Industrial process fault diagnosis method based on intervention enhanced invariant causal graph learning network

Abstract

The invention provides an industrial process fault diagnosis method based on an intervention-enhanced invariant causal graph learning network, and the method comprises the steps: 1, carrying out the data preprocessing of process data of process variables collected in a target industrial operation process, and constructing a variable association graph which reflects the topological relation between the variables; step 2, establishing an invariant causal graph learning network; 3, performing intervention enhancement, and constructing an intervention distribution environment; and step 4, based on a joint prediction mechanism, integrating a cause-and-fruit sub-graph prediction result and a false sub-graph prediction result, introducing an invariant intervention risk by using cause-and-fruit invariance, and identifying stable fault cause-and-fruit sub-graph structures in different intervention distribution environments. The fault cause subgraph structure plays a cause-effect decisive role on fault categories and describes a fault mechanism and a propagation path, and fault diagnosis is realized according to the fault cause subgraph structure. Compared with other methods, the fault diagnosis method provided by the invention can improve the reliability and interpretability of fault diagnosis in the industrial process.

Description

Technical Field

[0001] This invention belongs to the field of process monitoring and fault diagnosis, specifically relating to an industrial process fault diagnosis method based on an intervention-enhanced invariant causal graph learning network. Background Technology

[0002] Modern industrial processes are large-scale and have complex internal mechanisms. Faults often trigger chain reactions, leading to significant losses and safety risks. Therefore, industrial process fault diagnosis is crucial for ensuring industrial production safety and improving efficiency. Actual industrial processes involve a large amount of monitoring variable data, with complex interrelationships among these variables. Faults in industrial processes do not occur independently but propagate along fault paths, causing abnormal changes in multiple related variables. Graph Neural Networks (GNNs) can utilize neural networks to process graph-structured data, and their message passing mechanism allows each node to aggregate information from its neighbors to update the state of the central node. Fault diagnosis models based on GNNs can enhance the accuracy of fault diagnosis and perform excellently in graph classification tasks across multiple domains. Industrial process fault diagnosis is modeled as a graph classification task, where each fault corresponds to a different graph structure. Fault diagnosis is achieved by identifying and classifying the graph structures of different faults. Existing GNN-based industrial process fault diagnosis methods mostly describe the correlation between variables by building correlation graphs, but their weak generalization ability leads to decreased fault diagnosis performance. However, most existing research on causal discovery in industrial processes transforms multivariate sensor data into graph structures based on causal relationships defined by prediction, such as Granger causality analysis and transitive entropy. However, these methods largely remain at the level of statistical correlation, neglecting the mining and utilization of fault causal subgraph structures that play a causal deterministic role in fault categories and can describe the mechanisms and propagation paths of faults. The goal of these existing methods is usually to find statistical associations between the input graph and the true fault labels; however, such associations may not reflect true causal relationships, and the constructed causal graphs contain a large number of spurious correlations. Therefore, the model's fault diagnosis decisions are easily influenced by shortcut features in the training data, stemming from factors such as sample selection or environmental noise. Shortcut features lack cross-environment stability, resulting in weak out-of-distribution generalization ability and a sharp decline in performance when operating conditions change or new noise appears. Furthermore, because the decisions do not rely on causal features in the data, the interpretability of the fault diagnosis process is weak. In industrial processes, it is difficult to trace the propagation path and causes of faults. Therefore, a method should be adopted to extract the fault causal subgraph structure from the fault variable association graph structure, avoiding learning shortcut structures but learning the causal structures that truly determine the fault type, and using the fault causal subgraph structure for fault diagnosis to improve the generalization ability, stability and interpretability of the industrial process fault diagnosis model. Summary of the Invention

[0003] Objective of the Invention: The technical problem to be solved by this invention is to address the shortcomings of existing technologies by providing a method for fault diagnosis of industrial processes based on an intervention-enhanced invariant causal graph learning network, comprising the following steps:

[0004] Step 1: For the target industrial process, preprocess the process data of process variables collected during operation to obtain standardized time series data; generate dynamic time series data through a fixed-length sliding window, and construct a variable correlation graph reflecting the topological relationship between variables based on the Pearson correlation coefficient;

[0005] Step 2: Establish an Invariant Causal Graph Learning Network (IICGLN), which includes a causal generator, a graph encoder, and a classifier. Input the variable association graph into the causal generator to divide it into a fault causal subgraph and an environmental spurious subgraph. Use a shared graph encoder to extract subgraph features from the divided fault causal subgraph and environmental spurious subgraph respectively. The classifier includes a causal classifier and a spurious classifier, which are used to independently output causal prediction results and spurious prediction results.

[0006] Step 3: Based on the fault causal subgraph and the spurious environment subgraph divided by the causal generator, intervention enhancement is performed. By recombining the fault causal subgraph of the current training sample with the spurious environment subgraph of other training samples, an intervention distribution environment is actively constructed.

[0007] Step 4: Based on the joint prediction mechanism, the prediction results of causal subgraphs and spurious subgraphs are integrated. Causal invariance is used to introduce invariant intervention risk, and stable fault causal subgraph structures under different intervention distribution environments are identified, thereby improving the generalization ability and stability of the fault diagnosis model. The industrial process fault diagnosis model based on the intervention-enhanced invariant causal graph learning network finally uses only pure causal prediction when performing fault diagnosis, actively eliminating the influence of spurious environmental subgraphs identified during training, and ensuring that the fault diagnosis results are completely determined by stable fault causal subgraphs.

[0008] In step 1, the data preprocessing includes: Indicates inclusion The training data for each sample, where Let T represent the t-th sample containing n process variables, where T denotes transpose. Represent the real number space; perform robust normalization data preprocessing on the process data of the process variables, and for the first... Time series of 1 variable , ,in Indicates the first The Nth sample value of the nth variable, using the nth... Median of the variables Deviation from median Scaling is performed to obtain standardized data. ,in is the numerical stability constant; median denotes the median function;

[0009] definition ,in This represents the standardized data value of the nth variable in the tth sample. Let represent the standardized sample vector of the t-th sample; a sliding window is used to process the standardized time series data, and the input matrix of the t-th window is . ,in This indicates the size of the sliding window, and the corresponding fault label for the window is... .

[0010] In step 1, a variable relationship graph is built based on the standardized data within the sliding window. Each process variable is treated as a node in the graph, and a node set is defined. ,in This represents the nth process variable, and the i-th process variable is in the current window. The standardized sample sequence within is a node The attributes, using the Pearson correlation coefficient Quantization Node and Correlation strength within the current time window:

[0011] ,

[0012] in Represents a node and nodes Covariance within the current window. Represents a node The standard deviation within the current window. Represents a node The standard deviation within the current window. Representing variables The first in the current window One data point, Representing variables The mean within the current window;

[0013] Based on a preset threshold The association strength is filtered, and only strong associations above a threshold are retained as edges to construct a sparse adjacency matrix. :

[0014] ,

[0015] in This represents the element in the i-th row and j-th column of the sparse adjacency matrix A;

[0016] If and only if the correlation coefficient When, there are edges in the graph Construct an undirected graph , representing the relationship structure between variables within the window, serves as the input for the subsequent fault diagnosis model based on the Invariant Causal Graph Learning Network (IICGLN).

[0017] In step 2, the causal generator will generate an undirected graph. Divided into fault cause-effect subgraphs and environmental fake subgraph First, extract the node features of the variable association graph. Calculate the edge mask matrix based on node features , Represents a node and nodes The importance scores between them are expressed as:

[0018] ,

[0019] ,

[0020] in, and They represent the process. Extracted nodes Feature representation vectors and nodes The feature representation vector, This represents the Sigmoid activation function. It is a convolution of local extrema maps ( A graph feature extraction network consisting of stacked layers;

[0021] Based on mask matrix and the original adjacency matrix Determine the importance score for each edge to measure its contribution to fault diagnosis prediction;

[0022] According to the preset causal division ratio r, the edges with the highest scores in the top r ratios are selected to form the fault causal subgraph. edge set Select the remaining edges with lower scores as the false subgraph of the environment. edge set And find two sets of edge sets. and Construct fault cause-effect subgraphs for the corresponding nodes. and environmental fake subgraph , is represented as:

[0023] ,

[0024] ,

[0025] in, This refers to sorting the matrix elements according to the importance scores of all edges, and retaining the edges corresponding to the top r proportions with the highest scores. It refers to element-wise product.

[0026] In step 2, the fault causal subgraph is analyzed in the intervention environment. and environmental fake subgraph Encoding and independent classification force the model to capture cross-environment stable fault causal subgraph structures, where these subgraphs play a causal deterministic role in fault categories and describe fault mechanisms and propagation paths. This is achieved by first utilizing a shared graph encoder. Fault Cause-and-Effect Subgraphs Encode the spurious environmental subgraph s and use GNN to generate a fault cause-effect subgraph. Node feature representation Node feature representation of the spurious subgraph of the environment Then, global pooling is used to represent the node features. Aggregation into graph-level feature representation Representing node features Aggregation into graph-level feature representation Preserving the structure and key information of the subgraph, it is represented as follows:

[0027] ,

[0028] ,

[0029] in, It is a convolution of local extrema maps ( A graph feature extraction network consisting of stacked layers; Pooling refers to the global pooling operation.

[0030] Two independent classifiers are used to predict the fault causal subgraph c and the environmental spurious subgraph s, respectively. The causal classifier... Graph-level feature representation based on fault causal subgraph Perform causal prediction and output the predicted causal part of the fault. Learn the fundamental causal mechanisms of failures;

[0031] False classifier Graph-level representation based on environmental pseudo subgraphs Perform false predictions and output the predictions that are entirely false. This measures the ability of spurious environmental associations to predict fault labels.

[0032] In step 3, an intervention-enhanced training process is executed. This involves recombining the fault causal subgraph of the current training sample with the spurious environmental subgraph of other training samples to actively construct two or more intervention distribution environments. First, the spurious environmental subgraph is generated by dividing the fault variable association graph g after causal generation. Collect the data and construct a fake environmental sub-library; then perform intervention operations. ,in Represents environmental spurious subgraph variables. This represents the selected specific environment dummy subgraph instance sample, obtained by using the current sample. Fault Cause-and-Effect Graph The original fake environment subplots were fixed, and some of them were replaced with fake environment subplots from other samples in the library. Create new intervention samples .

[0033] In step 4, the joint prediction mechanism integrates the causal subgraph prediction results and the spurious subgraph prediction results, forcing the model to overcome the interference of spurious environmental subgraphs and focus on stable fault causal subgraph structures under different intervention environments; the model incorporates causal prediction during training. and false predictions Conduct joint forecasting:

[0034] ,

[0035] Among them, confidence level This indicates the reliability of the spurious environmental subgraph for fault diagnosis; if the spurious environmental association is reliable, then... , indicating confidence level The trend is towards 1, and the joint prediction preserves the causal subgraph prediction results. ;

[0036] If the spurious environmental subgraph is ineffective for fault diagnosis, then , indicating confidence level The trend towards 0 leads to the joint prediction results. Close to 0.

[0037] In step 4, the risk of invariable intervention remains unchanged. The calculation formula is:

[0038] ,

[0039] in, Indicates the actual fault label, This represents the fault diagnosis model. The balancing hyperparameters are used to adjust the weights of the variance term under the invariant intervention risk. Multiple different distribution environments are constructed under the intervention, resulting in a set of losses under different intervention environments. and These represent the mean and variance of the risk values, respectively. The risk loss function is expressed by the following formula:

[0040] ,

[0041] in Represents the cross-entropy loss function. Represent the distribution of data without intervention; calculate the expectation under a specific intervention distribution, in the uninterventional data distribution. In the middle, maintain the causal structure of the fault. Causal structure of the original sample The spurious environmental structure is a spurious environmental subgraph under specific interventions. ;

[0042] Causal prediction loss The calculation formula is:

[0043] ,

[0044] in, This represents the expectation over the original data distribution;

[0045] Design an independent environmental spurious loss function to train a spurious classifier. Identify spurious structures in the environment during training. Update only the fake classifier Parameters, environmental spurious loss function The calculation formula is:

[0046] ,

[0047] Total loss function for:

[0048] ,

[0049] After training, the model uses only pure causal prediction for fault diagnosis. First, the input graph g is processed by a causal generator. Extract the fault causal subgraph c, using only the encoder and causal classifier. The causal branches are based on the partitioned fault causal subgraph. To make the final causal prediction of the failure. By actively eliminating interference from spurious environmental sub-images identified during training, false classifiers are discarded. And a joint prediction mechanism to ensure that fault diagnosis decisions are entirely determined by a stable fault causal subgraph.

[0050] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0051] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0052] This method employs the Invariant Causal Graph Learning Network (IICGLN), which consists of a causal generator, a graph encoder, and a classifier. Based on this, intervention-based augmentation training is performed to proactively construct multiple intervention environments. Utilizing a joint prediction mechanism and invariant intervention risk optimization, a stable fault causal subgraph structure across environments is identified, and the influence of spurious environmental associations is proactively eliminated. Furthermore, the stable fault causal subgraph structure is used to achieve causal relationship-based industrial process fault diagnosis, improving the reliability and interpretability of industrial process monitoring.

[0053] The present invention has the following advantages: First, the industrial process fault diagnosis method based on intervention-enhanced invariant causal graph learning network of the present invention constructs a graph neural network model based on the fault causal subgraph structure. This causal structure describes the causal relationship between process variables, which conforms to the physical mechanism of industrial processes, making the industrial fault diagnosis model have high reliability and interpretability.

[0054] Second, the industrial process fault diagnosis method based on intervention-enhanced invariant causal graph learning network of the present invention has a network architecture including a causal generator, a graph encoder, and a classifier. The causal generator module can divide the variable association graph into a fault causal subgraph and an environmental spurious subgraph. The graph encoder module can uniformly extract features from the divided subgraphs and generate graph-level representations. The classifier module includes a causal classifier and a spurious classifier that can independently output causal prediction results and spurious prediction results. This network architecture is conducive to realizing causal discovery and fault detection.

[0055] Third, the industrial process fault diagnosis method based on intervention-enhanced invariant causal graph learning network of the present invention utilizes the intervention enhancement mechanism and the principle of causal invariance to extract the fault causal subgraph structure from the correlation of variable correlation graphs. It can proactively overcome the influence of spurious correlations in industrial data, improve the effectiveness of causal discovery and the interpretability of fault diagnosis. Attached Figure Description

[0056] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0057] Figure 1 This is a schematic diagram of a structural causal model.

[0058] Figure 2 This is a schematic diagram of the IICGLN, an intervention-enhanced invariant causal graph learning network.

[0059] Figure 3 This is a training accuracy graph for the Tennessee-Eastman process.

[0060] Figure 4 This is a process flow diagram of the Tennessee-Eastman process. Detailed Implementation

[0061] This embodiment provides an industrial process fault diagnosis method based on an intervention-enhanced invariant causal graph learning network, which includes the following steps S1 to S4. The implementation of each step is described in detail below.

[0062] S1. After determining the target industrial process, first understand the process flow to be modeled, and clarify the process variables and required process data involved in industrial process modeling. For the target industrial process, preprocess the process data of the process variables collected during operation to obtain standardized time-series data; generate dynamic time-series data through a fixed-length sliding window, and construct a variable correlation graph reflecting the topological relationship between variables based on the Pearson correlation coefficient; step S1 is as follows:

[0063] S11: The data preprocessing described above Indicates inclusion The training data for each sample, where Represents the nth process variable. There are 10 samples, and T represents the transpose. Represent the real number space; perform robust normalization data preprocessing on the process data of the process variables, and for the first... Time series of 1 variable , ,in Indicates the first The Nth sample value of a variable. Using the median of that variable. Deviation from median Scaling is performed to obtain standardized data. ,in is the numerical stability constant.

[0064] S12: Definition ,in This represents the standardized data value of the nth variable in the tth sample. Let represent the standardized sample vector of the t-th sample. A sliding window is used to process the standardized time series data; the input matrix for the t-th window is . ,in This indicates the size of the sliding window, and the corresponding fault label for the window is... .

[0065] S13: Build a variable relationship graph based on the standardized data within the sliding window, treating each individual process variable as a node in the graph, and define a node set. This represents n process variables, where the i-th process variable is in the current window. The standardized sample sequence within is its node The attributes, using the Pearson correlation coefficient Quantization Node and Correlation strength within the current time window:

[0066] ,

[0067] in Represents a node and nodes Covariance within the current window. Represents a node The standard deviation within the current window. Represents a node The standard deviation within the current window. Representing variables The first in the current window One data point, Representing variables The mean within the current window;

[0068] Based on a preset threshold The association strength is filtered, and only strong associations above a threshold are retained as edges, thus constructing a sparse adjacency matrix. :

[0069] ,

[0070] If and only if the correlation coefficient When, there are edges in the graph Constructing an undirected graph , representing the relationship structure between variables within the window, serves as the input for the subsequent fault diagnosis model based on the Invariant Causal Graph Learning Network (IICGLN).

[0071] S2. Establish an Invariant Causal Graph Learning Network (IICGLN), which includes a causal generator, a graph encoder, and a classifier. The variable association graph is input into the causal generator to divide it into a fault causal subgraph and an environmental spurious subgraph. The structure of the fault causal subgraph plays a causal deterministic role in the fault category and describes the fault mechanism and propagation path. A shared graph encoder is used to extract subgraph features from the divided fault causal subgraph and environmental spurious subgraph respectively. The classifier includes a causal classifier and a spurious classifier, used to independently output causal prediction results and spurious prediction results. Step S2 is detailed below:

[0072] S21: First, extract the node features of the variable association graph. Calculate the edge mask matrix based on node features , Represents a node and nodes The importance scores between them. The above steps are expressed as:

[0073] ,

[0074] ,

[0075] in, and They represent the process. Extracted nodes and nodes The feature representation vector, This represents the Sigmoid activation function. It is a convolution of local extrema maps ( A graph feature extraction network consisting of stacked layers.

[0076] Based on mask matrix and the original adjacency matrix Each edge is assigned an importance score to measure its contribution to fault diagnosis and prediction. This is done according to a pre-defined causal segmentation ratio. Select the top scorers The edges of the proportions form a fault causal subgraph edge set Select the remaining edges with lower scores as the false subgraph of the environment. edge set Find the nodes corresponding to the two sets of edges and construct fault cause-effect subgraphs respectively. and environmental fake subgraph The above steps are represented as follows:

[0077] ,

[0078] ,

[0079] in, This refers to sorting the matrix elements according to the importance scores of all edges, and retaining the elements with the highest scores. The side corresponding to the proportion, It refers to element-wise product.

[0080] S22: Shared graph encoder Fault Cause-and-Effect Subgraphs and environmental fake subgraph Encode, use Encode to generate their respective node feature representations and Then, global pooling is used to aggregate all node representations into graph-level feature representations. and This preserves the structure and key information of the subgraph. The above steps are represented as follows:

[0081] ,

[0082] ,

[0083] in, It is a convolution of local extrema maps ( A graph feature extraction network consisting of stacked layers; Pooling refers to the global pooling operation.

[0084] S23: Use two independent classifiers to classify the fault causal subgraph separately. and environmental fake subgraph Make predictions. Causal classifier. Graph-level representation based on fault-cause subgraph Perform causal prediction and output the predicted causal part of the fault. To learn the fundamental causal mechanism of failures. (Spurious classifier) Graph-level representation based on environmental pseudo subgraphs Perform false predictions and output the predictions that are entirely false. This measures the ability of false associations in the environment to predict fault labels.

[0085] S3. Based on the fault causal subgraph and the spurious environment subgraph divided by the causal generator, intervention enhancement is performed. This is achieved by recombining the fault causal subgraph of the current training sample with the spurious environment subgraph of other training samples, actively constructing multiple intervention distribution environments. Step S3 is detailed below:

[0086] S31: Execute an intervention-based enhancement training process by recombining the fault causal subgraph of the current training sample with the spurious environmental subgraph of other training samples to actively construct multiple intervention distribution environments; first, associate all fault variables with each other. Fake environmental subgraphs generated by causal generation Collect the data and construct a fake environmental sub-map library. Perform intervention operations. ,in Represents environmental spurious subgraph variables. This represents the selected specific environment dummy subgraph instance sample, obtained by using the current sample. Fault Cause-and-Effect Graph The original fake environment subplots were fixed, and some of them were replaced with fake environment subplots from other samples in the library. Create new intervention samples .

[0087] S4. Based on the joint prediction mechanism, the prediction results of the causal subgraph and the spurious subgraph are integrated. Then, the invariant intervention risk is introduced by utilizing causal invariance to identify the stable fault causal subgraph structure under different intervention distribution environments, thereby improving the generalization ability and stability of the fault diagnosis model. When performing fault diagnosis, the model only uses pure causal prediction, actively eliminating the influence of the spurious environmental subgraphs identified during training, ensuring that the fault diagnosis result is completely determined by the stable fault causal subgraph. Step S4 is as follows:

[0088] S41: The joint prediction mechanism integrates the prediction results of causal subgraphs and spurious subgraphs, forcing the model to overcome the interference of spurious environmental subgraphs and focus on stable fault causal subgraph structures under different intervention environments; the model incorporates causal prediction during training. and false predictions Joint forecasting is performed as follows:

[0089] ,

[0090] Among them, confidence level This indicates the reliability of the spurious environmental subgraph for fault diagnosis; if the spurious environmental association is reliable, then... Confidence level The trend is towards 1, and the joint prediction preserves the causal subgraph prediction results. Such samples offer limited help in learning causal structures and are therefore ignored by the model. If spurious environmental subgraphs are ineffective for fault diagnosis... Confidence level The trend towards 0 leads to the joint prediction results. As the error approaches zero, the joint prediction error increases, forcing the model to optimize the causal classifier. The model is forced to learn truly stable fault causal subgraph structures from environmental failure samples. The joint prediction mechanism compels the model to overcome the interference of spurious environmental subgraphs, focusing on maintaining stable fault causal subgraph structures under different intervention environments, thereby improving the generalization ability and stability of the fault diagnosis model.

[0091] S42: Introducing invariant intervention risk using causal invariance to identify stable fault causal subgraph structures under different intervention distributions; invariant intervention risk. The model's performance under intervention conditions is measured. The predictive loss for different intervention samples is calculated, and minimizing the mean and variance of this set of risk values ​​constitutes the invariant intervention risk loss. This simultaneously reduces the average risk and risk fluctuations under different interventions, ensuring the accuracy and stability of the fault diagnosis model's predictions. Therefore, the fault diagnosis decision-making is forced not to rely on spurious environmental structures that change with the environment, but rather to learn a fault causal subgraph structure that is stable under all intervention conditions. Invariant intervention risk The calculation is as follows:

[0092] ,

[0093] in, Indicates the actual fault label, This represents the fault diagnosis model. The balancing hyperparameters are used to adjust the weights of the variance term under the invariant intervention risk. Multiple different distribution environments are constructed under the intervention, resulting in a set of losses under different intervention environments. and These represent the mean and variance of the risk values, respectively. The risk loss function is expressed by the following formula:

[0094] ,

[0095] in Represents the cross-entropy loss function. Represent the distribution of data without intervention; calculate the expectation under a specific intervention distribution, in the uninterventional data distribution. In the middle, maintain the causal structure of the fault. Causal structure of the original sample The spurious environmental structure is a spurious environmental subgraph under specific interventions. .

[0096] Causal prediction loss The calculation formula is:

[0097] ,

[0098] in, This represents the expectation over the original data distribution.

[0099] Design an independent environmental spurious loss function to train a spurious classifier. Identify spurious structures in the environment during training. Update only the fake classifier Parameters, environmental spurious loss function The calculation formula is:

[0100] ,

[0101] Total loss function for:

[0102] ,

[0103] The trained model uses only pure causal prediction for fault diagnosis, first analyzing the input graph... Through causal generator Extracting the fault cause-effect subgraph Only using an encoder and a causal classifier The causal branches are based on the partitioned fault causal subgraph. To make the final causal prediction of the failure. By actively eliminating interference from spurious environmental sub-images identified during training, false classifiers are discarded. And a joint prediction mechanism to ensure that fault diagnosis decisions are entirely determined by a stable fault causal subgraph.

[0104] Based on the industrial process fault diagnosis method based on intervention-enhanced invariant causal graph learning network (IICGLN) shown in S1-S4 above, a new invariant causal graph learning network (IICGLN) is proposed, which improves the reliability and interpretability of industrial process monitoring. It should be noted that the industrial process fault diagnosis method based on intervention-enhanced invariant causal graph learning network is characterized by constructing a fault diagnosis model based on a fault causal subgraph structure through causal invariant learning. This model has significant advantages in both reliability and interpretability.

[0105] The above-mentioned industrial process fault diagnosis method based on intervention-enhanced invariant causal graph learning network will be applied to a specific example to demonstrate its implementation and technical effects.

[0106] In this embodiment, the Tennessee-Eastman process (TEP) is used as an example to illustrate the effectiveness of the present invention. The TEP involves 52 variables, including 22 continuous process measurements, 19 component measurements, and 11 manipulated variables. The TEP allows for the simulation of 21 types of faults, facilitating the evaluation of fault detection performance. The proposed method is evaluated using publicly available datasets from the TEP. The TEP process flow is as follows: Figure 4 As shown.

[0107] For TEP industrial processes, this embodiment provides a method for industrial process fault diagnosis based on an intervention-enhanced invariant causal graph learning network. The implementation steps are as follows:

[0108] Step 1: Data Processing

[0109] 1.1 Data Preprocessing

[0110] Eight variables representing different fault types in TEP were selected, using 6400 samples, with 800 samples for each fault. The training, validation, and test sets were divided at ratios of 0.7, 0.1, and 0.2. Indicates inclusion The training data for each sample, and the process data, are obtained after robust normalization preprocessing to obtain standardized data. The data is then reorganized using a sliding window of length w=20 to obtain the input matrix for the t-th window. The fault label corresponding to this window is .

[0111] 1.2 Graph Structure Construction

[0112] Build a variable correlation plot based on the standardized data within the sliding window, using the Pearson correlation coefficient. Quantize the correlation strength between nodes within the current time window. Based on a preset threshold. The association strength is filtered, and only strong associations above a threshold are retained as edges, thus constructing a sparse adjacency matrix. Finally, an undirected graph is constructed. , representing the relationship structure between variables within the window, serves as the input for the subsequent fault diagnosis model based on the Invariant Causal Graph Learning Network (IICGLN).

[0113] Step 2: Invariant Causal Graph Learning Network (IICGLN)

[0114] This embodiment establishes an Invariant Causal Graph Learning Network (IICGLN) including a causal generator, a graph encoder, and a classifier to find cross-environmentally stable fault causal subgraph structures from variable association graphs.

[0115] 2.1 Causal Generator

[0116] The causal generator extracts node features from the variable association graph and uses an MLP with ReLU activation to calculate edge importance scores. A preset causal edge partitioning ratio is established. Based on the importance score of the edges and the proportion of causal edges, Delineate the cause-effect subgraph of the fault and environmental fake subgraph .

[0117] 2.2 Master Predictor

[0118] The main predictor comprises a graph encoder and a classifier. The graph encoder extracts features through LEConv graph convolutional layer units to obtain graph-level feature representations. and In the classification and prediction phase, the model constructs independent binary branches: the causal branch learns a stable fault causal subgraph structure across environments, and the spurious branch independently learns spurious environmental structures. Training employs two independent Adam optimizers for model parameter updates, with a batch size of 64. The learning rate for the causal classifier optimizer is set to 0.001, and the learning rate for the spurious classifier optimizer is set to 0.003. An early stopping strategy with a tolerance of 10 is implemented. (Further details regarding causal intervention are needed.) The model incorporates causal prediction during training. and false predictions Joint forecasting This forces the model to optimize the causal classifier. The model is forced to learn a truly stable fault causal subgraph structure from samples of environmental failures. Environmental invariance loss weights. Set as .

[0119] Step 3: Model Evaluation and Validation

[0120] In this embodiment, process data from the publicly available TEP dataset is used to train and validate the process detection performance of the Invariant Causal Graph Learning Network (IICGLN) and its comparative models. Comparison methods include the K-Nearest Neighbors (KNN) algorithm, the Graph Attention Neural Network (GAT), and the industrial process fault diagnosis method based on intervention-enhanced Invariant Causal Graph Learning Network proposed in this invention. The accuracy of fault classification is selected as the performance evaluation criterion for fault diagnosis.

[0121] Table 1 presents the fault detection performance of all methods in TEP. IICGLN achieved the best overall performance with an average accuracy of 0.950, maintaining high detection precision. Benefiting from the discovery of the causal graph structure of the fault, IICGLN is the most reliable method for fault detection in TEP. GAT also performed excellently, with an average accuracy of 0.915, demonstrating the effectiveness of GNN-based industrial process fault diagnosis methods. In contrast, KNN, which relies on statistical correlation for fault detection, performed poorly, indicating the presence of spurious correlations in the data.

[0122] Table 1. Comparison of Fault Detection Performance of All Methods in TEP

[0123]

[0124] Figure 3This chart shows the accuracy results of the industrial process fault diagnosis method based on an intervention-enhanced invariant causal graph learning network during the training process. The accuracy on the training set is 0.999, and the accuracy on the test set is 0.950. Meanwhile, the prediction accuracy (Causal Acc) based on the fault causal subgraph structure is 0.950, and the prediction accuracy (ConfAcc) based on the spurious environmental structure is 0.846. It can be seen that the model can separate the fault causal subgraph structure and the spurious environmental subgraph structure, and the fault diagnosis decision made based on the fault causal subgraph structure has higher accuracy. Furthermore, it confirms that spurious environmental structures do exist in the variable association graph, which can predict fault labels but lack causal relationships.

[0125] As described above, the industrial process fault diagnosis method based on intervention-enhanced invariant causal graph learning network proposed in this invention has satisfactory reliability and interpretability, and can complete practical industrial process monitoring tasks.

[0126] This invention provides a method for fault diagnosis in industrial processes based on an intervention-enhanced invariant causal graph learning network. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for fault diagnosis of industrial processes based on intervention-enhanced invariant causal graph learning networks, characterized in that, Includes the following steps: Step 1: For the target industrial process, preprocess the process data of process variables collected during operation to obtain standardized time series data; generate dynamic time series data through a fixed-length sliding window, and construct a variable correlation graph reflecting the topological relationship between variables based on the Pearson correlation coefficient; Step 2: Establish an Invariant Causal Graph Learning Network (IICGLN), which includes a causal generator, a graph encoder, and a classifier; input the variable association graph into the causal generator to divide it into a fault causal subgraph and an environmental spurious subgraph; use the shared graph encoder to extract subgraph features from the divided fault causal subgraph and environmental spurious subgraph respectively; The classifier includes a causal classifier and a spurious classifier, which are used to independently output causal prediction results and spurious prediction results; Step 3: Based on the fault causal subgraph and the spurious environment subgraph divided by the causal generator, intervention enhancement is performed. By recombining the fault causal subgraph of the current training sample with the spurious environment subgraph of other training samples, an intervention distribution environment is actively constructed. Step 4: Based on the joint prediction mechanism, the prediction results of the causal subgraph and the spurious subgraph are combined. The invariant intervention risk is introduced by utilizing causal invariance to identify the stable fault causal subgraph structure under different intervention distribution environments. The industrial process fault diagnosis model based on the intervention-enhanced invariant causal graph learning network finally uses only pure causal prediction when performing fault diagnosis, actively eliminating the influence of the spurious environmental subgraphs identified in the training, and ensuring that the fault diagnosis result is completely determined by the stable fault causal subgraph.

2. The method as described in claim 1, characterized in that, In step 1, the data preprocessing includes: Indicates inclusion The training data for each sample, where Let T represent the t-th sample containing n process variables, where T denotes transpose. Represent the real number space; perform robust normalization data preprocessing on the process data of the process variables, and for the first... Time series of 1 variable , ,in Indicates the first The Nth sample value of the nth variable, using the nth... Median of the variables Deviation from median Scaling is performed to obtain standardized data. ,in is the numerical stability constant; median denotes the median function; definition ,in This represents the standardized data value of the nth variable in the tth sample. Let represent the standardized sample vector of the t-th sample; a sliding window is used to process the standardized time series data, and the input matrix of the t-th window is . ,in This indicates the size of the sliding window, and the corresponding fault label for the window is... .

3. The method as described in claim 2, characterized in that, In step 1, a variable relationship graph is built based on the standardized data within the sliding window. Each process variable is treated as a node in the graph, and a node set is defined. ,in This represents the nth process variable, and the i-th process variable is in the current window. The standardized sample sequence within is a node The attributes, using the Pearson correlation coefficient Quantization Node and Correlation strength within the current time window: ， in Represents a node and nodes Covariance within the current window. Represents a node The standard deviation within the current window. Represents a node The standard deviation within the current window. Representing variables The first in the current window One data point, Representing variables The mean within the current window; Based on a preset threshold The association strength is filtered, and only strong associations above a threshold are retained as edges to construct a sparse adjacency matrix. : ， in This represents the element in the i-th row and j-th column of the sparse adjacency matrix A; If and only if the correlation coefficient When, there are edges in the graph Construct an undirected graph , representing the relationship structure between variables within the window, serves as the input for the subsequent fault diagnosis model based on the Invariant Causal Graph Learning Network (IICGLN).

4. The method as described in claim 3, characterized in that, In step 2, the causal generator will generate an undirected graph. Divided into fault cause-effect subgraphs and environmental fake subgraph ; First, extract the node features of the variable association graph. Calculate the edge mask matrix based on node features , Represents a node and nodes The importance scores between them are expressed as: ， ， in, and They represent the process. Extracted nodes Feature representation vectors and nodes The feature representation vector, This represents the Sigmoid activation function. It is a graph feature extraction network composed of stacked local extremum map convolutional layers; Based on mask matrix and the original adjacency matrix Determine the importance score for each edge to measure its contribution to fault diagnosis prediction; According to the preset causal division ratio r, the edges with the highest scores in the top r ratios are selected to form the fault causal subgraph. edge set Select the remaining edges with lower scores as the false subgraph of the environment. edge set And find two sets of edge sets. and Construct fault cause-effect subgraphs for the corresponding nodes. and environmental fake subgraph , is represented as: ， ， in, This refers to sorting the matrix elements according to the importance scores of all edges, and retaining the edges corresponding to the top r proportions with the highest scores. It refers to element-wise product.

5. The method as described in claim 4, characterized in that, In step 2, the fault causal subgraph is analyzed in the intervention environment. and environmental fake subgraph Encoding and independent classification force the model to capture a cross-environmentally stable fault causal subgraph structure, firstly by utilizing a shared graph encoder. Fault Cause-and-Effect Subgraphs Encode the spurious environmental subgraph s and use GNN to generate a fault cause-effect subgraph. Node feature representation Node feature representation of the spurious subgraph of the environment Then, global pooling is used to represent the node features. Aggregation into graph-level feature representation Representing node features Aggregation into graph-level feature representation , is represented as: ， ， in, It is a convolution of local extrema maps ( A graph feature extraction network consisting of stacked layers; Pooling refers to the global pooling operation. Two independent classifiers are used to predict the fault causal subgraph c and the environmental spurious subgraph s, respectively. The causal classifier... Graph-level feature representation based on fault causal subgraph Perform causal prediction and output the predicted causal part of the fault. Learn the fundamental causal mechanisms of failures; False classifier Graph-level representation based on environmental pseudo subgraphs Perform false predictions and output the predictions that are entirely false. This measures the ability of spurious environmental associations to predict fault labels.

6. The method as described in claim 5, characterized in that, In step 3, an intervention-enhanced training process is executed. This involves recombining the fault causal subgraph of the current training sample with the spurious environmental subgraph of other training samples to actively construct two or more intervention distribution environments. First, the spurious environmental subgraph is generated by dividing the fault variable association graph g after causal generation. Collect the data and construct a fake environmental sub-library; then perform intervention operations. ,in Represents environmental spurious subgraph variables. This represents the selected specific environment dummy subgraph instance sample, obtained by using the current sample. Fault Cause-and-Effect Graph The original fake environment subplots were fixed, and some of them were replaced with fake environment subplots from other samples in the library. Create new intervention samples .

7. The method as described in claim 6, characterized in that, In step 4, the model incorporates causal prediction during training. and false predictions Conduct joint forecasting: ， Among them, confidence level This indicates the reliability of the spurious environmental subgraph for fault diagnosis; if the spurious environmental association is reliable, then... , indicating confidence level The trend is towards 1, and the joint prediction preserves the causal subgraph prediction results. ; If the spurious environmental subgraph is ineffective for fault diagnosis, then , indicating confidence level The trend towards 0 leads to the joint prediction results. Close to 0.

8. The method as described in claim 7, characterized in that, In step 4, the risk of invariable intervention remains unchanged. The calculation formula is: ， in, Indicates the actual fault label, This represents the fault diagnosis model. The balancing hyperparameters are used to adjust the weights of the variance term under the invariant intervention risk. Different distribution environments are constructed under the intervention, resulting in a set of losses under different intervention environments. and These represent the mean and variance of the risk value, respectively. The risk loss function is expressed by the following formula: ， in Represents the cross-entropy loss function. Represent the distribution of data without intervention; calculate the expectation under a specific intervention distribution, in the uninterventional data distribution. In the middle, maintain the causal structure of the fault. Causal structure of the original sample The spurious environmental structure is a spurious environmental subgraph under a specific intervention. ; Causal prediction loss The calculation formula is: ， in, This represents the expectation over the original data distribution; Design an independent environmental spurious loss function to train a spurious classifier. Identify spurious structures in the environment during training. Update only the fake classifier Parameters, environmental spurious loss function The calculation formula is: ， Total loss function for: ， After training, the model uses only pure causal prediction for fault diagnosis. First, the input graph g is processed by a causal generator. Extract the fault causal subgraph c, using only the encoder and causal classifier. The causal branches are based on the partitioned fault causal subgraph. To make the final causal prediction of the failure. By actively eliminating interference from spurious environmental sub-images identified during training, false classifiers are discarded. And a joint prediction mechanism to ensure that fault diagnosis decisions are entirely determined by a stable fault causal subgraph.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 8.

Images