolled products.

[0003] Continuous casting plays a connecting role in the steel production process. The complexity of the continuous casting process makes the produced continuous casting billets prone to surface defects such as surface cracks, slag inclusions, and gas pores, as well as internal defects such as center cracks, segregation, inclusions, porosity, and shrinkage cavities. The occurrence of internal quality defects in continuous casting billets is mainly related to multiple factors. Uneven cooling may lead to internal stress concentration, thereby generating cracks or other defects; secondly, improper control of the casting temperature will affect the fluidity of the molten metal, resulting in problems such as shrinkage cavities and gas pores. In addition, uneven alloy composition will also lead to differences in physical properties and increase the risk of defects. During the casting process, the inhalation of gases such as water vapor and oxygen by the molten metal may form gas pores, and unreasonable mold design may affect the filling and exhaust of the casting, causing defects such as cold shuts. Insufficient fluidity of the molten metal, mold material and design problems, and improper heat treatment process are also potential causes of internal defects. Accurately predicting the quality of continuous casting billets can not only avoid destructive experiments, but also help technicians adjust the process in time, reduce the defective rate, ensure the smooth progress of the subsequent rolling process, and improve product quality. However, due to the non-linear, strongly coupled characteristics and multi-disturbance factors of the continuous casting process, accurately predicting product quality poses great challenges. Therefore, the prediction of internal quality defects of continuous casting billets has always been an important research direction in the continuous casting production process.

[0004] Traditional methods mainly predict the quality of continuous casting billets by establishing mechanism models of the continuous casting process, but there are problems such as low prediction accuracy, complex models, and poor generalization ability. At present, methods for internal quality defects of continuous casting billets have achieved good performance, but they are all for single defect prediction tasks and cannot simultaneously predict multiple quality defects. However, in practical applications, due to the complexity and strong coupling of the production process, a continuous casting billet may have multiple internal quality defects, and predicting one internal quality defect cannot achieve the purpose of detecting defective products. Therefore, how to simultaneously predict multiple internal quality defects, that is, multi-task prediction, and improve the system's detection of defective products is of great significance. The research on multi-task learning started earlier and has formed a relatively mature theoretical system and application examples. The multi-task learning framework first proposed by Caruana laid the foundation for subsequent research; in recent years, with the development of deep learning, many multi-task learning models based on neural networks have been proposed, such as network structures using shared hidden layers and adaptive task weight adjustment strategies, multi-task learning models based on convolutional neural networks and recurrent neural networks, for Chinese text classification and image recognition; in addition, for specific application scenarios, multi-task learning methods combined with domain knowledge have improved the performance of the model in practical applications. At present, no multi-task learning model has been applied to the prediction of internal quality defects of continuous casting billets. Summary of the Invention

[0005] To solve the technical problem in the prior art that multiple internal quality defects of continuous casting billets cannot be predicted simultaneously, an embodiment of the present invention provides a multi-task continuous casting billet quality defect prediction method and device based on adaptive loss. The technical solution is as follows:

[0006] On the one hand, a multi-task continuous casting billet quality defect prediction method based on adaptive loss is provided. This method is implemented by a multi-task continuous casting billet quality defect prediction device based on adaptive loss, and the method includes:

[0007] S1. Obtain a continuous casting production sample data set;

[0008] S2. Preprocess the continuous casting production sample data set to obtain a preprocessed sample data set;

[0009] S3. Construct an initial multi-task learning model based on TabNet;

[0010] S4. Train the initial multi-task learning model based on TabNet according to the preprocessed sample data set to obtain a trained multi-task learning model based on TabNet;

[0011] S5. Obtain the continuous casting production data to be predicted; preprocess the continuous casting production data to be predicted to obtain the preprocessed continuous casting production data to be predicted; input the preprocessed continuous casting production data to be predicted into the trained multi-task learning model based on TabNet to obtain the prediction result of the internal quality of the continuous casting billet.

[0012] Optionally, the preprocessing of the continuous casting production sample data set in S2 to obtain the preprocessed sample data set includes:

[0013] S21. Use a data mining algorithm to remove the missing values and outliers in the continuous casting production sample data set to obtain the continuous casting production sample data set after removal processing;

[0014] S22. Use ONE-HOT encoding to numerically transform the processed continuous casting production sample data set, and normalize the numerically transformed data to obtain the preprocessed sample data set.

[0015] Optionally, the initial multi-task learning model based on TabNet includes:

[0016] A feature selection module, a TabNet shared layer, a splitting layer, and an adaptive loss module;

[0017] Among them, the feature selection module is used to perform feature selection on the input data;

[0018] Among them, the TabNet shared layer is used to extract high-level features;

[0019] Among them, the splitting layer includes two fully connected layers and an MLP classification head;

[0020] Among them, the adaptive loss module is used to balance the training progress of different tasks.

[0021] Optionally, the training of the initial multi-task learning model based on TabNet according to the preprocessed sample data set in S4 to obtain the trained multi-task learning model based on TabNet includes:

[0022] S41. Input the preprocessed sample data set into the initial multi-task learning model based on TabNet, perform feature selection through the feature selection module to obtain the selected feature set;

[0023] S42. Input the selected feature set into the TabNet shared layer to extract high-level features to obtain the feature representation of the high-level features;

[0024] S43. Input the feature representation of the advanced features into the splitting layer, perform calculations through two fully connected layers and an MLP classification head, output the calculation results, and perform task matching based on the calculation results to obtain the prediction result of the quality defects of the continuous casting billet.

[0025] S44. According to the prediction result of the quality defects of the continuous casting billet, train the initial multi-task learning model based on TabNet through the adaptive loss module to obtain the trained multi-task learning model based on TabNet.

[0026] Optionally, in step S41, input the preprocessed sample data set into the initial multi-task learning model based on TabNet, perform feature selection through the feature selection module to obtain the selected feature set, including:

[0027] S411. Input the preprocessed sample data set into the initial multi-task learning model based on TabNet, perform calculations using the random forest algorithm to obtain the feature contribution vector.

[0028] S412. According to the feature contribution vector, set the prior knowledge vector and threshold through the mechanism model, and calculate the feature weights.

[0029] S413. Perform calculations based on the feature weights and the preprocessed sample data set to obtain

[0030] the selected feature set.

[0031] Optionally, in step S44, according to the prediction result of the quality defects of the continuous casting billet, train the initial multi-task learning model based on TabNet through the adaptive loss module, including:

[0032] S441. Initialize the weights.

[0033] S442. According to the initialized weights, calculate the total loss in a single round; according to the total loss in a single round, calculate the gradient norm of a single task for the overall task; according to the gradient norm of a single task for the overall task, calculate the relative direction training speed of a single task.

[0034] S443. According to the gradient norm of a single task for the overall task, calculate the average gradient norm of all tasks.

[0035] S444. According to the average gradient norm of all tasks in a single round, the gradient norm of a single task for the overall task, and the relative direction training speed of a single task, calculate the gradient loss in a single round.

[0036] S445. Calculate the weights for a single round based on the gradient loss in the single round to obtain the weights for the next round; update the parameters by backpropagating the total loss in the single round, and update the weights by normalizing the weights for the next round.

[0037] S446. Determine whether the current situation meets the conditions that the number of training rounds reaches a preset value and the gradient of the total loss decrease is less than a set threshold. If it meets the conditions, stop the training and output the trained model; if it does not meet the conditions, return to step S441 to continue the training until the preset stop condition is met.

[0038] Optionally, inputting the preprocessed continuous casting production data to be predicted into the trained multi-task learning model based on TabNet to obtain the prediction result of the internal quality of the continuous casting billet includes:

[0039] S51. Input the preprocessed continuous casting production data to be predicted into the trained multi-task learning model based on TabNet, calculate the probability of the internal defect level in the data set, and obtain the probability value of the defect level.

[0040] S52. Select the defect level with the highest probability value as the final prediction result of the billet quality.

[0041] On the other hand, a multi-task billet quality defect prediction device based on adaptive loss is provided. The device is applied to the multi-task billet quality defect prediction method based on adaptive loss, and the device includes:

[0042] An acquisition unit, configured to acquire a continuous casting production sample data set;

[0043] A preprocessing unit, configured to preprocess the continuous casting production sample data set to obtain a preprocessed sample data set;

[0044] A construction unit, configured to construct an initial multi-task learning model based on TabNet;

[0045] A training unit, configured to train the initial multi-task learning model based on TabNet according to the preprocessed sample data set to obtain a trained multi-task learning model based on TabNet;

[0046] A prediction unit, configured to acquire continuous casting production data to be predicted; preprocess the continuous casting production data to be predicted to obtain preprocessed continuous casting production data to be predicted; input the preprocessed continuous casting production data to be predicted into the trained multi-task learning model based on TabNet to obtain the prediction result of the internal quality of the continuous casting billet.

[0047] Optionally, the preprocessing unit is configured to:

[0048] Using a data mining algorithm, missing values and outliers in the continuous casting production sample dataset are removed to obtain a processed continuous casting production sample dataset;

[0049] The processed continuous casting production sample dataset is numerically encoded using ONE-HOT encoding, and the numerically encoded data is normalized to obtain a preprocessed sample dataset.

[0050] Optionally, the initial TabNet-based multi-task learning model includes:

[0051] A feature selection module, a TabNet shared layer, a splitting layer, and an adaptive loss module;

[0052] Among them, the feature selection module is used to perform feature selection on the input data;

[0053] Among them, the TabNet shared layer is used to extract high-level features;

[0054] Among them, the splitting layer includes two fully connected layers and an MLP classification head;

[0055] Among them, the adaptive loss module is used to balance the training progress of different tasks.

[0056] Optionally, the training unit is used for:

[0057] Input the preprocessed sample dataset into the initial TabNet-based multi-task learning model, perform feature selection through the feature selection module to obtain a selected feature set;

[0058] Input the selected feature set into the TabNet shared layer to extract high-level features and obtain a feature representation of the high-level features;

[0059] Input the feature representation of the high-level features into the splitting layer, perform calculations through two fully connected layers and an MLP classification head, output the calculation results, and perform task matching according to the calculation results to obtain a prediction result of the quality defect of the billet;

[0060] According to the prediction result of the quality defect of the billet, the initial TabNet-based multi-task learning model is trained through the adaptive loss module to obtain a trained TabNet-based multi-task learning model.

[0061] Optionally, the step of inputting the preprocessed sample dataset into the initial TabNet-based multi-task learning model and performing feature selection through the feature selection module to obtain a selected feature set includes:

[0062] Input the preprocessed sample data set into the initial TabNet - based multi - task learning model, and use the random forest algorithm for calculation to obtain a feature contribution vector;

[0063] According to the feature contribution vector, set the prior knowledge vector and threshold through the mechanism model, and calculate the feature weights;

[0064] Perform calculations based on the feature weights and the preprocessed sample data set to obtain

[0065] A selected good feature set.

[0066] Optionally, training the initial TabNet - based multi - task learning model through the adaptive loss module according to the continuous casting billet quality defect prediction result includes:

[0067] Initialize the weights;

[0068] According to the initialized weights, calculate the total loss in a single round; according to the total loss in a single round, calculate the gradient norm of a single task for the overall task; according to the gradient norm of a single task for the overall task, calculate the relative direction training speed of a single task;

[0069] According to the gradient norm of a single task for the overall task, calculate the average gradient norm of all tasks;

[0070] According to the average gradient norm of all tasks in a single round, the gradient norm of a single task for the overall task, and the relative direction training speed of a single task, calculate the gradient loss in a single round;

[0071] Calculate the weights for a single round according to the gradient loss in a single round to obtain the weights for the next round; update the parameters by back - propagating according to the total loss in a single round, and update the weights by normalizing the weights for the next round;

[0072] Judge whether the current situation meets the conditions that the number of training rounds reaches the preset value and the descending gradient of the total loss is less than the set threshold. If it meets the conditions, stop training and output the trained model; if it does not meet the conditions, return to step S441 to continue training until the preset stop condition is met.

[0073] Optionally, the prediction unit is used for:

[0074] Input the preprocessed continuous casting production data to be predicted into the trained TabNet - based multi - task learning model, calculate the probability of the internal defect level of the data set, and obtain the probability value of the defect level;

[0075] Select the defect level with the highest probability value as the final continuous casting billet quality prediction result.

[0076] On the other hand, there is provided a multi-task slab quality defect prediction device based on adaptive loss. The multi-task slab quality defect prediction device based on adaptive loss includes: a processor; a memory storing computer-readable instructions thereon. When the computer-readable instructions are executed by the processor, any one of the methods in the above-mentioned multi-task slab quality defect prediction method based on adaptive loss is implemented.

[0077] On the other hand, there is provided a computer-readable storage medium storing at least one instruction, and the at least one instruction is loaded and executed by a processor to implement any one of the methods in the above-mentioned multi-task slab quality defect prediction method based on adaptive loss.

[0078] The beneficial effects brought by the technical solutions provided in the embodiments of the present invention at least include:

[0079] In the embodiments of the present invention, first, a continuous casting production sample data set is obtained; the continuous casting production sample data set is preprocessed to obtain a preprocessed sample data set; secondly, an initial multi-task learning model based on TabNet is constructed; the initial multi-task learning model based on TabNet is trained according to the preprocessed sample data set to obtain a trained multi-task learning model based on TabNet; the continuous casting production data to be predicted is obtained; finally, the continuous casting production data to be predicted is preprocessed to obtain preprocessed continuous casting production data to be predicted; the preprocessed continuous casting production data to be predicted is input into the trained multi-task learning model based on TabNet to obtain a prediction result of the internal quality of the continuous casting slab.

[0080] The embodiments of the present invention adopt a prior knowledge-based feature selection method, design a feature selection module, combine the feature contribution degree vectors obtained by prior knowledge and the random forest algorithm to guide feature selection, making the interpretability of feature selection stronger; the multi-task learning architecture based on TabNet can effectively integrate multiple related tasks, utilize the shared features and dynamic feature selection mechanism, thereby reducing the training time of the model and improving the generalization ability of the model. Design an adaptive loss module, which can automatically adjust the weights of different tasks in the total loss according to the training process, effectively optimize the training process of multi-task learning, improve the performance of the model on each task, and at the same time enhance the overall learning efficiency and robustness; a method for selecting the initial value of the multi-task weight is proposed, and the correlation between the target variable and the input features is used as the basis for weight allocation, making the initial value of the weight closer to the optimal solution, reducing the training process, and effectively avoiding the overfitting phenomenon. Using the present invention can improve the performance of predicting the internal quality defects of continuous casting slabs. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0082] Figure 1 is a schematic diagram of the overall structure of a multi-task continuous casting billet quality defect prediction method based on adaptive loss provided by an embodiment of the present invention;

[0083] Figure 2 is a flowchart of a multi-task continuous casting billet quality defect prediction method based on adaptive loss provided by an embodiment of the present invention;

[0084] Figure 3 is a schematic diagram of feature selection provided by an embodiment of the present invention;

[0085] Figure 4 is a schematic diagram of the TabNet structure provided by an embodiment of the present invention;

[0086] Figure 5 is a schematic diagram of weight update provided by an embodiment of the present invention;

[0087] Figure 6 is a block diagram of a multi-task continuous casting billet quality defect prediction device based on adaptive loss provided by an embodiment of the present invention;

[0088] Figure 7 is a schematic diagram of the structure of a multi-task continuous casting billet quality defect prediction device based on adaptive loss provided by an embodiment of the present invention. Detailed Embodiments

[0089] The following describes the technical solutions in the present invention with reference to the accompanying drawings.

[0090] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as "example" in the present invention should not be construed as more preferred or more advantageous than other embodiments or design solutions. Exactly speaking, the use of the word "example" is intended to present concepts in a specific way. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one of the two can be selected.

[0091] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when the difference is not emphasized, their intended meanings are the same. "Of", "corresponding", and "corresponding" can sometimes be used interchangeably. It should be noted that when the difference is not emphasized, their intended meanings are the same.

[0092] In the embodiments of the present invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, their intended meanings are the same.

[0093] To make the technical problems, technical solutions, and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments.

[0094] The embodiments of the present invention provide a multi-task continuous casting billet quality defect prediction method based on adaptive loss. This method can be implemented by a multi-task continuous casting billet quality defect prediction device based on adaptive loss. The multi-task continuous casting billet quality defect prediction device based on adaptive loss can be a terminal or a server. As Figure 1 shown in the overall structural schematic diagram of a multi-task continuous casting billet quality defect prediction method provided by the embodiments of the present invention; as Figure 2 shown in the flowchart of the multi-task continuous casting billet quality defect prediction method based on adaptive loss, the processing flow of this method can include the following steps:

[0095] S1. Obtain a continuous casting production sample data set.

[0096] S2. Preprocess the continuous casting production sample data set to obtain a preprocessed sample data set.

[0097] Optionally, the specific implementation process of S2 can include S21 - S22:

[0098] S21. Use a data mining algorithm to remove missing values and outliers in the continuous casting production sample data set to obtain a processed continuous casting production sample data set;

[0099] Among them, the data mining algorithm can include: decision tree, random forest algorithm, support vector machine, K-nearest neighbor algorithm, and principal component analysis algorithm, etc. Any one of these methods can be selected.

[0100] In a feasible implementation, box plots are used to detect outliers in the continuous casting production sample data set for deletion; a small number of missing data are filled using the grouped filling method, and parts with a large number of missing data are deleted.

[0101] Among them, the box plot and the grouped filling method are conventional technical means, and the present invention will not elaborate further.

[0102] S22. Numerically encode the processed continuous casting production sample data set using ONE-HOT encoding, and normalize the numerically encoded data to obtain a preprocessed sample data set.

[0103] Among them, normalizing the numerically encoded data can remove the interference of the dimension on the prediction of the multi-task learning model based on TabNet.

[0104] Among them, ONE-HOT encoding is a conventional technical means, and the present invention will not elaborate further.

[0105] S3. Construct an initial multi-task learning model based on TabNet.

[0106] Among them, the adaptive attention mechanism of TabNet can efficiently capture feature interactions, can extract the potential features of the data, can provide feature importance evaluation, and can enhance the interpretability of the multi-task learning model based on TabNet. TabNet is fast and accurate in processing sparse data and can be well extended on large-scale data sets. The fewer hyperparameter requirements of TabNet make model tuning simple and suitable for constructing a multi-task learning architecture.

[0107] Optionally, the initial multi-task learning model based on TabNet includes:

[0108] A feature selection module, a TabNet shared layer, a splitting layer, and an adaptive loss module;

[0109] Among them, the feature selection module is used to perform feature selection on the input data;

[0110] Among them, the TabNet shared layer is used to extract high-level features;

[0111] Among them, the splitting layer includes two fully connected layers and an MLP classification head;

[0112] Among them, the adaptive loss module is used to balance the training progress of different tasks.

[0113] S4. Train the initial multi-task learning model based on TabNet according to the preprocessed sample data set to obtain a trained multi-task learning model based on TabNet.

[0114] Among them, the ratio of the training set to the test set is 8:2.

[0115] In a feasible implementation manner, the embodiment of the present invention adopts the Pytorch 1.13.1 learning framework, uses the Adam optimizer, and sets the initial learning rate to 0.02.

[0116] Optionally, the specific implementation process of S4 may include S41 - S44:

[0117] S41. Input the pre - processed sample data set into the initial multi - task learning model based on TabNet, and perform feature selection through the feature selection module to obtain a selected feature set;

[0118] Among them, as Figure 3 shown is the feature selection schematic diagram provided by the embodiment of the present invention; the input features are selected through the feature selection module.

[0119] Optionally, the specific implementation process of S41 may include S411 - S413:

[0120] S411. Input the pre - processed sample data set into the initial multi - task learning model based on TabNet, and use the random forest algorithm for calculation to obtain a feature contribution vector;

[0121] Among them, the parameters of the random forest are searched by the grid search method, and the optimal result is retained.

[0122] Among them, the random forest algorithm is a conventional technical means and will not be further elaborated in the present invention.

[0123] S412. According to the feature contribution vector, set the prior knowledge vector and threshold through the mechanism model, and calculate the feature weights;

[0124] S413. Calculate according to the feature weights and the pre - processed sample data set to obtain

[0125] a selected feature set.

[0126] In a feasible implementation manner, the process of obtaining the selected feature set can be represented by the following formula (1) and formula (2):

[0127] (1)

[0128] (2)

[0129] Among them, FW represents the feature weight; FI represents the feature contribution degree vector; PK represents the prior knowledge vector; represents the set threshold, and the threshold set in the present invention is 15. input represents the pre - processed sample data set; Output represents the selected feature set.

[0130] In a feasible implementation manner, prior knowledge is combined with the feature contribution degree vector to jointly guide feature selection, which can reduce noise and redundant features and improve the interpretability of feature selection.

[0131] S42. Input the selected feature set into the TabNet shared layer to extract high-level features and obtain the feature representation of the high-level features.

[0132] In a feasible implementation manner, a TabNet shared layer is set, and a splitting layer with the number of tasks is set after the TabNet shared layer. As Figure 4 is the schematic diagram of the TabNet structure provided by the embodiment of the present invention; TabNet first encodes the input tabular data into a feature vector; the feature vector is sequentially input into multiple stacked Blocks for processing; among them, the number of Blocks is set to 3-5, and the present invention sets the number of Blocks to 3; among them, the output of the previous Block is used as the input of the next Block, and the structure of each Block is the same. The output of the last Block is input into the MLP classification head to obtain the classification result; among them, in each Block, the input features first pass through a masking layer, and the masking layer receives the information of the previous Block and automatically adjusts the selected features; the selected features pass through a feature transformer, splitting and activation in sequence, and continue to be calculated in the subsequent Blocks. Among them, part of the information after splitting passes through an attention transformer and is sent to the masking layer of the next Block for feature selection. The splitting layer uses 2 fully connected layers, matches the number of neurons with the task, predicts the quality of the continuous casting billet, and obtains the prediction results for different quality defects and defect levels.

[0133] S43. Input the feature representation of the high-level features into the splitting layer, perform calculations through two fully connected layers and the MLP classification head, output the calculation results, perform task matching according to the calculation results, and obtain the prediction results of the billet quality defects.

[0134] S44. According to the prediction results of the billet quality defects, train the initial TabNet-based multi-task learning model through the adaptive loss module to obtain the trained TabNet-based multi-task learning model.

[0135] Among them, during the training process, the adaptive loss module will update the weights at any time according to the gradients of the loss drops of different tasks.

[0136] In a feasible implementation manner, the adaptive loss module can ensure that the model adaptively focuses on the tasks with poor current performance during the training process and promotes the improvement of the overall performance by dynamically adjusting the weight of the single-task loss in the total loss. The adaptive loss module is not only applicable to the scenario of multi-task learning, but also applicable to other deep learning tasks that require dynamic adjustment of the loss weight.

[0137] Optionally, the specific implementation process of S44 may include S441 - S446:

[0138] S441. Initialize the weights;

[0139] In a feasible implementation manner, perform Pearson correlation analysis on the obtained input features and output features to obtain the correlation between each single-target variable and the input features, and determine the initial weights according to the weighted average correlation. Among them, the initial weights can be represented by the following formula (3):

[0140] (3)

[0141] Among them, represents the sum of the absolute values of the correlations between the i-th task target variable and the input features, and n represents the number of single tasks included in multi-task learning. represents the initial weights.

[0142] Among them, Pearson correlation analysis is a technical means mastered by those skilled in the art, and the present invention will not be further elaborated here.

[0143] Among them, by analyzing the mutual relationship between tasks, the setting of the initial weights is optimized to improve the convergence speed and overall performance of the model. Through the correlation analysis, appropriate initial weights can be provided for different tasks, promoting the effective progress of the training process.

[0144] S442. Calculate the total loss in a single round according to the initial weights; calculate the gradient norm of a single task for the overall task according to the total loss in a single round; calculate the relative direction training speed of a single task according to the gradient norm of a single task for the overall task.

[0145] In a feasible implementation manner, in the t-th round, calculate the total loss, which can be represented by the following formula (4):

[0146] (4)

[0147] Among them, represents the total loss in the t-th round; represents the loss of the i-th task; represents the weight of the loss of the i-th task in the total loss.

[0148] In a feasible implementation manner, the gradient norm of the i-th task with respect to the whole can be expressed by the following formula (5):

[0149] (5)

[0150] where represents the gradient norm of the i-th task with respect to the whole.

[0151] In a feasible implementation manner, the relative direction training speed of a single task can be expressed by the following formula (6):

[0152] (6)

[0153] where represents the loss ratio of task i at step t, represents the average value of the loss ratios of all tasks at step t; represents the relative backward training speed of task i; represents the initial loss of the i-th task.

[0154] S443. Calculate the average gradient norm of all tasks according to the gradient norm of a single task with respect to the whole task;

[0155] In a feasible implementation manner, the average gradient norm of all tasks can be expressed by the following formula (7):

[0156] (7)

[0157] where represents the average gradient norm of all tasks in the t-th round.

[0158] S444. Calculate the gradient loss in a single round according to the average gradient norm of all tasks in a single round, the gradient norm of a single task with respect to the whole task, and the relative direction training speed of a single task;

[0159] In a feasible implementation manner, the gradient loss in a single round can be expressed by the following formula (8):

[0160] (8)

[0161] where represents the gradient loss in the t-th round; represents an additional hyperparameter, which is set according to different tasks. Usually, a larger value is selected to enforce a stronger training rate balance. In the present invention, is set to 1.5.

[0162] S445. Calculate the weights for a single round based on the gradient loss in the single round to obtain the weights for the next round; update the parameters by backpropagating the total loss in the single round, and update the weights by normalizing the weights for the next round.

[0163] Such as Figure 5 FIG. is a schematic diagram of weight update provided by an embodiment of the present invention. In a feasible implementation, calculate the partial derivative of the gradient loss of the t-th round with respect to the weight of the i-th task at the t-th round to obtain the weight of the i-th task at the (t + 1)-th round.

[0164] In a feasible implementation, the normalization of the weights for the next round can be represented by the following formula (9):

[0165] (9)

[0166] S446. Determine whether the current training round reaches a preset value and the gradient of the total loss decrease is less than a set threshold. If satisfied, stop training and output the trained model; if not satisfied, return to step S441 to continue training until the preset stop condition is met.

[0167] S5. Obtain the continuous casting production data to be predicted; preprocess the continuous casting production data to be predicted to obtain the preprocessed continuous casting production data to be predicted; input the preprocessed continuous casting production data to be predicted into the trained multi-task learning model based on TabNet to obtain the prediction result of the internal quality of the continuous casting billet.

[0168] Among them, the continuous casting production data to be predicted is the actual production data during the continuous casting process of a steel plant. Among them, the output features include: the chemical composition of the continuous casting billet, the process parameters and equipment information during the continuous casting process. Select the two labels of central porosity and shrinkage cavity of the continuous casting billet as the target variables for training to verify the effectiveness of the model.

[0169] Optionally, inputting the preprocessed continuous casting production data to be predicted into the trained multi-task learning model based on TabNet to obtain the prediction result of the internal quality of the continuous casting billet in S5 includes:

[0170] S51. Input the preprocessed continuous casting production data to be predicted into the trained multi-task learning model based on TabNet, calculate the probability of the internal defect level in the data set, and obtain the probability value of the defect level;

[0171] S52. Select the defect level with the highest probability value as the final prediction result of the quality of the continuous casting billet.

[0172] In a feasible implementation, the true labels of the test set are obtained and compared with the continuous casting billet quality prediction results to calculate the accuracy of the multi-task learning model based on TabNet. The samples with incorrect predictions are analyzed to identify the reasons for the poor performance of the multi-task learning model based on TabNet, including problems such as data quality, feature selection, and model complexity. Through the analysis results, the misclassified samples are counted, and the common points of the misclassified samples are found through k-means clustering to identify the shortcomings of the model. The feature selection and model complexity are adjusted according to the shortcomings to further improve the prediction ability and accuracy of the model.

[0173] Among them, the present invention adopts the commonalities of different internal quality defect problems, can share feature representations, reduce redundant modeling, can maintain the prediction accuracy while effectively reducing the training cost and inference time, and provides an efficient and feasible technical path for solving the internal quality defect problems of continuous casting billets.

[0174] Among them, constructing a multi-task learning model based on TabNet has significant advantages, especially in processing tabular data; TabNet is a deep learning model based on the attention mechanism, which can achieve efficient representation learning in the feature selection and modeling process and is suitable for multi-task learning scenarios; by applying TabNet to the multi-task learning model, the prediction accuracy and reliability of different internal quality defects can be significantly improved. Constructing a multi-task learning model based on TabNet can not only provide efficient real-time monitoring, but also provide data support for subsequent quality improvement and control measures, promoting the double improvement of production efficiency and product quality.

[0175] In the embodiment of the present invention, first, a continuous casting production sample data set is obtained; the continuous casting production sample data set is preprocessed to obtain a preprocessed sample data set; secondly, an initial multi-task learning model based on TabNet is constructed; the initial multi-task learning model based on TabNet is trained according to the preprocessed sample data set to obtain a trained multi-task learning model based on TabNet; the continuous casting production data to be predicted is obtained; finally, the continuous casting production data to be predicted is preprocessed to obtain preprocessed continuous casting production data to be predicted; the preprocessed continuous casting production data to be predicted is input into the trained multi-task learning model based on TabNet to obtain the internal quality prediction result of the continuous casting billet.

[0176] The embodiment of the present invention adopts a feature selection method based on prior knowledge, designs a feature selection module, combines the feature contribution degree vectors obtained by prior knowledge and the random forest algorithm to guide feature selection, making the interpretability of feature selection stronger; the multi-task learning architecture based on TabNet can effectively integrate multiple related tasks, utilize shared features and dynamic feature selection mechanisms, thereby reducing the training time of the model and improving the generalization ability of the model. An adaptive loss module is designed, which can automatically adjust the weights of different tasks in the total loss according to the training process, effectively optimize the training process of multi-task learning, improve the performance of the model on each task, and enhance the overall learning efficiency and robustness at the same time; a method for selecting the initial value of multi-task weights is proposed, and the correlation between the target variable and the input features is used as the basis for allocating weights, making the initial value of the weights closer to the optimal solution, reducing the training process, and effectively avoiding the overfitting phenomenon. The use of the present invention can improve the performance of predicting internal quality defects of continuous casting billets.

[0177] Figure 6 It is a block diagram of a multi-task continuous casting billet quality defect prediction device based on adaptive loss shown according to an exemplary embodiment. This device is used for the multi-task continuous casting billet quality defect prediction method based on adaptive loss. Refer to Figure 6 , this device includes an acquisition unit 610, a preprocessing unit 620, a construction unit 630, a training unit 640, and a prediction unit 650. Among them:

[0178] The acquisition unit 610 is used to acquire a continuous casting production sample data set;

[0179] The preprocessing unit 620 is used to preprocess the continuous casting production sample data set to obtain a preprocessed sample data set;

[0180] The construction unit 630 is used to construct an initial multi-task learning model based on TabNet;

[0181] The training unit 640 is used to train the initial multi-task learning model based on TabNet according to the preprocessed sample data set to obtain a trained multi-task learning model based on TabNet;

[0182] The prediction unit 650 is used to acquire the continuous casting production data to be predicted; preprocess the continuous casting production data to be predicted to obtain preprocessed continuous casting production data to be predicted; input the preprocessed continuous casting production data to be predicted into the trained multi-task learning model based on TabNet to obtain the prediction result of the internal quality of the continuous casting billet.

[0183] Optionally, the preprocessing unit 620 is used for:

[0184] Adopt a data mining algorithm to remove missing values and outliers from the continuous casting production sample dataset, and obtain a processed continuous casting production sample dataset.

[0185] Use ONE-HOT encoding to numerically transform the processed continuous casting production sample dataset, and normalize the numerically transformed data to obtain a preprocessed sample dataset.

[0186] Optionally, the initial TabNet-based multi-task learning model includes:

[0187] A feature selection module, a TabNet shared layer, a splitting layer, and an adaptive loss module;

[0188] Among them, the feature selection module is used to perform feature selection on the input data;

[0189] Among them, the TabNet shared layer is used to extract high-level features;

[0190] Among them, the splitting layer includes two fully connected layers and an MLP classification head;

[0191] Among them, the adaptive loss module is used to balance the training progress of different tasks.

[0192] Optionally, the training unit 640 is used to:

[0193] Input the preprocessed sample dataset into the initial TabNet-based multi-task learning model, perform feature selection through the feature selection module, and obtain a selected feature set;

[0194] Input the selected feature set into the TabNet shared layer to extract high-level features and obtain a feature representation of the high-level features;

[0195] Input the feature representation of the high-level features into the splitting layer, perform calculations through two fully connected layers and an MLP classification head, output the calculation results, and perform task matching according to the calculation results to obtain a prediction result of the quality defect of the casting blank;

[0196] According to the prediction result of the quality defect of the casting blank, train the initial TabNet-based multi-task learning model through the adaptive loss module to obtain a trained TabNet-based multi-task learning model.

[0197] Optionally, the step of inputting the preprocessed sample dataset into the initial TabNet-based multi-task learning model, performing feature selection through the feature selection module, and obtaining a selected feature set includes:

[0198] Input the preprocessed sample data set into the initial TabNet-based multi-task learning model, and use the random forest algorithm for calculation to obtain a feature contribution vector;

[0199] According to the feature contribution vector, set the prior knowledge vector and threshold through the mechanism model, and calculate the feature weights;

[0200] Perform calculations based on the feature weights and the preprocessed sample data set to obtain

[0201] A selected set of good features.

[0202] Optionally, the training of the initial TabNet-based multi-task learning model through the adaptive loss module according to the continuous casting billet quality defect prediction result includes:

[0203] Initialize the weights;

[0204] According to the initialized weights, calculate the total loss in a single round; according to the total loss in a single round, calculate the gradient norm of a single task for the overall task; according to the gradient norm of a single task for the overall task, calculate the relative direction training speed of a single task;

[0205] According to the gradient norm of a single task for the overall task, calculate the average gradient norm of all tasks;

[0206] According to the average gradient norm of all tasks in a single round, the gradient norm of a single task for the overall task, and the relative direction training speed of a single task, calculate the gradient loss in a single round;

[0207] Calculate the weights for a single round according to the gradient loss in a single round to obtain the weights for the next round; backpropagate and update the parameters according to the total loss in a single round, and update the weights by normalizing the weights for the next round;

[0208] Judge whether the current situation meets the conditions that the number of training rounds reaches the preset value and the descending gradient of the total loss is less than the set threshold. If it is satisfied, stop training and output the trained model; if it is not satisfied, return to step S441 to continue training until the preset stop condition is met.

[0209] Optionally, the prediction unit 650 is used for:

[0210] Input the preprocessed continuous casting production data to be predicted into the trained TabNet-based multi-task learning model, calculate the probability of the internal defect level existing in the data set, and obtain the probability value of the defect level;

[0211] Select the defect level with the highest probability value as the final continuous casting billet quality prediction result.

[0212] In an embodiment of the present invention, a continuous casting production sample data set is first obtained; the continuous casting production sample data set is preprocessed to obtain a preprocessed sample data set; secondly, an initial multi-task learning model based on TabNet is constructed; the initial multi-task learning model based on TabNet is trained according to the preprocessed sample data set to obtain a trained multi-task learning model based on TabNet; the continuous casting production data to be predicted is obtained; finally, the continuous casting production data to be predicted is preprocessed to obtain preprocessed continuous casting production data to be predicted; the preprocessed continuous casting production data to be predicted is input into the trained multi-task learning model based on TabNet to obtain a prediction result of the internal quality of the continuous casting slab.

[0213] The embodiment of the present invention adopts a feature selection method with prior knowledge, designs a feature selection module, combines the feature contribution degree vectors obtained by prior knowledge and the random forest algorithm to guide feature selection, making the interpretability of feature selection stronger; the multi-task learning architecture based on TabNet can effectively integrate multiple related tasks, utilize shared features and dynamic feature selection mechanisms, thereby reducing the training time of the model and improving the generalization ability of the model. An adaptive loss module is designed, which can automatically adjust the weights of different tasks in the total loss according to the training process, effectively optimize the training process of multi-task learning, improve the performance of the model on each task, and enhance the overall learning efficiency and robustness at the same time; a method for selecting the initial value of multi-task weights is proposed, and the correlation between the target variable and the input features is used as the basis for allocating weights, making the initial value of the weights closer to the optimal solution, reducing the training process, and effectively avoiding the overfitting phenomenon. Using the present invention can improve the performance of predicting the internal quality defects of continuous casting slabs.

[0214] Figure 7 FIG. is a schematic structural diagram of a multi-task casting slab quality defect prediction device based on adaptive loss provided by an embodiment of the present invention, as Figure 7 shown, the multi-task casting slab quality defect prediction device based on adaptive loss may include the above-mentioned Figure 6 multi-task casting slab quality defect prediction device shown. Optionally, the multi-task casting slab quality defect prediction device 710 based on adaptive loss may include a first processor 2001.

[0215] Optionally, the multi-task casting slab quality defect prediction device 710 based on adaptive loss may further include a memory 2002 and a transceiver 2003.

[0216] Among them, the first processor 2001 is connected to the memory 2002 and the transceiver 2003, such as through a communication bus.

[0217] The following combines Figure 7Specifically introduce each component of the multi-task billet quality defect prediction device 710 based on adaptive loss:

[0218] Among them, the first processor 2001 is the control center of the multi-task billet quality defect prediction device 710 based on adaptive loss, which can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 is one or more central processing units (CPUs), or can be an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention, such as: one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs).

[0219] Optionally, the first processor 2001 can execute various functions of the multi-task billet quality defect prediction device 710 based on adaptive loss by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0220] In a specific implementation, as an embodiment, the first processor 2001 may include one or more CPUs, such as Figure 7 CPU0 and CPU1 shown in

[0221] In a specific implementation, as an embodiment, the multi-task billet quality defect prediction device 710 based on adaptive loss may also include multiple processors, such as Figure 7 the first processor 2001 and the second processor 2004 shown in

[0222] Among them, the memory 2002 is used to store the software program for executing the solution of the present invention and is controlled by the first processor 2001 for execution. The specific implementation manner can refer to the above method embodiments and will not be elaborated here.

[0223] Optionally, the memory 2002 may be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or may also be an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM), or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media, or other magnetic storage devices, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 2002 may be integrated with the first processor 2001 or may exist independently and be coupled to the first processor 2001 through an interface circuit ( Figure 7 not shown) of the multi-task billet quality defect prediction device 710 based on adaptive loss. The embodiments of the present invention do not make specific limitations on this.

[0224] The transceiver 2003 is used to communicate with network devices or with terminal devices.

[0225] Optionally, the transceiver 2003 may include a receiver and a transmitter ( Figure 7 not shown separately). Among them, the receiver is used to implement the receiving function, and the transmitter is used to implement the sending function.

[0226] Optionally, the transceiver 2003 may be integrated with the first processor 2001 or may exist independently and be coupled to the first processor 2001 through an interface circuit ( Figure 7 not shown) of the multi-task billet quality defect prediction device 710 based on adaptive loss. The embodiments of the present invention do not make specific limitations on this.

[0227] It should be noted that Figure 7 the structure of the multi-task billet quality defect prediction device 710 based on adaptive loss shown in does not constitute a limitation on the router. The actual knowledge structure recognition device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0228] In addition, the technical effects of the multi-task billet quality defect prediction device 710 based on adaptive loss may refer to the technical effects of the multi-task billet quality defect prediction method based on adaptive loss described in the above method embodiments, and will not be elaborated here.

[0229] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0230] It should also be understood that the memory in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0231] The above-described embodiments can be implemented in whole or in part by software, hardware (such as circuits), firmware, or any other combination. When implemented using software, the above-described embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present invention are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wired (such as infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that contains one or more collections of available media. The available media can be magnetic media (such as floppy disks, hard disks, magnetic tapes), optical media (such as DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0232] It should be understood that the term "and / or" in this document is merely a description of the association relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. Here, A and B can be singular or plural. In addition, the character " / " in this document generally represents an "or" relationship between the associated objects before and after, but it may also represent an "and / or" relationship, which can be specifically understood with reference to the context before and after.

[0233] In the present invention, "at least one" means one or more, and "a plurality" means two or more. "At least one of the following" or its similar expressions refer to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, or c can represent: a, b, c, a - b, a - c, b - c, or a - b - c, where a, b, and c can be single or multiple.

[0234] It should be understood that in various embodiments of the present invention, the magnitudes of the sequence numbers of the above processes do not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0235] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0236] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the above-described devices, apparatuses, and units can refer to the corresponding processes in the foregoing method embodiments, and will not be described herein again.

[0237] In several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there can be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0238] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they can be located in one place, or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0239] In addition, the functional units in various embodiments of the present invention can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit.

[0240] When the above-mentioned functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks, or optical discs.

[0241] As described above, the above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A multi-task continuous casting billet quality defect prediction method based on adaptive loss, characterized in that The method includes: S1. Obtain a continuous casting production sample data set; S2. Preprocess the continuous casting production sample data set to obtain a preprocessed sample data set; S3. Construct an initial multi-task learning model based on TabNet, including: A feature selection module, a TabNet shared layer, a splitting layer, and an adaptive loss module; Among them, the feature selection module is used to perform feature selection on the input data; Among them, the TabNet shared layer is used to extract high-level features; Among them, the splitting layer includes two fully connected layers and an MLP classification head; Among them, the adaptive loss module is used to balance the training progress of different tasks; S4. Train the initial multi-task learning model based on TabNet according to the preprocessed sample data set to obtain a trained multi-task learning model based on TabNet, including: S41. Input the preprocessed sample data set into the initial multi-task learning model based on TabNet, perform feature selection through the feature selection module to obtain a selected feature set; S42. Input the selected feature set into the TabNet shared layer to extract high-level features and obtain a feature representation of the high-level features; S43. Input the feature representation of the high-level features into the splitting layer, perform calculations through two fully connected layers and the MLP classification head, output the calculation results, and perform task matching according to the calculation results to obtain a prediction result of the quality defect of the continuous casting billet; S44. According to the prediction result of the quality defect of the continuous casting billet, train the initial multi-task learning model based on TabNet through the adaptive loss module to obtain a trained multi-task learning model based on TabNet, including: S441. Initialize the weights; S442. According to the initialized weights, calculate the total loss in a single round; according to the total loss in a single round, calculate the gradient norm of a single task with respect to the overall task; according to the gradient norm of a single task with respect to the overall task, calculate the relative direction training speed of a single task; S443. According to the gradient norm of a single task with respect to the overall task, calculate the average gradient norm of all tasks; S444. According to the average gradient norm of all tasks in a single round, the gradient norm of a single task with respect to the overall task, and the relative direction training speed of a single task, calculate the gradient loss in a single round; S445. Calculate the weights for a single round according to the gradient loss in a single round to obtain the weights for the next round; backpropagate and update the parameters according to the total loss in a single round, and update the weights by normalizing the weights for the next round; S446. Determine whether the current situation meets the conditions that the number of training rounds reaches a preset value and the gradient of the total loss decrease is less than a set threshold. If it meets the conditions, stop training and output the trained model; if it does not meet the conditions, return to step S441 to continue training until the preset stop condition is met; S5. Obtain the continuous casting production data to be predicted; preprocess the continuous casting production data to be predicted to obtain the preprocessed continuous casting production data to be predicted; input the preprocessed continuous casting production data to be predicted into the trained multi-task learning model based on TabNet to obtain the prediction result of the internal quality of the continuous casting billet.

2. The multi-task slab quality defect prediction method based on adaptive loss according to claim 1, wherein The preprocessing of the continuous casting production sample data set in S2 to obtain the preprocessed sample data set includes: S21. Use a data mining algorithm to remove the missing values and outliers in the continuous casting production sample data set to obtain the continuous casting production sample data set after the removal process; S22. Use ONE-HOT encoding to numerically transform the processed continuous casting production sample data set, and normalize the numerically transformed data to obtain the preprocessed sample data set.

3. The multi-task slab quality defect prediction method based on adaptive loss according to claim 1, wherein, The input of the preprocessed sample data set in S41 into the initial multi-task learning model based on TabNet to perform feature selection through the feature selection module to obtain the selected feature set includes: S411. Input the preprocessed sample data set into the initial multi-task learning model based on TabNet, and use the random forest algorithm to calculate to obtain the feature contribution vector; S412. According to the feature contribution vector, set the prior knowledge vector and threshold through the mechanism model, and calculate the feature weights; S413. Calculate according to the feature weights and the preprocessed sample data set to obtain The selected feature set.

4. The method for predicting multi-task slab quality defects based on adaptive loss according to claim 1, wherein The input of the preprocessed continuous casting production data to be predicted in S5 into the trained multi-task learning model based on TabNet to obtain the prediction result of the internal quality of the continuous casting billet includes: S51. Input the preprocessed continuous casting production data to be predicted into the trained multi-task learning model based on TabNet, calculate the probability of the internal defect level of the data set, and obtain the probability value of the defect level; S52. Select the defect level with the highest probability value as the final prediction result of the billet quality.

5. A multi-task continuous casting billet quality defect prediction device based on adaptive loss, which is used to implement the multi-task continuous casting billet quality defect prediction method based on adaptive loss as described in any one of claims 1-4, characterized in that, The device includes: An acquisition unit for acquiring the continuous casting production sample data set; A preprocessing unit for preprocessing the continuous casting production sample data set to obtain the preprocessed sample data set; A construction unit for constructing an initial multi-task learning model based on TabNet; A training unit for training the initial multi-task learning model based on TabNet according to the preprocessed sample data set to obtain the trained multi-task learning model based on TabNet; A prediction unit for acquiring the continuous casting production data to be predicted; preprocessing the continuous casting production data to be predicted to obtain the preprocessed continuous casting production data to be predicted; inputting the preprocessed continuous casting production data to be predicted into the trained multi-task learning model based on TabNet to obtain the prediction result of the internal quality of the continuous casting billet.

6. A multi-task billet quality defect prediction device based on adaptive loss, characterized in that, The multi-task billet quality defect prediction device based on adaptive loss includes: A processor; A memory, on which computer-readable instructions are stored, and when the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 4 is implemented.

7. A computer-readable storage medium, characterized in that, Program code is stored in the computer-readable storage medium, and the program code can be called by the processor to execute the method according to any one of claims 1 to 4.

Images