High-voltage cable multi-mode fault diagnosis method, system and device based on space-time heterogeneous graph neural network and medium

Abstract

The invention relates to the technical field of high-voltage cable intelligent monitoring, and discloses a high-voltage cable multi-mode fault diagnosis method, system and device based on a space-time heterogeneous graph neural network and a medium, and the method comprises the steps: constructing a space-time heterogeneous graph fusing the physical connection and electromagnetic-thermal field coupling relation; respectively extracting physical connection features and electromagnetic coupling features by adopting dual-channel convolution; modeling a time dependency using a timing gating unit; introducing heat conduction and Joule heat effect constraints into the loss function; and fault classification and three-dimensional positioning are realized based on an attention mechanism. According to the invention, by fusing multi-source monitoring data and physical rules, early warning, accurate positioning and dynamic risk assessment of cable faults are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for high-voltage cables, and in particular to a method, system, equipment, and medium for multimodal fault diagnosis of high-voltage cables based on spatiotemporal heterogeneous graph neural networks. Background Technology

[0002] With the rapid development of the national economy and the accelerated pace of urbanization, electricity demand has shown a continuous upward trend, placing unprecedented operational pressure on power equipment in the power transmission and distribution system. Under long-term operating conditions, the continuous heat dissipation of power equipment, especially high-voltage cables, has become increasingly prominent, leading to frequent cable fires. Statistics show that in the past five years, the average annual growth rate of fires caused by cable faults in my country's power system has reached 12.3%, resulting in alarming casualties and property losses. In 2024, a cable fault in the power supply system of a megacity's subway system caused a 14-hour shutdown, resulting in direct economic losses exceeding 23 million yuan and extremely negative social impact.

[0003] The essence of cable faults is a progressive degradation process: under abnormal operating conditions such as current fluctuations (defined by IEC 61000-3-7: fluctuation limit of 4% when voltage changes twice per hour) or high current overloads (specified by IEC 60038: voltage deviation limit of 10% for 10kV systems), the cable insulation layer undergoes irreversible damage. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a multimodal fault diagnosis method for high-voltage cables based on spatiotemporal heterogeneous graph neural networks, which can solve the signal attenuation problem of traditional monitoring technology in long-distance cable scenarios.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a multimodal fault diagnosis method for high-voltage cables based on a spatiotemporal heterogeneous graph neural network, comprising: constructing a spatiotemporal heterogeneous graph containing physical connectivity and electromagnetic-thermal coupling relationships to obtain graph structure data characterizing the multidimensional states of the cable system; processing physical edges and virtual edges in the heterogeneous graph separately using a dual-path convolution operator to obtain decoupled physical connectivity features and electromagnetic coupling features; performing temporal modeling on the fused features using gated temporal units to obtain a node state sequence containing time dependencies; introducing physical constraint terms based on heat conduction and Joule heating effects into the loss function to optimize the model parameters and obtain a prediction model that conforms to physical laws; and obtaining fault type classification and three-dimensional spatial localization results through node relationship reasoning and coordinate mapping based on an attention mechanism.

[0007] As a preferred embodiment of the high-voltage cable multimodal fault diagnosis method based on spatiotemporal heterogeneous graph neural network described in this invention, the spatiotemporal heterogeneous graph includes: abstracting cable segments into nodes and integrating multimodal sensor data to form node feature vectors; By defining physical edges based on cable topological impedance and virtual edges based on electromagnetic field coupling, a structured graph representation that simultaneously includes physical connections and field interactions is obtained.

[0008] As a preferred embodiment of the high-voltage cable multimodal fault diagnosis method based on spatiotemporal heterogeneous graph neural network described in this invention, the physical edge includes a physical edge weight matrix that reflects the strength of direct electrical connection between conductors by calculating the equivalent impedance between cable segment nodes and taking the reciprocal as the edge weight.

[0009] As a preferred embodiment of the high-voltage cable multimodal fault diagnosis method based on spatiotemporal heterogeneous graph neural network described in this invention, the virtual edge includes a virtual edge weight matrix that quantifies electromagnetic-thermal field interaction by combining the attenuation coefficient determined by the electromagnetic parameters of the cable material with the periodic change of the power frequency current to calculate the coupling weight between nodes.

[0010] As a preferred embodiment of the high-voltage cable multimodal fault diagnosis method based on spatiotemporal heterogeneous graph neural network described in this invention, the dual-path convolution operator includes extracting physical connection features and electromagnetic-thermal coupling features respectively, and proposing a heterogeneous graph convolution operator HetConv. Physical edge path calculation: in, Let i be the physical characteristic representation of node i. This is the physical edge weight matrix. Let be the input feature vector of node j. The learnable weight matrix for the physical pathway. Let i be the set of physical neighbors of node i. For activation functions; Virtual edge path calculation: in, Let i be the virtual feature representation of node i. Let i be the set of virtual neighbors of node i. For virtual edge weights, This is the learnable weight matrix for the virtual pathway; Final Feature Fusion Represented as: Among them, the fusion weight It can be trained or modulated by the intensity of the physical field.

[0011] As a preferred embodiment of the high-voltage cable multimodal fault diagnosis method based on spatiotemporal heterogeneous graph neural network described in this invention, the step of performing time-series modeling includes introducing a GTU gating mechanism to achieve time-series modeling, the structure of which is as follows: Let the current input be The historical status is The updated formula is: Gating coefficient calculation: Candidate state update: Final status update: in, To update the door, , , These are the input weight matrices for the update gate, reset gate, and candidate state, respectively. These are the cyclic weight matrices for the update gate, reset gate, and candidate state, respectively. To reset the door, Candidate state The hidden state at the current time step. It is the hyperbolic tangent function. This is element-wise multiplication.

[0012] As a preferred embodiment of the high-voltage cable multimodal fault diagnosis method based on spatiotemporal heterogeneous graph neural network described in this invention, the physical law constraint term includes the introduction of electro-thermal coupling constraint and the construction of physical regularization terms based on heat conduction and Joule heating effect. Fourier heat conduction equation: in, For density, For specific heat capacity, Thermal conductivity, For the heat source term, T represents the temperature field. Represents the heat conduction term; The Joule heat source term Q is expressed as: in, The conductivity of a conductor. Let be the potential function. For electric field strength, The potential gradient; Discrete-form embedding loss function, and discretize the above partial differential equation using the finite difference method: in, The physical loss function is in discrete form. The total number of nodes. Let i be the temperature at time t+1. Let i be the temperature at time t. Let be the potential gradient at node i. For time step, The spatial step size; Final total loss function for: in, To predict mission losses, This is the physical regularization coefficient.

[0013] This invention provides a high-voltage cable multimodal fault diagnosis system based on a spatiotemporal heterogeneous graph neural network.

[0014] As a preferred embodiment of the high-voltage cable multimodal fault diagnosis system based on spatiotemporal heterogeneous graph neural network described in this invention, it includes: deploying a distributed sensor network, relying on the actual cable topology, installing temperature sensors, current detection modules, partial discharge detection devices, and micro vibration accelerometers at key node locations, and constructing a multimodal data channel; By using a high-frequency sampling and time synchronization module, the above four types of signals are uniformly acquired and time-series calibrated to construct an observation vector sequence with accurate spatial location. Each node of the cable segment at each observation time point is regarded as a node in a graph neural network. The actual physical edges are defined by the connection impedance, and the virtual edges are defined by the electromagnetic coupling function. The heterogeneous graph structure is automatically generated. Each node contains four-dimensional state information and is input into the feature extraction module.

[0015] In the dual-path HetConv convolution module, the physical side path highlights the dominant role of conductor connections in signal propagation, while the virtual side path simulates the complex electromagnetic-thermal-mechanical coupling mechanism inside the cable. The fusion characteristics of the two types of paths are weighted through an attention mechanism. By using gated timing units (GTUs), the temperature rise trend or discharge signal growth trajectory across hours or even days can be captured to form a significant predictive signal. The system integrates feature inputs into the fault inference engine, models the fault propagation path through a graph attention mechanism, and outputs node risk scores, spatial location results, and fault type labels.

[0016] The present invention provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of a high-voltage cable multimodal fault diagnosis method based on a spatiotemporal heterogeneous graph neural network.

[0017] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of a high-voltage cable multimodal fault diagnosis method based on a spatiotemporal heterogeneous graph neural network are implemented.

[0018] The beneficial effects of this invention are as follows: It pioneers a heterogeneous graph topology structure with parallel physical and virtual edges, uniformly representing conductor connections and electromagnetic coupling effects; it proposes a feature decoupling mechanism to achieve separate processing of physical and electromagnetic signals, avoiding multi-source interference; it designs a GTU timing unit to replace the traditional LSTM, reducing the long-distance gradient decay rate from 95% to below 5%; and it develops gradient mask adaptive technology to dynamically balance the resource allocation for fault location and classification tasks, providing a new generation of diagnostic paradigms for smart grid construction. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of a high-voltage cable multimodal fault diagnosis method based on a spatiotemporal heterogeneous graph neural network, provided as an embodiment of the present invention. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0022] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a method for multimodal fault diagnosis of high-voltage cables based on spatiotemporal heterogeneous graph neural networks, including: S1: Construct a spatiotemporal heterogeneous graph that includes physical connections and electromagnetic-thermal field coupling to obtain graph structure data characterizing the multidimensional state of the cable system.

[0023] S2: The physical edges and virtual edges in the heterogeneous graph are processed separately by the dual-path convolution operator to obtain the decoupled physical connection features and electromagnetic coupling features.

[0024] S3: The fused features are modeled temporally using gated temporal units to obtain a sequence of node states containing time dependencies.

[0025] S4: Introduce physical constraint terms based on heat conduction and Joule heating effect into the loss function to optimize the model parameters and obtain a prediction model that conforms to physical laws.

[0026] S5: Through node relationship reasoning and coordinate mapping based on attention mechanism, the fault type classification and three-dimensional spatial localization results are obtained.

[0027] It should be noted that the core processing flow begins with dual-path convolutional parsing of heterogeneous graph data: the physical edge convolutional path focuses on extracting direct conductor features, revealing key information about the current conduction path; the virtual edge convolutional path captures the implicit correlation between electromagnetic induction and heat diffusion. The two feature paths are fused through an adaptive weighting mechanism to form a comprehensive feature matrix. The fused features are input into a gated temporal unit (GTU), which uses a dynamic gating mechanism to balance historical states with the current input, completely solving the signal attenuation problem in kilometer-level cable scenarios and maintaining a signal fidelity of over 95% within a length of 1500 meters. The time-series processed data enters the dynamic relational inference engine, which analyzes the correlation strength between nodes based on the graph attention algorithm, generates a fault propagation heatmap and locates the source, and finally outputs three-dimensional fault coordinates, fault type classification (6 categories including discharge / overheating / deterioration), and a risk score of 0-100. The physical law embedding mechanism is the theoretical cornerstone of this scheme: Fourier's law of heat conduction is embedded in the loss function to ensure that temperature prediction conforms to the principle of energy conservation; an electromagnetic verification layer is constructed based on Maxwell's equations to constrain the electromagnetic rationality of the prediction results. Taking the fault diagnosis of a 220kV submarine cable joint as an example, after the system detects an abnormal temperature rise (25.3℃→30.9℃) and a surge in leakage current (0.21mA→0.68mA) within 5 minutes, a heterogeneous graph containing 32 nodes, 48 ​​physical edges, and 217 virtual edges is constructed. Through dynamic relationship reasoning, node #17 is identified as the source of the fault (confidence level 89%), and the output diagnosis result is poor joint crimping (confidence level 98.2%), with three-dimensional coordinate positioning (128.7m, -35.2m, 22.3m) (error 0.3 meters) and a risk score of 86, guiding maintenance personnel to accurately handle the fault.

[0028] Example 2 is an embodiment of the present invention. Based on the above embodiment, a method for multimodal fault diagnosis of high-voltage cables based on spatiotemporal heterogeneous graph neural networks is provided.

[0029] Furthermore, in this embodiment, step S1 constructs a spatiotemporal heterogeneous graph containing physical connection relationships and electromagnetic-thermal field coupling relationships to obtain graph structure data characterizing the multidimensional state of the cable system. Specific steps include S101-S103: S101: Define the feature vector of the cable segment node: in, Let be the temperature at node i. Leakage current intensity, Represents the partial discharge energy integral. It represents the vibrational energy entropy.

[0030] S102: Establish the physical edge weight matrix : in, The equivalent impedance between node i and node j is obtained from the actual cable wiring structure and calculated through the impedance matrix, satisfying symmetry. .

[0031] S103: Establish the virtual edge weight matrix : Where t is time and π is the mathematical constant pi. The electromagnetic attenuation coefficient, The power frequency current frequency is... The distance between nodes. As the normalization factor, Permeability, Where is the dielectric constant. Conductivity of a conductor.

[0032] Furthermore, in this embodiment, step S2 processes the physical edges and virtual edges in the heterogeneous graph using a dual-path convolution operator to obtain the decoupled physical connection features and electromagnetic coupling features. Specific steps include S201-S203: S201: The multimodal data of the high-voltage cable system comes from devices such as temperature sensors, partial discharge detectors, vibration sensors, and current transformers. The data collected at each moment constitutes a node in the graph. To describe the physical connections and electromagnetic-thermal field interactions in the high-voltage cable system, a two-path graph structure is established: physical edges and virtual edges.

[0033] To extract physical connectivity features and electromagnetic-thermal coupling features respectively, a heterogeneous graph convolution operator, HetConv, is proposed.

[0034] S202: Physical edge path calculation: in, Let i be the physical characteristic representation of node i. Let be the input feature vector of node j. The learnable weight matrix for the physical pathway. Let i be the set of physical neighbors of node i. For activation functions; Virtual edge path calculation: in, Let i be the virtual feature representation of node i. Let i be the set of virtual neighbors of node i. This is the learnable weight matrix for the virtual pathway; Final Feature Fusion Represented as: in, For activation functions, fused weights It can be trained or modulated by the intensity of the physical field.

[0035] S203: HetConv Feature Decoupling Calculation: in, Virtual feature representation of node i, Describes the set of virtual edge neighbors of node i. For node conductivity, Activation function Let be the spatial coordinate vectors of nodes i and j, and k be the summation index variable used to traverse all virtual edge neighbor nodes of node i.

[0036] In an optional embodiment, feature fusion can also be achieved through channel attention-based weighted fusion. Specifically, the node feature vectors calculated by the physical side path are concatenated with the node feature vectors calculated by the virtual side path to form a longer joint feature vector. This joint feature vector is then input into the attention network. The first output weight value is multiplied by the physical feature vector to obtain the weighted physical feature; the second weight value is multiplied by the virtual feature vector to obtain the weighted virtual feature. Finally, the two weighted feature vectors are added together to obtain the final fused feature.

[0037] In another alternative embodiment, feature fusion can also be achieved through gating-based fusion. Specifically, physical edge features and virtual edge features are added together and a gating signal is generated using a sigmoid activation function. This gating signal is used to control the information flow from virtual edge features to physical edge features. The filtered virtual feature information is then added to the original physical features to achieve fusion.

[0038] Furthermore, in this embodiment, step S3 performs temporal modeling on the fused features using a gated temporal unit to obtain a node state sequence containing temporal dependencies, specifically including: To achieve timing modeling, a GTU (Gated Temporal Unit) gating mechanism is introduced, with the following structure: Let the current input be The historical status is The updated formula is: Gating coefficient calculation: Candidate state update: Final status update: in, To update the door, , , These are the input weight matrices for the update gate, reset gate, and candidate state, respectively. These are the cyclic weight matrices for the update gate, reset gate, and candidate state, respectively. To reset the door, Candidate state The hidden state at the current time step. It is the hyperbolic tangent function. This is element-wise multiplication.

[0039] The GTU unit adaptively fuses historical states with current inputs through a gating mechanism. Compared to traditional LSTM, it has a shorter computation path, reduces gradient decay, and can still maintain high-quality temporal modeling results in a 1500-meter-long cable system.

[0040] In an alternative embodiment, state updates can also be implemented through an update strategy based on direct replacement of candidate states. Specifically, the final state is determined by the update gate, which approximates whether to "completely retain the old state" or "completely adopt the candidate state." The output of the update gate is interpreted as the probability of selecting a candidate state. However, in the deterministic model, the saturation property of the Sigmoid function is used to approximate the effect of "partial retention and partial replacement," which is designed to simplify the update logic.

[0041] In another alternative embodiment, state updates can also be achieved through an update strategy based on a nonlinear transformation of historical states. Specifically, the old state is first subjected to a nonlinear transformation to obtain a "refined historical memory." The final hidden state is controlled by an update gate, which regulates the fusion ratio between this "refined historical memory" and the "candidate state."

[0042] Furthermore, in this embodiment, step S4 introduces physical constraint terms based on heat conduction and Joule heating effect into the loss function to optimize the model parameters and obtain a prediction model that conforms to physical laws. Specific steps include S401-S403: S401: In order to improve the reliability and physical consistency of model predictions, this invention introduces an electro-thermal coupling constraint into the loss function and constructs a physical regularization term based on thermal conduction and Joule heating effect.

[0043] Fourier heat conduction equation: in, For density, For specific heat capacity, Thermal conductivity, For the heat source term, T is the temperature. Represents the heat conduction term; S402: The Joule heat source term Q is expressed as: in, The conductivity of a conductor. Let be the potential function. For electric field strength, The potential gradient; S403: Discrete-form embedding loss function, using the finite difference method to discretize the above partial differential equation: in, The physical loss function is in discrete form. The total number of nodes. Let i be the temperature at time t+1. Let i be the temperature at time t. Let be the potential gradient at node i. For time step, The spatial step size; Final total loss function for: in, To predict mission losses, This is the physical regularization coefficient.

[0044] Furthermore, in this embodiment of the application, step S5 obtains the fault type classification and three-dimensional spatial localization results through node relationship reasoning and coordinate mapping based on the attention mechanism. Specific steps include S501-S502: S501: This invention uses a graph attention mechanism (GAT) to model the state transmission relationship between nodes, thereby realizing the reasoning of fault propagation paths and source location.

[0045] Node feature attention weight : Node coordinate mapping function: in, , Let be the feature vectors of node i and node j. The MLP is a three-layer fully connected network that outputs three-dimensional coordinates (x, y, z) to represent the spatial location of the fault point. This is the transpose of the parameter vector for the attention mechanism. This is the activation function, commonly used in attention mechanisms to avoid gradient vanishing. It is an exponential function. Let be the predicted 3D coordinates of node i, representing the spatial location of the fault point. It is a multilayer perceptron.

[0046] S502: Risk Scoring Function Design: in, To score risk, Rated temperature For the maximum allowable temperature, The maximum permissible leakage current threshold, , , The weighted coefficient is used to dynamically assess the risk level of the current node (0-100 points).

[0047] In an optional embodiment, risk scoring can also be implemented using a piecewise linear function. Specifically, multiple risk thresholds are set for each feature. For each feature, a preset base score is assigned based on the interval in which its value falls. The piecewise scoring results for each feature are weighted and summed using preset weighting coefficients to obtain a preliminary total risk score. The preliminary total score is then mapped to a standard scoring range of 0-100 using a linear scaling function, serving as the final risk score.

[0048] In another alternative embodiment, risk scoring can also be implemented using a small neural network. Specifically, the feature vector is input into a shallow, small, fully connected neural network. This network contains one or two hidden layers and uses activation functions such as ReLU. The output layer of this neural network consists of a single neuron, followed by a sigmoid activation function that compresses the output value to between 0 and 1. Multiplying the sigmoid output value by 100 yields the final risk score from 0 to 100.

[0049] Example 3 is the third embodiment of the present invention, which differs from the previous two embodiments in that: This embodiment also provides a high-voltage cable multimodal fault diagnosis system based on spatiotemporal heterogeneous graph neural networks, including: Deploy a distributed sensor network, relying on the actual cable topology, and install temperature sensors (such as fiber optic DTS), current detection modules, partial discharge detection devices, and miniature vibration accelerometers at key node locations to build a multimodal data channel.

[0050] The high-frequency sampling and time synchronization module performs unified acquisition and timing calibration of the above four types of signals, ensuring the construction of a highly consistent and spatially accurate observation vector sequence within a millisecond timescale. The sampling period can be configured from 1s to 30s, and can be flexibly adjusted according to the rate of change of the cable operating environment.

[0051] Each node of the cable segment at each observation time point is treated as a node in a graph neural network. Actual physical edges are defined by connection impedance, and virtual edges are defined by electromagnetic coupling functions, thus automatically generating the heterogeneous graph structure. Each node contains four-dimensional state information, which is then input into the feature extraction module.

[0052] In the dual-path HetConv convolution module, the physical sidepaths highlight the dominant role of conductor connections in signal propagation, while the virtual sidepaths simulate the complex electromagnetic-thermal-mechanical coupling mechanisms within the cable, such as harmonic waveform distortion accompanying partial discharge or temperature rise. The fusion features of the two types of paths are further weighted through an attention mechanism to ensure that multimodal information participates in decision-making in an optimal manner.

[0053] By using gated timing units (GTUs), the system can capture temperature rise trends or discharge signal growth trajectories across hours or even days, thus generating significant predictive signals before the fault is fully exposed.

[0054] The system integrates feature-based input into a fault inference engine, models the fault propagation path using a graph attention mechanism, and outputs node risk scores, spatial location results, and fault type labels (6 categories including overheating, partial discharge, abnormal mechanical vibration, and current surge). The entire process, from data acquisition to fault output, has a latency controlled within 5-10 seconds, meeting the rapid diagnostic needs of real-world power grids.

[0055] This system adopts a layered architecture design: Edge acquisition layer: The front-end data access and preliminary processing are completed by a low-power microcontroller (such as STM32H7) or an industrial-grade acquisition terminal. It has the functions of resuming interrupted transmission and local caching, and supports wireless or fiber optic backhaul.

[0056] Computational Analysis Layer: Deployed in data centers or edge servers, this layer integrates the STHGNN model described in this invention and is responsible for heterogeneous graph generation, feature extraction, and fault reasoning. The system supports modular updates and can perform hot updates and incremental training based on new fault modes.

[0057] Display and control layer: Through a web visualization platform (supporting B / S architecture), dynamically display the cable operation status heatmap, risk score distribution and alarm records, and realize the visualization of operation and maintenance strategies and data closure.

[0058] This embodiment also provides an electronic device, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the high-voltage cable multimodal fault diagnosis method based on spatiotemporal heterogeneous graph neural network proposed in the above embodiment.

[0059] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the high-voltage cable multimodal fault diagnosis method based on spatiotemporal heterogeneous graph neural network proposed in the above embodiment.

[0060] The storage medium proposed in this embodiment and the method for multimodal fault diagnosis of high-voltage cables based on spatiotemporal heterogeneous graph neural networks proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0061] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0062] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for multimodal fault diagnosis of high-voltage cables based on spatiotemporal heterogeneous graph neural networks, characterized in that: include, A spatiotemporal heterogeneous graph containing physical connectivity and electromagnetic-thermal field coupling is constructed to obtain graph structure data characterizing the multidimensional state of the cable system; By using a dual-path convolution operator to process the physical edges and virtual edges in the heterogeneous graph, the decoupled physical connection features and electromagnetic coupling features are obtained. Temporal modeling of the fused features is performed using gated temporal units to obtain a node state sequence containing time dependencies; By introducing physical constraints based on heat conduction and Joule heating into the loss function, the model parameters are optimized to obtain a prediction model that conforms to physical laws. By using node relationship reasoning and coordinate mapping based on attention mechanism, the results of fault type classification and three-dimensional spatial localization are obtained.

2. The high-voltage cable multimodal fault diagnosis method based on spatiotemporal heterogeneous graph neural network as described in claim 1, characterized in that: The spatiotemporal heterogeneous graph includes abstracting cable segments into nodes and integrating multimodal sensor data to form node feature vectors; By defining physical edges based on cable topological impedance and virtual edges based on electromagnetic field coupling, a structured graph representation that simultaneously includes physical connections and field interactions is obtained.

3. The high-voltage cable multimodal fault diagnosis method based on spatiotemporal heterogeneous graph neural network as described in claim 2, characterized in that: The physical edge includes a physical edge weight matrix that reflects the strength of the direct electrical connection between conductors, obtained by calculating the equivalent impedance between cable segment nodes and taking the reciprocal as the edge weight.

4. The high-voltage cable multimodal fault diagnosis method based on spatiotemporal heterogeneous graph neural network as described in claim 3, characterized in that: The virtual edge includes a virtual edge weight matrix that quantifies electromagnetic-thermal field interaction by combining the attenuation coefficient determined by the electromagnetic parameters of the cable material with the periodic variation of the power frequency current to calculate the coupling weight between nodes.

5. The high-voltage cable multimodal fault diagnosis method based on spatiotemporal heterogeneous graph neural network as described in claim 4, characterized in that: The dual-path convolution operator includes extracting physical connectivity features and electromagnetic-thermal coupling features respectively, and proposing a heterogeneous graph convolution operator HetConv; Physical edge path calculation: in, Let i be the physical characteristic representation of node i. This is the physical edge weight matrix. Let be the input feature vector of node j. The learnable weight matrix for the physical pathway. Let i be the set of physical neighbors of node i. For activation functions; Virtual edge path calculation: in, Let i be the virtual feature representation of node i. Let i be the set of virtual neighbors of node i. For virtual edge weights, This is the learnable weight matrix for the virtual pathway; Final Feature Fusion Represented as: Among them, the fusion weight It can be trained or modulated by the intensity of the physical field.

6. The high-voltage cable multimodal fault diagnosis method based on spatiotemporal heterogeneous graph neural network as described in claim 5, characterized in that: The timing modeling process includes introducing a GTU gating mechanism, the structure of which is as follows: Let the current input be The historical status is The updated formula is: Gating coefficient calculation: Candidate state update: Final status update: in, To update the door, , , These are the input weight matrices for the update gate, reset gate, and candidate state, respectively. These are the cyclic weight matrices for the update gate, reset gate, and candidate state, respectively. To reset the door, Candidate state The hidden state at the current time step. It is the hyperbolic tangent function. This is element-wise multiplication.

7. The high-voltage cable multimodal fault diagnosis method based on spatiotemporal heterogeneous graph neural network as described in claim 6, characterized in that: The physical law constraint terms include the introduction of electro-thermal coupling constraints and the construction of physical regularization terms based on heat conduction and Joule heating effect; Fourier heat conduction equation: in, For density, For specific heat capacity, Thermal conductivity, For the heat source term, T represents the temperature field. Represents the heat conduction term; The Joule heat source term Q is expressed as: in, The conductivity of a conductor. Let be the potential function. For electric field strength, The potential gradient; Discrete-form embedding loss function, and discretize the above partial differential equation using the finite difference method: in, The physical loss function is in discrete form. The total number of nodes. Let i be the temperature at time t+1. Let i be the temperature at time t. Let be the potential gradient at node i. For time step, The spatial step size; Final total loss function for: in, To predict mission losses, This is the physical regularization coefficient.

8. A high-voltage cable multimodal fault diagnosis system based on a spatiotemporal heterogeneous graph neural network, employing the high-voltage cable multimodal fault diagnosis method based on a spatiotemporal heterogeneous graph neural network as described in any one of claims 1 to 7, characterized in that, include: Deploy a distributed sensor network, relying on the actual cable topology, and install temperature sensors, current detection modules, partial discharge detection devices, and miniature vibration accelerometers at key node locations to build a multimodal data channel; By using a high-frequency sampling and time synchronization module, the above four types of signals are uniformly acquired and time-series calibrated to construct an observation vector sequence with accurate spatial location. Each node of the cable segment at each observation time point is regarded as a node in a graph neural network. The actual physical edges are defined by the connection impedance, and the virtual edges are defined by the electromagnetic coupling function. The heterogeneous graph structure is automatically generated. Each node contains four-dimensional state information and is input into the feature extraction module. In the dual-path HetConv convolution module, the physical side path highlights the dominant role of conductor connections in signal propagation, while the virtual side path simulates the complex electromagnetic-thermal-mechanical coupling mechanism inside the cable. The fusion characteristics of the two types of paths are weighted through an attention mechanism. By using gated timing units (GTUs), the temperature rise trend or discharge signal growth trajectory across hours or even days can be captured to form a significant predictive signal. The system integrates feature inputs into the fault inference engine, models the fault propagation path through a graph attention mechanism, and outputs node risk scores, spatial location results, and fault type labels.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the high-voltage cable multimodal fault diagnosis method based on spatiotemporal heterogeneous graph neural network as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the high-voltage cable multimodal fault diagnosis method based on spatiotemporal heterogeneous graph neural network as described in any one of claims 1 to 7.

Images