Industrial equipment fault early warning analysis method and system based on graph neural network

Abstract

The present invention discloses an industrial equipment fault warning analysis method and system based on a graph neural network, which relates to the technical field of intelligent fault diagnosis and equipment status monitoring. The method includes constructing a time segment data tensor to form a graph neural network input structure; based on the graph neural network input structure, performing feature learning and association analysis of the graph neural network to obtain a node-level embedded representation; based on the node-level embedded representation, calculating the fault correlation response distribution of the node, evaluating the propagation risk and locating the source of the equipment fault. The present invention solves the problem that traditional methods cannot effectively express the spatial structure or the collaborative relationship between measurement points in the early modeling stage, improves the spatial perception ability of the overall modeling, and thus lays a data foundation for subsequent feature extraction and anomaly analysis. By comparing the embedded reconstruction value with the original observation value, the node-level reconstruction error is calculated, and the fault impact range assessment and traceability positioning are realized, providing quantitative support for early warning response and maintenance decision-making.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent fault diagnosis and equipment status monitoring, and specifically to an industrial equipment fault early warning analysis method and system based on graph neural network. Background Art

[0002] Existing equipment anomaly detection technologies face the following challenges: Statistical model-based methods rely on statistical assumptions about normal data distributions and identify anomalies by measuring deviations from normal patterns. Common methods include the interquartile range (IQR) and local outlier factor (LOF). These methods offer low computational overhead and are easy to implement, making them suitable for detecting single-point mutations. However, they typically only handle single measurement points or simple global aggregations, struggle to capture dependencies between multiple sensor points, and are therefore limited in their effectiveness when dealing with coordinated anomalies at multiple points. Traditional machine learning-based methods utilize supervised or unsupervised algorithms to train and classify feature data. Common methods include one-class support vector machines (OC-SVMs), isolation forests, k-NNs, and decision trees. These methods can learn anomaly patterns from labeled data, exhibit a certain degree of robustness to data outside of Gaussian distributions, and offer good model interpretability. However, they require manual feature design, do not inherently account for time series dependencies, and struggle to model spatial topological relationships between sensor points. Deep learning-based methods employ deep models such as autoencoders, long short-term memory networks (LSTMs), and convolutional neural networks (CNNs) to automatically extract complex patterns from high-dimensional, multivariate time series. Deep models can automatically learn features from big data, have strong capabilities for modeling time series data, and improve detection accuracy. However, by focusing solely on the time series dimension, they lack explicit modeling of the topological structure between sensor nodes, making it difficult to capture the spatial correlations between equipment components. Graph neural network-based approaches explicitly represent the connectivity between multiple sensor points and components in the form of a graph structure, treating measurement points as nodes and the topological connections between sensors as edges, and modeling this using graph neural networks (GNNs). Research has shown that graph-based anomaly detection has great potential in complex industrial scenarios. By aggregating information and attributes of neighboring nodes, GNNs can learn high-quality embedded representations of nodes, which facilitates the discovery of implicit anomalies in node associations. However, most current GNN methods primarily target static or simple dynamic graphs, have limited capabilities for understanding the dynamic evolution of time series, and typically assume known and fixed topological relationships. In semi-supervised or unsupervised settings, they typically train only on normal samples, making it difficult to fully utilize anomaly information when labels are scarce.

[0003] While the aforementioned methods have achieved some success in their respective fields, they still face significant limitations in the complex scenarios of industrial equipment monitoring. They are unable to model the dependencies between multiple device nodes. Traditional anomaly detection models (such as time series models and traditional machine learning methods) typically model each device or sensor independently, ignoring the spatial dependencies or topological structure between them and failing to leverage existing device topological information. Inadequate modeling of topological and temporal relationships: Most statistical and machine learning methods ignore the spatial topological structure between device components. While deep learning methods can learn temporal features, they do not explicitly utilize the connectivity information of sensor points. Existing anomaly detection in dynamic graphs often relies on simple heuristic rules that only consider a single type of feature or temporal factor and lack flexible joint spatiotemporal modeling mechanisms. Therefore, when the spatial structure of equipment is complex or dynamically changing, these methods cannot accurately characterize spatiotemporal dependencies. They also fail to capture context-dependent anomalies. Some nodes may have normal values ​​on their own, but are classified as anomalies when combined or associated with adjacent component nodes. These "cooperative anomalies" or "combined anomalies" are often overlooked in traditional models because they can only be detected by considering the relationships between multiple nodes, which traditional methods cannot do. Existing technologies struggle to support downstream tasks such as fault propagation path analysis and fault tracing: In complex systems (such as industrial manufacturing, power systems, and the smart Internet of Things), an anomaly in a device or node often triggers a chain reaction of failures. Existing anomaly detection technologies are mostly based on single-point metrics (such as time series models, statistical testing, and machine learning classification), lacking the ability to model causal, dependency, and topological relationships between nodes. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: the existing methods based on rule thresholds or traditional time series modeling cannot fully utilize the topological correlation between measurement points, have insufficient anomaly detection accuracy, and have limited propagation path identification capabilities. It also solves the problem of how to construct a spatiotemporal fusion model based on a graph structure self-learning mechanism and realize anomaly identification, risk propagation analysis and source node positioning of industrial equipment failures.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: an industrial equipment fault early warning analysis method based on a graph neural network, comprising constructing a time segment data tensor to form a graph neural network input structure; based on the graph neural network input structure, performing feature learning and association analysis of the graph neural network to obtain a node-level embedding representation; based on the node-level embedding representation, calculating the fault correlation response distribution of the node, evaluating the propagation risk and locating the source of the equipment fault; the performing feature learning and association analysis of the graph neural network comprises extracting the node time features in the time segment data tensor, constructing a correlation score matrix between the nodes, generating a self-learning adjacency matrix based on the score matrix, fusing the self-learning adjacency matrix with the prior adjacency matrix and then normalizing it, performing graph convolution calculation using the normalized adjacency matrix and the node time features, and inputting the graph convolution result into a feedforward neural network to output a node-level embedding representation; the evaluating the propagation risk and locating the source of the equipment fault comprises calculating the fault correlation response distribution of the node, screening the fault source node according to a preset threshold, constructing an initial propagation state vector, iteratively propagating using the normalized adjacency matrix, generating the fault correlation response distribution of the node, and evaluating the propagation risk and locating the source of the equipment fault based on the fault correlation response distribution of the node.

[0007] As a preferred solution of the industrial equipment fault warning analysis method based on graph neural network described in the present invention, the construction of the time segment data tensor includes obtaining the original time series data of the measurement points of the components in the target equipment, preprocessing the measurement point time series data, constructing the time segment data tensor using a sliding window slicing method, and using the time segment data tensor and the prior adjacency matrix as the graph neural network input structure.

[0008] As a preferred solution of the industrial equipment fault early warning analysis method based on graph neural network described in the present invention, the prior adjacency matrix includes, if there is a device connection graph, parsing the device connection graph into a priori adjacency matrix.

[0009] If there is no device connection graph, the prior adjacency matrix calculated for the first time by the graph neural network is used.

[0010] As a preferred solution of the industrial equipment fault warning analysis method based on graph neural network described in the present invention, the feature learning and association analysis of the graph neural network is performed, including extracting the node time features in the time segment data tensor, constructing a correlation score matrix between nodes, retaining the Top-K scores of the correlation score matrix row by row to generate a sparse connection mask, and multiplying the sparse connection mask and the correlation score matrix element by element to obtain a self-learning adjacency matrix.

[0011] As a preferred solution of the industrial equipment fault warning analysis method based on graph neural network described in the present invention, the node-level embedding representation is obtained, including fusing the self-learning adjacency matrix and the prior adjacency matrix according to a preset ratio, iterating and normalizing the prior adjacency matrix to obtain a normalized adjacency matrix, performing graph convolution calculation on the normalized adjacency matrix and the node time feature, aggregating the representations of adjacent nodes and updating the time feature of the central node, inputting the graph convolution output result into the feedforward neural network layer, and outputting the node-level embedding representation.

[0012] As a preferred solution of the industrial equipment fault early warning analysis method based on graph neural network described in the present invention, wherein: the fault correlation response distribution of the computing node includes embedding the node level representation into the feedforward neural network decoder, performing reconstruction calculation on each node, obtaining the reconstructed value of the node in the time window, performing difference calculation between the reconstructed value of the node in the time window and the original input time series data of the sliding window, and outputting the node-level reconstruction error.

[0013] As a preferred solution of the industrial equipment fault early warning analysis method based on graph neural network described in the present invention, the method of evaluating the propagation risk and locating the source of equipment failure includes constructing the node whose node-level reconstruction error exceeds the preset threshold as the initial fault source, generating the initial propagation vector, performing iterative propagation calculation on the transpose of the normalized adjacency matrix through the initial propagation vector, stopping the propagation when the iterative propagation converges to the preset iteration threshold, obtaining the affected node set and evaluating the propagation risk, recording the node status activated in each round of propagation during the propagation iteration process, and locating the source of equipment failure.

[0014] Another object of the present invention is to provide an industrial equipment fault warning and analysis system based on graph neural networks, which can work together through three modules: graph model input construction, graph feature learning and association analysis, and fault propagation diagnosis, thereby solving the problem that the current equipment fault analysis technology based on a single time series model is unable to characterize spatial dependencies and propagation paths.

[0015] As a preferred solution of the artificial intelligence-based valve-side casing partial discharge identification system described in the present invention, it includes: a graph model input construction module, a graph feature learning and association analysis module, and a fault identification and propagation diagnosis module; the graph model input construction module includes a sliding window tensor construction unit and an adjacency matrix generation unit, the sliding window tensor construction unit is used to receive time series data of multiple measuring points of industrial equipment, and slice the original data using a sliding window method to generate a time segment data tensor, the adjacency matrix generation unit is used to parse the equipment structure diagram and construct an initial prior adjacency matrix; the graph feature learning and association analysis module includes a graph convolution encoding unit and a feedforward neural decoding unit, the graph convolution encoding unit is used to extract the spatial and temporal joint features of the nodes, and the feedforward neural decoding unit is used to restore the time series of each node and generate a reconstruction value; the fault identification and propagation diagnosis module includes an anomaly scoring unit and a propagation diagnosis unit, the anomaly scoring unit is used to calculate the reconstruction error of the node and generate an anomaly score, and the propagation diagnosis unit is used to perform propagation simulation, identify affected nodes and locate the fault source.

[0016] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a step of an industrial equipment fault early warning analysis method based on a graph neural network.

[0017] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of an industrial equipment fault early warning analysis method based on a graph neural network.

[0018] The beneficial effects of the present invention are as follows: by constructing a time segment data tensor and jointly constructing an input structure with the adjacency matrix corresponding to the equipment structure, the problem that traditional methods cannot effectively express spatial structures or collaborative relationships between measurement points in the early modeling stage is solved, and a highly adaptable structured input is provided for the graph neural network, which improves the spatial perception ability of the overall modeling, thereby laying a data foundation for subsequent feature extraction and anomaly analysis. By introducing the graph neural network structure and combining it with the feedforward neural network to complete the nested fusion and expression extraction of spatial-temporal information, the problem of insufficient ability of flat vector modeling in expressing the interactive dynamics of measurement points is solved. By comparing the embedded reconstructed value with the original observation value, calculating the node-level reconstruction error, and constructing the response distribution, the graph propagation mechanism is further used to simulate the anomaly diffusion process, evaluate the influence conduction path between nodes and identify the propagation end point, realize the fault impact range assessment and traceability positioning, realize the spatial extension analysis of the fault from "point" to "domain", and improve the system's understanding and response capabilities for multi-point coupled faults of complex industrial equipment, providing quantitative support for early warning response and maintenance decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 This is an overall flow chart of the industrial equipment fault early warning analysis method based on graph neural network provided by the first embodiment of the present invention.

[0021] Figure 2 A self-learning structure diagram of the industrial equipment fault early warning analysis method based on graph neural network provided in the second embodiment of the present invention.

[0022] Figure 3 A comparison chart of anomaly detection accuracy under different algorithms and structural configurations of the industrial equipment fault early warning analysis method based on graph neural network provided in the second embodiment of the present invention. DETAILED DESCRIPTION

[0023] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0024] Example 1, with reference to Figure 1 , as one embodiment of the present invention, provides an industrial equipment fault early warning analysis method based on graph neural network, comprising:

[0025] S1: Construct a time segment data tensor to form the graph neural network input structure.

[0026] Furthermore, constructing the time-slice data tensor includes obtaining the original time series data of the measurement points of the components in the target device, preprocessing the measurement point time series data, constructing the time-slice data tensor using a sliding window slicing method, and using the time-slice data tensor and the prior adjacency matrix as the input structure of the graph neural network.

[0027] It should be noted that the original time series data of the measurement points of the components in the target device is expressed as:

[0028] ;

[0029] in, Represents the original time series data of the measurement points, that is, the overall matrix composed of the time series data of all measurement points in the target device, Indicates the device Time series data of each measurement point, Indicates the total number of measurement points of the device. Indicates the total time length of the original data of the measuring point. Representation matrix One of the real numbers dimensional matrix.

[0030] The preprocessing of the measurement point time series data includes normalizing the original time series data of all measurement points into:

[0031] ;

[0032] in, Indicates the measurement points at time step Normalized measurement point time series data on Indicates the measurement points at time step The original time series data on Indicates the The mean of the measurement points in all time steps, Indicates the The standard deviation of a measurement point over all time steps.

[0033] The normalized measurement point time series data is constructed by sliding window slicing, and the obtained time segment data tensor is expressed as:

[0034] ;

[0035] in, Indicates the The measuring point is Normalized measurement point time series data within a sliding window, Indicates the sliding window length.

[0036] It should also be noted that by representing the original time series data of all measurement points of the target device as a two-dimensional matrix and performing normalization processing on each row, that is, the time series of each measurement point, the consistency calibration of the data distribution between different measurement points is achieved, thereby avoiding the bias problem caused by the difference in the numerical scale of the measurement points on model learning. By introducing a fixed-length sliding window on the basis of the normalized measurement point time series data and slicing the data along the time dimension, the purpose of converting long time series into short-time local segments is achieved, thereby enhancing the model's sensitivity to local fluctuations and mutation trends in time.

[0037] Furthermore, the prior adjacency matrix includes, if a device connection graph exists, parsing the device connection graph into a prior adjacency matrix;

[0038] If there is no device connection graph, the prior adjacency matrix calculated for the first time by the graph neural network is used.

[0039] It should be noted that if there is an equipment connection diagram (such as P&ID diagram, electrical connection diagram, process diagram), it is parsed as a priori adjacency matrix ,in, It represents the prior adjacency matrix and can be directly used as the initial graph structure input of the graph neural network. If the target device has an unstructured connection graph, the initial adjacency relationship is obtained based on the first round of graph convolution results during the graph neural network training process and constructed as a data-driven prior adjacency matrix to replace the manually specified connection structure.

[0040] It should also be noted that by parsing and constructing the prior adjacency matrix, compatible modeling capabilities are achieved for both actual industrial situations where structural graphs exist and where structural graphs do not exist. This enables graph neural networks to be stably initialized under different conditions of industrial equipment information completeness, avoiding the problem of structure-dependent modeling failure, and thus overcoming the limitation of traditional topology-based modeling methods that cannot model in scenarios where graph information is missing. At the same time, it avoids the risks of slow convergence and large expression deviation brought about by learning graph structures completely from scratch.

[0041] S2: Based on the graph neural network input structure, perform feature learning and association analysis of the graph neural network to obtain node-level embedding representation.

[0042] Furthermore, performing feature learning and association analysis of graph neural networks includes extracting node time features from the time segment data tensor, constructing a correlation score matrix between nodes, generating a sparse connection mask by retaining the Top-K scores of the correlation score matrix row by row, and multiplying the sparse connection mask by the correlation score matrix element by element to obtain a self-learning adjacency matrix.

[0043] It should be noted that two unbiased one-dimensional time convolutions are used to extract the node time features in the time segment data tensor. After feature extraction, dot product is performed to construct the correlation score matrix between nodes, which is expressed as:

[0044] ;

[0045] in, represents the correlation score matrix, represents the activation function, represents the node time feature, 、 represents an unbiased one-dimensional convolution module, Represents the scaling factor.

[0046] Unbiased one-dimensional convolution module 、 , the number of kernels is 1, and the hidden layer size is The unbiased one-dimensional convolution layer of the two modules does not share weights and can be learned. For each row vector of the correlation score matrix, the Top-K maximum value is retained and the rest is set to zero to obtain the mask matrix , and finally the mask matrix and the correlation score matrix Multiply and shield the nodes with lower scores, and the self-learning adjacency matrix is ​​expressed as:

[0047] ;

[0048] in, represents the self-learning adjacency matrix, represents the mask matrix, Represents the correlation score matrix.

[0049] It should also be noted that by using two one-dimensional unbiased convolution modules that do not share parameters, the time series features of each node are projected in different directions, and then scaled dot products are performed and Sigmoid activation is used to generate a correlation score matrix, thereby realizing unsupervised modeling of potential correlations between nodes, thereby effectively replacing manually defined graph structures or fixed topologies. By executing the maximum retention strategy on the correlation score matrix, only the most significant K connections of each node are retained, a sparse connection mask matrix is ​​constructed, and a self-learning adjacency matrix is ​​obtained, which realizes active shielding of weakly correlated connections, thereby compressing redundant edges in the graph structure, suppressing noise propagation, and reducing the computational complexity of the graph structure.

[0050] Furthermore, obtaining the node-level embedding representation includes fusing the self-learning adjacency matrix with the prior adjacency matrix in a preset ratio, iterating and normalizing the prior adjacency matrix to obtain the normalized adjacency matrix, performing graph convolution calculation on the normalized adjacency matrix and the node time features, aggregating the representations of adjacent nodes and updating the central node time features, inputting the graph convolution output result into the feedforward neural network layer, and outputting the node-level embedding representation.

[0051] It should be noted that the self-learning adjacency matrix and the prior adjacency matrix are fused in a preset ratio and expressed as:

[0052] ;

[0053] in, represents the adjacency matrix after fusion in the current iteration round, represents the fusion coefficient, represents the prior adjacency matrix of the previous round, Indicates the current iteration coefficient, represents the prior adjacency matrix.

[0054] when When , the previous round of fusion results are used as the new prior, and then fused with the self-learning adjacency matrix to form a gradual structure update mechanism.

[0055] when When there is a priori graph structure, the given prior adjacency matrix and the data self-learning graph structure are weighted and fused for the first time as the model input structure.

[0056] when When there is no prior graph structure, the adjacency matrix learned in the current round is used as the initial graph structure.

[0057] is the prior adjacency matrix and the self-learning adjacency matrix learning rate. Each time after the fusion matrix is ​​calculated, the initial prior adjacency matrix is ​​updated to obtain the iterative prior adjacency matrix, which is expressed as:

[0058] ;

[0059] in, represents the prior adjacency matrix after iteration.

[0060] The construction degree matrix is ​​expressed as:

[0061] ;

[0062] in, Degree matrix No. diagonal elements, indicating the The connectivity of the nodes, Representation node With node The connection weights of the prior adjacency matrix after iteration, represents the identity matrix, Indicates the total number of nodes, that is, the total number of measurement points of the device.

[0063] The normalized adjacency matrix is ​​expressed as:

[0064] ;

[0065] in, represents the normalized adjacency matrix.

[0066] The graph convolution calculation of the normalized adjacency matrix and the node time feature is expressed as:

[0067] ;

[0068] in, Indicates the Layer graph convolution output result, for activation function, It is the output data of the previous spatial graph convolution module. The overall spatial graph convolution module can be repeatedly superimposed. Second-rate, Indicates the Layer graph convolution weights, represents the convolution kernel size, Represents a transpose operation.

[0069] After spatial convolution performs spatial feature extraction, the time dimension feature of the hidden layer is extracted in depth in the connected feedforward neural network layer and expressed as:

[0070] ;

[0071] in, represents the feature output after graph convolution, , are all learnable weight matrices, is the number of features in the middle hidden layer, which is much larger than the length of the sliding window , used to convert data into Map to high-dimensional space, and then use Map to low-dimensional space to extract deeper features, , For bias.

[0072] The feature output after graph convolution is connected to the initial input data using the residual connection method, and the batch normalization operation is used to process the final result to avoid gradient explosion or gradient disappearance. The final output result is expressed as:

[0073] ;

[0074] in, Indicates the The result of the layer graph neural network decoding output is the node-level embedding representation. Indicates the The initial input of the layer, Indicates that batch normalization is performed on the residual connection results to speed up training convergence and suppress gradient explosion or disappearance.

[0075] It should also be noted that by introducing an iterative prior fusion mechanism, the model can flexibly adjust the degree of structural dependence. It not only retains the stable connection relationship in the existing engineering knowledge graph, but also dynamically updates the adjacency relationship based on the data, thereby improving the model's adaptability to incomplete graph structure or noise disturbances. By multiplying the normalized adjacency matrix with the node temporal feature matrix and combining the graph convolution kernel weights to extract node embedding features, the joint modeling of temporal behavior and structural relationships between multiple nodes is achieved, which enhances the expression ability of the context between nodes, making the node embedding more refined and semantically rich, and providing a reliable foundation for subsequent anomaly detection and fault prediction.

[0076] S3: Based on the node-level embedding representation, the fault correlation response distribution of the nodes is calculated to evaluate the propagation risk and locate the source of the equipment failure.

[0077] Furthermore, the fault correlation response distribution of the computing node includes: inputting the node-level embedding representation into the feedforward neural network decoder, performing reconstruction calculation on each node, obtaining the reconstructed value of the node in the time window, performing difference calculation between the reconstructed value of the node in the time window and the original input time series data of the sliding window, and outputting the node-level reconstruction error.

[0078] It should be noted that the node-level embedding representation is input into the feedforward neural network decoder as:

[0079] ;

[0080] in, represents the reconstructed value of the node within the time window, represents a feedforward neural network decoder.

[0081] The node-level reconstruction error is calculated as:

[0082] ;

[0083] in, represents the node-level reconstruction error, Indicates the The original input time series data of the sliding window of nodes, Indicates the The reconstructed value of a node in the time window, Represents the one-norm of a vector.

[0084] It should also be noted that reconstruction through learned embedded representations has stronger feature expression capabilities than directly fitting the original input, and can better capture nonlinear change trends under the time and space levels, providing a precise and dynamic basis for subsequent fault location. By combining node reconstruction errors and structural sparsity, it not only has abnormality perception capabilities, but also has structural self-constraint capabilities, ensuring that the representations trained under unlabeled conditions are stable and reliable, with good interpretability and generalization capabilities. The time window mechanism is integrated with the graph neural network decoder. Compared with the anomaly detection method of static node features, it can realize time-sensitive perception of node fault status, which is especially suitable for complex fault scenarios with delays or evolution patterns in industrial equipment, and can achieve accurate propagation path perception and source positioning.

[0085] Furthermore, evaluating the propagation risk and locating the source of equipment failure includes constructing the node whose node-level reconstruction error exceeds a preset threshold as the initial fault source, generating an initial propagation vector, performing iterative propagation calculation on the transpose of the normalized adjacency matrix through the initial propagation vector, stopping the propagation when the iterative propagation converges to a preset iteration threshold, obtaining the affected node set and evaluating the propagation risk, recording the node status activated in each round of propagation during the propagation iteration process, and locating the source of the equipment failure.

[0086] It should be noted that a preferred solution for the preset threshold is 2 times the standard deviation of the reconstruction error set of the training set. The node set of is taken as the fault source candidate, where It is twice the standard deviation of the reconstruction error set of the training set.

[0087] Generate the initial propagation vector as:

[0088] ;

[0089] in, represents the initial propagation vector, Represents the indicator function, which takes 1 if the fault source candidate conditions are met and takes 0 if they are not met. Represents element-wise multiplication.

[0090] The iterative propagation calculation performed on the normalized adjacency matrix through the initial propagation vector is expressed as:

[0091] ;

[0092] in, Indicates the The propagation probability distribution vector of each node after the round of propagation, represents the transpose of the normalized adjacency matrix, Indicates the The propagation probability result after round propagation is: ∈(0,1) is the attenuation coefficient, which controls the propagation attenuation speed.

[0093] Preset iteration threshold A preferred solution is 1e-3. When , the iteration ends and the affected node set is defined as:

[0094] ;

[0095] in, Represents the set of affected nodes, After the transmission ends, The final impact probability of nodes, Indicates the impact judgment threshold.

[0096] Evaluating the risk of transmission involves recording the probability of transmission in each round of transmission greater than The node set of all rounds of propagation is aggregated to form an activation node set, and the activation node sets of all rounds of propagation are summarized to form an activation trajectory sequence. The evolution process of the propagation path is visualized to determine the diffusion trend of the propagation impact and the key paths, and locate the source of the equipment failure.

[0097] It should also be noted that by recording the nodes whose propagation probability is greater than the preset impact judgment threshold in each round of propagation iteration and constructing an activated node set, the status of key nodes affected by faults in the potential propagation path can be captured in real time, breaking through the limitation of traditional methods that only identify risks based on the final state of propagation, and enhancing the visibility and traceability of the fault propagation path; by accumulating and summarizing the activated node set in each round, a propagation activation trajectory sequence is generated, which can be used to dynamically evolve the propagation diffusion process, and then identify the key paths and node clusters in the propagation path, which helps to achieve real-time perception and trend prediction of the propagation situation, and provide an accurate basis for subsequent equipment fault location and tracing; by introducing a propagation termination judgment mechanism in the propagation process, it not only improves the algorithm operation efficiency and avoids invalid propagation iterations, but also ensures that the propagation final state data has stability and statistical significance, providing data support for subsequent risk aggregation and fault location based on the activation state.

[0098] Example 2 is an embodiment of the present invention, which provides an industrial equipment fault warning analysis method based on graph neural network. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0099] This experiment collects multi-dimensional sensor data such as pressure (kPa), temperature (°C), and vibration frequency (Hz) at each measuring point during the operation of an industrial equipment. The measuring points are numbered 1 to 7. 10,000 normal training samples are recorded as shown in Table 1, and 500 test samples are recorded as shown in Table 2.

[0100] Table 1 Normal training samples of equipment component operation data

[0101]

[0102] Table 2 Equipment component operation data verification sample

[0103]

[0104] In the training phase, the unbiased convolutional structure and the prior adjacency matrix are used to guide the network to learn spatial relationships. The output adjacency matrix after learning is shown in Table 3, which shows the spatial dependency weights between different nodes. The structural coupling relationship is automatically identified from the original data, indicating that measurement points 3 and 6 are important structural upstream or downstream of measurement point 1. Compared with traditional fixed topology or pure data-driven models, the adjacency matrix shows that this model can identify implicit dependency paths across variables, especially the 0.3932 structural dependency between measurement points 6 and measurement points 1, which has no obvious physical location adjacency but is accurately captured. Figure 2 , is a self-learning structure diagram, which shows the spatial dependency path between nodes. Nodes 0–6 represent 7 measurement points. The direction of the arrow indicates the dependency direction. The existence of an edge indicates that the connection is judged by the model to have significant structural association after training.

[0105] Table 3. Output adjacency matrix of graph neural network after training

[0106]

[0107] During each propagation iteration, the set of activated nodes is compared with the “anomaly mark” field in the actual verification sample to determine whether the propagation path touches the real abnormal node. The propagation time steps required from the initial propagation source to the target node where the abnormality occurs are counted. The fault propagation calculation is shown in Table 4.

[0108] Table 4 Fault propagation calculation table

[0109]

[0110] The data in Table 4 shows that the anomaly at measuring point 5 can be accurately identified as propagating to measuring point 4, confirming the effectiveness of the "structural propagation path prediction." The propagation path hits the actual abnormal node, and the fault propagates from measuring point 6 to downstream points 4 and 5 in just 50 steps. This indicates that the model path is shortened, the alarm is issued earlier, and the model can be traced to the end of the indirect abnormal path (node ​​0), demonstrating the path structure perception capability. Even with multi-source input, the anomaly at node 4 is still successfully and synchronously located, demonstrating fault tolerance and multi-point concurrent identification capabilities.

[0111] The accuracy of abnormality monitoring of different algorithms and the impact of different parameters on the accuracy are as follows: Figure 3As shown in the figure, under all top-k settings (k=1 to 7), the model introducing the prior structure is always better than the model without introducing the prior structure in terms of accuracy, which shows that the introduction of prior knowledge in the present invention can effectively improve the robustness and detection ability of graph structure construction; for the case without introducing the prior, the overall accuracy is distributed in the range of 0.85 to 0.87, with certain fluctuations, while the accuracy distribution of the scheme introducing the prior is concentrated between 0.89 and 0.91 as a whole, with higher stability, indicating that it can still maintain excellent performance when the number of neighbors changes; when the top-k is set to 6, the accuracy reaches the highest value, which further illustrates that under the reasonable setting of the number of neighbors, the prior structure constructed by the present invention is of key significance to improving model performance.

[0112] Example 3 is an embodiment of the present invention, which provides an industrial equipment fault warning and analysis system based on graph neural network, including a graph model input construction module, a graph feature learning and association analysis module, and a fault identification and propagation diagnosis module.

[0113] The graphical model input construction module includes a sliding window tensor construction unit and an adjacency matrix generation unit. The sliding window tensor construction unit is used to receive time series data from multiple measurement points of industrial equipment, and slice the original data using a sliding window method to generate a time segment data tensor. The adjacency matrix generation unit is used to parse the equipment structure diagram and construct an initial prior adjacency matrix.

[0114] The graph feature learning and association analysis module includes a graph convolutional encoding unit and a feedforward neural decoding unit. The graph convolutional encoding unit is used to extract the spatial and temporal joint features of the nodes, and the feedforward neural decoding unit is used to restore the time series of each node and generate a reconstructed value.

[0115] The fault identification and propagation diagnosis module includes an anomaly scoring unit and a propagation diagnosis unit. The anomaly scoring unit is used to calculate the reconstruction error of the node and generate an anomaly score. The propagation diagnosis unit is used to perform propagation simulation, identify the affected nodes and locate the fault source.

Claims

1. An industrial equipment fault early warning analysis method based on graph neural network, characterized by: include: Construct time segment data tensors to form the graph neural network input structure; Based on the graph neural network input structure, perform feature learning and association analysis of the graph neural network to obtain node-level embedding representation; Based on the node-level embedding representation, the fault correlation response distribution of the node is calculated to evaluate the propagation risk and locate the source of the equipment failure; The execution of feature learning and association analysis of the graph neural network includes extracting node time features from the time segment data tensor, constructing a correlation score matrix between nodes, generating a self-learning adjacency matrix based on the score matrix, fusing the self-learning adjacency matrix with the prior adjacency matrix and performing normalization, performing graph convolution calculation using the normalized adjacency matrix and the node time features, and inputting the graph convolution result into a feedforward neural network to output a node-level embedding representation; The method of evaluating the propagation risk and locating the source of the equipment fault includes calculating the fault correlation response distribution of the nodes, screening the fault source nodes according to a preset threshold, constructing an initial propagation state vector, performing iterative propagation using a normalized adjacency matrix, generating the fault correlation response distribution of the nodes, and evaluating the propagation risk and locating the source of the equipment fault based on the fault correlation response distribution of the nodes.

2. The industrial equipment fault early warning analysis method based on graph neural network according to claim 1, characterized in that: The construction of the time-slice data tensor includes obtaining the original time series data of the measurement points of the components in the target device, preprocessing the measurement point time series data, constructing the time-slice data tensor using a sliding window slicing method, and using the time-slice data tensor and the prior adjacency matrix as the graph neural network input structure.

3. The industrial equipment fault early warning analysis method based on graph neural network according to claim 2, characterized in that: The a priori adjacency matrix includes, if a device connection graph exists, parsing the device connection graph into a a priori adjacency matrix; If there is no device connection graph, the prior adjacency matrix calculated for the first time by the graph neural network is used.

4. The industrial equipment fault early warning analysis method based on graph neural network according to claim 3 is characterized by: The feature learning and association analysis of the execution graph neural network includes extracting node time features in the time segment data tensor, constructing a correlation score matrix between nodes, retaining the Top-K scores of the correlation score matrix row by row to generate a sparse connection mask, and multiplying the sparse connection mask and the correlation score matrix element by element to obtain a self-learning adjacency matrix.

5. The industrial equipment fault early warning analysis method based on graph neural network according to claim 4 is characterized in that: The node-level embedding representation is obtained by fusing the self-learning adjacency matrix with the prior adjacency matrix in a preset ratio, iterating and normalizing the prior adjacency matrix to obtain a normalized adjacency matrix, performing graph convolution calculation on the normalized adjacency matrix and the node time feature, aggregating the representations of adjacent nodes and updating the central node time feature, inputting the graph convolution output result into the feedforward neural network layer, and outputting the node-level embedding representation.

6. The industrial equipment fault early warning analysis method based on graph neural network according to claim 5, characterized in that: The fault correlation response distribution of the computing node includes: embedding the node level representation into a feedforward neural network decoder, performing reconstruction calculation on each node to obtain a reconstructed value of the node within a time window, performing difference calculation between the reconstructed value of the node within the time window and the original input time series data of the sliding window, and outputting a node-level reconstruction error.

7. The industrial equipment fault early warning analysis method based on graph neural network according to claim 6, characterized in that: The method of evaluating the propagation risk and locating the source of the equipment failure includes constructing a node whose node-level reconstruction error exceeds a preset threshold as an initial fault source, generating an initial propagation vector, performing an iterative propagation calculation on the transpose of the normalized adjacency matrix using the initial propagation vector, stopping the propagation when the iterative propagation converges to a preset iteration threshold, obtaining a set of affected nodes and evaluating the propagation risk, recording the status of the nodes activated in each round of propagation during the propagation iteration process, and locating the source of the equipment failure.

8. An industrial equipment fault early warning analysis system based on a graph neural network, using the industrial equipment fault early warning analysis method based on a graph neural network according to any one of claims 1 to 7, characterized in that: It includes a graph model input construction module, a graph feature learning and association analysis module, and a fault identification and propagation diagnosis module; The graph model input construction module includes a sliding window tensor construction unit and an adjacency matrix generation unit. The sliding window tensor construction unit is used to receive time series data of multiple measurement points of industrial equipment and slice the original data using a sliding window method to generate a time slice data tensor. The adjacency matrix generation unit is used to parse the equipment structure diagram and construct an initial prior adjacency matrix. The graph feature learning and association analysis module includes a graph convolutional encoding unit and a feedforward neural decoding unit. The graph convolutional encoding unit is used to extract the spatial and temporal joint features of the nodes, and the feedforward neural decoding unit is used to restore the time series of each node and generate a reconstructed value. The fault identification and propagation diagnosis module includes an anomaly scoring unit and a propagation diagnosis unit. The anomaly scoring unit is used to calculate the reconstruction error of the node and generate an anomaly score. The propagation diagnosis unit is used to perform propagation simulation, identify affected nodes and locate the fault source.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the industrial equipment fault early warning analysis method based on graph neural network described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the industrial equipment fault early warning analysis method based on graph neural network described in any one of claims 1 to 7 are implemented.

Images