AI model training method and system based on LF refining process

Abstract

The present application provides an AI model training method and system based on LF refining process, wherein the AI ​​model contains an LF refining process anomaly diagnosis network, which is composed of x graph convolution construction units (x≥2) and y graph convolution restoration units (y≥3) linked in sequence, and each unit contains a heuristic search function layer and a graph convolution function layer linked in sequence. The method first obtains the template refining process monitoring data containing process parameters, raw material characteristics and equipment status data, performs feature processing through the graph convolution construction unit and graph convolution restoration through the graph convolution restoration unit, estimates the template anomaly diagnosis result based on the graph convolution restoration result, and then trains the network neuron weight information according to the loss function value of the labeled anomaly diagnosis result and the template anomaly diagnosis result. The method can effectively process complex data and improve the accuracy of anomaly diagnosis of LF refining process.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to an AI model training method and system based on LF refining process. Background Art

[0002] In the steel production process, the LF refining process plays a vital role. It can effectively improve the quality of molten steel and optimize the performance of steel. However, the LF refining process is complex, involving many links, and is affected by a variety of factors, such as fluctuations in process parameters, differences in raw material properties, and changes in equipment status, which makes the process prone to abnormalities during actual operation.

[0003] At present, in the field of abnormal diagnosis of LF refining process, traditional methods mainly rely on manual experience and simple threshold judgment. Operators rely on their long-term experience and some fixed indicator thresholds to analyze the data in the process to determine whether there are abnormalities. This method has great limitations. On the one hand, manual experience is highly subjective and has individual differences. Different operators may come to different judgment results, making it difficult to ensure the accuracy of diagnosis; on the other hand, simple threshold judgment cannot cope with the complex and changeable actual production situation. Some subtle abnormal changes are often not detected in time, which can easily cause production accidents and affect product quality and production efficiency.

[0004] With the development of information technology, some existing technologies have tried to introduce some basic data analysis technologies to perform abnormality diagnosis, such as methods based on statistical analysis. These methods identify abnormalities by statistically analyzing historical data and establishing corresponding models. However, most of them can only handle simple data relationships. It is difficult to effectively mine the inherent correlation between complex multi-source heterogeneous data in the LF refining process, such as process parameter data, raw material characteristic data, and equipment status data. It is impossible to fully and accurately reflect the true state of the process, and the diagnostic effect is unsatisfactory.

[0005] In addition, there are some preliminary attempts based on machine learning, but the model structure adopted is relatively simple, usually only processing a single type of data, and failing to fully consider the diversity and complexity of LF refining process data. Moreover, these models lack effective feature extraction and data restoration mechanisms, making it difficult to extract valuable information from massive amounts of data, resulting in low diagnostic accuracy and reliability in the face of complex abnormal situations, and unable to meet the high-quality diagnostic needs in actual production.

[0006] In other words, the existing technology has obvious deficiencies in accuracy, comprehensiveness, and the ability to process complex data. There is an urgent need for a technical approach that can effectively integrate multi-source data, deeply mine data features, and accurately diagnose abnormal situations in order to improve the stability of the LF refining process and product quality. Summary of the invention

[0007] In view of the above-mentioned problems, in combination with the first aspect of the present application, an embodiment of the present application provides an AI model training method based on an LF refining process, wherein the AI ​​model includes an LF refining process abnormality diagnosis network, wherein the LF refining process abnormality diagnosis network includes x sequentially linked graph convolution construction units and y sequentially linked graph convolution restoration units, each of the graph convolution construction units and each of the graph convolution restoration units includes a sequentially linked heuristic search function layer and a graph convolution function layer, x is a positive integer not less than 2, and y is a positive integer not less than 3, and the method includes:

[0008] Acquiring template refining process monitoring data, wherein the template refining process monitoring data includes process parameter data, raw material characteristic data, and equipment status data;

[0009] The template refining process monitoring data is loaded into the x graph convolution construction units, and each of the x graph convolution construction units is used to perform feature processing respectively to generate graph convolution construction results extracted by the x graph convolution construction units; wherein each of the graph convolution construction units is used to perform feature processing respectively using the heuristic search function layer and the graph convolution function layer, and to process the graph convolution encoding result of the graph convolution construction unit using the supplementary links of the forward access node and the backward export node of the heuristic search function layer;

[0010] The graph convolution construction result is loaded into the y graph convolution restoration units, and each of the y graph convolution restoration units is used to perform graph convolution restoration respectively, to generate graph convolution restoration results output by the y graph convolution restoration units; wherein each of the graph convolution restoration units is used to perform graph convolution restoration respectively using the heuristic search function layer and the graph convolution function layer, and to process the graph convolution restoration result of the graph convolution restoration unit using the supplementary link;

[0011] estimating a template anomaly diagnosis result of the template refining process monitoring data based on the graph convolution restoration result;

[0012] Based on the labeled abnormality diagnosis result of the template refining process monitoring data and the loss function value of the template abnormality diagnosis result, the neuron weight information of the LF refining process abnormality diagnosis network is trained.

[0013] On the other hand, an embodiment of the present application further provides an artificial intelligence system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0014] Based on the above aspects, by setting up a LF refining process anomaly diagnosis network containing a specific number of graph convolution construction units (x, x is a positive integer not less than 2) and graph convolution restoration units (y, y is a positive integer not less than 3), and each unit is composed of a heuristic search function layer and a graph convolution function layer sequentially linked, and the processing results are optimized by using the supplementary links of the heuristic search function layer, it is possible to perform comprehensive and in-depth feature processing and graph convolution restoration on the template refining process monitoring data. Specifically, after obtaining the template refining process monitoring data covering process parameter data, raw material characteristic data and equipment status data, multiple graph convolution construction units can effectively extract data features respectively and generate graph convolution construction results; subsequent multiple graph convolution restoration units further restore the construction results, and can more accurately estimate the template abnormality diagnosis results based on the graph convolution restoration results, greatly improving the accuracy and reliability of LF refining process abnormality diagnosis, and can effectively identify abnormal situations in the LF refining process, providing strong support for timely measures to avoid production accidents and ensure production quality; at the same time, training the network neuron weight information based on the loss function value can continuously optimize the network performance and improve its adaptability and effectiveness in LF refining process abnormality diagnosis, thereby significantly improving the intelligence level and production efficiency of the entire LF refining process. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the execution flow of the AI ​​model training method based on the LF refining process provided in an embodiment of the present application.

[0016] Figure 2 It is a schematic diagram of the hardware architecture of the artificial intelligence system provided in the embodiment of the present application. DETAILED DESCRIPTION

[0017] The present application will be described in detail below with reference to the accompanying drawings. Figure 1It is a flow chart of an AI model training method based on LF refining process provided by an embodiment of the present application. The AI ​​model training method based on LF refining process is introduced in detail below, wherein the AI ​​model includes an LF refining process abnormality diagnosis network, and the LF refining process abnormality diagnosis network includes x sequentially linked graph convolution construction units and y sequentially linked graph convolution restoration units, each of the graph convolution construction unit and each of the graph convolution restoration unit includes a sequentially linked heuristic search function layer and a graph convolution function layer, x is a positive integer not less than 2, and y is a positive integer not less than 3.

[0018] Step S110, acquiring template refining process monitoring data, wherein the template refining process monitoring data includes process parameter data, raw material characteristic data and equipment status data.

[0019] In this embodiment, in a large steel production enterprise, the LF (Ladle Furnace) refining process is an important link in the steel production process. In order to ensure the normal operation of the LF refining process and to detect possible abnormal situations in a timely manner, it is necessary to monitor it. For the acquisition of template refining process monitoring data, the process parameter data contains a lot of numerical information related to the refining process. For example, the temperature changes during the refining process, from the temperature when the molten steel is initially poured into the LF furnace, to the heating and cooling rates at different stages of the refining process. For example, at the beginning of refining, the temperature of the molten steel is 1550 degrees Celsius. After a specific heating operation, the temperature rises by 10-15 degrees Celsius every 5 minutes. These data accurately reflect the heat transfer during the refining process.

[0020] Raw material characteristic data also plays a key role. The raw materials for steel production are mainly iron ore, scrap steel, etc. The composition, purity and other characteristics of these raw materials will affect the refining effect. Taking iron ore as an example, its iron content, impurity content (such as sulfur, phosphorus, etc.) and other data are recorded in the raw material characteristic data. If the sulfur content in the iron ore is too high, special desulfurization operations are required during the refining process, which is closely related to the normal operation of the refining process.

[0021] The equipment status data covers the operating status information of each part of the LF refining equipment. For example, the electrode status of the LF furnace, the electrode consumption rate, whether there is a short circuit risk, etc. If the electrode is consumed too quickly, it may mean that the arc between the electrode and the molten steel is unstable, which will affect the heating efficiency and the refining quality of the molten steel. There are also data such as the refractory condition of the furnace body, the degree of wear of the refractory material, the remaining thickness, etc. If the refractory material is severely worn, it may cause serious problems such as excessive heat loss or molten steel leakage. These process parameter data, raw material characteristic data, and equipment status data together constitute the template refining process monitoring data, which provides a data basis for subsequent abnormal diagnosis.

[0022] Step S120, loading the template refining process monitoring data into the x graph convolution construction units, using each of the x graph convolution construction units to perform feature processing respectively, and generating graph convolution construction results extracted by the x graph convolution construction units. Each of the graph convolution construction units is used to perform feature processing using the heuristic search function layer and the graph convolution function layer respectively, and to process the graph convolution encoding result of the graph convolution construction unit using the supplementary links of the forward access node and the backward export node of the heuristic search function layer.

[0023] In this embodiment, it is assumed that in this LF refining process abnormality diagnosis network, x=3, that is, there are 3 graph convolution construction units. First, the template refining process monitoring data obtained in step S110 is loaded into the first graph convolution construction unit.

[0024] In this graph convolution building unit, the heuristic search function layer starts working. For example, for the temperature data in the process parameter data, the dynamic feature path search mechanism of the heuristic search function layer constructs a search space containing multiple candidate paths. One of the candidate paths may be an analysis of the correlation between the temperature change rate and other parameters (such as the impurity content in the raw materials), and the other candidate path may be an analysis of the relationship between temperature and the electrode state of the equipment. The initial priority score of each candidate path is calculated based on the characteristic distribution information of the temperature data. It is assumed that the initial priority score of the candidate path for the relationship between temperature and electrode state is 0.8, and the initial priority score of the candidate path for the relationship between temperature and impurity content is 0.6. According to the descending order of the initial priority score, the top N (assuming N=2) candidate paths are selected from the search space. During the training process, the gradient contribution of the two candidate paths is dynamically monitored. If the gradient contribution of the candidate path for the relationship between temperature and electrode state increases in a certain training round, its priority score may be updated to 0.9, and the priority score of the other candidate path is adjusted accordingly. The output features of the two candidate paths are weighted and summed, and the weighted weight is determined by the joint function of the priority score and the gradient contribution to obtain the preliminary search features containing the key feature path.

[0025] Next, the path weight dynamic allocation operation is performed in the heuristic search function layer based on the preliminary search features. For example, for the temperature-related key feature paths, the preliminary search features are nonlinearly weighted fused according to the contribution of each feature path. If a feature path shows that the temperature rise is strongly associated with the rapid consumption of the electrode, and the contribution of the path is 0.7, then in the nonlinear weighted fusion process, this part of the feature will be given a higher weight to generate an optimized search feature.

[0026] At the same time, the temperature data in the template refining process monitoring data is synchronously loaded into the multi-directional self-converging network layer in the graph convolution function layer. The multi-directional self-converging network layer constructs multiple parallel branches, such as the dilated convolution branch, the depth-separable convolution branch and the global average pooling branch. In the dilated convolution branch, a larger convolution kernel size and a suitable expansion rate are used to extract the features of the temperature data, which may extract the changing features of the temperature data at different time scales; the depth-separable convolution branch focuses on mining the deep features of the temperature data; the global average pooling branch obtains the global features of the temperature data. After the output features of each branch are spliced ​​along the channel dimension, dynamic weights are assigned to the spliced ​​features of each channel dimension through the adaptive channel attention mechanism. For example, it is found that the channel features related to the temperature change trend have a higher weight, while some channel features related to the details of the temperature fluctuation have a lower weight. The spliced ​​features of each channel dimension are weighted based on the assigned dynamic weights, and then the weighted spliced ​​features are reduced in dimension to match the input dimensions of the subsequent processing stage to generate initial aggregated features containing local features and global features.

[0027] In the multi-directional self-convergence network layer, cross-channel feature interaction operations are performed on the initial aggregated features, and the weight distribution of each channel feature is adjusted through the adaptive channel attention mechanism. For example, for the channel features related to different heating stages in the temperature data, the weights are redistributed according to their importance in the refining process to generate optimized aggregated features.

[0028] The optimized search features and optimized aggregation features are input to the forward access node of the supplementary link to perform feature dimension alignment and dynamic splicing operations. Assuming that the dimension of the optimized search feature is [10,20] and the dimension of the optimized aggregation feature is [10,15], they are adjusted to the same dimension through the forward access node of the supplementary link and then dynamically spliced ​​to generate fused intermediate features.

[0029] The fused intermediate features are passed to the feature normalization module through the backward export node of the supplementary link to perform batch normalization and nonlinear activation processing. For example, batch normalization adjusts the numerical distribution of the fused intermediate features to an appropriate range, and then processes them through a nonlinear activation function (such as the ReLU function) to generate normalized fused features.

[0030] The normalized fusion feature is residually connected to the original temperature data in the template refining process monitoring data, and the feature graphs of the two are added element by element through the jump connection mechanism. For example, if a value in the original temperature data is 1560 degrees Celsius, the corresponding value in the normalized fusion feature is calculated and added to it to generate the feature processing result of the first graph convolution construction unit.

[0031] The feature processing results are input into the feedback adjustment module of the multi-directional self-converging network layer. In the back-propagation stage, the gradient amplitude information of the multi-directional self-converging network layer is collected. It is assumed that the gradient amplitude of the dilated convolution branch is large, while the gradient amplitude of the depth-separable convolution branch is small. The learning rate coefficient of each convolution branch is dynamically adjusted according to the gradient contribution of each convolution branch. If the gradient contribution of the depth-separable convolution branch is lower than the set contribution threshold, it is temporarily frozen. In the parameter update stage, the weight parameters of each convolution branch are adjusted using a hierarchical adaptive optimization algorithm to generate an updated multi-directional self-converging network layer.

[0032] The updated multi-directional self-convergence network layer is used to perform secondary multi-scale feature extraction on the feature processing results to generate enhanced aggregate features. For example, the characteristic information of temperature data at different time and space scales can be further mined.

[0033] The enhanced aggregation features and optimized search features are input into the iterative optimization module of the heuristic search function layer. A feature importance evaluation model based on random forest is constructed to rank the importance of the iteratively optimized search features at each forward propagation. It is assumed that among the temperature-related search features, the features related to temperature control in the key stage of refining are of higher importance. Low-importance features are truncated according to the preset feature retention ratio threshold, and the remaining retained search features are subjected to local linear embedding dimensionality reduction processing. The reduced-dimensional search features are dynamically weighted and combined with the historically retained features to generate iteratively optimized search features.

[0034] The iteratively optimized search features and enhanced aggregate features are cross-modally fused in a complementary link. For example, the spatial dimensions of the two are aligned, the correlation matrix between them is calculated through a bilinear interactive attention mechanism, and a spatial attention mask and a channel attention mask are generated based on the correlation matrix. The spatial attention mask is used to enhance the original search features and enhanced aggregate features in the spatial region to obtain the first enhanced feature; the channel attention mask is used to selectively enhance the feature channel to obtain the second enhanced feature. The first enhanced feature and the second enhanced feature are mixed by element-by-element multiplication and addition to generate a cross-modal fusion feature.

[0035] The spatial pyramid pooling operation is performed on the cross-modal fusion features, and the spatial context information of different scales is extracted through multi-level pooling windows to generate multi-scale pooling features. For example, for the spatial distribution of temperature data at different refining stages, features are extracted through pooling windows of different sizes.

[0036] The multi-scale pooling features are input into the feature compression module, and the feature channel dimension is reduced through the point-by-point convolution layer to generate a compressed low-dimensional feature representation. Assume that the channel dimension of the original multi-scale pooling feature is 30, which is reduced to 10 dimensions after processing by the point-by-point convolution layer.

[0037] The low-dimensional feature representation is gated and fused with the original temperature data in the template refining process monitoring data. Both are input into the fully connected layer to generate the initial gating weights, and the sigmoid activation function is applied to the initial gating weights to normalize them to the interval [0, 1] to obtain the normalized gating weights. The normalized gating weights are applied to the corresponding low-dimensional feature representation and each spatial position of the original temperature data through the broadcast mechanism to calculate the weighted linear combination as the final feature encoding result. This final feature encoding result is passed to the next processing stage as the output of the first graph convolution construction unit and as the input data source for the subsequent graph convolution function layer.

[0038] Then, the output result of the first graph convolution construction unit is loaded into the second graph convolution construction unit, and the above-mentioned similar processing process is repeated, including various operations of the heuristic search function layer, the multi-directional self-convergence network layer of the graph convolution function layer and the multi-layer perceptron (if any), and the related operations of the supplementary link, to generate the graph convolution construction result of the second graph convolution construction unit. Similarly, the output result of the second graph convolution construction unit is loaded into the third graph convolution construction unit, and the corresponding processing is performed again, and finally the graph convolution construction results extracted by the three graph convolution construction units are obtained.

[0039] Step S130, loading the graph convolution construction result into the y graph convolution restoration units, using each of the y graph convolution restoration units to perform graph convolution restoration respectively, and generating graph convolution restoration results output by the y graph convolution restoration units. Each of the graph convolution restoration units is used to perform graph convolution restoration using the heuristic search function layer and the graph convolution function layer respectively, and use the supplementary link to process the graph convolution restoration result of the graph convolution restoration unit.

[0040] Assume that y=4, that is, there are 4 graph convolution restoration units. First, the graph convolution construction result obtained in step S120 is loaded into the first graph convolution restoration unit.

[0041] In this graph convolution reduction unit, the heuristic search function layer begins to process the graph convolution construction results. For example, for the temperature-related feature data processed by the previous graph convolution construction unit, the heuristic search function layer will perform multi-dimensional feature path traversal again to build a new search space. This search space may contain a different combination of candidate paths from the previous one, because the tasks in the graph convolution reduction phase are different from those in the construction phase. Based on the feature distribution of the graph convolution construction results, the initial priority score of each candidate path is calculated. Assume that the initial priority score of a candidate path related to temperature and the final quality of refining is 0.7. According to the initial priority score, the top N (assuming N=3) candidate paths are selected, and the gradient contribution of these candidate paths is monitored during the training process and the priority score is updated. The output features of the selected candidate paths are weighted and summed to obtain the preliminary search features, and then the path weights are dynamically allocated to generate optimized search features.

[0042] At the same time, the first multi-directional self-convergence network layer in the graph convolution functional layer starts working. For example, for the temperature-related feature data in the graph convolution construction result, feature restoration operations are performed by constructing multiple parallel branches (such as dilated convolution branches, depthwise separable convolution branches, and global average pooling branches). Each branch uses different convolution kernel sizes and expansion rates for feature extraction, splices the output features of each branch along the channel dimension, and assigns dynamic weights to the spliced ​​features of each channel dimension through the adaptive channel attention mechanism to generate initial aggregate features containing local features and global features. Perform cross-channel feature interaction operations on the initial aggregate features, adjust the weight distribution of each channel feature, and generate optimized aggregate features.

[0043] The optimized search features and optimized aggregation features are input to the forward access node of the supplementary link, and feature dimension alignment and dynamic splicing operations are performed to generate fused intermediate features. The fused intermediate features are passed to the feature normalization module through the backward export node of the supplementary link, and batch normalization and nonlinear activation processing are performed to generate normalized fused features.

[0044] The original temperature-related features in the graph convolution construction result are residually connected with the normalized fusion features, and the feature maps of the two are element-by-element added through the jump connection mechanism to generate the feature processing results of the first graph convolution restoration unit.

[0045] In a similar manner, the processing result of the first graph convolution restoration unit is loaded into the second graph convolution restoration unit. In the second graph convolution restoration unit, the heuristic search function layer and the second multi-directional self-convergence network layer in the graph convolution function layer perform corresponding operations. For example, the heuristic search function layer searches and optimizes new feature paths for the results processed by the first graph convolution restoration unit, and the second multi-directional self-convergence network layer of the graph convolution function layer performs a more detailed restoration operation on temperature-related features, including adjusting parameters such as the convolution kernel size and expansion rate in a multi-branch structure to better restore the characteristics of temperature data during the refining process. The relevant features are fused through supplementary links to generate the processing result of the second graph convolution restoration unit.

[0046] Next, the processing result of the second graph convolution restoration unit is loaded into the third graph convolution restoration unit, and the above-mentioned heuristic search function layer and graph convolution function layer (here is the multi-layer perceptron) operations are repeated, as well as the processing of the supplementary link, to obtain the processing result of the third graph convolution restoration unit. Finally, the processing result of the third graph convolution restoration unit is loaded into the fourth graph convolution restoration unit, and the corresponding operations are performed again, and finally the graph convolution restoration results output by the four graph convolution restoration units are obtained.

[0047] Step S140, estimating a template anomaly diagnosis result of the template refining process monitoring data based on the graph convolution restoration result.

[0048] The template abnormality diagnosis result of the template refining process monitoring data is estimated based on the graph convolution restoration result obtained in step S130. Taking temperature data as an example, if the graph convolution restoration result shows that at a certain stage in the refining process, the temperature change trend deviates greatly from the normal refining process temperature curve, and the correlation with other related parameters (such as the impurity content in the raw materials, the electrode status of the equipment, etc.) does not conform to the normal mode, then it can be judged that there may be an abnormality at this stage.

[0049] For example, under normal circumstances, in the middle stage of refining, the temperature should be stable between 1600-1620 degrees Celsius, and as the impurity removal operation proceeds, the temperature and impurity content should show a certain negative correlation. If the graph convolution reduction results show that the temperature suddenly rises to 1650 degrees Celsius at this stage, and the relationship with the impurity content becomes positively correlated, this suggests that there may be problems such as out-of-control heating operation or abnormal raw material composition, thus obtaining a diagnostic result that there is an abnormality in the template refining process monitoring data.

[0050] A similar approach is used to analyze the performance of other process parameter data, raw material characteristic data, and equipment status data in the graph convolution restoration results. If the performance of multiple parameters deviates from the normal mode, the possibility of abnormality is higher, and the type of abnormality can be preliminarily determined based on the degree and characteristics of the deviation, such as equipment failure, raw material quality problems, or operating process errors.

[0051] Step S150 , training neuron weight information of the LF refining process abnormality diagnosis network based on the labeled abnormality diagnosis result of the template refining process monitoring data and the loss function value of the template abnormality diagnosis result.

[0052] Assume that the anomaly diagnosis result of the template refining process monitoring data has been configured, and this anomaly diagnosis result is predetermined by expert experience or other accurate diagnosis methods. For example, for a batch of refining process monitoring data, the expert determines that there is an abnormality caused by equipment electrode failure at a certain time point in the refining process based on detailed production records and on-site inspections. This is the anomaly diagnosis result.

[0053] Calculate the loss function value between the template abnormality diagnosis result and the labeled abnormality diagnosis result. If the loss function adopts the mean square error (MSE) function, for each corresponding parameter (such as temperature, impurity content, etc.) diagnosis result, calculate the square of the difference between the two, and then sum the square differences of all parameters to obtain the loss function value.

[0054] The neuron weight information of the LF refining process abnormality diagnosis network is trained according to this loss function value. For example, if a neuron in the graph convolution construction unit or the graph convolution reduction unit is highly related to the processing of temperature data, and it is found that the output of the neuron leads to an increase in the loss function value, then the weight of the neuron will be adjusted during the training process so that it can more accurately process the features related to the temperature data in subsequent processing.

[0055] A similar adjustment method is also used for the neuron weights related to other parameters. In the entire network, the weight information of each neuron is gradually adjusted according to the gradient information of the loss function value through the back propagation algorithm. For example, the neuron weights of the heuristic search function layer in the graph convolution construction unit will adjust the priority of path search, the weight of feature fusion, etc. according to the loss function value; the neuron weights of the multi-directional self-convergence network layer and the multi-layer perceptron in the graph convolution function layer will adjust the parameters of the convolution kernel, the weight of the channel attention mechanism, etc., so that the entire LF refining process abnormality diagnosis network can perform abnormality diagnosis more accurately.

[0056] Based on the above steps, the embodiment of the present application sets up a LF refining process anomaly diagnosis network including a specific number of graph convolution construction units (x, x is a positive integer not less than 2) and graph convolution restoration units (y, y is a positive integer not less than 3), and each unit is composed of a heuristic search function layer and a graph convolution function layer sequentially linked, and the processing results are optimized by using the supplementary links of the heuristic search function layer, so as to perform comprehensive and in-depth feature processing and graph convolution restoration on the template refining process monitoring data. Specifically, after obtaining the template refining process monitoring data covering process parameter data, raw material characteristic data and equipment status data, multiple graph convolution construction units can effectively extract data features respectively and generate graph convolution construction results; subsequent multiple graph convolution restoration units further restore the construction results, and can more accurately estimate the template abnormality diagnosis results based on the graph convolution restoration results, greatly improving the accuracy and reliability of LF refining process abnormality diagnosis, and can effectively identify abnormal situations in the LF refining process, providing strong support for timely measures to avoid production accidents and ensure production quality; at the same time, training the network neuron weight information based on the loss function value can continuously optimize the network performance and improve its adaptability and effectiveness in LF refining process abnormality diagnosis, thereby significantly improving the intelligence level and production efficiency of the entire LF refining process.

[0057] In a possible implementation, the LF refining process abnormality diagnosis network includes w feature extraction subnetworks linked sequentially, each of the feature extraction subnetworks includes a plurality of the graph convolution construction units linked sequentially in the x graph convolution construction units, and w is a positive integer not less than 2. Step S120 includes:

[0058] Step S121, loading the template refining process monitoring data into the first feature extraction sub-network, using each of the graph convolution construction units in the first feature extraction sub-network to perform feature processing respectively, and generating a processing result of the first feature extraction sub-network.

[0059] Step S122, load the processing result of the qth feature extraction subnetwork into the q+1th feature extraction subnetwork, use each of the graph convolution construction units in the q+1th feature extraction subnetwork to perform feature processing respectively, and generate the processing result of the q+1th feature extraction subnetwork, where q is a positive integer and q+1 is not greater than w.

[0060] In this embodiment, in the LF refining process scenario of a large steel enterprise, the LF refining process abnormality diagnosis network includes w sequentially linked feature extraction subnetworks, where w is assumed to be 3. Each feature extraction subnetwork consists of x sequentially linked graph convolution building units, where x is assumed to be 2.

[0061] First, the template refining process monitoring data is loaded into the first feature extraction subnetwork. The first graph convolution building unit in this feature extraction subnetwork starts to process data. For the process parameter data in the template refining process monitoring data, such as the molten steel temperature data during the refining process, the heuristic search function layer in the first graph convolution building unit constructs a search space based on the numerical distribution of the molten steel temperature. The candidate paths in the search space cover a variety of temperature-related feature transformation operation combinations, such as the association between temperature and heating time, and the association between temperature and electrode voltage. The priority score of each candidate path is calculated based on the initial feature distribution of the temperature data. For example, the initial priority score of the candidate path associated with the electrode voltage is 0.8. The top candidate paths are selected according to the priority, and their gradient contributions are dynamically monitored and the priority scores are updated during the processing. At the same time, the multi-directional self-convergence network layer in the graph convolution function layer processes the temperature data. Its multi-branch topology structure includes a dilated convolution branch, a depth-separable convolution branch, and a global average pooling branch. The dilated convolution branch extracts the features of temperature data at different scales with a specific convolution kernel size and expansion rate. The depthwise separable convolution branch mines the deep features of temperature data. The global average pooling branch obtains the global features of temperature data. The output features of each branch are spliced ​​along the channel dimension, and then assigned dynamic weights and weighted processing by the adaptive channel attention mechanism, and then the initial aggregate features are obtained by dimensionality reduction. The preliminary search features obtained by the heuristic search function layer and the initial aggregate features obtained by the graph convolution function layer are fused in the supplementary link, and after operations such as feature dimension alignment, dynamic splicing, normalization, and nonlinear activation, they are residually connected with the original temperature data to generate the processing result of the first graph convolution construction unit.

[0062] Then, this result enters the second graph convolution construction unit of the feature extraction subnetwork. The heuristic search function layer in the second graph convolution construction unit traverses the multi-dimensional feature path of the data again to build a new search space, such as searching for the association path between temperature and other parameters that is different from the previous one. The multi-layer perceptron in the graph convolution function layer further extracts and transforms the processing results of the first graph convolution construction unit, and fuses the relevant features according to the supplementary link operation similar to the first graph convolution construction unit, and finally generates the processing results of the first feature extraction subnetwork.

[0063] Then, the processing result of the first feature extraction subnetwork is loaded into the second feature extraction subnetwork. The first graph convolution construction unit in the second feature extraction subnetwork starts working. Taking the iron ore impurity content data in the raw material characteristic data as an example, the heuristic search function layer reconstructs the search space and searches for characteristic paths related to the impurity content, such as the correlation between the impurity content and the amount of refining agent added, the correlation between the impurity content and the fluidity of molten steel, etc. The multi-directional self-convergence network layer of the graph convolution function layer adjusts the parameters of the multi-branch topology structure according to the new data characteristics, extracts multi-scale features of the impurity content data, and obtains the processing result of the graph convolution construction unit after fusing related features through the supplementary link and connecting with the residual of the original impurity content data. The second graph convolution construction unit continues to process this result, and generates the processing result of the second feature extraction subnetwork through the new path search of the heuristic search function layer and the feature conversion of the multi-layer perceptron, as well as the fusion operation of the supplementary link.

[0064] The LF refining process abnormality diagnosis network includes w sequentially linked feature restoration sub-networks, each of which includes at least three of the y graph convolution restoration units sequentially linked. Step S130 includes:

[0065] Step S131, loading the processing result of the w-th feature extraction subnetwork into the first feature restoration subnetwork, using each of the graph convolution restoration units in the first feature restoration subnetwork to perform graph convolution restoration respectively, to generate the processing result of the first feature restoration subnetwork.

[0066] Step S132: load the processing results of the w-th feature extraction subnetwork and the processing results of the r-th feature restoration subnetwork into the r+1-th feature restoration subnetwork, and use each of the graph convolution restoration units in the r+1-th feature restoration subnetwork to perform graph convolution restoration respectively to generate the processing results of the r+1-th feature restoration subnetwork, where r is a positive integer and r+1 is not greater than w.

[0067] Similarly, the processing result of the second feature extraction subnetwork is loaded into the third feature extraction subnetwork, and processed according to the above process to finally obtain the processing result of the third feature extraction subnetwork.

[0068] For the feature reduction subnetwork, the LF refining process abnormality diagnosis network contains three sequentially linked feature reduction subnetworks, each of which contains y sequentially linked graph convolutional reduction units, assuming y=3.

[0069] The processing results of the third feature extraction subnetwork are loaded into the first feature restoration subnetwork. The first graph convolution restoration unit in the first feature restoration subnetwork starts graph convolution restoration. Taking the electrode consumption data in the equipment status data as an example, the heuristic search function layer constructs a search space to search for feature paths related to electrode consumption, such as the correlation between electrode consumption and refining time, the correlation between electrode consumption and molten steel temperature fluctuation, etc. The first multi-directional self-convergence network layer in the graph convolution function layer restores the features of the electrode consumption data through its multi-branch structure. The output features of each branch are weighted and fused by the channel attention mechanism to obtain the initial aggregated features, which are fused with the optimized search features of the heuristic search function layer through supplementary link processing, including feature dimension alignment, dynamic splicing, normalization, nonlinear activation and other operations, and then connected with the residual of the original electrode consumption data to obtain the processing result of the first graph convolution restoration unit.

[0070] The second graph convolution restoration unit in the first feature restoration subnetwork continues processing. The heuristic search function layer re-searches the feature paths related to the processing results of the first graph convolution restoration unit. The second multi-directional self-convergence network layer in the graph convolution function layer adjusts the multi-branch structure parameters according to the new data situation to restore the features. After fusing the related features through the supplementary links, the processing results of the second graph convolution restoration unit are obtained. The multi-layer perceptron in the third graph convolution restoration unit performs the final feature restoration processing on the processing results of the second graph convolution restoration unit. After the path search of the heuristic search function layer and the fusion operation of the supplementary links, the processing results of the first feature restoration subnetwork are generated.

[0071] The processing results of the third feature extraction subnetwork and the processing results of the first feature restoration subnetwork are loaded into the second feature restoration subnetwork. The first graph convolution restoration unit in the second feature restoration subnetwork processes these data. For the comprehensive refining process data, the heuristic search function layer constructs a new search space to search for various feature paths related to the comprehensive data. The first multi-directional self-convergence network layer in the graph convolution function layer performs feature restoration on the comprehensive data through the adjusted multi-branch structure, and obtains the processing result of the graph convolution restoration unit after the relevant features are fused through the supplementary links. The second graph convolution restoration unit and the third graph convolution restoration unit follow a similar process to perform new feature path search, feature restoration operations of the multi-directional self-convergence network layer or the multi-layer perceptron, and fusion operations of the supplementary links, respectively, to generate the processing result of the second feature restoration subnetwork.

[0072] Finally, the processing results of the third feature extraction subnetwork and the processing results of the second feature restoration subnetwork are loaded into the third feature restoration subnetwork, and processed according to the above process to finally obtain the processing results of the third feature restoration subnetwork, which is also the graph convolution restoration result output by the entire y graph convolution restoration units. Through such a process, in the LF refining process abnormality diagnosis network, the feature extraction subnetwork and the feature restoration subnetwork are used to perform comprehensive feature processing and restoration operations on the template refining process monitoring data, laying the foundation for subsequent abnormality diagnosis.

[0073] In a possible implementation, each of the feature extraction subnetworks includes a first graph convolution construction unit and a second graph convolution construction unit that are sequentially connected, wherein the graph convolution functional layer of the first graph convolution construction unit is a multi-directional self-convergence network layer, and the graph convolution functional layer of the second graph convolution construction unit is a multi-layer perceptron.

[0074] Step S122 may include:

[0075] Step S1221: Load the processing result of the qth feature extraction subnetwork into the first graph convolution construction unit of the q+1th feature extraction subnetwork.

[0076] Step S1222: perform feature processing on the processing result of the qth feature extraction subnetwork using the heuristic search function layer and the multi-directional self-convergence network layer of the first graph convolution construction unit respectively.

[0077] Step S1223, the processing results of the qth feature extraction subnetwork and the processing results of the heuristic search function layer of the first graph convolution construction unit are merged with the processing results of the multi-directional self-convergence network layer of the first graph convolution construction unit using the supplementary link to generate the processing results of the first graph convolution construction unit.

[0078] Step S1224: load the processing result of the first graph convolution construction unit into the second graph convolution construction unit of the q+1th feature extraction subnetwork.

[0079] Step S1225, performing feature processing using the heuristic search function layer and the multilayer perceptron of the second graph convolution construction unit respectively.

[0080] Step S1226, the processing results of the first graph convolution construction unit and the processing results of the heuristic search function layer of the second graph convolution construction unit are fused with the processing results of the multi-layer perceptron of the second graph convolution construction unit using the supplementary link to generate the processing results of the q+1th feature extraction subnetwork.

[0081] In a possible implementation, step S1223 includes:

[0082] Step S1223-1, determining a first fusion result of a processing result of the qth feature extraction subnetwork and a first influencing factor, and determining a second fusion result of a processing result of the heuristic search function layer of the first graph convolution construction unit and a second influencing factor.

[0083] Step S1223-2: The first fusion result and the second fusion result are fused using the supplementary link and the processing result of the multi-directional self-convergence network layer of the first graph convolution construction unit.

[0084] Step S1226 includes:

[0085] Step S1226-1, determining a third fusion result of the processing result of the first graph convolution construction unit and the third influencing factor, and determining a fourth fusion result of the processing result of the heuristic search function layer of the second graph convolution construction unit and the fourth influencing factor.

[0086] Step S1226-2: The third fusion result and the fourth fusion result are fused using the supplementary link and the processing result of the multi-layer perceptron of the second graph convolution construction unit.

[0087] In a possible implementation, the method further includes:

[0088] At least one of the first influencing factor, the second influencing factor, the third influencing factor and the fourth influencing factor is optimized based on the loss function value of the labeled abnormality diagnosis result of the template refining process monitoring data and the template abnormality diagnosis result.

[0089] In this embodiment, in the LF refining process scenario in steel production, the feature extraction subnetwork structure and related processing flow in the LF refining process abnormality diagnosis network are taken into consideration. Each feature extraction subnetwork includes a first graph convolution construction unit and a second graph convolution construction unit that are sequentially linked, wherein the graph convolution function layer of the first graph convolution construction unit is a multi-directional self-convergence network layer, and the graph convolution function layer of the second graph convolution construction unit is a multi-layer perceptron.

[0090] When the processing result of the qth feature extraction subnetwork is loaded into the q+1th feature extraction subnetwork, the entire processing flow is as follows:

[0091] First, the processing result of the qth feature extraction subnetwork is loaded into the first graph convolution construction unit of the q+1th feature extraction subnetwork. Taking the molten steel temperature data in the LF refining process as an example, the processing result of the molten steel temperature data output by the qth feature extraction subnetwork enters the first graph convolution construction unit of the q+1th feature extraction subnetwork.

[0092] Then, the processing results of the qth feature extraction subnetwork are processed by the heuristic search function layer and the multi-directional self-convergence network layer of the first graph convolution construction unit. In the heuristic search function layer, a search space is constructed for the molten steel temperature data, which contains many candidate paths related to the molten steel temperature. For example, there may be a relationship path between the molten steel temperature and the electrode heating power, a relationship path between the molten steel temperature and the refining time, etc. According to the characteristic distribution information of the molten steel temperature data, the initial priority score of each candidate path is calculated. It is assumed that the initial priority score of the relationship path between the molten steel temperature and the electrode heating power is 0.8, and the initial priority score of the relationship path between the molten steel temperature and the refining time is 0.6. According to the descending order of the initial priority score, the top candidate paths are selected, and the gradient contribution of these candidate paths is dynamically monitored during the training process. If the gradient contribution of the relationship path between the molten steel temperature and the refining time increases at a certain training stage, its priority score will be updated accordingly. By weighted summing up the output features of the screened candidate paths, the preliminary search features containing the key feature paths are obtained. Then, the path weight dynamic allocation operation is performed based on the preliminary search features. The preliminary search features are nonlinearly weighted fused according to the contribution of each feature path to generate optimized search features.

[0093] At the same time, the molten steel temperature processing result of the qth feature extraction subnetwork is synchronously loaded into the multi-directional self-converging network layer. The multi-branch topological structure of the multi-directional self-converging network layer includes a dilated convolution branch, a depth-wise separable convolution branch, and a global average pooling branch. In the dilated convolution branch, for the molten steel temperature data, a specific convolution kernel size and dilation rate are used for feature extraction, which may extract the variation characteristics of the molten steel temperature at different time scales; the depth-wise separable convolution branch focuses on mining the deep features of the molten steel temperature data; the global average pooling branch obtains the global features of the molten steel temperature data. After the features output by each branch are spliced ​​along the channel dimension, a dynamic weight is assigned to the spliced ​​features of each channel dimension through an adaptive channel attention mechanism. For example, a higher weight is assigned to the channel features related to the rapid rise stage of the molten steel temperature, while a lower weight is assigned to the channel features related to the temperature stabilization stage. The spliced ​​features of each channel dimension are weighted based on the assigned dynamic weights, and then the weighted spliced ​​features are subjected to dimensionality reduction to match the input dimensions of the subsequent processing stage, generating initial aggregate features containing local features and global features. In the multi-directional self-convergence network layer, cross-channel feature interaction operations are performed on the initial aggregated features, and the weight distribution of each channel feature is adjusted again through the adaptive channel attention mechanism to generate optimized aggregated features.

[0094] Next, the processing result of the qth feature extraction subnetwork and the processing result of the heuristic search function layer of the first graph convolution construction unit are fused with the processing result of the multi-directional self-convergence network layer of the first graph convolution construction unit using a supplementary link to generate the processing result of the first graph convolution construction unit. Specifically, the first fusion result of the processing result of the qth feature extraction subnetwork and the first influencing factor is determined, and the second fusion result of the processing result of the heuristic search function layer of the first graph convolution construction unit and the second influencing factor is determined. For example, assuming that the processing result of the qth feature extraction subnetwork on the molten steel temperature is a feature vector A, and the first influencing factor is α, then the first fusion result is α*A; the processing result of the heuristic search function layer of the first graph convolution construction unit on the molten steel temperature is a feature vector B, the second influencing factor is β, and the second fusion result is β*B. The first fusion result and the second fusion result are fused with the processing result of the multi-directional self-convergence network layer of the first graph convolution construction unit (assuming it is a feature vector C) using a supplementary link. In the supplementary link, operations such as feature dimension alignment, dynamic splicing, normalization, and nonlinear activation may be involved, and the fused results are used to generate the processing results of the first graph convolution construction unit.

[0095] Afterwards, the processing result of the first graph convolution construction unit is loaded into the second graph convolution construction unit of the q+1th feature extraction subnetwork.

[0096] The heuristic search function layer and multi-layer perceptron of the second graph convolution construction unit are used for feature processing. In the heuristic search function layer of the second graph convolution construction unit, the search space is reconstructed for the molten steel temperature data results processed by the first graph convolution construction unit, and a feature path different from the previous one is searched. For example, the relationship path between the molten steel temperature and the input amount of other raw materials may be searched. According to the operation similar to the heuristic search function layer of the first graph convolution construction unit, the optimized search feature is obtained. The multi-layer perceptron further converts and extracts the features of the processing results of the first graph convolution construction unit.

[0097] Finally, the processing result of the first graph convolution construction unit and the processing result of the heuristic search function layer of the second graph convolution construction unit are fused with the processing result of the multi-layer perceptron of the second graph convolution construction unit by using the supplementary link to generate the processing result of the q+1th feature extraction sub-network. Determine the third fusion result of the processing result of the first graph convolution construction unit and the third influencing factor, and determine the fourth fusion result of the processing result of the heuristic search function layer of the second graph convolution construction unit and the fourth influencing factor. Assume that the processing result of the first graph convolution construction unit on the molten steel temperature is the feature vector D, the third influencing factor is γ, and the third fusion result is γ*D; the processing result of the heuristic search function layer on the molten steel temperature of the second graph convolution construction unit is the feature vector E, the fourth influencing factor is δ, and the fourth fusion result is δ*E. The third fusion result and the fourth fusion result are fused with the processing result of the multi-layer perceptron of the second graph convolution construction unit (assuming it is the feature vector F) by using the supplementary link, and the processing result of the q+1th feature extraction sub-network is generated after performing relevant operations in the supplementary link.

[0098] In the entire LF refining process abnormality diagnosis process, it also includes optimizing at least one of the first influencing factor, the second influencing factor, the third influencing factor and the fourth influencing factor based on the loss function value of the anomaly diagnosis result of the template refining process monitoring data and the template anomaly diagnosis result. For example, in the actual LF refining process, for molten steel temperature data, if the anomaly diagnosis result of the anomaly diagnosis result of the template refining process monitoring data indicates that the molten steel temperature should be between 1600 and 1620 degrees Celsius at a certain refining stage, but the template anomaly diagnosis result shows that the predicted temperature exceeds this range, the loss function value between the two is calculated (such as using a mean square error function). According to this loss function value, if it is found that the prediction deviation is caused by the unreasonableness of the first influencing factor, the second influencing factor, the third influencing factor or the fourth influencing factor, these influencing factors are optimized. Assuming that the deviation is caused by the first influencing factor, by adjusting the value of the first influencing factor, the above-mentioned feature processing and fusion operations are re-performed to improve the diagnostic accuracy of the LF refining process abnormality diagnosis network for the molten steel temperature data, thereby improving the accuracy of the entire LF refining process abnormality diagnosis. The processing of other process parameter data, raw material characteristic data and equipment status data in the feature extraction sub-network is also carried out according to a similar process, so that the LF refining process anomaly diagnosis network can effectively process and diagnose the entire LF refining process monitoring data.

[0099] In a possible implementation, step S1222 may include:

[0100] Step S1222-1, load the processing result of the qth feature extraction subnetwork as the first input data to the heuristic search function layer, perform multi-dimensional feature path traversal on the first input data through the dynamic feature path search mechanism in the heuristic search function layer, and generate preliminary search features containing key feature paths.

[0101] Step S1222-2, based on the preliminary search features, a path weight dynamic allocation operation is performed in the heuristic search function layer, and the preliminary search features are nonlinearly weighted fused according to the contribution of each feature path to generate an optimized search feature.

[0102] Step S1222-3, synchronously load the processing result of the qth feature extraction subnetwork to the multi-directional self-convergence network layer, perform parallel multi-scale feature extraction on the first input data through the multi-branch topology structure in the multi-directional self-convergence network layer, and generate initial aggregation features including local features and global features.

[0103] Step S1222-4, performing cross-channel feature interaction operations on the initial aggregated features in the multi-directional self-convergence network layer, adjusting the weight distribution of each channel feature through an adaptive channel attention mechanism, and generating optimized aggregated features.

[0104] Step S1222-5, input the optimized search feature and the optimized aggregation feature to the forward access node of the supplementary link, perform feature dimension alignment and dynamic splicing operations, generate fused intermediate features, pass the fused intermediate features to the feature normalization module through the backward export node of the supplementary link, perform batch normalization and nonlinear activation processing, and generate normalized fused features.

[0105] Step S1222-6, perform residual connection on the normalized fusion feature and the processing result of the qth feature extraction sub-network, perform element-by-element addition operation on the feature graphs of the two through a jump connection mechanism, generate the feature processing result of the first graph convolution construction unit, input the feature processing result into the feedback adjustment module of the multi-directional self-convergence network layer, dynamically adjust the parameter configuration of the multi-branch topology structure based on the gradient back propagation information of the current training round, and generate an updated multi-directional self-convergence network layer.

[0106] Step S1222-7, using the updated multi-directional self-convergence network layer to perform secondary multi-scale feature extraction on the feature processing results to generate enhanced aggregation features, input the enhanced aggregation features and the optimized search features into the iterative optimization module of the heuristic search function layer, screen out a subset of significant distinguishing features through a recursive feature selection mechanism, generate iteratively optimized search features, cross-modally fuse the iteratively optimized search features and the enhanced aggregation features in the supplementary link, calculate the joint weight matrix of the iteratively optimized search features and the enhanced aggregation features through a multimodal attention mechanism, and generate cross-modal fusion features.

[0107] Step S1222-8, performing a spatial pyramid pooling operation on the cross-modal fusion feature, extracting spatial context information of different scales through a multi-level pooling window, generating a multi-scale pooling feature, inputting the multi-scale pooling feature into a feature compression module, reducing the feature channel dimension through a point-by-point convolution layer, generating a compressed low-dimensional feature representation, performing gated fusion on the low-dimensional feature representation and the processing result of the qth feature extraction subnetwork, dynamically adjusting the fusion ratio of the two through a learnable gating weight, and generating a final feature encoding result.

[0108] Step S1222-9, passing the feature encoding result to the next processing stage as the output of the first graph convolution construction unit and as the input data source of the subsequent graph convolution functional layer.

[0109] For example, the execution steps of the dynamic feature path search mechanism include:

[0110] A search space containing multiple candidate paths is constructed, where each candidate path corresponds to a different combination of feature transformation operations.

[0111] An initial priority score of each candidate path is calculated based on the feature distribution information of the first input data.

[0112] According to the descending order of the initial priority scores, the top N candidate paths are screened out from the search space.

[0113] During the training process, the gradient contribution of each of the top N candidate paths is dynamically monitored and their priority scores are updated.

[0114] The output features of the top N candidate paths in the ranking are weighted and summed, wherein the weights of the weightings are determined by a joint function of the priority score and the gradient contribution.

[0115] For example, the operation steps of the multi-branch topology structure include:

[0116] Constructing multiple parallel branches, the multiple branches include a dilated convolution branch, a depthwise separable convolution branch, and a global average pooling branch.

[0117] Different convolution kernel sizes and dilation rates are used for feature extraction in each branch, and after the output features of each branch are spliced ​​along the channel dimension, dynamic weights are assigned to the spliced ​​features of each channel dimension through an adaptive channel attention mechanism, and the spliced ​​features of each channel dimension are weighted based on the assigned dynamic weights, and a dimensionality reduction operation is performed on the weighted spliced ​​features to match the input dimension of the subsequent processing stage.

[0118] For example, the step of cross-modally fusing the iteratively optimized search features and the enhanced aggregated features in the supplementary link, calculating the joint weight matrix of the iteratively optimized search features and the enhanced aggregated features through a multimodal attention mechanism, and generating a cross-modal fusion feature includes:

[0119] The iteratively optimized search features and the enhanced aggregate features are spatially aligned respectively, and the correlation matrix between the iteratively optimized search features and the enhanced aggregate features is calculated through a bilinear interactive attention mechanism. After generating a spatial attention mask and a channel attention mask based on the correlation matrix, the spatial attention mask is used to perform spatial region enhancement on the original search features and the enhanced aggregate features to obtain a first enhanced feature, and the channel attention mask is used to selectively enhance the feature channel to obtain a second enhanced feature. The first enhanced feature and the second enhanced feature are mixed by element-by-element multiplication and addition to generate a cross-modal fusion feature.

[0120] For example, the step of performing gated fusion on the low-dimensional feature representation and the processing result of the qth feature extraction subnetwork, dynamically adjusting the fusion ratio of the two by learnable gated weights, and generating the final feature encoding result includes:

[0121] The low-dimensional feature representation and the processing result of the qth feature extraction subnetwork are input into the fully connected layer to generate initial gating weights, and a sigmoid activation function is applied to the initial gating weights to normalize them to the interval [0, 1] to obtain normalized gating weights, and the normalized gating weights are applied to the corresponding low-dimensional feature representation and each spatial position of the processing result of the qth feature extraction subnetwork through a broadcast mechanism to calculate the weighted linear combination of the low-dimensional feature representation and the processing result of the qth feature extraction subnetwork as the final feature encoding result.

[0122] For example, the step of inputting the feature processing result into the feedback adjustment module of the multi-directional self-converging network layer, dynamically adjusting the parameter configuration of the multi-branch topology structure based on the gradient back propagation information of the current training round, and generating an updated multi-directional self-converging network layer includes:

[0123] In the back-propagation stage, the gradient amplitude information of the multi-directional self-convergence network layer is collected, and the learning rate coefficient of each convolution branch is dynamically adjusted according to the gradient contribution of each convolution branch. The convolution branches whose contribution is lower than the set contribution threshold are temporarily frozen. In the parameter updating stage, the weight parameters of each convolution branch are adjusted using a hierarchical adaptive optimization algorithm.

[0124] For example, the step of inputting the enhanced aggregation feature and the optimized search feature into the iterative optimization module of the heuristic search function layer, selecting a subset of significant distinguishing features through a recursive feature selection mechanism, and generating an iteratively optimized search feature includes:

[0125] A feature importance evaluation model based on random forest is constructed, and the iteratively optimized search features are ranked by importance during each forward propagation. Low-importance features in the importance ranking results are truncated according to a preset feature retention ratio threshold, and the remaining retained search features are subjected to local linear embedding dimensionality reduction processing, and the reduced-dimensionality search features are dynamically weighted combined with the historical retained features.

[0126] In this embodiment, in the heuristic search function layer, the dynamic feature path search mechanism is enabled. Taking the molten steel temperature data as an example, a search space containing multiple candidate paths is constructed, and these candidate paths correspond to different combinations of feature transformation operations. For example, one candidate path may be to analyze the relationship between the molten steel temperature and the electrode heating power, which involves performing a specific mathematical transformation operation on the molten steel temperature data and the electrode power data; another candidate path may be to study the relationship between the molten steel temperature and the refining agent dosage added during the refining process, which requires another form of association operation on the molten steel temperature data and the refining agent dosage data. The initial priority score of each candidate path is calculated based on the feature distribution information of the molten steel temperature data (i.e., the first input data). Assume that the initial priority score of the path of the relationship between the molten steel temperature and the electrode heating power is 0.8, because in the prior knowledge of the LF refining process, the electrode heating power has a more direct and significant impact on the molten steel temperature; while the initial priority score of the path of the relationship between the molten steel temperature and the refining agent dosage is 0.6, and the effect of the refining agent dosage on the molten steel temperature may be more indirect.

[0127] According to the descending order of the initial priority score, the top N (assuming N=3) candidate paths are selected from the search space. During the training process, the gradient contribution of these three candidate paths is dynamically monitored. For example, as the training progresses, it is found that the relationship path between molten steel temperature and refining agent dosage has a greater impact on the final diagnosis result than initially expected at certain specific refining stages, and its gradient contribution increases, so its priority score will be updated accordingly. The output features of the top N candidate paths in the three rankings are weighted and summed, and the weighted weight is determined by the joint function of the priority score and the gradient contribution, thereby generating a preliminary search feature containing the key feature path.

[0128] Next, the path weight dynamic allocation operation is performed in the heuristic search function layer based on the preliminary search features. For the preliminary search features related to the molten steel temperature data, nonlinear weighted fusion is performed on them according to the contribution of each feature path. For example, if a feature path shows that there is a strong linear relationship between the molten steel temperature and the electrode heating power in a specific refining time period, and this path is determined to have a high contribution in the previous analysis, then in the nonlinear weighted fusion process, this part of the feature will be given a higher weight, thereby generating an optimized search feature.

[0129] At the same time, the processing results of the qth feature extraction subnetwork on the molten steel temperature are synchronously loaded into the multi-directional self-converging network layer. The multi-branch topology in the multi-directional self-converging network layer begins to perform parallel multi-scale feature extraction on the molten steel temperature data. The multi-branch topology constructs multiple parallel branches, including the dilated convolution branch, the depth-separable convolution branch, and the global average pooling branch. In the dilated convolution branch, for the molten steel temperature data, a larger convolution kernel size and a suitable expansion rate are used for feature extraction. For example, a larger convolution kernel size can capture the trend of molten steel temperature changes over a longer time scale, and may find temperature fluctuation characteristics every 10-15 minutes; the depth-separable convolution branch focuses on mining the deep features of molten steel temperature data, and may analyze the implicit information of temperature data at different depth levels; the global average pooling branch obtains the global features of molten steel temperature data, such as the average temperature level during the entire refining process. After the output features of each branch are spliced ​​along the channel dimension, a dynamic weight is assigned to the spliced ​​features of each channel dimension through the adaptive channel attention mechanism. For example, a higher weight is assigned to the channel features related to the stage of rapid temperature rise of molten steel, because the temperature change in this stage may be more critical to the refining process, while a lower weight is assigned to the channel features related to the temperature stabilization stage. The concatenated features of each channel dimension are weighted based on the assigned dynamic weight, and then the weighted concatenated features are subjected to dimensionality reduction to match the input dimensions of the subsequent processing stage, generating initial aggregated features containing local features and global features.

[0130] In the multi-directional self-convergence network layer, the initial aggregated features are subjected to cross-channel feature interaction operations, and the weight distribution of each channel feature is adjusted through the adaptive channel attention mechanism. For the initial aggregated features of the molten steel temperature data, the weights are redistributed according to the feature importance of the molten steel temperature represented by different channels at different refining stages or different regions (if there is spatial distribution information). For example, if a channel feature is highly correlated with the critical heating stage of the molten steel temperature at the initial stage of refining, the weight of the channel will be increased to generate an optimized aggregated feature.

[0131] The optimized search features and optimized aggregation features are input to the forward access node of the supplementary link to perform feature dimension alignment and dynamic splicing operations. Assuming that the dimension of the optimized search feature is [10,20] and the dimension of the optimized aggregation feature is [10,15], they are adjusted to the same dimension through the forward access node of the supplementary link and then dynamically spliced ​​to generate fused intermediate features.

[0132] The fused intermediate features are passed to the feature normalization module through the backward export node of the supplementary link to perform batch normalization and nonlinear activation processing. Batch normalization adjusts the numerical distribution of the fused intermediate features to an appropriate range, such as mapping the values ​​to a specific mean and standard deviation range, and then processed through a nonlinear activation function (such as the ReLU function) to generate normalized fused features.

[0133] The normalized fusion feature is residually connected to the processing result of the qth feature extraction subnetwork on the molten steel temperature, and the feature graphs of the two are added element by element through the jump connection mechanism. For example, if a certain value in the molten steel temperature processing result of the qth feature extraction subnetwork is 1560 degrees Celsius, the corresponding value in the normalized fusion feature is calculated and added to it to generate the feature processing result of the first graph convolution construction unit.

[0134] The feature processing results are input into the feedback adjustment module of the multi-directional self-converging network layer. The gradient amplitude information of the multi-directional self-converging network layer is collected in the back propagation stage. For example, the gradient amplitude of the hole convolution branch is large, indicating that this branch has a greater influence on adjusting the temperature characteristics of the molten steel; while the gradient amplitude of the depth-separable convolution branch is small. The learning rate coefficient of each convolution branch is dynamically adjusted according to the gradient contribution of each convolution branch. If the gradient contribution of the depth-separable convolution branch is lower than the set contribution threshold, it is temporarily frozen to reduce unnecessary computing resource consumption. In the parameter update stage, the layered adaptive optimization algorithm is used to adjust the weight parameters of each convolution branch to generate an updated multi-directional self-converging network layer.

[0135] The updated multi-directional self-converging network layer is used to perform secondary multi-scale feature extraction on the feature processing results to generate enhanced aggregate features. For example, further mining the feature information of molten steel temperature data at different time and spatial scales (if there is spatial information) may reveal temperature fluctuation details that were not fully captured before or new correlation features with other process parameters.

[0136] The enhanced aggregation features and optimized search features are input into the iterative optimization module of the heuristic search function layer. A feature importance evaluation model based on random forest is constructed to rank the importance of the iteratively optimized search features at each forward propagation. For search features related to molten steel temperature, for example, features related to temperature control in the key stage of refining are of high importance. Low-importance features are truncated according to the preset feature retention ratio threshold, and the remaining retained search features are subjected to local linear embedding dimensionality reduction processing. The reduced-dimensional search features are dynamically weighted and combined with the historical retained features to generate iteratively optimized search features.

[0137] The iteratively optimized search features and enhanced aggregate features are cross-modally fused in the supplementary link. The spatial dimensions of the two are aligned respectively, and the correlation matrix between them is calculated through the bilinear interactive attention mechanism. After generating the spatial attention mask and the channel attention mask based on the correlation matrix, the spatial attention mask is used to enhance the original search features and enhanced aggregate features in the spatial region to obtain the first enhanced feature. For example, if there is a local anomaly in the spatial distribution of the molten steel temperature (if there is spatial distribution information), the spatial attention mask will enhance the feature representation of this area; the channel attention mask is used to selectively enhance the feature channel to obtain the second enhanced feature, and the first enhanced feature and the second enhanced feature are mixed by element-by-element multiplication and addition to generate a cross-modal fusion feature.

[0138] A spatial pyramid pooling operation is performed on the cross-modal fusion features, and spatial context information of different scales is extracted through multi-level pooling windows. For molten steel temperature data, if there is spatial distribution information, pooling windows of different sizes can capture the temperature change characteristics from local to global. For example, a small pooling window captures the temperature fluctuation of a small area on the surface of the molten steel, and a large pooling window captures the average temperature trend in the entire molten steel container, generating multi-scale pooling features.

[0139] The multi-scale pooling features are input into the feature compression module, and the feature channel dimension is reduced through the point-by-point convolution layer to generate a compressed low-dimensional feature representation. Assuming that the channel dimension of the original multi-scale pooling feature is 30, it is reduced to 10 dimensions after processing by the point-by-point convolution layer, reducing the amount of data while retaining key feature information.

[0140] The low-dimensional feature representation and the processing result of the qth feature extraction subnetwork on the molten steel temperature are gated and fused. Both are input into the fully connected layer to generate the initial gating weight, and the sigmoid activation function is applied to the initial gating weight to normalize it to the interval [0, 1] to obtain the normalized gating weight. The normalized gating weight is applied to the corresponding low-dimensional feature representation and the processing result of the qth feature extraction subnetwork in each spatial position to calculate the weighted linear combination as the final feature encoding result through the broadcast mechanism. This final feature encoding result is passed to the next processing stage as the output of the first graph convolution construction unit and as the input data source of the subsequent graph convolution function layer, so as to continue the subsequent processing operations in the LF refining process abnormal diagnosis network, thereby realizing effective monitoring and abnormal diagnosis of the LF refining process.

[0141] In a possible implementation, each of the feature restoration subnetworks includes a first graph convolution restoration unit, a second graph convolution restoration unit, and a third graph convolution restoration unit that are sequentially connected, the graph convolution functional layer of the first graph convolution restoration unit is a first multi-directional self-convergence network layer, the graph convolution functional layer of the second graph convolution restoration unit is a second multi-directional self-convergence network layer, and the graph convolution functional layer of the third graph convolution restoration unit is a multi-layer perceptron.

[0142] Step S132 may include:

[0143] Step S1321: Load the processing result of the rth feature restoration sub-network into the first graph convolution restoration unit of the r+1th feature restoration sub-network.

[0144] Step S1322: perform feature processing on the processing result of the rth feature restoration subnetwork by using the heuristic search function layer and the first multi-directional self-convergence network layer of the first graph convolution restoration unit respectively.

[0145] Step S1323, the processing result of the rth feature restoration subnetwork and the processing result of the heuristic search function layer of the first graph convolution restoration unit are merged with the processing result of the first multidirectional self-convergence network layer of the first graph convolution restoration unit by using the supplementary link to generate the processing result of the first graph convolution restoration unit.

[0146] Step S1324: Load the processing result of the w-th feature extraction subnetwork and the processing result of the first graph convolution restoration unit into the second graph convolution restoration unit of the r+1-th feature restoration subnetwork.

[0147] Step S1325: perform feature processing using the heuristic search function layer and the second multi-directional self-convergence network layer of the second graph convolutional restoration unit respectively.

[0148] Step S1326, the processing results of the first graph convolution restoration unit and the processing results of the heuristic search function layer of the second graph convolution restoration unit are fused using the supplementary link and the processing results of the second multi-directional self-convergence network layer of the second graph convolution restoration unit to generate the processing results of the second graph convolution restoration unit.

[0149] Step S1327: Load the processing result of the second graph convolution restoration unit into the third graph convolution restoration unit of the r+1th feature restoration subnetwork.

[0150] Step S1328, performing feature processing using the heuristic search function layer and the multi-layer perceptron of the third graph convolutional restoration unit respectively.

[0151] Step S1329, the processing results of the second graph convolution restoration unit and the processing results of the heuristic search function layer of the third graph convolution restoration unit are fused with the processing results of the multi-layer perceptron of the third graph convolution restoration unit using the supplementary link to generate the processing results of the r+1th feature restoration subnetwork.

[0152] In this embodiment, it is assumed that the processing result of the rth feature restoration subnetwork on the molten steel temperature data includes a variety of feature information related to the molten steel temperature obtained in the previous restoration process, and this information is loaded into the first graph convolution restoration unit of the r+1th feature restoration subnetwork.

[0153] In the first graph convolutional reduction unit, the processing results of the rth feature reduction subnetwork are processed by the heuristic search function layer and the first multi-directional self-convergence network layer. For the heuristic search function layer, taking the molten steel temperature data as an example, it will construct a search space for the processing results of the rth feature reduction subnetwork, which contains many candidate paths related to the molten steel temperature. For example, there may be a relationship path between the molten steel temperature and the electrode heating strategy in a specific time period during the refining process, or a relationship path between the molten steel temperature and the effect of a certain refining additive at a specific stage. Based on the feature distribution information of the rth feature reduction subnetwork on the molten steel temperature processing results, the initial priority score of each candidate path is calculated. Assume that the initial priority score of the relationship path between the molten steel temperature and the electrode heating strategy is 0.7, because the electrode heating strategy has a direct impact on the control of the molten steel temperature; and the initial priority score of the relationship path between the molten steel temperature and the effect of the refining additive is 0.6. According to the descending order of the initial priority score, the top candidate paths are selected, and the gradient contribution of these candidate paths is dynamically monitored during the training process. If, at a certain training stage, it is found that the gradient contribution of the path of the relationship between the temperature of the molten steel and the effect of the refining additive increases, for example, because the effect of the additive on the temperature becomes more critical during the refining process of some special steel grades, then its priority score will be updated accordingly. By weighted summing the output features of the screened candidate paths, a preliminary search feature containing the key feature path is obtained, and then a path weight dynamic allocation operation is performed based on the preliminary search feature. The preliminary search features are nonlinearly weighted fused according to the contribution of each feature path to generate an optimized search feature.

[0154] At the same time, the processing results of the rth feature restoration subnetwork on the molten steel temperature are synchronously loaded into the first multi-directional self-converging network layer. The multi-branch topological structure in the first multi-directional self-converging network layer begins to perform parallel multi-scale feature restoration operations on the molten steel temperature data. This multi-branch topological structure constructs multiple parallel branches, such as a dilated convolution branch, a depth-separable convolution branch, and a global average pooling branch. In the dilated convolution branch, a specific convolution kernel size and expansion rate are used to restore the features of the molten steel temperature data. For example, a larger convolution kernel size can restore the original change characteristics of the molten steel temperature on a longer time scale, which may be the temperature fluctuation characteristics every 10-15 minutes during the entire refining cycle; the depth-separable convolution branch focuses on mining the deep features of the molten steel temperature data and restoring the implicit temperature information at different depth levels; the global average pooling branch obtains the global features of the molten steel temperature data, such as the average temperature level during the entire refining process. After the output features of each branch are spliced ​​along the channel dimension, a dynamic weight is assigned to the spliced ​​features of each channel dimension through an adaptive channel attention mechanism. For example, higher weights are assigned to channel features related to the stage of rapid temperature rise of molten steel, because the temperature change in this stage is of great significance to the control of the refining process and the final quality of molten steel, while lower weights are assigned to channel features related to the stage of temperature stabilization. The concatenated features of each channel dimension are weighted based on the assigned dynamic weights, and then the weighted concatenated features are reduced to match the input dimensions of the subsequent processing stage to generate initial aggregate features containing local features and global features. In the first multi-directional self-convergence network layer, cross-channel feature interaction operations are performed on the initial aggregate features, and the weight distribution of each channel feature is adjusted again through the adaptive channel attention mechanism to generate optimized aggregate features.

[0155] The processing result of the r-th feature restoration subnetwork and the processing result of the heuristic search function layer of the first graph convolution restoration unit are fused with the processing result of the first multi-directional self-convergence network layer of the first graph convolution restoration unit using a supplementary link to generate the processing result of the first graph convolution restoration unit. Specifically, the first fusion result of the processing result of the r-th feature restoration subnetwork and the first influencing factor (assuming α) is determined, that is, α multiplied by the processing result of the r-th feature restoration subnetwork (assuming feature vector A) to obtain α*A; the second fusion result of the processing result of the heuristic search function layer of the first graph convolution restoration unit (assuming feature vector B) and the second influencing factor (assuming β) is determined, that is, β*B. The first fusion result and the second fusion result are fused with the processing result of the first multi-directional self-convergence network layer of the first graph convolution restoration unit (assuming feature vector C) using a supplementary link. In the supplementary link, operations such as feature dimension alignment, dynamic splicing, normalization, and nonlinear activation may be involved to finally generate the processing result of the first graph convolution restoration unit.

[0156] Next, the processing result of the w-th feature extraction subnetwork and the processing result of the first graph convolution restoration unit are loaded into the second graph convolution restoration unit of the r+1-th feature restoration subnetwork. The processing result of the w-th feature extraction subnetwork may contain other information related to the molten steel temperature obtained from the original monitoring data through multiple rounds of feature extraction, and this information enters the second graph convolution restoration unit together with the processing result of the first graph convolution restoration unit.

[0157] In the second graph convolutional reduction unit, the heuristic search function layer and the second multi-directional self-convergence network layer are used for feature processing. The heuristic search function layer constructs a new search space for the molten steel temperature related results from the w-th feature extraction subnetwork and the first graph convolutional reduction unit, and searches for new feature paths. For example, new relationship paths between molten steel temperature and equipment operating parameters at different refining stages may be searched, such as the association path between molten steel temperature and furnace stirring speed in a specific refining period. According to operations similar to the heuristic search function layer of the first graph convolutional reduction unit, the initial priority score of each candidate path is calculated, the top candidate paths are screened out, the gradient contribution is monitored, and weighted summation, dynamic path weight allocation and other operations are performed to generate optimized search features.

[0158] At the same time, the processing results of the w-th feature extraction subnetwork and the first graph convolution restoration unit on the molten steel temperature are loaded into the second multi-directional self-convergence network layer. The second multi-directional self-convergence network layer adjusts the parameters of the multi-branch topology structure according to the new data situation, such as changing the convolution kernel size and expansion rate of the hole convolution branch, the structural parameters of the depth-separable convolution branch, etc., to perform feature restoration operations on the molten steel temperature data. After the output features of each branch are spliced ​​along the channel dimension, the adaptive channel attention mechanism is used to assign dynamic weights, weighting, dimensionality reduction and other operations to generate initial aggregate features containing local features and global features, and then perform cross-channel feature interaction operations to adjust the weight distribution and generate optimized aggregate features.

[0159] The processing result of the first graph convolution restoration unit and the processing result of the heuristic search function layer of the second graph convolution restoration unit are fused with the processing result of the second multi-directional self-convergence network layer of the second graph convolution restoration unit using a supplementary link to generate the processing result of the second graph convolution restoration unit. Determine the third fusion result of the processing result of the first graph convolution restoration unit and the third influence factor (assuming γ), that is, γ multiplied by the processing result of the first graph convolution restoration unit (assuming the feature vector D) to obtain γ*D; determine the fourth fusion result of the processing result of the heuristic search function layer of the second graph convolution restoration unit (assuming the feature vector E) and the fourth influence factor (assuming δ), that is, δ*E. The third fusion result and the fourth fusion result are fused with the processing result of the second multi-directional self-convergence network layer of the second graph convolution restoration unit (assuming the feature vector F) using a supplementary link, and the processing result of the second graph convolution restoration unit is generated after performing relevant operations in the supplementary link.

[0160] Then, the processing result of the second graph convolution restoration unit is loaded into the third graph convolution restoration unit of the r+1th feature restoration subnetwork.

[0161] In the third graph convolution restoration unit, feature processing is performed using a heuristic search function layer and a multi-layer perceptron. The heuristic search function layer constructs a search space for the processing results of the second graph convolution restoration unit on the molten steel temperature, searching for, for example, the relationship path between the molten steel temperature and the quality index of the final refined product. According to the previous operation method, the optimized search features are obtained. The multi-layer perceptron performs further feature restoration and conversion operations on the processing results of the second graph convolution restoration unit, such as nonlinear transformation of the molten steel temperature related features, feature dimension adjustment, and other operations.

[0162] The processing result of the second graph convolution restoration unit and the processing result of the heuristic search function layer of the third graph convolution restoration unit are fused with the processing result of the multi-layer perceptron of the third graph convolution restoration unit by using a supplementary link to generate the processing result of the r+1th feature restoration sub-network. The fifth fusion result of the processing result of the second graph convolution restoration unit and the fifth influence factor (assuming ε) is determined, that is, ε multiplied by the processing result of the second graph convolution restoration unit (assuming the feature vector G) to obtain ε*G; the sixth fusion result of the processing result of the heuristic search function layer of the third graph convolution restoration unit (assuming the feature vector H) and the sixth influence factor (assuming ζ) is determined, that is, ζ*H. The fifth fusion result and the sixth fusion result are fused with the processing result of the multi-layer perceptron of the third graph convolution restoration unit (assuming the feature vector I) by using a supplementary link, and the processing result of the r+1th feature restoration sub-network is generated after performing relevant operations in the supplementary link. This result contains more accurate and comprehensive restored information about the molten steel temperature data after processing by multiple graph convolutional reduction units, which is of great significance for the abnormal diagnosis of LF refining process. In the entire LF refining process abnormal diagnosis network, other process parameter data, raw material characteristic data and equipment status data are processed according to a similar process, thereby realizing comprehensive monitoring and abnormal diagnosis of the LF refining process.

[0163] In a possible implementation, the method further includes:

[0164] When the number of cyclic training rounds of the LF refining process abnormality diagnosis network is not greater than the first set number of rounds, the optimization parameter amount of the neuron weight information optimized in each round of cyclic training of the LF refining process abnormality diagnosis network is set to be no greater than the set parameter amount.

[0165] When the cyclic training round number of the LF refining process abnormality diagnosis network is not less than a second set round number, the setting of the set parameter quantity is canceled, and the second set round number is greater than the first set round number.

[0166] In this embodiment, in this embodiment, first, about the situation when the number of cyclic training rounds of the LF refining process abnormality diagnosis network is not greater than the first set round. In the actual application scenario of the LF refining process, for example, in the LF refining link of steel production, the first set round (assuming 50 rounds) is a pre-set training stage boundary. In this initial training stage, the optimization parameter amount of the neuron weight information optimized by each round of cyclic training of the LF refining process abnormality diagnosis network is set to be no greater than the set parameter amount (assuming the set parameter amount is 100). This is because in the early stages of network training, too much weight optimization may cause the network to overfit. Taking molten steel temperature data as an example, when the network just starts training, if too much neuron weight information is optimized at one time, the network may pay too much attention to the local features in the training data, such as only overfitting the temperature fluctuation characteristics of a specific batch of molten steel at a certain refining stage, while ignoring the overall temperature change law and the relationship with other process parameters. Therefore, by limiting the number of parameters optimized in each round, the network can gradually learn the basic characteristics of the data, so that the network can converge more robustly in the initial training.

[0167] When the number of cyclic training rounds of the LF refining process abnormality diagnosis network is not less than the second set round (assuming that the second set round is 150 rounds), cancel the setting of the set parameter amount. This is because as the number of training rounds increases, the network has a certain understanding of the data, and at this time it is necessary to adjust the neuron weight information more comprehensively to further improve the performance of the network. For example, after 150 rounds of training, the network has a more comprehensive feature understanding of the molten steel temperature data and is no longer limited to the initial simple feature learning. At this time, canceling the parameter amount limit allows the network to optimize more neuron weights based on the global error information, thereby better capturing the complex relationship between the molten steel temperature and other process parameters (such as the amount of refining agent added, electrode power, etc.), and further improving the accuracy of the network in diagnosing LF refining process abnormalities.

[0168] In a possible implementation, the method further includes:

[0169] Obtain candidate refining process monitoring data.

[0170] The candidate refining process monitoring data is loaded into the x graph convolution construction units, and each of the x graph convolution construction units is used to perform feature processing respectively to generate graph convolution construction results extracted by the x graph convolution construction units. Each of the graph convolution construction units is used to perform feature processing respectively using the heuristic search function layer and the graph convolution function layer, and to process the graph convolution encoding result of the graph convolution construction unit using the supplementary links of the forward access node and the backward export node of the heuristic search function layer.

[0171] The graph convolution construction result is loaded into the y graph convolution restoration units, and each of the y graph convolution restoration units is used to perform graph convolution restoration respectively, to generate graph convolution restoration results output by the y graph convolution restoration units. Each of the graph convolution restoration units is used to perform graph convolution restoration respectively using the heuristic search function layer and the graph convolution function layer, and to process the graph convolution restoration results of the graph convolution restoration units using the supplementary link.

[0172] An estimated abnormality diagnosis result of the candidate refining process monitoring data is estimated based on the graph convolution restoration result.

[0173] In the LF refining process scenario, the candidate refining process monitoring data contains various information related to the refining process. Taking the LF refining furnace of a steel plant as an example, the process parameter data in the candidate refining process monitoring data include the real-time temperature of molten steel, the pressure value during the refining process, flow information, etc. The molten steel temperature data is accurately collected by the temperature sensor installed in the refining furnace. The pressure value reflects the pressure environment in the refining furnace, and the flow information involves the flow of various refining agents, protective gases, etc. The raw material characteristic data covers the composition information of iron ore, scrap steel, etc. put into the refining furnace, such as the iron content, impurity (sulfur, phosphorus, etc.) content in the iron ore, and the material and source of the scrap steel. The equipment status data includes the electrode status of the LF refining furnace (such as the degree of electrode wear, the current and voltage of the electrode, etc.), the refractory material status of the furnace body (such as the remaining thickness of the refractory material, the wear rate, etc.), and the operating status information of other auxiliary equipment (such as the speed of the stirring device, the operating stability, etc.).

[0174] Then, the candidate refining process monitoring data is loaded into x graph convolution building units. Assume x=3, that is, the candidate refining process monitoring data is loaded into these three graph convolution building units. Taking the first graph convolution building unit as an example, it starts to perform feature processing on the candidate refining process monitoring data. For the molten steel temperature data, the heuristic search function layer in the graph convolution building unit starts to work. The heuristic search function layer constructs a search space, which contains various candidate paths related to the molten steel temperature. For example, one candidate path is the relationship analysis between the molten steel temperature and the pressure value in the refining process, and the other may be the relationship analysis between the molten steel temperature and the impurity content in the iron ore. The initial priority score of each candidate path is calculated based on the feature distribution information of the molten steel temperature data. Assume that the initial priority score of the path between the molten steel temperature and the pressure value is 0.8, because in the LF refining process, the pressure change may have a more direct impact on the molten steel temperature; and the initial priority score of the path between the molten steel temperature and the impurity content in the iron ore is 0.6. According to the descending order of the initial priority score, the top candidate paths are selected, and the gradient contributions of these candidate paths are dynamically monitored during the training process. If the gradient contribution of the path of the relationship between molten steel temperature and iron ore impurity content is found to increase at a certain training stage, for example, because the influence of specific impurity content on molten steel temperature becomes more critical in the refining process of certain special steel grades, then its priority score will be updated accordingly. By weighted summing the output features of the selected candidate paths, a preliminary search feature containing key feature paths is obtained, and then a path weight dynamic allocation operation is performed based on the preliminary search features. The preliminary search features are nonlinearly weighted fused according to the contribution of each feature path to generate an optimized search feature.

[0175] At the same time, the molten steel temperature data is synchronously loaded into the graph convolution function layer. The multi-directional self-converging network layer in the graph convolution function layer processes the molten steel temperature data. The multi-directional self-converging network layer constructs multiple parallel branches, such as the dilated convolution branch, the depth-separable convolution branch, and the global average pooling branch. In the dilated convolution branch, a specific convolution kernel size and expansion rate are used to extract features for the molten steel temperature data. For example, a larger convolution kernel size can capture the changing trend of the molten steel temperature over a longer time scale, and may find the temperature fluctuation characteristics every 10-15 minutes; the depth-separable convolution branch focuses on mining the deep features of the molten steel temperature data, and may analyze the implicit information of the temperature data at different depth levels; the global average pooling branch obtains the global features of the molten steel temperature data, such as the average temperature level during the entire refining process. After splicing the output features of each branch along the channel dimension, the adaptive channel attention mechanism is used to assign dynamic weights to the spliced ​​features of each channel dimension. For example, higher weights are assigned to channel features related to the stage of rapid temperature rise of molten steel, because the temperature change in this stage may be more critical to the refining process, while lower weights are assigned to channel features related to the stage of temperature stabilization. The concatenated features of each channel dimension are weighted based on the assigned dynamic weights, and then the weighted concatenated features are reduced to match the input dimensions of the subsequent processing stage to generate initial aggregate features containing local features and global features. Cross-channel feature interaction operations are performed on the initial aggregate features in the multi-directional self-convergence network layer, and the weight distribution of each channel feature is adjusted through the adaptive channel attention mechanism to generate optimized aggregate features.

[0176] The optimized search feature and the optimized aggregation feature are input to the forward access node of the heuristic search function layer to perform feature dimension alignment and dynamic splicing operations. Assuming that the dimension of the optimized search feature is [10,20] and the dimension of the optimized aggregation feature is [10,15], they are adjusted to the same dimension through the forward access node of the heuristic search function layer and then dynamically spliced ​​to generate fused intermediate features. The fused intermediate features are passed to the feature normalization module through the backward export node of the heuristic search function layer to perform batch normalization and nonlinear activation processing. Batch normalization adjusts the numerical distribution of the fused intermediate features to an appropriate range, such as mapping the values ​​to a specific mean and standard deviation range, and then processes them through a nonlinear activation function (such as the ReLU function) to generate normalized fused features. The normalized fused features are residually connected to the original molten steel temperature data, and the feature maps of the two are element-by-element added through the jump connection mechanism to generate the feature processing results of the first graph convolution construction unit.

[0177] In the same way, the second and third graph convolution construction units also perform feature processing on the molten steel temperature data and other data (such as pressure value, flow information, raw material characteristic data, equipment status data, etc.) in the candidate refining process monitoring data, and finally generate the graph convolution construction results extracted by the three graph convolution construction units.

[0178] Next, the graph convolution construction result is loaded into y graph convolution reduction units. Assume that y=3, that is, the graph convolution construction result is loaded into these three graph convolution reduction units. Taking the first graph convolution reduction unit as an example, it performs graph convolution reduction on the graph convolution construction result. For the molten steel temperature related features in the graph convolution construction result, the heuristic search function layer in the graph convolution reduction unit starts to work. The heuristic search function layer constructs a search space, which contains new candidate paths related to the molten steel temperature. For example, one candidate path is the relationship analysis between the molten steel temperature and the current of the refining furnace electrode, and the other may be the relationship analysis between the molten steel temperature and the speed of the stirring device. The initial priority score of each candidate path is calculated based on the feature distribution information of the molten steel temperature related features in the graph convolution construction result. Assume that the initial priority score of the path between the molten steel temperature and the electrode current is 0.7, because the electrode current has an important influence on the control of the molten steel temperature; and the initial priority score of the path between the molten steel temperature and the speed of the stirring device is 0.6. According to the descending order of the initial priority score, the top candidate paths are selected, and the gradient contributions of these candidate paths are dynamically monitored during the training process. If the gradient contribution of the path of the relationship between molten steel temperature and stirring device speed is found to increase at a certain training stage, for example, because the influence of stirring device speed on the heat transfer inside molten steel becomes more critical under certain refining conditions, then its priority score will be updated accordingly. By weighted summing the output features of the selected candidate paths, a preliminary search feature containing key feature paths is obtained, and then a path weight dynamic allocation operation is performed based on the preliminary search features. The preliminary search features are nonlinearly weighted fused according to the contribution of each feature path to generate an optimized search feature.

[0179] At the same time, the molten steel temperature-related features in the graph convolution construction result are synchronously loaded into the first multi-directional self-converging network layer in the graph convolution reduction unit. The first multi-directional self-converging network layer processes the molten steel temperature-related features. Its multi-branch topological structure constructs multiple parallel branches, such as the dilated convolution branch, the depth-separable convolution branch, and the global average pooling branch. In the dilated convolution branch, a specific convolution kernel size and expansion rate are used for feature restoration. For example, the original change characteristics of the molten steel temperature at a specific time scale are restored by a suitable convolution kernel size; the depth-separable convolution branch focuses on mining the deep features of the molten steel temperature-related features and restores the implicit temperature information at different depth levels; the global average pooling branch obtains the global features of the molten steel temperature-related features, such as the average level of the molten steel temperature-related features in the entire graph convolution construction result. After the output features of each branch are spliced ​​along the channel dimension, a dynamic weight is assigned to the spliced ​​features of each channel dimension through an adaptive channel attention mechanism. For example, a higher weight is given to the channel features related to the key change stage of the molten steel temperature, while a lower weight is given to the channel features related to less important temperature fluctuations. The concatenated features of each channel dimension are weighted based on the assigned dynamic weights, and then the weighted concatenated features are reduced to match the input dimensions of the subsequent processing stage to generate initial aggregated features containing local features and global features. In the first multi-directional self-convergence network layer, cross-channel feature interaction operations are performed on the initial aggregated features, and the weight distribution of each channel feature is adjusted through the adaptive channel attention mechanism to generate optimized aggregated features.

[0180] The optimized search feature and the optimized aggregation feature are input to the forward access node of the supplementary link to perform feature dimension alignment and dynamic splicing operations. Assuming that the dimension of the optimized search feature is [8,18] and the dimension of the optimized aggregation feature is [8,16], they are adjusted to the same dimension through the forward access node of the supplementary link and then dynamically spliced ​​to generate fused intermediate features. The fused intermediate features are passed to the feature normalization module through the backward export node of the supplementary link to perform batch normalization and nonlinear activation processing. Batch normalization adjusts the numerical distribution of the fused intermediate features to an appropriate range, such as mapping the values ​​to a specific mean and standard deviation range, and then processed through a nonlinear activation function (such as the ReLU function) to generate normalized fused features. The normalized fused features are residually connected to the original molten steel temperature-related features in the graph convolution construction result, and the feature maps of the two are element-by-element added through the jump connection mechanism to generate the feature processing results of the first graph convolution reduction unit.

[0181] In the same way, the second and third graph convolution restoration units also perform graph convolution restoration on the molten steel temperature-related features and other data (such as pressure values, flow information, raw material characteristic data, equipment status data, etc.) in the graph convolution construction results, and finally generate the graph convolution restoration results output by the three graph convolution restoration units.

[0182] Finally, the estimated abnormal diagnosis results of the candidate refining process monitoring data are estimated based on the graph convolution restoration results. Taking the molten steel temperature data as an example, if the graph convolution restoration results show that at a certain stage in the refining process, the change trend of the molten steel temperature deviates greatly from the normal refining process temperature curve, and the correlation with other related parameters (such as electrode current, stirring device speed, impurity content in raw materials, etc.) does not conform to the normal mode, then it can be judged that there may be abnormal conditions at this stage. For example, under normal circumstances, in the middle stage of refining, the molten steel temperature should be stable between 1600-1620 degrees Celsius, and as the refining proceeds, the molten steel temperature and the electrode current should show a certain negative correlation (because the reduction of electrode current helps to maintain temperature stability). If the graph convolution restoration results show that the molten steel temperature suddenly rises to 1650 degrees Celsius at this stage, and the relationship with the electrode current becomes positively correlated, this suggests that there may be problems such as electrode failure or refining operation errors, thereby obtaining a diagnosis result that there is an abnormality in this candidate refining process monitoring data. A similar approach is used to conduct a comprehensive analysis of other process parameter data, raw material characteristic data, and equipment status data in the graph convolution restoration results. If the data performance of multiple parameters deviates from the normal mode, the possibility of anomaly is higher, and the type of anomaly can be preliminarily determined based on the degree and characteristics of the deviation, such as equipment failure, raw material quality problems, or operational process errors. In this way, the entire candidate refining process monitoring data is comprehensively analyzed to obtain accurate estimated anomaly diagnosis results.

[0183] Figure 2 The hardware structure of the artificial intelligence system 100 provided in the embodiment of the present application is shown as follows: Figure 2 As shown, the artificial intelligence system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0184] In one possible design, the artificial intelligence system 100 may be a single server or a server group. The server group may be centralized or distributed (e.g., the artificial intelligence system 100 may be a distributed system). In some embodiments, the artificial intelligence system 100 may be local or remote. For example, the artificial intelligence system 100 may access information and / or data stored in the machine-readable storage medium 120 via a network. For another example, the artificial intelligence system 100 may be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the artificial intelligence system 100 may be implemented on an artificial intelligence system. By way of example only, the artificial intelligence system may include a private cloud, a semantically relevant cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, or the like, or any aggregation thereof.

[0185] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the artificial intelligence system 100 uses to execute or use to complete the exemplary methods described in this application.

[0186] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the AI ​​model training method based on the LF refining process in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0187] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned artificial intelligence system 100. The implementation principles and technical effects are similar and will not be repeated in this embodiment.

[0188] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are set. When the processor executes the computer executable instructions, the above-mentioned AI model training method based on the LF refining process is implemented.

[0189] It should be noted that, according to the above embodiments of the present invention, those skilled in the art can fully realize the full scope of the independent claims and dependent rights of the present invention, and the implementation process and method are the same as the above embodiments; and the part not described in detail in the present invention belongs to the well-known technology in the field. However, the protection scope of the present invention is not limited to this, and any changes or substitutions that can be easily thought of by any person familiar with the field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. An AI model training method based on LF refining process, characterized in that: The AI ​​model includes an LF refining process abnormality diagnosis network, the LF refining process abnormality diagnosis network includes x graph convolution construction units linked in sequence and y graph convolution reduction units linked in sequence, each of the graph convolution construction units and each of the graph convolution reduction units includes a heuristic search function layer and a graph convolution function layer linked in sequence, x is a positive integer not less than 2, y is a positive integer not less than 3, and the method includes: Acquiring template refining process monitoring data, wherein the template refining process monitoring data includes process parameter data, raw material characteristic data, and equipment status data; The template refining process monitoring data is loaded into the x graph convolution construction units, and each of the x graph convolution construction units is used to perform feature processing respectively to generate graph convolution construction results extracted by the x graph convolution construction units; wherein each of the graph convolution construction units is used to perform feature processing respectively using the heuristic search function layer and the graph convolution function layer, and to process the graph convolution encoding result of the graph convolution construction unit using the supplementary links of the forward access node and the backward export node of the heuristic search function layer; The graph convolution construction result is loaded into the y graph convolution restoration units, and each of the y graph convolution restoration units is used to perform graph convolution restoration respectively, to generate graph convolution restoration results output by the y graph convolution restoration units; wherein each of the graph convolution restoration units is used to perform graph convolution restoration respectively using the heuristic search function layer and the graph convolution function layer, and to process the graph convolution restoration result of the graph convolution restoration unit using the supplementary link; estimating a template anomaly diagnosis result of the template refining process monitoring data based on the graph convolution restoration result; Based on the labeled abnormality diagnosis result of the template refining process monitoring data and the loss function value of the template abnormality diagnosis result, the neuron weight information of the LF refining process abnormality diagnosis network is trained.

2. The AI ​​model training method based on LF refining process according to claim 1 is characterized in that: The LF refining process abnormality diagnosis network includes w sequentially linked feature extraction subnetworks, each of the feature extraction subnetworks includes a plurality of the graph convolution construction units sequentially linked in the x graph convolution construction units, w is a positive integer not less than 2; the template refining process monitoring data is loaded into the x graph convolution construction units, and each of the x graph convolution construction units is used to perform feature processing respectively to generate the graph convolution construction results extracted by the x graph convolution construction units, including: Loading the template refining process monitoring data into a first feature extraction subnetwork, performing feature processing using each of the graph convolution construction units in the first feature extraction subnetwork, and generating a processing result of the first feature extraction subnetwork; Loading the processing result of the qth feature extraction subnetwork into the q+1th feature extraction subnetwork, using each of the graph convolution construction units in the q+1th feature extraction subnetwork to perform feature processing respectively, generating the processing result of the q+1th feature extraction subnetwork, where q is a positive integer and q+1 is not greater than w; The LF refining process abnormality diagnosis network includes w sequentially linked feature restoration subnetworks, each of which includes at least three of the y graph convolution restoration units sequentially linked; the graph convolution construction result is loaded into the y graph convolution restoration units, and each of the y graph convolution restoration units is used to perform graph convolution restoration respectively, to generate the graph convolution restoration results output by the y graph convolution restoration units, including: Loading the processing result of the w-th feature extraction subnetwork into the first feature restoration subnetwork, using each of the graph convolution restoration units in the first feature restoration subnetwork to perform graph convolution restoration respectively, to generate the processing result of the first feature restoration subnetwork; The processing results of the w-th feature extraction subnetwork and the processing results of the r-th feature restoration subnetwork are loaded into the r+1-th feature restoration subnetwork, and each of the graph convolution restoration units in the r+1-th feature restoration subnetwork is used to perform graph convolution restoration respectively to generate the processing results of the r+1-th feature restoration subnetwork, where r is a positive integer and r+1 is not greater than w.

3. The AI ​​model training method based on LF refining process according to claim 2 is characterized in that: Each of the feature extraction subnetworks comprises a first graph convolution construction unit and a second graph convolution construction unit which are sequentially connected, wherein the graph convolution function layer of the first graph convolution construction unit is a multi-directional self-convergence network layer, and the graph convolution function layer of the second graph convolution construction unit is a multi-layer perceptron; The processing result of the qth feature extraction subnetwork is loaded into the q+1th feature extraction subnetwork, and each of the graph convolution construction units in the q+1th feature extraction subnetwork is used to perform feature processing respectively to generate the processing result of the q+1th feature extraction subnetwork, including: Loading the processing result of the qth feature extraction subnetwork into the first graph convolution construction unit of the q+1th feature extraction subnetwork; Using the heuristic search function layer and the multi-directional self-convergence network layer of the first graph convolution construction unit to perform feature processing on the processing result of the qth feature extraction subnetwork respectively; The processing result of the qth feature extraction subnetwork and the processing result of the heuristic search function layer of the first graph convolution construction unit are fused with the processing result of the multi-directional self-convergence network layer of the first graph convolution construction unit by using the supplementary link to generate the processing result of the first graph convolution construction unit; Loading the processing result of the first graph convolution construction unit into the second graph convolution construction unit of the q+1th feature extraction subnetwork; Using the heuristic search function layer and the multi-layer perceptron of the second graph convolution construction unit to perform feature processing respectively; The processing result of the first graph convolution construction unit and the processing result of the heuristic search function layer of the second graph convolution construction unit are fused with the processing result of the multi-layer perceptron of the second graph convolution construction unit by using the supplementary link to generate the processing result of the q+1th feature extraction subnetwork.

4. The AI ​​model training method based on LF refining process according to claim 3 is characterized in that: The step of fusing the processing result of the qth feature extraction subnetwork and the processing result of the heuristic search function layer of the first graph convolution construction unit with the processing result of the multidirectional self-convergence network layer of the first graph convolution construction unit by using the supplementary link to generate the processing result of the first graph convolution construction unit includes: Determine a first fusion result of a processing result of the qth feature extraction subnetwork and a first influencing factor, and determine a second fusion result of a processing result of the heuristic search function layer of the first graph convolution construction unit and a second influencing factor; The first fusion result and the second fusion result are merged by using the supplementary link and the processing result of the multi-directional self-convergence network layer of the first graph convolution construction unit; The processing result of the first graph convolution construction unit and the processing result of the heuristic search function layer of the second graph convolution construction unit are fused with the processing result of the multi-layer perceptron of the second graph convolution construction unit by using the supplementary link to generate the processing result of the q+1th feature extraction subnetwork, including: Determine a third fusion result of the processing result of the first graph convolution construction unit and the third influencing factor; and determine a fourth fusion result of the processing result of the heuristic search function layer of the second graph convolution construction unit and the fourth influencing factor; The third fusion result and the fourth fusion result are fused with the processing result of the multilayer perceptron of the second graph convolution construction unit by using the supplementary link.

5. The AI ​​model training method based on LF refining process according to claim 4 is characterized in that: The method further comprises: At least one of the first influencing factor, the second influencing factor, the third influencing factor and the fourth influencing factor is optimized based on the loss function value of the labeled abnormality diagnosis result of the template refining process monitoring data and the template abnormality diagnosis result.

6. The AI ​​model training method based on LF refining process according to claim 3 is characterized in that: The step of performing feature processing on the processing result of the qth feature extraction subnetwork by respectively utilizing the heuristic search function layer and the multidirectional self-convergence network layer of the first graph convolution construction unit comprises: Loading the processing result of the qth feature extraction subnetwork as the first input data into the heuristic search function layer, performing multi-dimensional feature path traversal on the first input data through the dynamic feature path search mechanism in the heuristic search function layer, and generating preliminary search features including key feature paths; Based on the preliminary search features, a path weight dynamic allocation operation is performed in the heuristic search function layer, and the preliminary search features are nonlinearly weighted fused according to the contribution of each feature path to generate an optimized search feature; Synchronously loading the processing result of the qth feature extraction subnetwork into the multi-directional self-converging network layer, performing parallel multi-scale feature extraction on the first input data through the multi-branch topological structure in the multi-directional self-converging network layer, and generating initial aggregated features including local features and global features; Performing cross-channel feature interaction operations on the initial aggregated features in the multi-directional self-convergence network layer, adjusting the weight distribution of each channel feature through an adaptive channel attention mechanism, and generating optimized aggregated features; Inputting the optimized search feature and the optimized aggregation feature into the forward access node of the supplementary link, performing feature dimension alignment and dynamic splicing operations, and generating a fused intermediate feature; The fused intermediate features are transmitted to the feature normalization module through the backward derivation node of the supplementary link, batch normalization and nonlinear activation processing are performed to generate normalized fused features; Performing a residual connection between the normalized fusion feature and the processing result of the qth feature extraction subnetwork, performing an element-by-element addition operation on the feature graphs of the two through a skip connection mechanism, and generating a feature processing result of the first graph convolution construction unit; Inputting the feature processing result into the feedback adjustment module of the multi-directional self-converging network layer, dynamically adjusting the parameter configuration of the multi-branch topology structure based on the gradient back propagation information of the current training round, and generating an updated multi-directional self-converging network layer; Performing secondary multi-scale feature extraction on the feature processing result by using the updated multi-directional self-converging network layer to generate enhanced aggregation features; Inputting the enhanced aggregation features and the optimized search features into the iterative optimization module of the heuristic search function layer, screening out a subset of significant distinguishing features through a recursive feature selection mechanism, and generating an iteratively optimized search feature; Cross-modally fuse the iteratively optimized search features and the enhanced aggregated features in the supplementary link, calculate the joint weight matrix of the iteratively optimized search features and the enhanced aggregated features through a multimodal attention mechanism, and generate a cross-modal fusion feature; Performing a spatial pyramid pooling operation on the cross-modal fusion feature, extracting spatial context information of different scales through a multi-level pooling window, and generating a multi-scale pooling feature; Inputting the multi-scale pooling features into a feature compression module, reducing the feature channel dimension through a point-by-point convolution layer, and generating a compressed low-dimensional feature representation; Perform gated fusion on the low-dimensional feature representation and the processing result of the qth feature extraction subnetwork, dynamically adjust the fusion ratio of the two through learnable gating weights, and generate a final feature encoding result; The feature encoding result is passed to the next processing stage as the output of the first graph convolution construction unit and as the input data source of the subsequent graph convolution functional layer.

7. The AI ​​model training method based on LF refining process according to any one of claims 2 to 5, characterized in that: Each of the feature restoration subnetworks comprises a first graph convolution restoration unit, a second graph convolution restoration unit and a third graph convolution restoration unit which are sequentially connected, wherein the graph convolution function layer of the first graph convolution restoration unit is a first multi-directional self-convergence network layer, the graph convolution function layer of the second graph convolution restoration unit is a second multi-directional self-convergence network layer, and the graph convolution function layer of the third graph convolution restoration unit is a multi-layer perceptron; The processing result of the w-th feature extraction subnetwork and the processing result of the r-th feature restoration subnetwork are loaded into the r+1-th feature restoration subnetwork, and each of the graph convolution restoration units in the r+1-th feature restoration subnetwork is used to perform graph convolution restoration respectively to generate the processing result of the r+1-th feature restoration subnetwork, including: Loading the processing result of the rth feature restoration sub-network into the first graph convolution restoration unit of the r+1th feature restoration sub-network; Using the heuristic search function layer and the first multi-directional self-convergence network layer of the first graph convolution restoration unit to perform feature processing on the processing result of the rth feature restoration sub-network respectively; The processing result of the r-th feature restoration subnetwork and the processing result of the heuristic search function layer of the first graph convolution restoration unit are fused with the processing result of the first multi-directional self-convergence network layer of the first graph convolution restoration unit by using the supplementary link to generate the processing result of the first graph convolution restoration unit; Loading the processing result of the w-th feature extraction subnetwork and the processing result of the first graph convolution restoration unit into the second graph convolution restoration unit of the r+1-th feature restoration subnetwork; Using the heuristic search function layer and the second multi-directional self-convergence network layer of the second graph convolutional restoration unit to perform feature processing respectively; The processing result of the first graph convolution restoration unit and the processing result of the heuristic search function layer of the second graph convolution restoration unit are fused with the processing result of the second multi-directional self-convergence network layer of the second graph convolution restoration unit by using the supplementary link to generate the processing result of the second graph convolution restoration unit; Loading the processing result of the second graph convolution restoration unit to the third graph convolution restoration unit of the r+1th feature restoration subnetwork; Using the heuristic search function layer and the multi-layer perceptron of the third graph convolutional restoration unit to perform feature processing respectively; The processing results of the second graph convolution restoration unit and the processing results of the heuristic search function layer of the third graph convolution restoration unit are fused with the processing results of the multi-layer perceptron of the third graph convolution restoration unit using the supplementary link to generate the processing results of the r+1th feature restoration subnetwork.

8. The AI ​​model training method based on LF refining process according to any one of claims 1 to 5, characterized in that: The method further comprises: When the number of cyclic training rounds of the LF refining process abnormality diagnosis network is not greater than the first set number of rounds, the optimization parameter amount of the neuron weight information optimized by each cyclic training of the LF refining process abnormality diagnosis network is set to be not greater than the set parameter amount; When the cyclic training round number of the LF refining process abnormality diagnosis network is not less than a second set round number, the setting of the set parameter quantity is canceled, and the second set round number is greater than the first set round number.

9. The AI ​​model training method based on LF refining process according to any one of claims 1 to 5, characterized in that: The method further comprises: Obtain candidate refining process monitoring data; The candidate refining process monitoring data is loaded into the x graph convolution construction units, and each of the x graph convolution construction units is used to perform feature processing respectively to generate graph convolution construction results extracted by the x graph convolution construction units; wherein each of the graph convolution construction units is used to perform feature processing respectively using the heuristic search function layer and the graph convolution function layer, and to process the graph convolution encoding result of the graph convolution construction unit using the supplementary links of the forward access node and the backward export node of the heuristic search function layer; The graph convolution construction result is loaded into the y graph convolution restoration units, and each of the y graph convolution restoration units is used to perform graph convolution restoration respectively, to generate graph convolution restoration results output by the y graph convolution restoration units; wherein each of the graph convolution restoration units is used to perform graph convolution restoration respectively using the heuristic search function layer and the graph convolution function layer, and to process the graph convolution restoration result of the graph convolution restoration unit using the supplementary link; An estimated abnormality diagnosis result of the candidate refining process monitoring data is estimated based on the graph convolution restoration result.

10. An artificial intelligence system, characterized in that: The artificial intelligence system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the AI ​​model training method based on the LF refining process as described in any one of claims 1 to 9 above.

Images