Electric power discharge detection method and system based on artificial intelligence

Abstract

The invention discloses a power discharge detection method and system based on artificial intelligence, and relates to the technical field of power equipment safety monitoring, and the method comprises the steps: collecting multi-modal data, and executing adaptive decomposition to extract a discharge characteristic signal; constructing a depth map joint learning model, converting the discharge feature signal into a map data structure, and generating a candidate feature set; based on the candidate feature set, training a discharge mode recognizer by adopting a comparative learning framework, establishing a mapping relationship between discharge features and discharge types and discharge positions, and completing recognition and positioning of a single discharge source; distinguishing and positioning a plurality of discharge sources with overlapped time domains by using node embedding and community detection of the graph neural network; and outputting a discharge risk assessment result in combination with the dynamic evolution of the discharge characteristics. According to the method, the accuracy and the signal-to-noise ratio of discharge feature extraction are remarkably improved through a multi-modal data adaptive decomposition technology, and the modeling capability of discharge spatial-temporal features is remarkably enhanced through a depth map combined learning framework.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment safety monitoring, and particularly to an artificial intelligence-based power discharge detection method and system. Background Art

[0002] During the operation of power equipment, power discharge is a key factor leading to equipment failures and safety hazards. Traditional power discharge detection methods mainly rely on manual experience and single-modal detection means, and have the following technical limitations: First, the signal processing ability of traditional detection methods is limited, making it difficult to accurately extract discharge characteristics from complex electromagnetic environments, with a low signal-to-noise ratio and a high misjudgment rate. The multi-modal data fails to be effectively fused, resulting in insufficient accuracy in extracting discharge characteristics. Second, existing technologies are difficult to capture the complex spatio-temporal characteristics of power discharge. Traditional methods usually can only perform static discharge type identification and cannot achieve precise positioning of the discharge source and dynamic evolution analysis. Moreover, for multiple overlapping discharge sources in the time domain, traditional technologies can hardly effectively distinguish and reconstruct them, restricting the intelligent level of power equipment fault diagnosis. Summary of the Invention

[0003] The present invention provides an artificial intelligence-based power discharge detection method and system for solving the technical problems of inaccurate signal extraction, insufficient feature learning, and inaccurate discharge source identification in power equipment discharge detection.

[0004] In view of this, in the first aspect of the present invention, an artificial intelligence-based power discharge detection method is provided, including: collecting multi-modal data of power equipment and performing adaptive decomposition to extract discharge feature signals; Constructing a deep graph joint learning model, converting the discharge feature signals into a graph data structure, and generating a candidate feature set through a multi-scale spatio-temporal graph convolutional network; Based on the candidate feature set, training a discharge pattern recognizer using a contrastive learning framework, establishing a mapping relationship between discharge features and discharge types and discharge positions, and completing the identification and positioning of a single discharge source; Using node embedding and community detection of graph neural networks to distinguish and locate multiple overlapping discharge sources in the time domain; According to the discharge type and discharge position, combining the dynamic evolution of discharge features, outputting a discharge risk assessment result.

[0005] Optionally, performing adaptive decomposition to extract discharge feature signals includes: Constructing a multi-scale analysis framework based on discrete wavelet transform and performing preliminary decomposition on the multi-modal data; Using a matrix decomposition algorithm to separate the decomposed sub-band signals into background noise components and potential discharge components; Constructing a discharge feature enhancement model to process the potential discharge components and obtaining a time-frequency domain feature representation; Design the weights for multi-modal feature fusion, perform weighted combination on the time-frequency domain feature representation to form an enhanced discharge feature signal.

[0006] Optionally, constructing a deep graph joint learning model includes: In the deep graph joint learning model, convert the enhanced discharge feature signal into a graph data structure and establish the connection weights between nodes; Apply the self-attention mechanism to the graph data structure, calculate the correlation between different modal features, and adjust the node connection weights; Perform dual-path encoding on the graph data structure. The first encoder processes the signal attribute features of nodes, and the second encoder processes the spatio-temporal position relationship of nodes. Integrate the information of the two paths through a fusion module to generate a unified node representation; Input the unified node representation into a multi-scale spatio-temporal graph convolutional network for processing to generate a preliminary candidate feature set; where the deformable convolution is used in the time dimension to capture the time dynamic features, and the spectral domain and spatial domain graph convolutions are combined in the spatial dimension to extract the spatial topological features; Perform feature clustering and classification on the preliminary candidate feature set to identify and separate the single discharge source candidate feature subset and the multi-discharge source overlapping candidate feature subset.

[0007] Optionally, training the discharge pattern recognizer using a contrastive learning framework includes: Use the single discharge source candidate feature subset as the input feature for contrastive learning to construct a multi-level contrast sample library; Construct a two-branch contrastive learning network. The first branch processes discharge type recognition, and the second branch processes discharge position localization. The two branches share the underlying feature extraction network; Extract the time dynamic features and spatial topological features from the single discharge source candidate feature subset, input the time dynamic features into the first branch, and input the spatial topological features into the second branch; Calculate the contrastive loss for the two branches respectively. The first branch calculates the type recognition loss, and the second branch calculates the position localization loss, and combine the two weighted to form a total loss function; Through joint optimization of the total loss function, complete the synchronous training of discharge type recognition and position localization to obtain the network parameters of the discharge pattern recognizer; Initialize the discharge pattern recognizer with the network parameters. The discharge pattern recognizer receives the single discharge source candidate feature subset as the input and outputs the discharge type probability distribution and the discharge position coordinates.

[0008] Optionally, using the node embedding and community detection of the graph neural network to distinguish and locate multiple overlapping discharge sources in the time domain includes: Construct a dynamic graph network according to the multi-discharge source overlapping candidate feature subset, and perform node embedding on the dynamic graph network using the graph neural network; Design a multi-scale community detection algorithm based on spectral clustering, cluster nodes in the embedded space, and identify clusters of power source nodes with overlapping time domains; Through a feature decoupling model, separate and reconstruct the spatio-temporal features of each power source to determine the spatio-temporal boundaries of the power source; Construct a multi-power source feature mapping model to obtain the probability distribution of discharge types and the coordinate of discharge positions for each power source; Output the reconstruction results of multiple power sources, including the number of power sources, the spatio-temporal boundaries, position coordinates, and discharge types of each power source.

[0009] Optionally, the output of the discharge risk assessment results includes: Construct a discharge risk assessment model, use the discharge type and discharge position as basic parameters, and establish a parameter mapping mechanism for risk assessment; Set the basic risk levels according to different discharge types, and adjust the risk weights in combination with the sensitivity of the discharge position to form a differentiated risk assessment strategy; Based on the dynamic evolution of discharge characteristics, dynamically adjust the risk levels of discharge types to quantify the potential equipment damage risks; Establish an association model between the discharge position and the risk diffusion path to evaluate the impact degree of discharges at different positions on the integrity of power equipment; Generate a discharge risk assessment report to comprehensively present the risk analysis results.

[0010] Optionally, the multi-modal data includes electrical parameter data, acoustic feature data, electromagnetic radiation data, and environmental parameter data.

[0011] The second aspect of the present invention provides an artificial intelligence-based power discharge detection system, including: a feature extraction module for collecting multi-modal data of power equipment and performing adaptive decomposition to extract discharge feature signals; a depth map construction module for constructing a depth map joint learning model, converting the discharge feature signals into a graph data structure, and generating a candidate feature set through a multi-scale spatio-temporal graph convolutional network; a single source localization module for training a discharge pattern recognizer based on the candidate feature set using a contrastive learning framework, establishing a mapping relationship between discharge features and discharge types and discharge positions, and completing the identification and localization of a single power source; a multi-source localization module for distinguishing and localizing multiple power sources with overlapping time domains by using node embedding and community detection of a graph neural network; a risk assessment module for outputting discharge risk assessment results according to discharge types and discharge positions, in combination with the dynamic evolution of discharge features.

[0012] The beneficial effects of the present invention are as follows: Through the multi-modal data adaptive decomposition technology, the present invention significantly improves the accuracy and signal-to-noise ratio of discharge feature extraction; the depth map joint learning framework significantly enhances the modeling ability of discharge spatio-temporal features; the contrastive learning framework realizes the precise coordination of discharge type and location recognition; the graph neural network technology significantly improves the intelligent analysis ability of the multi-discharge source overlap scenario; the dynamic risk assessment model transforms the traditional static assessment into an intelligent early warning based on feature evolution. The present invention not only significantly improves the intelligent level of power equipment safety monitoring, transforms the power equipment fault diagnosis from a static and experience-driven mode to a dynamic and data-driven intelligent early warning paradigm, but also provides a systematic and forward-looking technical solution for fault diagnosis in complex electromagnetic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0014] Figure 1 It is a framework flowchart of a power discharge detection method based on artificial intelligence.

[0015] Figure 2 It is a processing flowchart of a depth map joint learning model for a power discharge detection method based on artificial intelligence.

[0016] Figure 3 It is a flowchart for the identification and location of the discharge source of a power discharge detection method based on artificial intelligence. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] In order to make the objectives, features, and advantages of the present invention more obvious and understandable, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0018] Embodiment 1, referring to Figures 1 to 3 , is the first embodiment of the present invention. This embodiment provides a power discharge detection method based on artificial intelligence, and the framework flowchart is as Figure 1 shown, including: S1: Collect multi-modal data of power equipment and perform adaptive decomposition to extract discharge feature signals.

[0019] Among them, multimodal data includes at least two of electrical parameter data, acoustic feature data, electromagnetic radiation data and environmental parameter data. Traditional power discharge detection methods rely only on single modal data, which makes it difficult to fully capture the complex physical process of discharge, and the feature extraction accuracy and anti-interference ability are weak. In contrast, the joint collection of multimodal data provides a more comprehensive and richer information basis for discharge feature extraction.

[0020] Further, performing adaptive decomposition to extract the discharge characteristic signal includes the following steps: S1.1: Construct a multiscale analysis framework based on discrete wavelet transform to perform preliminary decomposition of multimodal data.

[0021] In this embodiment, by selecting a high-matching wavelet basis function combination and an adaptive decomposition strategy, the complex power discharge signal is accurately decomposed. Specifically, it includes: building a wavelet basis function library, dynamically selecting the optimal wavelet basis according to the signal entropy value and kurtosis index; determining the decomposition level according to the complexity of the signal; introducing the signal reconstruction error minimization criterion, and accurately controlling the wavelet decomposition parameters.

[0022] S1.2: Using low-rank and sparse matrix decomposition algorithms, the decomposed sub-band signal is separated into background noise component and potential discharge component.

[0023] The core of the algorithm is to construct a constraint model that takes into account both low rank and sparsity, and through multi-dimensional constraints, it can achieve accurate identification of background noise and extraction of effective discharge signals.

[0024] S1.3: Construct a discharge feature enhancement model, process the potential discharge components, and obtain the time-frequency domain feature representation.

[0025] Specifically, a deep feature extraction network based on a multi-head attention mechanism is designed to capture the complex nonlinear characteristics of discharge signals; a metric learning framework is constructed to optimize the feature space mapping strategy, which significantly improves the aggregation of similar discharge features and the separability of heterogeneous features; a self-supervised learning mechanism is introduced to deeply mine implicit feature patterns through a comparative learning strategy.

[0026] S1.4: Based on the mutual information theory, the multimodal feature fusion weights are designed to perform weighted combination of the time-frequency domain feature representations to form an enhanced discharge feature signal.

[0027] Preferably, the present invention breaks through the technical limitations of traditional power discharge detection methods, and compared with the existing technology, significantly improves the accuracy of discharge feature extraction, signal-to-noise ratio and anti-interference ability, and the time response speed and recognition accuracy of detection are significantly improved.

[0028] S2: Construct a deep graph joint learning model, convert the discharge feature signal into a graph data structure, and generate a candidate feature set through a multi-scale spatio-temporal graph convolutional network.

[0029] Specifically, the processing flow chart of the deep graph joint learning model is as Figure 2 shown, including the following steps: S2.1: In the deep graph joint learning model, convert the enhanced discharge feature signal into a graph data structure and establish the connection weights between nodes.

[0030] In this embodiment, map the enhanced discharge feature signal to a set of nodes, and the connection weights between nodes are dynamically calculated through the self-attention mechanism, reflecting the correlation strength between different modal features.

[0031] S2.2: Apply the self-attention mechanism to the graph data structure, calculate the correlation between different modal features, and adjust the node connection weights.

[0032] In this embodiment, aiming at the noise interference and feature importance differences in the power discharge signal, design a multi-head graph attention mechanism. The specific implementation includes: constructing a query matrix, a key matrix, and a value matrix, and obtaining the node feature representation through a learnable linear transformation; when calculating the attention coefficient, for nodes i and j, obtain the attention weight through a concatenation operation and normalize it through the LeakyReLU activation function; introduce attention coefficient sparsification processing, use the softmax operation and threshold truncation to retain the most important node connections; dynamically adjust the node connection weights according to the attention coefficient to enhance the propagation efficiency of key features and effectively suppress the information interference of low-importance nodes.

[0033] S2.3: Perform dual-path encoding on the graph data structure. The first encoder processes the signal attribute features of nodes, and the second encoder processes the spatio-temporal position relationship of nodes. Integrate the information of the two paths through a fusion module to generate a unified node representation.

[0034] In this embodiment, to capture both the physical attributes and spatio-temporal evolution laws of discharge features simultaneously, design an attribute-spatio-temporal dual-path encoding architecture and achieve information complementarity through a gated fusion mechanism. Among them, the first encoder uses a graph convolutional neural network to extract the inherent attribute features of nodes through multiple graph convolutional layers, focusing on the internal relationship between the sensor signal waveform features and the device state parameters; the second encoder uses a temporal graph network to combine temporal convolution and spatial graph structure to capture the temporal evolution and spatial topological relationship of discharge events; the fusion module uses a gated fusion mechanism to integrate the information of the two paths through adaptive weights to generate a unified node representation.

[0035] S2.4: Input the unified node representation into a multi-scale spatio-temporal graph convolutional network for processing to generate a preliminary candidate feature set.

[0036] In this embodiment, the time dimension adopts deformable convolution. By learning the dynamic sampling offset, it can effectively capture the non-uniform time dynamic characteristics of the discharge signal. The spatial dimension design combines a hybrid architecture of spectral domain and spatial domain graph convolution, introducing Chebyshev spectral convolution and spatial local aggregation operators to extract the spatial topological characteristics of the graph structure from multiple angles. By setting graph convolution layers with multiple different receptive fields, a multi-scale feature extraction network is constructed to achieve a hierarchical representation of the discharge characteristics. Finally, the feature maps in the time and space dimensions are integrated to generate a candidate feature set with global semantic information.

[0037] S2.5: Perform feature clustering and classification on the preliminary candidate feature set to identify and separate the candidate feature subset of a single discharge source and the candidate feature subset of overlapping multiple discharge sources.

[0038] Specifically, design a feature overlap measurement model. By analyzing the spatio-temporal feature cross-entropy and spatial correlation of the candidate feature set, evaluate the overlap degree between discharge sources; construct a discharge source feature decoupling algorithm to separate signals for candidate feature subsets with highly overlapping spatio-temporal features; establish a multi-scale feature mapping mechanism to distinguish the candidate feature subset of a single discharge source and the candidate feature subset of overlapping multiple discharge sources; implement a feature clustering algorithm to divide the candidate feature set into two subsets based on the overlap degree; design a feature subset marking module to add structured marks to the candidate feature subsets of single discharge sources and overlapping multiple discharge sources.

[0039] S2.6: Use a generative adversarial network to generate boundary discharge feature signals and train a deep graph joint learning model in combination with the candidate feature subsets.

[0040] In this embodiment, the generative adversarial network simulates complex discharge boundary features through the adversarial training of the generator and the discriminator, effectively alleviating the problem of unbalanced boundary samples in the training data. The generator synthesizes boundary discharge signals with high authenticity based on the candidate feature subsets, and the discriminator learns to distinguish real and generated boundary features, enhancing the model's ability to identify weak discharge sources through iterative optimization.

[0041] Preferably, the present invention realizes the accurate identification and decoupling of power discharge feature signals through a deep graph joint learning model. Compared with traditional methods, the present invention has obvious improvements in aspects such as the accuracy of discharge source identification, false alarm rate control, and weak discharge source detection. The feature extraction ability of the model in a complex electromagnetic interference environment is improved, and the feature decoupling processing in the scenario of overlapping multiple discharge sources is more accurate. At the same time, the computational efficiency of the model is improved, providing a more practical technical solution for on-line monitoring of power equipment.

[0042] S3: Based on the candidate feature set, train a discharge pattern recognizer using a contrastive learning framework, establish the mapping relationship between discharge features, discharge types, and discharge positions, and complete the identification and positioning of a single discharge source.

[0043] Specifically, the flowchart for identifying and locating the power source is as Figure 3 shown, including the following steps: S3.1: Use the single power source candidate feature subset as the input feature for contrastive learning to construct a multi-level contrast sample library.

[0044] Among them, the multi-level contrast sample library includes: inter-class contrast sample pairs of different physical features, intra-class contrast sample pairs of the same discharge type but different fine features, and position correlation sample pairs of different spatial positions but with similar feature patterns.

[0045] S3.2: Construct a two-branch contrastive learning network. The first branch processes discharge type recognition, and the second branch processes discharge position location. The two branches share the underlying feature extraction network.

[0046] Specifically, the two-branch contrastive learning network consists of a feature extraction backbone network with residual connections, a discharge type recognition branch, and a discharge position location branch. The design goal of the network is to accurately extract and identify discharge features from complex power discharge signals and achieve precise positioning of the discharge type and position.

[0047] Among them, the feature extraction backbone network uses multi-layer graph attention convolutional layers, which are composed of a feature mapping unit and an attention weight adaptive adjustment module, and can adaptively capture complex feature information; the discharge type recognition branch introduces a multi-head attention classifier (used to capture complex features of the discharge type from multiple perspectives) and a softmax loss function with a hard sample mining strategy, and sets a type uncertainty quantification module to improve the learning ability for boundary samples; the discharge position location branch enhances the positioning accuracy through a position regression module based on spatio-temporal graph convolution (which can consider both time evolution and spatial topology features), a coordinate system error correction unit (to correct the positioning deviation), and a position prediction confidence evaluation unit (to provide a reliability evaluation of the positioning result).

[0048] Furthermore, the two branches perform feature interaction through a cross-branch feature attention fusion mechanism, a dynamic weight sharing strategy, and a multi-task learning framework. The cross-branch feature attention fusion mechanism allows information exchange between different branches, the dynamic weight sharing strategy can adjust the focus of feature extraction, and the multi-task learning framework realizes the joint optimization of discharge type recognition and position location.

[0049] Preferably, this design not only improves the feature expression ability of the network but also significantly enhances the understanding and processing ability of complex power discharge signals, providing strong technical support for accurate identification and positioning.

[0050] S3.3: Extract the time dynamic features and spatial topological features from the single discharge source candidate feature subset, input the time dynamic features into the first branch, and input the spatial topological features into the second branch.

[0051] S3.4: Calculate the contrast loss for the two branches respectively. The first branch calculates the type recognition loss, and the second branch calculates the position localization loss, and combine the two with weights to form the total loss function.

[0052] Specifically, embed a regularization term in the loss function to suppress overfitting and enhance the generalization ability of the model: ; where, is the L2 norm regularization term, W is the learnable parameter tensor, is the L2 regularization coefficient (controlling the parameter penalty strength), is the KL divergence regularization coefficient, KL is the KL divergence (used to measure the difference between the predicted distribution and the reference distribution), p is the probability distribution of the discharge type predicted by the model, and q is the true type probability distribution of the training data. It should be noted that by introducing the KL divergence into the loss function, the overfitting of the model to the training data can be reduced, the generalization ability of the model on unseen data can be improved, and the prediction bias of the model can be balanced.

[0053] Furthermore, for the type recognition loss of the first branch, an improved contrast cross-entropy loss function is adopted, and its calculation formula is: ; where, y i represents the true discharge type label, p i represents the predicted type probability, w i is the dynamic type weight coefficient, which is calculated in the following way: ; where, d i is the Euclidean distance between the current sample and the type center, is the smoothing factor.

[0054] Furthermore, for the position localization loss of the second branch, a regression loss function combined with local sensitive hashing constraints is defined as: ; where, y loc is the true position coordinate, p loc is the predicted position coordinate, LSH penalty dynamically adjusts the penalty strength of the localization loss through the hash coding similarity, and is.

[0055] Furthermore, the total loss function is defined as a weighted combination of weights: ; Among them, L total is the total loss function, L type is the type recognition loss, and L loc is the position localization loss. is the loss weight coefficient, which is used to control the relative importance of the type recognition loss L type and the position localization loss L loc . The initial value is set to 0.5. By adjusting , the learning objectives of the two branches can be flexibly balanced, enabling the model to achieve a balance between type recognition and position localization.

[0056] Furthermore, a loss balance strategy based on meta-learning is introduced. By learning the meta-network of the loss weights, the relative importance of L type and L loc is automatically adjusted. This meta-network uses gradient decoupling and the correlation between loss gradient directions as the optimization objectives.

[0057] Preferably, by introducing a contrast cross-entropy loss with dynamic type weights, the loss calculation can be adaptively adjusted according to the distance between the sample and the type center; a position localization loss with local sensitive hashing constraints is adopted to enhance the accuracy of position prediction; an adaptive weight fusion mechanism based on the training stage and sample complexity is designed to achieve the dynamic balance of the loss function; and the meta-learning strategy is combined to automatically learn the loss weights, reducing the dependence on manual parameter tuning. At the same time, through a carefully designed regularization term, overfitting of the model is effectively suppressed, and the generalization performance of the model in complex power discharge detection scenarios is improved.

[0058] S3.5: By jointly optimizing the total loss function, the synchronous training of discharge type recognition and position localization is completed, and the network parameters of the discharge pattern recognizer are obtained.

[0059] S3.6: Initialize the discharge pattern recognizer with the network parameters. The discharge pattern recognizer receives a single candidate feature subset of the discharge source as input and outputs the discharge type probability distribution and the discharge position coordinates to complete the recognition and localization of a single discharge source.

[0060] Preferably, the present invention innovatively solves the technical problem of difficult accurate identification of discharge types and positions in power discharge detection through a contrastive learning framework. Compared with traditional single identification methods, this solution designs a dual-branch contrastive learning network to achieve collaborative intelligent identification of discharge types and positions. By constructing a multi-level contrastive sample library and introducing a cross-branch feature attention fusion mechanism, the network's feature extraction and understanding capabilities for complex discharge signals are significantly improved. The innovative loss function design, including the contrastive cross-entropy loss with dynamic type weights and the position localization loss with local sensitive hashing constraints, effectively balances the learning objectives of type identification and position localization. Introducing a regularization strategy and meta-learning method further enhances the model's generalization performance in complex electromagnetic environments, providing a more intelligent and efficient technical solution for accurate identification of power equipment discharge sources.

[0061] S4: Use node embedding and community detection of graph neural networks to distinguish and locate multiple discharge sources with overlapping time domains.

[0062] Specifically, it includes the following steps: S4.1: Construct a dynamic graph network based on the multi-discharge source overlapping candidate feature subset and define the association rules between nodes and edges.

[0063] Specifically, design a node mapping strategy based on feature correlation, where: map the feature segments of the multi-discharge source overlapping candidate feature subset to the nodes of the graph network, and based on the timestamps, spectral correlations, and spatial distributions of the feature segments, use an improved cosine similarity to calculate the association degree between nodes. The calculation formula is: ; where is the timestamp difference between two node feature segments, T max is the maximum time span of the observation window, F1 and F2 are the feature vectors of two nodes, is a tuning parameter used to control the intensity of the exponential norm term, is the square of the Euclidean distance between feature vectors, used to measure the difference between feature vectors. This formula comprehensively considers feature similarity, time correlation, and non-linear changes, and can more accurately describe the complex association relationships between multi-discharge source nodes. Among them, the cosine similarity measures the degree of closeness of the directions of feature vectors (the closer to 1, the more similar the directions), the time decay term reflects the law of the decreasing association between nodes over time, and the exponential norm term is used to adjust the non-linear change of the association degree.

[0064] Furthermore, based on the cross-correlation coefficient of node spectral features, a frequency-domain correlation threshold is set, and a time decay function is introduced to construct a multi-dimensional dynamic edge association rule; an adaptive edge dynamic generation and pruning mechanism is designed. By comprehensively considering the node feature similarity, the mutual information entropy of information transfer between nodes, and the consistency of spectral and spatio-temporal features, the network topology structure is dynamically adjusted, and an edge weight threshold is set to accurately prune low-correlation edges.

[0065] S4.2: Use a graph neural network to perform node embedding on the dynamic graph network and learn the spatio-temporal dependence relationship between nodes.

[0066] S4.3: Design a multi-scale community detection algorithm based on spectral clustering to cluster the node embedding space and identify the clusters of power source nodes with overlapping time domains.

[0067] Specifically, perform node embedding space mapping to convert the power source nodes into low-dimensional feature space vectors while maintaining the original topological relationship between nodes; construct a node similarity metric matrix to calculate the similarity and distance relationships of node vectors in the embedding space; design a multi-scale clustering discrimination criterion to evaluate the tightness and separability of node clustering and adaptively determine the number of clustering clusters; implement node grouping based on spectral clustering, perform eigen-decomposition on the similarity matrix, and select the eigenvectors as the basis for clustering; output the mapping of the clusters of power source nodes with overlapping time domains, mark the spatio-temporal boundaries and features of each node cluster, and extract the temporal correlation features between node clusters.

[0068] S4.4: Separate and reconstruct the spatio-temporal features of each power source through a feature decoupling model to determine the spatio-temporal boundaries of the power source.

[0069] Furthermore, input the temporal correlation features of the node cluster mapping as the initial input of the feature decoupling model; construct a feature decoupling network based on a variational autoencoder, and gradually separate the mixed spatio-temporal features through reconstruction loss and regularization constraints; use an attention mechanism to adaptively learn the time dynamic features and spatial distribution features of each power source; based on the reconstruction results of feature decoupling, extract the independent spatio-temporal feature vectors of each power source; through the boundary detection algorithm of spatio-temporal feature vectors, accurately locate and mark the start and end points of time and the spatial distribution areas of each power source.

[0070] S4.5: Based on the independent spatio-temporal feature vectors of feature decoupling, construct a multi-power source feature mapping model to obtain the discharge type probability distribution and discharge position coordinates of each power source.

[0071] Furthermore, design a multi-instance feature association network to convert the decoupled spatio-temporal feature vectors into a unified feature representation space, construct an association mapping mechanism for the feature of the power source, and capture the potential correlation between multiple power sources; apply a graph attention encoder to learn the cross-dimensional correlation pattern of the feature vectors of multiple power sources, dynamically adjust the weight distribution between features through the self-attention mechanism, and enhance the discriminative ability of feature representation; implement a multi-branch parallel inference strategy to simultaneously perform type recognition and location positioning on each power source feature vector, and configure independent type recognition and location regression branches based on the shared feature extraction backbone.

[0072] S4.6: Output the reconstruction results of multiple power sources, including the number of power sources, the spatio-temporal boundaries, position coordinates, and discharge types of each power source.

[0073] Preferably, the present invention effectively solves the technical problem of difficult recognition in the scenario of overlapping multiple power sources in traditional power discharge detection through the graph neural network method. Compared with the prior art, this solution realizes the precise decoupling and reconstruction of time-domain overlapping power sources, and improves the analysis ability of multi-source discharge signals through technologies such as constructing a dynamic graph network, adaptive node association, and multi-scale community detection. Innovatively introducing a feature decoupling model and a multi-instance feature association network can accurately separate and reconstruct the spatio-temporal features of each power source, and simultaneously obtain its type and position information, realizing the intelligent recognition of multi-source discharge signals.

[0074] S5: According to the discharge type and discharge position, combined with the dynamic evolution of discharge characteristics, output the discharge risk assessment result.

[0075] Specifically, it includes the following steps: S5.1: Construct a discharge risk assessment model, taking the discharge type and discharge position as basic parameters, and establish a parameter mapping mechanism for risk assessment.

[0076] S5.2: Set the basic risk level according to different discharge types, and adjust the risk weight in combination with the sensitivity of the discharge position to form a differentiated risk assessment strategy.

[0077] Specifically, by classifying different discharge types in detail, according to historical data and expert experience, assign an initial risk level to each discharge type. For different power equipment positions, comprehensively consider factors such as equipment importance, geographical location, and surrounding environment, and make differential adjustments to the risk weight. Formulate risk level adjustment rules, and comprehensively consider the discharge type and position sensitivity to generate multi-dimensional risk assessment results.

[0078] S5.3: Based on the dynamic evolution of discharge characteristics, dynamically adjust the risk level of the discharge type to quantify the potential equipment damage risk.

[0079] Specifically, continuously track the change process of discharge characteristics. For each type of discharge, establish a correlation mechanism between feature evolution and risk level, and analyze the degree of influence of different feature changes on equipment damage. According to the time-series change of discharge characteristics, dynamically adjust the risk level, and quantify the potential damage probability and expected loss degree of various discharges to the equipment.

[0080] S5.4: Establish a correlation model between the discharge location and the risk diffusion path, and evaluate the degree of influence of discharges at different locations on the integrity of power equipment.

[0081] Specifically, draw the spatial layout and connection relationship of power equipment, and analyze the physical and electrical connections between each equipment. Trace and record the location where the discharge occurs, and calculate the potential influence range of the discharge at this location on adjacent equipment and key equipment. According to the criticality and connection tightness of the equipment, formulate the propagation weight of the risk diffusion path, and evaluate the influence of discharges at different locations on the integrity of the overall system.

[0082] S5.5: Generate a discharge risk assessment report to comprehensively present the results of risk analysis.

[0083] Furthermore, this embodiment also provides an artificial intelligence-based power discharge detection system, including: a feature extraction module for collecting multimodal data of power equipment and performing adaptive decomposition to extract discharge feature signals; a depth map construction module for constructing a depth map joint learning model, converting the discharge feature signals into a graph data structure, and generating a candidate feature set through a multi-scale spatio-temporal graph convolutional network; a single-source localization module for training a discharge pattern recognizer based on the candidate feature set using a contrastive learning framework, establishing a mapping relationship between discharge features and discharge types and locations, and completing the identification and localization of a single discharge source; a multi-source localization module for using node embedding and community detection of a graph neural network to distinguish and locate multiple overlapping discharge sources in the time domain; a risk assessment module for outputting a discharge risk assessment result according to the discharge type and location, combined with the dynamic evolution of discharge features.

[0084] In summary, through the multimodal data adaptive decomposition technology, the present invention significantly improves the accuracy and signal-to-noise ratio of discharge feature extraction; the depth map joint learning framework greatly enhances the modeling ability of discharge spatio-temporal features; the contrastive learning framework realizes the precise coordination of discharge type and location identification; the graph neural network technology significantly improves the intelligent analysis ability of the multi-discharge source overlapping scenario; the dynamic risk assessment model transforms the traditional static assessment into an intelligent early warning based on feature evolution. The present invention not only significantly improves the intelligent level of power equipment safety monitoring, transforms the power equipment fault diagnosis from a static and experience-driven mode to a dynamic and data-driven intelligent early warning paradigm, but also provides a systematic and forward-looking technical solution for fault diagnosis in a complex electromagnetic environment.

[0085] Example 2, refer toFigures 1 to 3 , which is the second embodiment of the present invention. This embodiment provides an artificial intelligence-based power discharge detection method. To verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0086] In the actual operation environment of a 220 kV UHV transmission line, important nodes are selected as the research object to verify the multi-modal power discharge detection method based on artificial intelligence. The experimental scenario is located in a substation in East China. This site is located in the Yangtze River Delta region, with a complex electromagnetic environment and diverse equipment operating states.

[0087] The multi-modal data acquisition system is configured with a high-precision sensor array. The sensors are deployed on key equipment insulators and connection points. The data sampling frequency is 128 kHz, and the sampling window is 1024 ms. The sensor array consists of 4 electromagnetic sensors, 3 acoustic sensors, and 2 infrared thermal imaging sensors, realizing multi-dimensional and accurate capture of discharge characteristics.

[0088] In the data preprocessing and feature extraction stage, improved discrete wavelet transform is adopted. The Daubechies wavelet basis (db4) is selected, and the decomposition level is 6 layers. The wavelet basis selection is dynamically optimized through signal entropy value and kurtosis index to effectively suppress background noise. During the experiment, the background noise suppression rate reaches 87.6%, and the signal reconstruction error is controlled within the range of ±2.8%, laying a solid foundation for subsequent feature extraction.

[0089] The construction of the depth map joint learning model adopts a multi-head graph attention mechanism. The network structure includes 3 attention heads, the node feature dimension is 64, and the number of graph convolutional layers is 4 layers. Experiments are carried out on 20 different types of discharge samples. The recognition accuracy of the model for a single discharge source is 94.3%, and the feature decoupling accuracy in the multi-discharge source overlapping scenario is 91.7%. The generative adversarial network simulates the boundary discharge characteristics, and the boundary sample recognition rate is increased to 82.5%, effectively improving the problem of unbalanced boundary samples in the training data.

[0090] To verify the practicality of the method, continuous monitoring of this 220 kV transmission line is carried out for 6 months. During this period, a total of 52 discharge events are captured and analyzed. Compared with the traditional single-modal-based detection method, the positioning accuracy of the discharge source of this method is improved by 24.8%, and the risk prediction accuracy is increased to 94.2%. To intuitively show the technological innovation, the following comparative analysis is given, as shown in Table 1.

[0091] Table 1 Performance comparison between the present invention and the prior art Evaluation index Traditional unimodal method Method of the present invention Identification accuracy of discharge source 72.5% 94.3% Decoupling ability of multiple discharge sources 63.2% 91.7% Improvement of signal-to-noise ratio 12.6 dB 35.4 dB Recognition rate of boundary samples 48.6% 82.5% Accuracy of risk prediction 69.4% 94.2% This method demonstrates significant technical advantages in the field of on-line monitoring of power equipment, providing a more accurate detection means for the safe operation of power systems. Data shows that through multi-modal data fusion, adaptive feature extraction, and deep learning technologies, the present invention has achieved a substantial breakthrough in the field of power discharge detection.

[0092] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements 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.

Claims

1. A power discharge detection method based on artificial intelligence, characterized in that: include: Collect multi-modal data of power equipment and perform adaptive decomposition to extract discharge characteristic signals; Constructing a deep graph joint learning model, converting the discharge feature signal into a graph data structure, and generating a candidate feature set through a multi-scale spatiotemporal graph convolutional network; Based on the candidate feature set, a contrastive learning framework is used to train a discharge pattern recognizer, a mapping relationship between discharge features and discharge types and discharge positions is established, and identification and positioning of a single discharge source is completed; Use node embedding and community detection of graph neural networks to distinguish and locate multiple discharge sources that overlap in time domains; According to the discharge type and discharge location, combined with the dynamic evolution of discharge characteristics, the discharge risk assessment results are output.

2. The power discharge detection method based on artificial intelligence according to claim 1 is characterized in that: The performing of adaptive decomposition to extract the discharge characteristic signal comprises: A multiscale analysis framework is constructed based on discrete wavelet transform to perform preliminary decomposition of multimodal data; Using matrix decomposition algorithm, the decomposed sub-band signal is separated into background noise component and potential discharge component; Constructing a discharge feature enhancement model, processing the potential discharge component, and obtaining a time-frequency domain feature representation; The multimodal feature fusion weights are designed to perform weighted combination on the time-frequency domain feature representations to form an enhanced discharge feature signal.

3. The power discharge detection method based on artificial intelligence according to claim 1 is characterized in that: The construction of the deep graph joint learning model includes: In the deep graph joint learning model, the enhanced discharge feature signal is converted into a graph data structure, and the connection weights between nodes are established; Applying a self-attention mechanism to the graph data structure, calculating the correlation between different modal features, and adjusting the node connection weights; The graph data structure is encoded in two ways, wherein the first encoder processes the signal attribute characteristics of the nodes, and the second encoder processes the spatiotemporal position relationship of the nodes, and the two information are integrated through a fusion module to generate a unified node representation; The unified node representation is input into a multi-scale spatiotemporal graph convolutional network for processing to generate a preliminary candidate feature set; wherein the time dimension uses deformable convolution to capture temporal dynamic features, and the spatial dimension combines spectral domain and spatial domain graph convolution to extract spatial topological features; The preliminary candidate feature set is subjected to feature clustering and classification, and a single discharge source candidate feature subset and a multi-discharge source overlapping candidate feature subset are identified and separated.

4. The power discharge detection method based on artificial intelligence according to claim 1 is characterized in that: The use of contrastive learning framework to train the discharge pattern recognizer includes: A single discharge source candidate feature subset is used as the input feature of contrastive learning to construct a multi-level contrast sample library; Construct a dual-branch contrastive learning network, where the first branch processes the discharge type recognition and the second branch processes the discharge position location. The two branches share the underlying feature extraction network. Extracting time dynamic features and space topological features from the single discharge source candidate feature subset, inputting the time dynamic features into the first branch, and inputting the space topological features into the second branch; The contrast loss is calculated for the two branches respectively. The first branch calculates the type recognition loss, and the second branch calculates the position positioning loss. The two are weighted and combined to form the total loss function. By jointly optimizing the total loss function, the synchronous training of discharge type recognition and position positioning is completed to obtain the network parameters of the discharge pattern recognizer; The network parameters are used to initialize a discharge pattern recognizer, which receives a single discharge source candidate feature subset as input and outputs a discharge type probability distribution and a discharge position coordinate.

5. The power discharge detection method based on artificial intelligence according to claim 1 is characterized in that: The node embedding and community detection using graph neural network to distinguish and locate multiple discharge sources overlapping in time domain includes: Constructing a dynamic graph network based on the overlapping candidate feature subsets of multiple sources, and embedding nodes in the dynamic graph network using a graph neural network; A multi-scale community detection algorithm based on spectral clustering is designed to embed nodes into spatial clusters and identify clusters of discharge source nodes that overlap in the time domain. Through the feature decoupling model, the spatiotemporal characteristics of each discharge source are separated and reconstructed, and the spatiotemporal boundaries of the discharge source are determined; Construct a multi-discharge source feature mapping model to obtain the discharge type probability distribution and discharge position coordinates of each discharge source; Output multiple discharge source reconstruction results, including the number of discharge sources, the temporal and spatial boundaries of each discharge source, location coordinates, and discharge type.

6. The power discharge detection method based on artificial intelligence according to claim 1 is characterized in that: The output discharge risk assessment results include: Construct a discharge risk assessment model, take discharge type and discharge location as basic parameters, and establish a parameter mapping mechanism for risk assessment; Set basic risk levels according to different discharge types, and adjust risk weights based on the sensitivity of the discharge location to form a differentiated risk assessment strategy; Based on the dynamic evolution of discharge characteristics, the risk level of the discharge type is dynamically adjusted to quantify the potential risk of equipment damage; Establish a correlation model between discharge location and risk diffusion path to evaluate the impact of discharge at different locations on the integrity of power equipment; Generate a discharge risk assessment report and comprehensively present the risk analysis results.

7. The power discharge detection method based on artificial intelligence according to claim 1 is characterized in that: The multimodal data includes electrical parameter data, acoustic characteristic data, electromagnetic radiation data and environmental parameter data.

8. An artificial intelligence-based power discharge detection system, characterized in that: include: A feature extraction module is used to collect multi-modal data of power equipment and perform adaptive decomposition to extract discharge feature signals; A deep graph construction module, used to construct a deep graph joint learning model, convert the discharge feature signal into a graph data structure, and generate a candidate feature set through a multi-scale spatiotemporal graph convolutional network; A single source positioning module is used to train a discharge pattern recognizer based on the candidate feature set using a contrastive learning framework, establish a mapping relationship between discharge features and discharge types and discharge positions, and complete the identification and positioning of a single discharge source; Multi-source localization module, which uses node embedding and community detection of graph neural networks to distinguish and locate multiple discharge sources that overlap in the time domain; The risk assessment module is used to output the discharge risk assessment result according to the discharge type and discharge location combined with the dynamic evolution of the discharge characteristics.

Images