Heterogeneous graph-based power equipment fault prediction method, device and equipment

Abstract

This invention provides a method, apparatus, and device for predicting power equipment faults based on heterogeneous graphs, relating to the field of power equipment monitoring technology. The method includes: constructing a heterogeneous graph of power equipment, using each power device in the power system as a node and determining the edges of each node based on the relationships between the power devices; acquiring multimodal data of each power device; determining the initial features of each node based on the multimodal data; determining the aggregated features of each node based on the initial features of each node through at least one aggregation process; wherein, in each aggregation process, the current features of all neighboring nodes of each node are aggregated to update the features of each node; generating a spatiotemporal representation vector of each power device based on the continuous time series of the aggregated features of each node, and predicting the fault identification results of each device. This invention can improve the accuracy of power equipment fault prediction and provide technical support for the safe operation of power systems.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, and in particular to a method, apparatus and equipment for predicting power equipment faults based on heterogeneous graphs. Background Technology

[0002] The power system is a critical infrastructure of the national economy, and equipment failures can lead to large-scale power outages and economic losses. Traditional equipment failure prediction methods mainly rely on a single data source and human experience, making it difficult to fully utilize the multi-source heterogeneous data in the power system, including equipment operating parameters, environmental monitoring data, and historical maintenance records.

[0003] Existing technologies for predicting power equipment faults by integrating multi-source data often analyze data from different modes separately to obtain the analysis results for a single mode, and then couple the results from multiple modes to obtain a comprehensive analysis result. In addition, existing technologies usually treat power equipment as an independent entity for condition assessment, making it difficult to comprehensively analyze data from different devices and ignoring the propagation characteristics and associated impacts of equipment faults in the power network.

[0004] Therefore, there is an urgent need for a power equipment fault prediction method that can make full use of heterogeneous correlation information and multi-source data in the power system to improve the accuracy of power equipment fault prediction and provide technical support for the safe operation of the power system. Summary of the Invention

[0005] This invention provides a method, apparatus, and device for predicting power equipment faults based on heterogeneous graphs, in order to improve the accuracy of power equipment fault prediction and provide technical support for the safe operation of power systems.

[0006] In a first aspect, embodiments of the present invention provide a method for predicting power equipment faults based on heterogeneous graphs, including: Using each power device in the power system as a node, and determining the edges of each node based on the relationships between the power devices, a heterogeneous graph of power devices is constructed. Acquire multimodal data of each power device; determine the initial features of each node in the heterogeneous graph of the power devices based on the multimodal data; Based on the initial features of each node, the aggregated features of each node are determined through at least one aggregation process; wherein, in each aggregation process, the current features of all neighboring nodes of each node are aggregated to update the features of each node. Based on the continuous time series of the aggregated features of each node, a spatiotemporal representation vector of each power device is generated, and the fault identification result of each device is predicted.

[0007] In one possible implementation, the current features of all neighboring nodes of each node are aggregated during each aggregation process to update the features of each node, including: For each node, perform the following steps: Get the set of neighbor nodes of the given node under each edge type; We perform weighted aggregation of the current features of all neighbor nodes in each set of neighbor nodes to determine the relation features of the node under each edge type. The node's features are updated by integrating the relational features of the node under various edge types.

[0008] In one possible implementation, the step of weighted aggregation of the current features of all neighbor nodes in each set of neighbor nodes to determine the relational features of the node under each edge type includes: For each edge type of this node, perform the following steps: Based on the feature transformation matrix and attention parameter vector of this edge type, the aggregation weights of each neighboring node of this node under this edge type are determined; wherein, the feature transformation matrix and the attention parameter vector are obtained through pre-training optimization; Based on the aggregation weight, the current features of each neighboring node of the node under the edge type are weighted and aggregated to determine the relation features of the node under each edge type.

[0009] In one possible implementation, acquiring multimodal data of each power device and determining the initial features of each node in the heterogeneous graph of the power devices based on the multimodal data includes: Based on the time-varying metric matrix, dynamic time warping is performed on the multimodal data of each power device; different modal features of each node are extracted from the time-warped multimodal data; Based on the feature projection matrix, the different modal features of each node are mapped to the joint embedding space to obtain the embedding features of each node; A topological consistency constraint is established for the heterogeneous graph of the power equipment, and the embedding features of each node are adjusted based on a cross-modal attention mechanism to obtain the initial features of each node.

[0010] In one possible implementation, the time-varying metric matrix is ​​obtained from metric mapping parameters and equipment operating condition characteristics; wherein, the metric mapping parameters are parameters obtained by pre-training and optimization using historical data, and the equipment operating condition characteristics are output by the first gated loop unit based on the real-time load, real-time ambient temperature, and data acquisition time interval of the equipment.

[0011] In one possible implementation, the construction of a heterogeneous graph of power devices, using each power device in the power system as a node and determining the edges of each node based on the relationships between the power devices, includes: Based on the power equipment types and relationships between power equipment in the power system, the types of each node and the types of connecting edges are defined, and the adjacency matrix of each edge type is determined to construct the heterogeneous graph of the power equipment.

[0012] In one possible implementation, generating the spatiotemporal representation vector of each power device based on the continuous time series of the aggregated features of each node, and predicting the fault identification result of each device, includes: The continuous time series of the aggregated features of each node are input into the pre-trained second gated recurrent unit to obtain the spatiotemporal representation vector of each power device; Based on the spatiotemporal representation vectors of each power device, a multilayer perceptron is used to predict fault identification results; wherein, the fault identification results include fault probability, fault type, and a list of affected devices.

[0013] In one possible implementation, the method further includes: When the first device fails and causes the second device to fail, in the heterogeneous graph of the power equipment, a causal edge is added between the first node corresponding to the first device and the second node corresponding to the second device, and the aggregation weight of the second node under the causal edge of the first node is adjusted.

[0014] Secondly, embodiments of the present invention provide a power equipment fault prediction device based on heterogeneous graphs, comprising: The heterogeneous graph construction module is used to construct a heterogeneous graph of power equipment, taking each power device in the power system as a node and determining the edges of each node based on the relationship between the power devices. A multimodal feature extraction module is used to acquire multimodal data of various power devices; and to determine the initial features of each node in the heterogeneous graph of the power devices based on the multimodal data. The feature aggregation module is used to determine the aggregated features of each node based on the initial features of each node through at least one aggregation process; wherein, in each aggregation process, the current features of all neighboring nodes of each node are aggregated to update the features of each node. The fault prediction module is used to generate a spatiotemporal representation vector of each power device based on the continuous time series of the aggregated features of each node, and to predict the fault identification result of each device.

[0015] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect or any possible implementation thereof.

[0016] This invention uses power equipment as nodes and constructs a heterogeneous graph based on the relationships between equipment to establish a comprehensive data framework that reflects heterogeneous relationships. This facilitates fault prediction using information about interactions between equipment. Initial node features are determined based on multimodal data, fully mining multi-dimensional state information to comprehensively reflect real-time operating conditions and health status, compensating for the deficiencies of single-modal data and improving the accuracy and robustness of fault identification. Features of neighboring nodes are aggregated through multi-layer adaptive heterogeneous graph convolution to achieve adaptive modeling of heterogeneous relationships. This comprehensively learns the influence of different relationship types and neighboring equipment, enabling aggregated features to integrate their own multimodal information and multi-hop neighborhood structure semantics, deeply characterizing the global context information of equipment in the system and providing more discriminative feature representations for fault assessment. A spatiotemporal representation vector is generated based on the time series of aggregated features to predict fault identification results, capturing the temporal evolution and spatial interaction of states, reflecting the gradual fault progression and neighborhood collaborative changes, achieving early warning and accurate location, and significantly improving system safety and reliability.

[0017] The embodiments of the present invention form a complete technical chain from data analysis, feature identification and risk prediction by organically combining the construction of heterogeneous graphs, the fusion of multimodal data, adaptive graph convolution and spatiotemporal sequence modeling, and finally realizes efficient and accurate prediction of power equipment failures. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the implementation of the power equipment fault prediction method based on heterogeneous graphs provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the structure of the power equipment fault prediction device based on heterogeneous graphs provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0020] In actual power system operation, there is significant heterogeneity in the multi-source monitoring data of power equipment. For example, sensors collect numerical time-series data such as oil temperature and current at high frequencies on the order of seconds / minutes, while textual information such as equipment maintenance records and inspection reports are only generated monthly, and infrared images are often acquired through irregular shooting. Directly stitching or processing these data of different modalities and frequencies separately will lead to fragmented feature semantics, making it impossible to accurately represent the true state of power equipment. In addition, the functional dependency graph obtained from historical fault statistics may be inconsistent with the physical topology of the power grid (for example, power equipment without direct electrical connections may frequently experience simultaneous faults due to environmental factors). If only physical connection relationships are considered, the results of fault tracing and impact range determination are often inaccurate. Therefore, developing a power equipment fault prediction method that can combine the operating conditions of different power equipment and the relationships between power equipment has important practical significance and application value.

[0021] like Figure 1 As shown, this embodiment of the invention provides a method for predicting power equipment faults based on heterogeneous graphs, detailed below: Step S101: Using each power device in the power system as a node, and determining the edges of each node based on the association between the power devices, a heterogeneous graph of power devices is constructed.

[0022] Specifically, taking each power device in the power system as a node, and determining the edges of each node based on the relationships between the power devices, a heterogeneous graph of power devices is constructed, including: Based on the types of power equipment and the relationships between them in a power system, this paper defines the types of each node and the types of connecting edges, determines the adjacency matrix of each edge type, and constructs a heterogeneous graph of power equipment. The types of each node can include generator nodes, transformer nodes, transmission line nodes, distribution equipment nodes, and monitoring sensor nodes, and a set of node types is determined. The types of connection edges can include physical connection edges, electrical coupling edges, spatial proximity edges, functionally dependent edges, and historically associated edges, and the set of edge types is determined. The adjacency matrix for each edge type can be represented as follows: ,in, r This indicates the edge type of the adjacency matrix.

[0023] Interrelationships between various power devices refer to the various physical or logical dependencies that exist between different devices in a power system. These relationships can include: physical connections (such as a transformer directly connected to a transmission line), electrical coupling (such as devices powered by the same busbar), spatial proximity (such as devices located in the same substation being affected by the same environment), functional dependencies (such as a main transformer failure causing downstream distribution equipment outages), and historical relationships (such as pairs of devices that have historically experienced frequent shared failures). These relationships reflect the inherent connections between devices in terms of operation, structure, or fault propagation.

[0024] An adjacency matrix is ​​used to mathematically represent the connection structure under a certain type of association. For each edge type... r (e.g., physical connections), and their corresponding adjacency matrices It is a two-dimensional matrix, where the elements Represents a node i With nodes j There exists a type of r The edge, otherwise the element (Or learnable weight values). By defining adjacency matrices for each edge type, the multi-relationship structure of heterogeneous graphs can be fully characterized.

[0025] Heterogeneous graphs of power equipment can simultaneously model multiple types of power equipment (such as generators, transformers, sensors, etc.) and the relationships between different power equipment, thereby reconstructing the complex topological relationships and interaction patterns of the power system. Compared with homogeneous graphs, heterogeneous graphs can distinguish the impact of different relationships on power fault propagation, laying the foundation for adaptive feature aggregation based on edge type.

[0026] In this embodiment of the invention, the power system is transformed from a collection of multiple isolated devices into a heterogeneous graph-structured network, which can model the multi-dimensional relationships between power devices. This provides accurate graph structure support for integrating multi-source data and capturing the propagation patterns of power faults on different types of paths, significantly improving the interpretability and accuracy of fault prediction.

[0027] Step S102: Obtain multimodal data of each power device; determine the initial features of each node in the heterogeneous graph of the power devices based on the multimodal data.

[0028] In one possible implementation, multimodal data of each power device is acquired; based on the multimodal data, initial characteristics of each node in the heterogeneous graph of the power devices are determined, including: Based on the time-varying metric matrix, dynamic time warping is performed on the multimodal data of each power device; different modal features of each node are extracted from the time-warped multimodal data; Based on the feature projection matrix, the different modal features of each node are mapped to the joint embedding space to obtain the embedding features of each node; A topological consistency constraint is established for the heterogeneous graph of power equipment, and the embedding features of each node are adjusted based on the cross-modal attention mechanism to obtain the initial features of each node.

[0029] Dynamic Time Warping (DTW) is an algorithm for aligning two time series of different lengths or sampling frequencies. It achieves semantic alignment of the two series along the time axis by finding the optimal nonlinear bending path. For example, it establishes a reasonable temporal correspondence between high-frequency sensor data (acquired every minute) and low-frequency maintenance records (generated daily). This invention embodiment incorporates a time-varying metric matrix. This invention improves upon traditional DTW: Traditional DTW uses Euclidean distance (i.e., the metric matrix is ​​an identity matrix), while this embodiment introduces a time-varying metric matrix that dynamically changes with the device's operating state. This makes the time alignment more consistent with the actual operating patterns of power equipment.

[0030] The different modal features of each node are feature vectors extracted from different modal data, corresponding to each power device. These can include numerical features (such as time-series statistics from sensors like oil temperature, current, and voltage), text features (such as semantic vectors encoded by the BERT model in maintenance reports), and image features (such as visual representations extracted from infrared thermal images using ResNet). These features have different dimensions and semantic spaces. In this embodiment of the invention, they are spatially aligned using a feature projection matrix. The feature projection matrix is ​​a set of updatable linear transformation matrices that act on the features of different modalities, mapping these features to a unified joint embedding space to obtain embedded features, thereby eliminating scale and distribution differences between different modalities. The joint embedding space is a unified high-dimensional vector space. The features of each modality are comparable and additive in the joint embedding space after projection. The joint embedding space enables the consistent expression of the semantics corresponding to different modal features such as "oil temperature rise," "report description of overheating," and "infrared image reddening." Embedded features are vectors formed by projecting and concatenating (or weighting and fusing) the features of each modality, serving as the initial semantic representation of the nodes, but heterogeneous graph structure information has not yet been considered.

[0031] Topological consistency constraints are a type of regularization mechanism; specifically, they can be based on the edge type corresponding to physical connectivity relationships. adjacency matrix Edge types corresponding to functional dependencies Adjacency matrix The Frobenius norm between them guides the direction of cross-modal feature adjustment, avoiding conflicts between semantic alignment results and the actual topological relationships of the power system.

[0032] Cross-modal attention mechanisms are used to calculate the correlation weights between different modal features (such as the attention score of text features to numerical features), and dynamically adjust the contribution of each modal feature to the initial features of a node based on these correlation weights. Embodiments of this invention can further calibrate multimodal semantics in the joint embedding space, for example, bringing "high oil temperature" and "overheating description" closer together in the vector space, thereby improving the consistency and discriminative power of the initial features.

[0033] The initial features of each node generated in the embodiments of the present invention not only retain the richness of multimodal information, but also have semantic consistency and structural rationality, providing high-quality input for feature aggregation and updating based on heterogeneous graphs, and can significantly improve the accuracy and robustness of fault prediction.

[0034] Step S103: Based on the initial features of each node, determine the aggregated features of each node through at least one aggregation process; wherein, in each aggregation process, the current features of all neighboring nodes of each node are aggregated to update the features of each node.

[0035] The aggregation process represents an iterative update of node features. Each time aggregation is completed, a node's features incorporate information from its neighboring nodes within a one-hop radius; if performed twice, information from two hop radius nodes (i.e., neighbors of neighbors) is incorporated, and so on. In this embodiment of the invention, three-stage aggregation can be employed, enabling each node to ultimately perceive device information within its three-hop neighborhood, thereby capturing a wider range of fault propagation.

[0036] When aggregating neighbor features, average aggregation or fixed weighting methods cannot reflect the differences in importance of different neighbor nodes to the central node, nor can they perceive changes in the importance of the same neighbor node under different operating states. Therefore, aggregating the current features of all neighbor nodes of each node in each aggregation process to update the features of each node can include: For each node, perform the following steps: Get the set of neighbor nodes of the given node under each edge type; We perform weighted aggregation of the current features of all neighbor nodes in each set of neighbor nodes to determine the relation features of the node under each edge type. The node's relationship features across different edge types are merged, and the node's features are updated.

[0037] In this embodiment of the invention, weighted aggregation dynamically adjusts the contribution of each neighbor. The weights of the weighted aggregation are not preset but learned from historical fault data, automatically identifying which nodes are more indicative of fault prediction under specific relationships. Therefore, this embodiment of the invention can more precisely characterize the dynamic influence relationships between devices, effectively improving the quality of feature representation and the sensitivity of fault identification. Through a multi-layered, relation-based, and weighted aggregation mechanism, multi-layered, adaptive information propagation between heterogeneous graph nodes is achieved, thereby accurately capturing fault propagation paths without interference between different types of information. Simultaneously, it highlights the contribution of key neighbors, suppresses interference from irrelevant or normal neighbor nodes, amplifies the influence of abnormal or high-risk neighbor nodes, and enhances feature discrimination. Furthermore, through multi-layered aggregation, node features integrate local and global device states, providing data rich in inter-device correlation information for spatiotemporal modeling. This embodiment of the invention enables the heterogeneous graph information aggregation process to have adaptive perception capabilities for heterogeneous correlations and dynamic operating conditions of the power system, significantly improving the accuracy and interpretability of fault prediction.

[0038] Step S104: Based on the continuous time series of the aggregated features of each node, generate the spatiotemporal representation vector of each power device and predict the fault identification result of each device.

[0039] In one possible implementation, based on the continuous time series of the aggregated features of each node, a spatiotemporal representation vector of each power device is generated, and the fault identification result of each device is predicted, including: The continuous time series of aggregated features of each node are input into the pre-trained second gated recurrent unit to obtain the spatiotemporal representation vector of each power device; Based on the spatiotemporal representation vectors of each power device, the fault identification results are predicted using a multilayer perceptron; the fault identification results include the fault probability, fault type, and a list of affected devices.

[0040] A continuous time series refers to a sequence formed by arranging the aggregated features obtained from multiple consecutive time steps for each power device (i.e., each node in a heterogeneous graph) in chronological order. It contains both the evolutionary information of the device's own state and the dynamic influence of its neighborhood over time.

[0041] The spatiotemporal representation vector is a fixed-dimensional vector output by encoding the continuous time series of aggregated features through a temporal modeling network. It integrates the multi-hop correlation information of power equipment in space, as well as the trend and abrupt change features reflected by the evolution of historical states in time. The spatiotemporal representation vector can be regarded as a comprehensive representation of the current health status of the equipment and its short-term evolution trend.

[0042] In one possible implementation, the temporal modeling network that outputs the spatiotemporal representation vector can be a pre-trained second-gated recurrent unit, which can be trained using historical monitoring data and fault labels before the implementation of this scheme. The second-gated recurrent unit can effectively capture long-term dependencies, periodic patterns, and anomalous mutations in the aggregated feature sequence.

[0043] The failure probability is the probability of the device failing within a preset time window, as output by the model, and is used for risk ranking and early warning threshold determination; the failure type is the specific failure category predicted, such as winding overheating, insulation degradation, etc., to help maintenance personnel quickly locate problems; the list of affected devices is a set of other power devices that may be affected by this failure, inferred from the path of the edge propagation of the failure in the heterogeneous graph, and is used to analyze the scope of the failure impact and assist technicians in formulating isolation strategies.

[0044] This invention improves upon traditional methods by moving from single-moment data extraction to historical evolution trajectories. By encoding the sequence of aggregated features through a pre-trained second-gated recurrent unit, it not only preserves the spatial correlation information provided by the heterogeneous graph but also uncovers the temporal patterns of fault development. Based on this, the multilayer perceptron transforms the high-dimensional spatiotemporal representation into corresponding fault identification results. The prediction output is no longer limited to whether a fault has occurred but provides comprehensive decision support including probability quantification, type diagnosis, and impact range. This effectively overcomes the limitations of traditional methods that rely solely on current state analysis, enhancing the foresight and practicality of fault prediction and providing reliable technical support for power system operation and maintenance.

[0045] The embodiments of the present invention form a complete technical chain from data analysis, feature identification and risk prediction by organically combining the construction of heterogeneous graphs, the fusion of multimodal data, adaptive graph convolution and spatiotemporal sequence modeling, and finally realizes efficient and accurate prediction of power equipment failures.

[0046] In some embodiments, edge pairs of different types in a heterogeneous graph propagate information with different semantics. Using a uniform aggregation rule would confuse relation specificity, leading to distorted feature fusion. Therefore, it is necessary to provide adaptive weights for each edge type to achieve refined neighbor aggregation with relation awareness. Weighted aggregation of the current features of all neighbor nodes in each set of neighbor nodes to determine the relation features of that node under each edge type can include: For each edge type of this node, perform the following steps: Based on the feature transformation matrix and attention parameter vector of this edge type, the aggregation weights of each neighbor node of this node under this edge type are determined; wherein, the feature transformation matrix and attention parameter vector are obtained by pre-training optimization; Based on the aggregation weight, the current features of each neighboring node of the node under the edge type are weighted and aggregated to determine the relation features of the node under each edge type.

[0047] The feature transformation matrix is ​​a learnable weight matrix pre-trained for each edge type. Its function is to map the features of neighboring nodes to a semantic space that is more suitable for information interaction under that relationship. For example, under electrical coupling, the feature transformation matrix can increase the emphasis on electrical quantities such as current and voltage; while under spatial proximity, it can focus on features corresponding to environmental data such as temperature and humidity.

[0048] The attention parameter vector is also a learnable vector specific to edge types, used to calculate the relevance score between the center node and each neighbor node in the transformed feature space. The features of the center node and neighbor nodes, transformed by the feature transformation matrix respectively, are concatenated together, and then multiplied with the attention parameter vector. After non-linear activation and normalization, the aggregate weights are obtained. This mechanism enables the model to learn to determine which neighbors are more worthy of attention in this type of relationship under the current device state.

[0049] This invention achieves customized dynamic aggregation of relationships by equipping each edge type with independent feature transformation and attention parameters. This not only distinguishes the semantics of different relationships but also adaptively adjusts the importance of neighbors based on the real-time operating status of the devices. Compared to schemes using fixed weights, this invention significantly improves the ability of aggregated features to model complex correlations and fault propagation patterns in power systems.

[0050] In some embodiments, since the fault sensitivity characteristics of power equipment change dynamically with operating conditions, the dynamic time warping of traditional fixed metrics is difficult to adapt to seasonal or load-related alignment requirements. Therefore, the time-varying metric matrix can be obtained from metric mapping parameters and equipment operating condition characteristics. The metric mapping parameters can be parameters optimized through pre-training on historical data, and the equipment operating condition characteristics can be output by the pre-trained first gated loop unit based on the real-time load, real-time ambient temperature and data acquisition time interval of the equipment.

[0051] Metric mapping parameters Used to convert the timing encoding of equipment operating conditions (such as the hidden state output by a gated loop unit) into a time-varying metric matrix. A set of learnable weights. Specifically, It can be a matrix, whose function is to map the equipment condition features output by the first gated loop unit to a matrix space, thereby generating a feature distance metric that can be dynamically adjusted. The metric mapping parameters can be pre-trained and optimized using historical monitoring data and fault labels. They determine the relative importance of each feature dimension (such as temperature and load) in time alignment under specific operating conditions, enabling dynamic time warping to adaptively strengthen the alignment weights of key features based on conditions such as season and load.

[0052] The time-varying metric matrix obtained in the embodiments of the present invention Used to calculate weighted distance, it can automatically strengthen the alignment weight of temperature-related features during high-temperature periods, while focusing on load fluctuation features in winter, making time alignment more consistent with the actual operating patterns of power equipment.

[0053] In some embodiments, since the propagation path of a fault in a power system can be understood as the causal relationship between faults occurring between power devices, and this causal relationship can more accurately reflect the fault transmission mechanism, the method may further include: When the first device fails and causes the second device to fail, in the heterogeneous graph of power equipment, a causal edge is added between the first node corresponding to the first device and the second node corresponding to the second device, and the aggregation weight of the second node under the causal edge of the first node is adjusted.

[0054] A causal edge is a directed relation edge used to represent a causal dependency where a failure in one device causes a failure or malfunction in another device.

[0055] This invention improves the accuracy of fault tracing and impact range prediction by modeling the causal edges of fault propagation and dynamically adjusting their aggregation weights. This enables the aggregation features obtained from heterogeneous graph aggregation to more accurately capture the real fault impact paths between devices.

[0056] In one possible implementation, extracting different modal features of each node from time-warped multimodal data may include: For numerical operating parameter data, a sliding time window is used to extract statistical features, including mean, variance, peak value, and rate of change. For text-based maintenance records, a pre-trained language model is used to extract semantic embedding vectors; For image-based monitoring data, a pre-trained convolutional neural network is used to extract visual features.

[0057] In one possible implementation, dynamic time warping of the multimodal data of various power devices may include: The Dynamic Time Warping (DTW) algorithm is applied to high-dimensional sensor data to establish a time mapping relationship with low-frequency maintenance records;

[0058] Where π is the alignment path and the time-varying metric matrix. for:

[0059] in, For the measurement mapping parameters, the inputs of the first gated loop unit include real-time load, temperature, and time interval.

[0060] In one possible implementation, based on the feature projection matrix, the different modal features of each node are mapped to the joint embedding space to obtain the embedding features of each node, which may include: Constructing a multimodal joint embedding space:

[0061] in, , , For each modality, there is a feature projection matrix. ⊕ denotes the Hadamard product, and ⊕ denotes feature concatenation. The update method for the feature projection matrix corresponding to each modality is as follows:

[0062] in, The feature projection matrix at the current time; This is the feature projection matrix from the previous time step; This method identifies the feature differences between multimodal data at the current time, enabling the quantification of semantic differences between different modalities (e.g., the correlation strength between a partial discharge image and its corresponding text data). Based on the feature projection matrix from the previous time step and the feature differences between the multimodal data at the current time step, a gated recurrent unit function is used. Update the feature projection matrix at the current time step It can preserve the time-dependent operation of power equipment and capture the gradual changes in the state of power equipment (such as the cumulative changes in insulation aging over several months).

[0063] In one possible implementation, the topology consistency constraint is: ;in, Edge type corresponding to physical connection relationship The adjacency matrix, Edge type corresponding to functional dependency relationship Adjacency matrix, This represents the Frobenius norm.

[0064] Based on the cross-modal attention mechanism, the embedded features of each node are adjusted to obtain the initial features of each node, which may include: Based on the attention coefficient, determine the correlation matrix between different modal features of each node. (N is the total number of nodes):

[0065] and These are the learnable query matrix and the learnable key matrix, respectively. Modal mapping of a node to the joint embedding space m The characteristics, i.e., modes m Specific embedding; For the modalities mapped to the joint embedding space n The characteristics, i.e., modes n Specific embedding.

[0066] Based on the correlation matrix, cross-modal feature fusion is performed on the modality-specific embeddings of each node's embedding features:

[0067] in, As a residual connection, it can preserve the original specific embedding. To prevent information loss; The fusion weight matrix is ​​dynamically selected based on the fusion modality type (e.g., text → image mapping selection). ).

[0068] Normalize the specific embeddings after cross-modal fusion:

[0069] Where LayerNorm is layer normalization; TimeNorm is time-axis normalization, which can suppress interference signals outside the maintenance cycle of power equipment; normalization weights... λ The output is learned by the gated loop unit.

[0070] The initial features of each node are updated based on the normalized modality-specific embeddings.

[0071] In one possible implementation, each aggregation process may include: For each relation type r Design a relationship-aware feature aggregation mechanism: For nodes i To obtain its relation type r The set of neighboring nodes ; The importance weights of neighboring nodes are calculated using a relation-specific attention mechanism:

[0072] in, For relationshipr The characteristic transformation matrix, Let be the attention parameter vector, and || denote the concatenation operation.

[0073] Execution of relation-aware feature aggregation:

[0074] in, l Indicates the first l Secondary polymerization process.

[0075] Aggregation results that combine multiple relationship types:

[0076] Among them, AGG can adopt weighted summation, attention weighting, or gating mechanisms.

[0077] In one possible implementation, the fault identification results may include core output items and auxiliary output items; Core output items may include: Failure probability scalar (continuous value): This represents the probability of a failure occurring within a preset time window (e.g., 72 hours). The failure probability can be output using the Sigmoid activation function, and the alarm threshold for the failure probability can be adjusted according to actual needs (e.g., 0.7 as the alarm threshold).

[0078] Multi-level state coding (discrete classification): A three-level state system is adopted (corresponding to DL / T 596-2021 Preventive Testing Procedures for Power Equipment): 0 indicates normal ( ); 1 indicates a warning ( ), it is recommended to strengthen monitoring; 2 indicates high risk ( (This part needs immediate repair.)

[0079] Auxiliary output items may include: Fault type vector: This corresponds to the probabilities of subtypes such as insulation fault, mechanical fault, and overheating fault. Impact Scope Identification: Through heterogeneous graph propagation path analysis, output a list of affected device IDs (e.g., a fault in transformer T1 will affect lines L1-L3). In one possible implementation, based on end-to-end training, pre-trained first gated recurrent units, second gated recurrent units, feature transformation matrices, attention parameter vectors, metric mapping parameters, language models, and convolutional neural networks are obtained. Training set construction methods include: For each historical fault, slide forward to retrieve the data before the fault occurred. Data slices for different time periods are labeled as follows: t-72h to t-24h → Status 1 (Warning) From t-24h to the time of failure → Status 2 (High Risk) If a failure in device A causes device B to trip, a virtual fault label (state 2) is added to B in the heterogeneous graph, and a causal edge "A → [causal edge] → B" is constructed, with propagation weights assigned. (DTWD stands for Dynamic Time Warped Distance).

[0080] End-to-end training methods include: Construct positive and negative sample pairs; positive samples: device node pairs under the same meta-path (e.g., nodes on the path "transformer-oil chromatograph sensor-fault record"); negative samples: randomly replacing faulty nodes in the meta-path (e.g., replacing a faulty transformer with a normal one). The loss function is... Where GAE is the loss function of the graph autoencoder; ORL is the ordinal regression loss, which can guarantee the orderliness of the state levels.

[0081] Specifically, the technical effect of the power equipment fault prediction method based on heterogeneous graphs provided in this embodiment of the invention can be illustrated by the following example: For a 110kV transformer substation in a certain power grid, a heterogeneous graph is constructed, containing 50 transformer nodes, 120 line nodes, and 200 sensor nodes. Edge types include electrical connections (transformer-line), monitoring relationships (sensor-device), and spatial proximity (devices within 100m).

[0082] The collected data includes: (1) Transformer operating parameters: load rate, oil temperature, winding temperature, etc., with a sampling frequency of 15 minutes; (2) Sensor data: partial discharge, vibration signal, ambient temperature and humidity, etc.; (3) Maintenance records: text data such as historical fault reports and maintenance records.

[0083] Model configuration: Heterogeneous graph convolutional layers: 3; Hidden layer dimension: 128; Number of attention heads: 8; Time window: 24 hours (96 time steps); Prediction window: 72 hours in the future.

[0084] The training used data from 18 months of historical data, including 45 real-world fault cases. Test results showed that the accuracy rate of early warnings reached 89.3% 48 hours before a fault occurred, with a false alarm rate controlled below 5.2%, representing a 22% improvement in accuracy compared to traditional threshold-based early warning methods.

[0085] This can also be illustrated through the following example: For urban power distribution networks, a large-scale heterogeneous graph with more than 500 nodes is constructed, and the node types are expanded to include user load nodes, distributed power generation nodes, etc. Special attention is paid to the cascading propagation effect of faults.

[0086] During heterogeneous graph convolution, physically connected edges are assigned higher weights (1.5 times by default) because direct electrical connections have a greater impact on fault propagation. A graph attention pooling layer is also introduced to identify key node sets.

[0087] In actual operation, the system successfully predicted three equipment anomalies that could have triggered cascading failures, and took isolation measures in advance, thus avoiding a large-scale power outage. The predicted scope of the fault impact matched the actual situation with over 85% accuracy.

[0088] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0089] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0090] Figure 2 A schematic diagram of the power equipment fault prediction device based on heterogeneous graphs provided in an embodiment of the present invention is shown. For ease of explanation, only the parts related to the embodiment of the present invention are shown, and are described in detail below: like Figure 2 As shown, the power equipment fault prediction device 2 based on heterogeneous graphs includes: Heterogeneous graph construction module 21 is used to construct a heterogeneous graph of power equipment by taking each power device in the power system as a node and determining the edges of each node based on the relationship between each power device. The multimodal feature extraction module 22 is used to acquire multimodal data of each power device; and to determine the initial features of each node in the heterogeneous graph of the power devices based on the multimodal data. The feature aggregation module 23 is used to determine the aggregated features of each node based on the initial features of each node through at least one aggregation process; wherein, in each aggregation process, the current features of all neighboring nodes of each node are aggregated to update the features of each node. The fault prediction module 24 is used to generate the spatiotemporal representation vector of each power device based on the continuous time series of the aggregated features of each node, and to predict the fault identification result of each device.

[0091] In one possible implementation, feature aggregation module 23 is used to perform the following steps for each node: Get the set of neighbor nodes of the given node under each edge type; We perform weighted aggregation of the current features of all neighbor nodes in each set of neighbor nodes to determine the relation features of the node under each edge type. The node's relationship features across different edge types are merged, and the node's features are updated.

[0092] In one possible implementation, the feature aggregation module 23 is further configured to perform the following steps for each edge type of the node: Based on the feature transformation matrix and attention parameter vector of this edge type, the aggregation weights of each neighbor node of this node under this edge type are determined; wherein, the feature transformation matrix and attention parameter vector are obtained by pre-training optimization; Based on the aggregation weight, the current features of each neighboring node of the node under the edge type are weighted and aggregated to determine the relation features of the node under each edge type.

[0093] In one possible implementation, the multimodal feature extraction module 22 is used to dynamically time-warp the multimodal data of each power device based on the time-varying metric matrix; and extract different modal features of each node from the time-warped multimodal data. Based on the feature projection matrix, the different modal features of each node are mapped to the joint embedding space to obtain the embedding features of each node; A topological consistency constraint is established for the heterogeneous graph of power equipment, and the embedding features of each node are adjusted based on the cross-modal attention mechanism to obtain the initial features of each node.

[0094] In one possible implementation, the heterogeneous graph construction module 21 is used to define the type of each node and the type of connecting edges based on the power equipment type and the relationship between power equipment in the power system, and to determine the adjacency matrix of each edge type to construct a heterogeneous graph of power equipment.

[0095] In one possible implementation, the fault prediction module 24 is used to input the continuous time series of the aggregated features of each node into the pre-trained second gated recurrent unit to obtain the spatiotemporal representation vector of each power device. Based on the spatiotemporal representation vectors of each power device, the fault identification results are predicted using a multilayer perceptron; the fault identification results include the fault probability, fault type, and a list of affected devices.

[0096] Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. For example... Figure 3As shown, the electronic device 3 of this embodiment includes a processor 30 and a memory 31. The memory 31 stores a computer program 32. When the processor 30 executes the computer program 32, it implements the steps in the various method embodiments described above. Alternatively, when the processor 30 executes the computer program 32, it implements the functions of each module / unit in the various device embodiments described above.

[0097] For example, computer program 32 may be divided into one or more modules / units, which are stored in memory 31 and executed by processor 30 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 32 in electronic device 3.

[0098] Electronic device 3 may include, but is not limited to, processor 30 and memory 31. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 3 may also include input / output devices, network access devices, buses, etc.

[0099] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.

[0100] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0101] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for predicting power equipment faults based on heterogeneous graphs, characterized in that, include: Using each power device in the power system as a node, and determining the edges of each node based on the relationships between the power devices, a heterogeneous graph of power devices is constructed. Acquire multimodal data from various power devices; The initial characteristics of each node in the heterogeneous graph of the power equipment are determined based on the multimodal data. Based on the initial features of each node, the aggregated features of each node are determined through at least one aggregation process; wherein, in each aggregation process, the current features of all neighboring nodes of each node are aggregated to update the features of each node. Based on the continuous time series of the aggregated features of each node, a spatiotemporal representation vector of each power device is generated, and the fault identification result of each device is predicted.

2. The heterogeneous graph-based power equipment fault prediction method according to claim 1, characterized in that, The aggregation process in each step aggregates the current features of all neighboring nodes of each node to update the features of each node, including: For each node, perform the following steps: Get the set of neighbor nodes of the given node under each edge type; We perform weighted aggregation of the current features of all neighbor nodes in each set of neighbor nodes to determine the relation features of the node under each edge type. The node's features are updated by integrating the relational features of the node under various edge types.

3. The heterogeneous graph-based power equipment fault prediction method according to claim 2, characterized in that, The step of weighted aggregation of the current features of all neighbor nodes in each neighbor node set to determine the relation features of the node under each edge type includes: For each edge type of this node, perform the following steps: Based on the feature transformation matrix and attention parameter vector of this edge type, the aggregation weights of each neighboring node of this node under this edge type are determined; wherein, the feature transformation matrix and the attention parameter vector are obtained through pre-training optimization; Based on the aggregation weight, the current features of each neighboring node of the node under the edge type are weighted and aggregated to determine the relation features of the node under each edge type. 4.The heterogeneous graph-based power equipment fault prediction method according to claim 1, characterized in that, The process involves acquiring multimodal data from various power devices. The initial characteristics of each node in the heterogeneous graph of the power equipment are determined based on the multimodal data, including: Based on the time-varying metric matrix, dynamic time warping is performed on the multimodal data of each power device; different modal features of each node are extracted from the time-warped multimodal data; Based on the feature projection matrix, the different modal features of each node are mapped to the joint embedding space to obtain the embedding features of each node; A topological consistency constraint is established for the heterogeneous graph of the power equipment, and the embedding features of each node are adjusted based on a cross-modal attention mechanism to obtain the initial features of each node.

5. The heterogeneous graph-based power equipment fault prediction method according to claim 4, characterized in that, The time-varying metric matrix is ​​obtained from the metric mapping parameters and the equipment operating condition characteristics; wherein, the metric mapping parameters are parameters obtained by pre-training and optimization through historical data, and the equipment operating condition characteristics are output by the first gated loop unit based on the real-time load, real-time ambient temperature and data acquisition time interval of the equipment. 6.The heterogeneous graph-based power equipment fault prediction method according to claim 1, characterized in that, The process of constructing a heterogeneous graph of power devices, using each power device in the power system as a node and determining the edges of each node based on the relationships between the power devices, includes: Based on the power equipment types and relationships between power equipment in the power system, the types of each node and the types of connecting edges are defined, and the adjacency matrix of each edge type is determined to construct the heterogeneous graph of the power equipment. 7.The heterogeneous graph-based power equipment fault prediction method according to claim 1, characterized in that, The step of generating a spatiotemporal representation vector for each power device based on the continuous time series of the aggregated features of each node, and predicting the fault identification result for each device, includes: The continuous time series of the aggregated features of each node are input into the pre-trained second gated recurrent unit to obtain the spatiotemporal representation vector of each power device; Based on the spatiotemporal representation vectors of each power device, a multilayer perceptron is used to predict fault identification results; wherein, the fault identification results include fault probability, fault type, and a list of affected devices.

8. The method for predicting power equipment faults based on heterogeneous graphs according to any one of claims 1 to 7, characterized in that, The method further includes: When the first device fails and causes the second device to fail, in the heterogeneous graph of the power equipment, a causal edge is added between the first node corresponding to the first device and the second node corresponding to the second device, and the aggregation weight of the second node under the causal edge of the first node is adjusted.

9. A power equipment fault prediction device based on heterogeneous graphs, characterized in that, include: The heterogeneous graph construction module is used to construct a heterogeneous graph of power equipment, taking each power device in the power system as a node and determining the edges of each node based on the relationship between the power devices. A multimodal feature extraction module is used to acquire multimodal data of various power devices; and to determine the initial features of each node in the heterogeneous graph of the power devices based on the multimodal data. The feature aggregation module is used to determine the aggregated features of each node based on the initial features of each node through at least one aggregation process; wherein, in each aggregation process, the current features of all neighboring nodes of each node are aggregated to update the features of each node. The fault prediction module is used to generate a spatiotemporal representation vector of each power device based on the continuous time series of the aggregated features of each node, and to predict the fault identification result of each device.

10. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 8.

Images