Method for identifying and diagnosing fault propagation path of batch process based on graph neural network
Patent Information
- Application Number
- CN202611051427.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-15
- Publication Date
- 2026-10-16
AI Technical Summary
[0004]（1）大多假设变量间关系静态不变，忽略了间歇过程多阶段操作中变量耦合关系的动态演化特性；
[0020]本发明实施例提供的上述技术方案具体有益效果至少包括：
Smart Images
Figure CN122926112A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intermittent process fault diagnosis technology, and specifically to a method for identifying and diagnosing intermittent process fault propagation paths based on graph neural networks. Background Technology
[0002] Batch processes involve complex material circulation and energy transfer, with close coupling between production units. When an anomaly occurs in one link, the fault often propagates through the network via material, energy, and information flows, triggering a chain reaction that ultimately leads to product quality degradation or even safety accidents. Traditional fault diagnosis methods mainly focus on fault detection and classification, lacking effective analytical tools for understanding how faults propagate between process variables and how influence paths evolve.
[0003] Existing data-driven fault propagation analysis methods have the following shortcomings:
[0004] (1) Most of them assume that the relationship between variables is static and invariant, ignoring the dynamic evolution characteristics of the coupling relationship between variables in the multi-stage operation of the intermittent process; (2) Using correlation analysis or Granger causality test to identify the influence relationship between variables makes it difficult to capture nonlinear and high-order dependency structures; (3) The lack of visualization and interpretability analysis of the fault propagation path makes it difficult for the diagnostic results to guide operators to locate the root cause.
[0005] Therefore, how to provide a method that can improve the accuracy of fault propagation path identification and root cause localization is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0006] In view of the above problems, the present invention is proposed to provide a method for identifying and diagnosing intermittent process fault propagation paths based on graph neural networks to overcome or at least partially solve the above problems, thereby realizing intuitive visualization of fault propagation paths and accurate location of root causes.
[0007] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a method for identifying and diagnosing fault propagation paths in intermittent processes based on graph neural networks, comprising the following steps: S1: Collect time-series data from multiple sources of sensors during intermittent processes, and construct a dynamic variable relationship graph based on the statistical correlation between process variables in the time-series data from multiple sources of sensors. The nodes of the dynamic variable relationship graph represent process variables, and the edge weights represent the coupling strength between variables. S2: Extract spatial topological features based on the dynamic variable relationship diagram, extract time series features based on the multi-source sensor time series data, fuse the spatial topological features and the time series features, construct a spatiotemporal coupled encoder, and obtain a deep hidden state representation of the process variables; S3: Based on the deep hidden state representation, a dual attention filtering module is constructed using an attention mechanism to filter the set of key variables and key time windows from the variable dimension and the time dimension, and output the fault type prediction results; S4: Calculate the structural similarity between the current dynamic variable relationship diagram and each historical fault pattern diagram in the historical fault case library. If there is a historical fault pattern diagram with a structural similarity exceeding a preset threshold, then the fault type corresponding to the historical fault pattern diagram is taken as the diagnosis result, and the corresponding root cause prior information is transferred as the diagnostic root cause; otherwise, proceed to step S5. S5: Based on the deep hidden state representation, the direction of influence propagation between process variables is traced backward from the moment the fault occurs through the reverse interpretation mechanism of the graph neural network, generating a directed graph of fault propagation, locating the root cause variable of the fault as the diagnostic root cause, and using the fault type prediction result as the diagnostic result.
[0008] Preferably, in step S1, the statistical correlation includes the maximum information coefficient and the transit entropy, and the adjacency matrix A of the dynamic variable relationship graph is jointly constructed based on the maximum information coefficient and the transit entropy:
[0009] in, process variables to process variables Transfer entropy between them process variables to process variables Transfer entropy between them process variables and process variables The maximum information coefficient between them.
[0010] Preferably, in step S2, the spatiotemporal coupling encoder is a graph convolution gated cyclic encoder, which embeds the graph convolution operation into the gated mechanism of the cyclic encoder, including the following steps: At each time step of the graph convolutional gated recurrent unit encoder, the current input process variable and the hidden state of the previous time step are respectively input to the graph convolutional layer to generate spatial topological features. The spatial topological features are then input to the gating mechanism of the gated recurrent unit to output the hidden state of the current time step.
[0011] Preferably, in step S3, the dual attention filtering module includes: Variable-level attention is used to calculate the attention weight of each process variable at each time step and to filter the set of key variables. Temporal attention is used to calculate the attention weights at each time step and to filter the key time windows.
[0012] Preferably, the historical fault pattern diagram is pre-stored in the historical fault case library and associated with fault type labels and root cause variable sets.
[0013] Preferably, in step S4, the structural similarity is calculated using a graph kernel method, including jointly calculating node feature similarity and edge weight similarity:
[0014] in, This is a historical fault mode diagram. This is a dynamic variable relationship diagram. and They are respectively and Nodes in i and nodes j Corresponding process variable characteristics, The Gaussian kernel function; and For nodes i and nodes j The edge weight vector, It is a linear kernel function.
[0015] Preferably, the historical fault mode diagram is a dynamic variable relationship diagram obtained by performing S1 on the historical intermittent process.
[0016] Preferably, in step S5, the reverse interpretation mechanism is a reverse interpretation mechanism based on hidden state gradient attribution, including: Constructing process variables Sensitivity measurement S i :
[0017] in, The prediction result for the fault type, process variables The hidden state vector; At the moment of failure, the root cause variable of the failure is located based on the comparison results of the sensitivity measures of each process variable.
[0018] Preferably, in step S5, the reverse interpretation mechanism uses the set of key variables and the key time window as spatial and temporal constraints for gradient calculation.
[0019] Preferably, step S5 further includes: Sensitivity measurement S i The node with the largest process variable is identified as the root cause variable of the fault; For each variable in the set of key variables, calculate its sensitivity measure at each time step within the key time window. S i ; Compare the sensitivity measures of each variable at the time of failure. S i ,Will S i The variable with the largest value is identified as the root cause variable. Starting with the aforementioned root cause variables of the fault, tracing back along the time axis, based on each variable... S i The order in which values appear determines the direction of influence propagation between variables, and the directed graph of fault propagation is constructed.
[0020] The specific beneficial effects of the above-mentioned technical solutions provided in the embodiments of the present invention include at least the following: This invention constructs a dynamic variable relationship graph by combining transfer entropy and maximum information coefficient, and comprehensively measures the directed causal strength and nonlinear correlation coupling degree between variables through the edge weight formula. This invention can effectively capture the nonlinear, directed, and time-varying coupling relationships between variables in multi-stage operations of intermittent processes, overcoming the limitations of traditional methods that assume linear and static variable relationships.
[0021] This invention proposes a graph convolution gated recurrent unit hybrid encoder, which embeds graph convolution operations into the reset gate, update gate, and candidate state calculation of the gated recurrent unit, achieving deep fusion of spatial topological features and temporal series features rather than simple serial superposition. The hybrid encoder of this invention enables the hidden state update at each time step to simultaneously depend on the adjacency matrix and the hidden states of all variables at the previous time step. The output hidden state matrix can more fully learn the spatiotemporal correlation information between variables. Experiments show that this design significantly contributes to the accuracy of fault identification.
[0022] This invention utilizes a gradient attribution mechanism to calculate the influence gradient matrix between variables. Combined with a key variable subset and key time window identified by a dual-attention screening module, it traces backward layer by layer from the moment of fault occurrence along the direction of maximum influence to the root cause variable, generating a directed fault propagation graph with the root cause as the source point. This invention achieves intuitive visualization of the fault propagation path and accurate location of the root cause. Experimental results on a penicillin fermentation process dataset show that the fault identification accuracy of this invention is significantly improved compared to the average accuracy of existing methods based on static graph neural networks, fully demonstrating the superiority of this invention. Attached Figure Description
[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0024] Figure 1 A flowchart illustrating the intermittent process fault propagation path identification and diagnosis method based on graph neural networks provided in this embodiment of the invention; Figure 2 This is a diagram of the network structure of a convolutional gated recurrent unit provided in an embodiment of the present invention; Figure 3 The image shows the visualization results of fault diagnosis provided in the embodiments of the present invention. Detailed Implementation
[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] This invention specifically discloses a method for identifying and diagnosing intermittent process fault propagation paths based on graph neural networks, combined with, for example... Figure 1 The flowchart shown illustrates the steps of the method in this embodiment of the invention as follows: S1: Collect time-series data from multiple sources of sensors during intermittent processes. Construct a dynamic variable relationship graph based on the statistical correlation between process variables in the time-series data from multiple sources of sensors. The nodes of the dynamic variable relationship graph represent process variables, and the edge weights represent the coupling strength between variables. S2: Extract spatial topological features from the dynamic variable relationship diagram, extract time series features from multi-source sensor time series data, fuse spatial topological features and time series features, construct a spatiotemporal coupled encoder, and obtain the deep hidden state representation of process variables; S3: Based on the deep hidden state representation, a dual attention filtering module is constructed using the attention mechanism to filter the set of key variables and key time windows from the variable dimension and the time dimension, and output the failure type prediction results; S4: Calculate the structural similarity between the current dynamic variable relationship diagram and each historical fault pattern diagram in the historical fault case library. If there is a historical fault pattern diagram with a structural similarity exceeding the preset threshold, then the fault type corresponding to the historical fault pattern diagram is taken as the diagnosis result, and the corresponding root cause prior information is transferred as the diagnostic root cause; otherwise, proceed to step S5. S5: Based on the deep hidden state representation, the reverse interpretation mechanism of the graph neural network is used to trace the direction of influence propagation between process variables from the moment the fault occurs, generate a directed graph of fault propagation, locate the root cause variable of the fault as the root cause of diagnosis, and use the fault type prediction result as the diagnosis result.
[0027] In one embodiment, in step S1, statistical correlation includes the maximum information coefficient (MIC) and the transit entropy (TE). A method combining MIC and TE is used to measure the nonlinear dependency and causal transmission strength between variables. The adjacency matrix A of the dynamic variable relationship graph is constructed based on the MIC and TE.
[0028] in, process variables to process variables Transfer entropy between them process variables to process variables Transfer entropy between them process variables and process variables The maximum information coefficient between them.
[0029] In this embodiment, transfer entropy is used to quantify the directed information transfer between variables, enabling the capture of nonlinear causal relationships. For variables... and The propagation entropy is defined as:
[0030] in, and Variables and The historical state, k and l For the embedded dimension.
[0031] In this embodiment, the maximum information coefficient The calculation process is as follows: Pair the variables... The sample point scatter plot is divided into different grid scales. The mutual information value under each division is calculated and normalized. The maximum value is taken as the MIC to measure the symmetric nonlinear correlation between variables.
[0032] In one embodiment, in step S2, the spatiotemporal coupled encoder is a graph convolution gated recurrent unit encoder (GCGRU). The GCGRU embeds graph convolution operations within the gating mechanism of the gated recurrent unit (GRU), including the following steps: At each time step of the graph convolutional gated recurrent unit encoder, the current input process variable and the hidden state of the previous time step are respectively input into the graph convolutional layer to generate spatial topological features. The spatial topological features are then input into the gating mechanism of the gated recurrent unit to output the hidden state of the current time step.
[0033] The specific fusion method is as follows: at each time step of GCGRU, the current input is... The hidden state of the previous time step The features are fed into the graph convolutional layer to generate features containing spatial topological information. These features are then fed into the gating mechanism of GRU to achieve joint learning of spatial and temporal features.
[0034] Combination Figure 2 To further explain the update process of the graph convolution gated recurrent unit, it is defined as follows:
[0035] Among them, the reset door Update door and candidate states All calculations are incorporated into graph convolution operations:
[0036] in, for t time N 1 variable F 3D feature input, This is the hidden state from the previous moment. D To hide the state dimension, ⊙ represents element-wise multiplication. Graph convolution operation. ,in, , For degree matrix, These are the learnable convolutional kernel parameters. The output data of this hybrid encoder is the hidden state matrix at each time step. The first [item] in this matrix i Row vectors are variables X i At any moment t The deep hidden state representation, whose update process also depends on the adjacency matrix. A It contains the hidden states of all variables from the previous moment, and therefore includes spatial relationship information between variables.
[0037] In one embodiment, in step S3, the dual attention filtering module includes: Variable-level attention is used to calculate the attention weight of each process variable at each time step and to filter the set of key variables. Temporal attention is used to calculate the attention weights at each time step and to filter key time windows.
[0038] The specific process of this embodiment is as follows: The dual attention filtering module operates on the hidden state sequence output in step S2. ,in, T This represents the length of the time window. The specific construction steps are as follows: Step 31: Variable-level attention. For each time step... t Calculate variables i attention weights :
[0039]
[0040] in, for t Time variable i The hidden state vector. , , These are learnable parameters. The final variable-level attention output is... .
[0041] Step 32: Temporal Attention. Attention at the variable level for all time steps. Stacked into a matrix Calculate time steps t weight β t :
[0042]
[0043] in, For the first t The output vector at each time step. The temporal attention output is... .
[0044] The results identified through the dual attention mechanism will be used for fault propagation path tracing in the subsequent step S5.
[0045] In one embodiment, the historical failure pattern diagram is pre-stored in a historical failure case library and associated with failure type labels and a set of root cause variables.
[0046] In one embodiment, the historical failure mode diagram is a dynamic variable relationship diagram obtained by performing S1 on the historical intermittent process.
[0047] It should be noted that the fault mode diagram is a specific instantiation of the dynamic variable relationship diagram under different data conditions. The historical fault mode diagram is constructed based on historical batch data according to the method in step S1 and is pre-stored in the fault case library. The current dynamic variable relationship diagram is used as the current fault mode diagram, and structural similarity is calculated between it and the historical fault mode diagram.
[0048] In one embodiment, step S4 uses a graph kernel method to calculate structural similarity, including jointly calculating node feature similarity and edge weight similarity:
[0049] in, This is a historical fault mode diagram. This is a dynamic variable relationship diagram. and They are respectively and Nodes in i and nodes j Corresponding process variable characteristics, The Gaussian kernel function; and For nodes i and nodes j The edge weight vector, It is a linear kernel function.
[0050] In one embodiment, in step S4, based on the historical fault case library ,in Label the fault type. Calculate the current fault mode diagram. With each historical map Kernel similarity .like ( δ If the threshold is set, the corresponding fault type is selected as the preliminary diagnosis result, and the root cause variable set marked in the historical case is migrated as prior information; otherwise, it is determined to be a new fault type, and the process proceeds to step S5 for detailed tracing.
[0051] In one embodiment, in step S5, the reverse interpretation mechanism is a reverse interpretation mechanism based on hidden state gradient attribution, including: For a trained graph convolutional memory network model, construct process variables. Sensitivity measurement S i Fault type prediction output The hidden state vector of this variable Gradient magnitude:
[0052] in, For the fault type prediction results, process variables The hidden state vector; S i The larger the value, the more sensitive the variable is to the prediction of fault type, and the more likely it is to be a key node in the fault propagation path.
[0053] At the moment of failure, the root cause variable of the failure is located based on the comparison results of the sensitivity measures of each process variable.
[0054] In one embodiment, in step S5, the back-interpretation mechanism uses a set of key variables and a key time window as spatial and temporal constraints for gradient computation. The back-interpretation mechanism implicitly receives the hidden state sequence stored by the GCGRU encoder during forward propagation in step S2, and a subset of key variables from the dual attention module in step S3. The key time window is input to the back-interpretation mechanism in the form of interval endpoint pairs (or an array of indices containing start and end times) consisting of the start and fault occurrence times, constraining gradient computation to consider only the hidden states at each time step within this window.
[0055] In one embodiment, step S5 further includes: Sensitivity measurement S i The node with the largest process variable is identified as the root cause variable of the fault. For each variable in the set of key variables, calculate its sensitivity measure at each time step within the key time window. S i ; Compare the sensitivity measures of each variable at the time of failure. S i ,Will S i The variable with the largest value is identified as the root cause variable. Starting with the root cause variable of the failure, backtracking along the time axis, based on each variable... S i The order in which values appear determines the direction of influence propagation between variables, and a directed graph of fault propagation is constructed.
[0056] The specific implementation method is explained below using the penicillin fermentation process as an example: Penicillin fermentation is a typical time-varying batch process involving complex biological and chemical changes from raw materials to the final product. Numerous control loops and variables exist during fermentation, and failures frequently occur, severely impacting production performance and product quality. An industrial-scale penicillin fermentation platform, in addition to modeling necessary online and offline variables, considers the growth, metabolism, and degradation of penicillin during large-scale fermentation. Multiple control strategies are specified to reduce system fluctuations and maximize penicillin yield. This study selected 50 batches of data under different control strategies, including normal batches and batches with various failure types, with a fermentation reaction time of 230 hours and data collected every 12 minutes. Twenty-two key process variables were selected, as shown in Table 1.
[0057] Table 1. Description of process variables in industrial-scale penicillin fermentation.
[0058] The specific steps are as follows: S10: Constructing a dynamic variable relationship graph based on transitive entropy: Multi-source sensor data are collected under normal and fault conditions during intermittent processes. A dynamic variable relationship graph is constructed based on the maximum information coefficient and transfer entropy between process variables. For variables... and The transitive entropy is defined as:
[0059] The adjacency matrix A of the dynamic variable relationship graph is constructed by combining transitive entropy and maximum information coefficient. Nodes in the graph represent process variables, and edge weights represent the coupling strength between variables. A graph convolutional neural network is used to extract the spatial topological features between variables.
[0060] S20: Constructing a graph convolutional gated recurrent unit hybrid encoder: The extracted spatial topological features are fused with temporal series features to construct a hybrid encoder with graph convolutional gated recurrent units. The update process of the graph convolutional gated recurrent unit is defined as follows:
[0061] Among them, the reset door Update door and candidate states All calculations are incorporated into graph convolution operations. A graph convolution operation is defined as... ,in, , For degree matrix, These are learnable convolutional kernel parameters. This hybrid encoder jointly models the spatiotemporal coupling dynamics of the process variables, learning a deep hidden state representation containing variable correlation information.
[0062] S30: Constructing a dual-attention filtering module: An attention mechanism is used to calculate the contribution weights of each variable to the fault state at different time steps, constructing a dual-attention screening module at both the variable and time-series levels. Variable-level attention identifies the key variables most relevant to the fault, while time-series attention locates the critical time window where the fault occurs. Through this dual-attention mechanism, a subset of key variables and a critical time window are identified. This identification result is used in step S50: the gradient attribution matrix is calculated only within the subset of key variables and the critical time window to reduce computational complexity and focus on the fault-related region.
[0063] S40: Fault type matching based on graph kernel: Based on a historical failure case database, a graph kernel method is used to calculate the structural similarity between the current failure mode and historical failure modes. The graph kernel function is defined as follows: This approach combines transfer learning strategies to achieve rapid matching and preliminary diagnosis of fault types. When the similarity between the current fault mode and a historical fault mode in the graph kernel exceeds a set threshold, that historical fault type is used as the preliminary diagnostic result.
[0064] S50: Gradient Attribution-Based Fault Propagation Path Tracing The reverse interpretation mechanism of a graph neural network is used to trace the direction of influence propagation between variables backward from the moment the fault occurs. The direction of fault propagation is determined by calculating the gradient matrix of influence between variables.
[0065] in, h i and h j Don't be a variable X i and X j The corresponding hidden state vector. A directed graph of fault propagation is constructed, with the root cause variable as the source and the reverse direction of the tracing path as the propagation direction.
[0066] The above steps describe the application of this invention in identifying fault propagation paths by collecting key variables during typical intermittent processes. The parameters of the graph convolutional gated recurrent unit are as follows: the number of graph convolutional layers is set to 2, with output feature dimensions of 32 and 16 for each layer; the hidden layer dimension of the gated recurrent unit is set to 64; in the graph kernel method, the node kernel function uses a Gaussian kernel, and the edge kernel function uses a linear kernel; the gradient attribution tracing depth is set to 5 layers. To visualize the effectiveness of the model, the fault identification accuracy is used to evaluate the diagnostic effect, and the results are as follows: Figure 3As shown, the horizontal and vertical axes represent the first and second principal component dimensions after t-SNE dimensionality reduction, respectively, used to project high-dimensional fault feature data onto the first and second spatial dimensions of a two-dimensional plane. If each category can be clearly separated and not mixed in the two-dimensional space after t-SNE dimensionality reduction, it directly proves that the high-dimensional features extracted by the model have good discriminative power, with a discrimination accuracy of over 98%.
[0067] The network model proposed in this invention accurately identifies fault propagation paths, fully considers the spatiotemporal coupling characteristics of intermittent processes, and utilizes transfer entropy to construct a dynamic variable relationship graph to capture the nonlinear causal relationships between variables. It extracts spatiotemporal features using graph convolutional gated recurrent units, constructs a dual attention mechanism to screen key variables and time windows, uses a graph kernel method to achieve rapid fault type matching, and combines gradient attribution to achieve visual tracking of fault propagation paths. When it is necessary to identify fault propagation paths, it traces the direction of influence propagation backward from the fault moment to generate a directed fault propagation graph, accurately locating root cause variables and improving the interpretability of fault diagnosis. It monitors the fault state of complex intermittent processes in real time, accurately identifying fault propagation paths and root causes.
[0068] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0069] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for identifying and diagnosing fault propagation paths in intermittent processes based on graph neural networks, characterized in that, Includes the following steps: S1: Collect time-series data from multiple sources of sensors during intermittent processes, and construct a dynamic variable relationship graph based on the statistical correlation between process variables in the time-series data from multiple sources of sensors. The nodes of the dynamic variable relationship graph represent process variables, and the edge weights represent the coupling strength between variables. S2: Extract spatial topological features based on the dynamic variable relationship diagram, extract time series features based on the multi-source sensor time series data, fuse the spatial topological features and the time series features, construct a spatiotemporal coupled encoder, and obtain a deep hidden state representation of the process variables; S3: Based on the deep hidden state representation, a dual attention filtering module is constructed using an attention mechanism to filter the set of key variables and key time windows from the variable dimension and the time dimension, and output the fault type prediction results; S4: Calculate the structural similarity between the current dynamic variable relationship diagram and each historical fault pattern diagram in the historical fault case library. If there is a historical fault pattern diagram with a structural similarity exceeding a preset threshold, then the fault type corresponding to the historical fault pattern diagram is taken as the diagnosis result, and the corresponding root cause prior information is transferred as the diagnostic root cause; otherwise, proceed to step S5. S5: Based on the deep hidden state representation, the direction of influence propagation between process variables is traced backward from the moment the fault occurs through the reverse interpretation mechanism of the graph neural network, generating a directed graph of fault propagation, locating the root cause variable of the fault as the diagnostic root cause, and using the fault type prediction result as the diagnostic result.
2. The method for identifying and diagnosing intermittent process fault propagation paths based on graph neural networks as described in claim 1, characterized in that, In step S1, the statistical correlation includes the maximum information coefficient and the transit entropy, and the adjacency matrix A of the dynamic variable relationship graph is jointly constructed based on the maximum information coefficient and the transit entropy: in, process variables to process variables Transfer entropy between them process variables to process variables Transfer entropy between them process variables and process variables The maximum information coefficient between them.
3. The method for identifying and diagnosing intermittent process fault propagation paths based on graph neural networks as described in claim 1, characterized in that, In step S2, the spatiotemporal coupling encoder is a graph convolution gated cyclic encoder. The graph convolution gated cyclic encoder embeds the graph convolution operation into the gated mechanism of the gated cyclic unit, including the following steps: At each time step of the graph convolutional gated recurrent unit encoder, the current input process variable and the hidden state of the previous time step are respectively input to the graph convolutional layer to generate spatial topological features. The spatial topological features are then input to the gating mechanism of the gated recurrent unit to output the hidden state of the current time step.
4. The method for identifying and diagnosing intermittent process fault propagation paths based on graph neural networks as described in claim 1, characterized in that, In step S3, the dual attention filtering module includes: Variable-level attention is used to calculate the attention weight of each process variable at each time step and to filter the set of key variables. Temporal attention is used to calculate the attention weights at each time step and to filter the key time windows.
5. The method for identifying and diagnosing intermittent process fault propagation paths based on graph neural networks as described in claim 1, characterized in that, The historical fault pattern diagram is pre-stored in the historical fault case library and is associated with fault type labels and root cause variable sets.
6. The method for identifying and diagnosing intermittent process fault propagation paths based on graph neural networks as described in claim 1, characterized in that, In step S4, the structural similarity is calculated using the graph kernel method, including jointly calculating node feature similarity and edge weight similarity: in, This is a historical fault mode diagram. This is a dynamic variable relationship diagram. and They are respectively and Nodes in i and nodes j Corresponding process variable characteristics, The Gaussian kernel function; and For nodes i and nodes j The edge weight vector, It is a linear kernel function.
7. The method for identifying and diagnosing intermittent process fault propagation paths based on graph neural networks as described in claim 1, characterized in that, The historical fault mode diagram is a dynamic variable relationship diagram obtained by executing S1 on the historical intermittent process.
8. The method for identifying and diagnosing intermittent process fault propagation paths based on graph neural networks as described in claim 1, characterized in that, In step S5, the reverse interpretation mechanism is a reverse interpretation mechanism based on hidden state gradient attribution, including: Constructing process variables Sensitivity measurement S i : in, The prediction result for the fault type, process variables The hidden state vector; At the moment of failure, the root cause variable of the failure is located based on the comparison results of the sensitivity measures of each process variable.
9. The method for identifying and diagnosing intermittent process fault propagation paths based on graph neural networks as described in claim 8, characterized in that, In step S5, the reverse interpretation mechanism uses the set of key variables and the key time window as spatial and temporal constraints for gradient calculation.
10. The method for identifying and diagnosing intermittent process fault propagation paths based on graph neural networks as described in claim 9, characterized in that, Step S5 also includes: Sensitivity measurement S i The node with the largest process variable is identified as the root cause variable of the fault; For each variable in the set of key variables, calculate its sensitivity measure at each time step within the key time window. S i ; Compare the sensitivity measures of each variable at the time of failure. S i ,Will S i The variable with the largest value is identified as the root cause variable. Starting with the aforementioned root cause variables of the fault, tracing back along the time axis, based on each variable... S i The order in which values appear determines the direction of influence propagation between variables, and the directed graph of fault propagation is constructed.