Locating method, system and equipment for partial discharge of switch cabinet and medium

Abstract

The invention relates to the technical field of intelligent monitoring, and discloses a switch cabinet partial discharge positioning method, system and device and a medium, and the method comprises the steps: collecting switch cabinet partial discharge data in real time, carrying out the first processing, and outputting partial discharge signal features and gas component features; after the partial discharge signal features and the gas component features are fused through a first fusion neural network, the mode type of partial discharge is output; constructing a spatial three-dimensional distribution diagram of the gas component concentration through a second neural network; based on the mode type of partial discharge and the spatial three-dimensional distribution diagram, minimizing the spatial difference between the partial discharge signal and the gas component concentration through a first fusion algorithm, and obtaining the accurate positioning of the partial discharge source; evaluating the health state of the switch cabinet according to the accurate positioning, and outputting a health state evaluation report. Through the multi-modal graph neural network and the dynamic graph optimization algorithm, the local discharge source positioning precision and the state evaluation reliability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and positioning technology, and in particular to a method, system, device and medium for locating partial discharge in switchgear. Background Technology

[0002] As a critical piece of equipment in power systems, the accurate detection and location of partial discharge in switchgear is of great significance for ensuring the safe operation of the power grid. Traditional partial discharge detection technologies mainly rely on single-mode data such as ultra-high frequency, ultrasonic, or chemical gas detection, and achieve preliminary diagnosis through threshold judgment or simple statistical models. These technologies suffer from weak anti-interference capabilities and low location accuracy. With the development of artificial intelligence technology, signal classification algorithms based on convolutional neural networks and gas diffusion inversion models have been gradually introduced into this field. Researchers have begun to explore using multi-sensor spatial correlation and gas concentration gradient distribution to infer the location of the discharge source.

[0003] However, existing technologies still have significant shortcomings in multimodal data collaborative analysis and spatial modeling: on the one hand, heterogeneous features such as UHF signals, ultrasonic waveforms and gas concentration data are difficult to deeply integrate, and traditional linear weighting or simple feature splicing methods cannot effectively capture the nonlinear interaction between modes; on the other hand, gas concentration distribution modeling often uses conventional interpolation algorithms, which fail to fully consider the spatial topological characteristics and fluid dynamics of sensor networks, resulting in a large deviation between the reconstructed three-dimensional concentration field and the real physical distribution.

[0004] Existing technologies have significant limitations on improving the accuracy of partial discharge localization, necessitating the development of novel localization methods that can deeply integrate multi-source heterogeneous data and accurately model the gas diffusion process. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] Therefore, the present invention provides a method, system, device and medium for locating partial discharge in switchgear, which solves the problem of large location error of partial discharge in switchgear caused by insufficient fusion of heterogeneous features of multimodal data and insufficient accuracy of spatial modeling of gas concentration.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a method for locating partial discharge in a switchgear, comprising:

[0009] Real-time acquisition of partial discharge data from the switchgear and initial processing are performed to output partial discharge signal characteristics and gas composition characteristics.

[0010] After fusing the partial discharge signal features and gas component features using a first fusion neural network, the partial discharge mode type is output.

[0011] A spatial three-dimensional distribution map of gas component concentrations is constructed using a second neural network;

[0012] Based on the mode type and spatial three-dimensional distribution map of the partial discharge, the spatial difference between the partial discharge signal and the gas component concentration is minimized by the first fusion algorithm to obtain the precise location of the partial discharge source.

[0013] The health status of the switchgear is assessed based on the precise location, and a health status assessment report is output.

[0014] In a preferred embodiment of the partial discharge location method for switchgear according to the present invention, the output partial discharge mode types include:

[0015] Using the partial discharge signal features and gas component features as nodes in a graph, and capturing the similarity and interaction relationships between features through edge weights, a multimodal graph structure is constructed.

[0016] The interaction weights of nodes are dynamically calculated using the first attention mechanism, and an interaction weight matrix is ​​generated.

[0017] The partial discharge signal features and gas component features are fused based on the interaction weight matrix to output fused features;

[0018] The fusion features are input into the fully connected layer of the first fusion neural network, and the partial discharge mode type is output.

[0019] The beneficial effects of this preferred technical solution are: by using graph attention mechanism to dynamically learn the cross-modal correlation between partial discharge signal features and gas component features, the physical interpretability of feature representation is enhanced.

[0020] In a preferred embodiment of the partial discharge location method for switchgear described in this invention, the generation of the interaction weight matrix includes:

[0021] Calculate the first attention weight and the second attention weight of the multimodal graph structure;

[0022] An interaction weight matrix is ​​generated by combining the first attention weight and the second attention weight.

[0023] As a preferred embodiment of the partial discharge location method for switchgear according to the present invention, the step of constructing a spatial three-dimensional distribution map of gas component concentration through a second neural network includes:

[0024] A graph network structure is constructed by defining edges based on the spatial positional relationships of gas component sensor nodes;

[0025] By using a second neural network to fuse gas component data and spatial relationships, the gas components of unsampled areas are predicted to construct a three-dimensional spatial distribution map.

[0026] The beneficial effects of this preferred technical solution are: it integrates the spatial topological relationship of sensor nodes into gas concentration prediction, and achieves sub-meter resolution modeling of unsampled areas.

[0027] In a preferred embodiment of the partial discharge location method for switchgear described in this invention, the precise location of the partial discharge source includes:

[0028] Dynamic graph embedding optimization is used to dynamically embed the partial discharge signal and gas component concentration into the graph structure;

[0029] By combining adaptive graph convolutional networks to capture the spatial correlation of multimodal data, optimized node features are output.

[0030] Based on the dynamic graph structure and the optimized node characteristics, the spatial difference between the partial discharge signal and the gas component concentration is minimized, and the precise location of the partial discharge power source is output.

[0031] In a preferred embodiment of the partial discharge location method for switchgear according to the present invention, the step of assessing the health status of the switchgear based on the precise location includes:

[0032] The health status of the switchgear can be divided into normal, slightly abnormal, and severely abnormal.

[0033] The health status of the switchgear is classified by combining the fusion features, and the health status level is output. When the health status is slightly abnormal or seriously abnormal, an alarm is triggered and information is pushed.

[0034] By combining the analytic hierarchy process (AHP) and fault tree analysis, a health status assessment report is generated.

[0035] As a preferred embodiment of the partial discharge location method for switchgear according to the present invention, the step of real-time acquisition of partial discharge data of switchgear and performing a first processing includes:

[0036] Real-time acquisition of partial discharge signals and gas composition;

[0037] Wavelet transform denoising method is used to remove high-frequency noise and low-frequency interference from partial discharge signals;

[0038] Kalman filtering is used to remove noise from gas component data;

[0039] The partial discharge signal features and gas component features are output by combining normalization and feature extraction.

[0040] In a second aspect, the present invention provides a partial discharge locating system for switchgear, comprising:

[0041] The data acquisition module is used to acquire partial discharge data of the switchgear in real time and perform the first processing, outputting partial discharge signal characteristics and gas composition characteristics;

[0042] The feature fusion module is used to fuse the partial discharge signal features and gas component features using a first fusion neural network, and then output the partial discharge mode type.

[0043] The distribution modeling module is used to construct a spatial three-dimensional distribution map of gas component concentrations through a second neural network.

[0044] The precise positioning module is used to obtain the precise location of the partial discharge source by minimizing the spatial difference between the partial discharge signal and the gas component concentration through a first fusion algorithm based on the mode type and spatial three-dimensional distribution map of the partial discharge.

[0045] The status assessment module is used to assess the health status of the switchgear based on the precise location and output a health status assessment report.

[0046] Thirdly, the present invention provides an electronic device, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor executes the computer-executable instructions to implement the steps of a partial discharge positioning method for a switchgear.

[0047] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of a partial discharge location method for a switchgear.

[0048] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a method, system, device, and medium for locating partial discharge in switchgear. Through multimodal graph neural networks and dynamic graph optimization algorithms, it significantly improves the accuracy of partial discharge source location and the reliability of status assessment. It constructs a feature fusion network based on a graph structure, utilizing a graph attention mechanism to dynamically learn the cross-modal correlation between partial discharge signal features and gas component features, enhancing the physical interpretability of feature representation. Combining graph neural networks and Kriging interpolation algorithms, it integrates the spatial topological relationships of sensor nodes into gas concentration prediction, achieving sub-meter resolution modeling of unsampled areas. Furthermore, by using an adaptive graph convolutional network to fuse multi-source features, it enables dynamic hierarchical early warning of equipment status, providing a reliable decision-making basis for predictive maintenance of switchgear. Attached Figure Description

[0049] 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.

[0050] Figure 1 This is a schematic diagram of the overall process logic of the partial discharge location method for switchgear according to an embodiment of the present invention.

[0051] Figure 2 This is a flowchart of a partial discharge location method for switchgear according to an embodiment of the present invention.

[0052] Figure 3 This is a flowchart of data preprocessing and feature extraction for a partial discharge location method for switchgear according to an embodiment of the present invention. Detailed Implementation

[0053] 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.

[0054] Example 1, referring to Figure 1 As one embodiment of the present invention, a method for locating partial discharge in a switchgear is provided, such as... Figure 1 The specific steps shown are as follows:

[0055] S100: Real-time acquisition of partial discharge data from the switchgear and first processing, outputting partial discharge signal characteristics and gas composition characteristics;

[0056] S200: After fusing the partial discharge signal features and gas component features using the first fusion neural network, the partial discharge mode type is output.

[0057] S300: Constructs a spatial three-dimensional distribution map of gas component concentrations using a second neural network;

[0058] S400: Based on the mode type and spatial three-dimensional distribution map of partial discharge, the spatial difference between the partial discharge signal and the gas component concentration is minimized through the first fusion algorithm to obtain the precise location of the partial discharge source;

[0059] S500: Assess the health status of the switchgear based on precise positioning and output a health status assessment report.

[0060] It should be noted that, to address the issue of large localization errors in switchgear partial discharge sources caused by insufficient fusion of heterogeneous features from multimodal data and inadequate accuracy in spatial modeling of gas concentration, steps S100–S500 significantly improve the localization accuracy and reliability of status assessment of partial discharge sources through multimodal graph neural networks and dynamic graph optimization algorithms. A graph-based feature fusion network is constructed, utilizing graph attention mechanisms to dynamically learn the cross-modal correlation between partial discharge signal features and gas component features, enhancing the physical interpretability of feature representation. By combining graph neural networks and Kriging interpolation algorithms, the spatial topological relationships of sensor nodes are integrated into gas concentration prediction, achieving sub-meter resolution modeling of unsampled areas. Finally, through an adaptive graph convolutional network, multi-source features are fused to enable dynamic hierarchical early warning of equipment status, providing a reliable decision-making basis for predictive maintenance of switchgear.

[0061] Example 2, refer to Figure 2 and Figure 3 Based on the previous embodiment, this embodiment provides a specific implementation method for a partial discharge location method for switchgear, which will be used to illustrate the technical means employed in this method.

[0062] In this embodiment of the application, the above step S100, which involves real-time acquisition of partial discharge data from the switchgear and first processing to output partial discharge signal characteristics and gas composition characteristics, includes the following sub-steps A1 to A4:

[0063] In A1: Real-time acquisition of partial discharge signals and gas composition;

[0064] Specifically, such as Figure 2 As shown, a high-frequency current transformer is installed at the grounding wire of the switchgear to collect partial discharge signals. An electrochemical gas sensor and a photoacoustic spectroscopy sensor are used to detect gas components. The high-frequency current transformer has strong high-frequency signal capture capability, interference resistance and engineering adaptability. The electrochemical gas sensor has high sensitivity and selectivity, fast response and stability, low power consumption and miniaturization. The photoacoustic spectroscopy sensor has ultra-high selectivity and anti-interference capability, trace gas detection capability, calibration-free and long life. An FPGA controller is used to synchronize the timestamps of multiple sensors, so that the discharge event and the gas concentration change have spatiotemporal correlation.

[0065] Preferably, the combination of high-frequency current transformer, electrochemical gas sensor and photoacoustic spectroscopy sensor, through electro-chemical-optical multimodal data fusion, constructs a high-sensitivity, high-selectivity, strong anti-interference capability and early warning for partial discharge detection, providing a reliable technical solution for power equipment condition monitoring. The three complement each other's advantages, significantly improving the accuracy of fault diagnosis and system reliability.

[0066] It should be noted that, unlike traditional fixed-function integrated circuits or general-purpose processors, FPGA controllers offer unique advantages in achieving multi-sensor synchronization, including flexibility, parallel processing capabilities, and low latency.

[0067] In A2: Wavelet transform denoising method is used to remove high-frequency noise and low-frequency interference from partial discharge signal;

[0068] Specifically, such as Figure 3 The signal is decomposed into five layers of wavelets, with high-frequency coefficients truncated using a hard thresholding method and low-frequency coefficients retained, thereby improving the signal-to-noise ratio and obtaining a partial discharge signal with waveform fidelity.

[0069] It should be noted that wavelet transform denoising has significant advantages in partial discharge signal processing. Its multi-scale analysis capability and time-frequency localization characteristics can effectively separate noise from effective signals, preserve waveform fidelity, and balance time-frequency resolution, providing a high-quality data foundation for subsequent discharge pattern recognition, localization, and health assessment.

[0070] In A3: Kalman filtering is used to remove noise from the gas component data;

[0071] Specifically, Kalman filtering is used to remove noise from gas concentration data. In partial discharge monitoring, the concentration data of gas components are easily affected. Kalman filtering can effectively eliminate noise and improve data reliability, providing high-quality input for subsequent multimodal fusion and improving the accuracy, stability and real-time performance of gas concentration data. The noise suppression of Kalman filtering improves the signal-to-noise ratio and reduces measurement error. Improved compatibility makes standardization reasonable and multimodal fusion friendly.

[0072] In A4: Normalization and feature extraction are combined to output partial discharge signal features and gas composition features;

[0073] Specifically, segmented normalization is used to process partial discharge signals to reduce the impact of differences in discharge intensity. Standard deviation is used to standardize the gas components. Feature extraction is performed on the processed data to improve the data classification efficiency and anti-interference ability, and output the partial discharge signal features and gas component features.

[0074] Specifically, the extracted features include time-domain features, frequency-domain features, and time-frequency-domain features.

[0075] It should be noted that normalization improves stability and generalization ability by eliminating dimensional differences and data distribution bias. Standard deviation standardization unifies data from different gas sensors into dimensionless data, reducing the bias towards high-dimensional features caused by magnitude differences. Piecewise normalization eliminates the influence of different discharge intensities on features. Extraction of time-domain features improves the accuracy of pulse characteristic characterization and optimizes real-time performance. Extraction of frequency-domain features reduces the misjudgment rate of frequency band energy distribution and strengthens the anti-aliasing interference capability. Joint time-frequency domain analysis improves the accuracy of non-stationary signal analysis and positioning. Through the synergistic optimization of normalization and feature extraction, the accuracy, efficiency, and robustness of the partial discharge detection system are improved.

[0076] In this embodiment of the application, the above step S200, after fusing the partial discharge signal features and gas component features using a first fusion neural network, outputs the partial discharge mode type, including the following sub-steps B1 to B4:

[0077] In B1: Partial discharge signal features and gas component features are used as nodes of the graph, and the similarity and interaction between features are captured by edge weights to construct a multimodal graph structure;

[0078] Specifically, partial discharge signal features and gas component features are used as nodes in a graph. Each node contains a feature vector, and an initial graph structure containing the two types of nodes is constructed. By utilizing the similarity and interaction relationships between nodes, connecting edges between nodes are defined, resulting in a graph structure containing nodes and edges. The edge weights between nodes are calculated using cosine similarity based on the similarity and interaction relationships between node features, and a graph structure containing weight values ​​is output. Threshold filtering is used for optimization to remove redundant edges and reduce graph complexity. Physical constraints are then introduced to optimize edge weights and enhance the rationality of the graph structure. Finally, the node features are standardized to eliminate dimensional differences, and an optimized multimodal graph structure is output.

[0079] It should be noted that the node definition realizes a unified representation of multimodal data, while preserving the original feature information of partial discharge signals and gas components, thus ensuring data integrity. The edge definition expresses the correlation between features through edge weights, enhancing the semantic information of the graph structure, reducing redundant connections, and improving the computational efficiency of the graph structure. Edge weight calculation enhances the expressive power of the graph structure, captures the nonlinear relationships between multimodal features, and optimizes the graph structure, improving its computational efficiency and robustness while reducing computational complexity.

[0080] In B2: The interaction weights of nodes are dynamically calculated using the first attention mechanism, and an interaction weight matrix is ​​generated; the specific steps include:

[0081] A multimodal graph structure was constructed using the characteristics of partial discharge signals and gas composition.

[0082] Calculate the first attention weight and the second attention weight of the multimodal graph structure;

[0083] An interaction weight matrix is ​​generated by combining the first attention weight and the second attention weight.

[0084] Specifically, the partial discharge signal features and gas component features are defined as nodes. Each node of the discharge signal node contains a phase-resolved partial discharge feature vector within a time window. Each gas component node corresponds to one node for each sensor, and the feature vector is a statistical measure of the multi-component concentration time series data. A modality-specific encoder is used to map the features to a unified dimension, and the aligned node feature matrix is ​​output to construct a multimodal graph structure.

[0085] Specifically, multi-head self-attention calculation is performed on the multimodal graph structure to calculate the intramodal attention weights, i.e., the first attention weights; combined with cross-modal cross-attention, a dynamic correlation between the signal and gas features is established, and the cross-modal attention weights, i.e., the second attention weights, are output.

[0086] In this embodiment, the expression for the dynamic attention weight optimization function is as follows:

[0087] W′=W+α·tanh(G);

[0088] Where W′ represents the optimized weight matrix, W is the initial weight matrix, α represents the adaptive step size coefficient, α∈((0,0.5], α determines the magnitude of each weight adjustment, takes a smaller value in the early stage of training and gradually increases in the later stage, tanh is the hyperbolic tangent function, which maps the original gradient to a reasonable range and preserves the gradient direction information, and G represents the gradient of the loss function with respect to the weights, G∈(-1,1).

[0089] In an optional embodiment, the first attention weight can also be a graph attention mechanism based on physical constraints. This mechanism constructs a physical constraint matrix by introducing prior knowledge of discharge physics and jointly optimizes it with data-driven attention weights. Specifically, the correlation strength between modes is first predefined based on the typical physical characteristics of partial discharge types, and then the residual weights between the actual data and the physical model are learned through a differentiable graph network.

[0090] In another optional embodiment, the first attention weight can also be based on spatiotemporally synchronized dynamic graph convolution, which explicitly models the spatiotemporal correlation of sensor nodes as a spatiotemporal graph structure, and synchronously captures the spatial dependence and temporal evolution of features through a spatiotemporal graph convolutional network. Specifically, the temporal dynamic patterns of node features are extracted using a temporal convolution module, and the real-time interaction weights between nodes are calculated in combination with a spatial graph attention module. Finally, the spatiotemporal weights are fused through a gating mechanism.

[0091] In this embodiment of the application, the expression for the cross-modal attention weight function is as follows:

[0092] w = max(0,r)·g(d);

[0093] Where w represents the final weight value of the cross-modal edge, with a value range of [0,+∞), r is the correlation coefficient between the discharge signal and the gas characteristics, with a value range of [-1,1], g(d) represents the spatial decay function, which monotonically decreases with increasing distance, with a value range of (0,1], d is the spatial distance between sensor nodes, in meters, with a value range of [0,+∞), and max(0,r) ensures that the weights are non-negative, and negatively correlated features are automatically filtered out.

[0094] In an optional embodiment, the second attention weight can also be based on the energy transfer efficiency between sensor nodes by establishing a coupled physical model of partial discharge energy propagation and gas diffusion. Specifically, firstly, the propagation loss coefficient of the UHF signal in the cabinet is calculated based on the electromagnetic wave attenuation model. Then, the concentration gradient change rate of the gas diffusion equation is combined to construct an energy-concentration joint transfer matrix. Finally, the deviation weight between the physical model and the actual observation is dynamically adjusted through learnable parameters.

[0095] In another alternative embodiment, the second attention weight can also be introduced into a generative adversarial network framework, where a discriminator dynamically evaluates the spatial distribution consistency of different modal features to generate domain-invariant attention weights. Specifically, a generator is used to simulate the distribution of multimodal features under ideal, interference-free conditions, while the discriminator quantifies the difference between the actual data and the ideal distribution using an adversarial loss function, and finally normalizes the inverse of the difference into cross-modal attention weights.

[0096] It should be noted that by strengthening the feature dependencies within the same modality through the self-attention mechanism, the generated attention weights can intuitively show which modalities and features contribute the most to the diagnostic results, assisting maintenance personnel in analyzing the fault mechanism. When sensors fail or data is missing, the attention mechanism can automatically adjust the dependency path to maintain system reliability.

[0097] In this embodiment of the application, the intramodal self-attention weights and cross-modal attention weights are dynamically fused, the fusion coefficients are calculated and the weights are fused to output a preliminary fused weight matrix, which is then combined with the physical rule base for weight correction, and K-value truncation sparsity and symmetric normalization are used to generate an interactive weight matrix.

[0098] It should be noted that the expression for the interaction weight matrix generation function is as follows:

[0099] R = a·M + (1-a)·B

[0100] Where R represents the final interaction weight matrix with a value range of [0,1], M is the intramodal attention weight matrix, reflecting the correlation strength between features of the same modality with a value range of [0,1], B represents the cross-modal attention weight matrix, reflecting the interaction relationship between features of different modalities with a value range of [0,1], a is the intramodal weight adjustment coefficient with a value range of (0,1), and 1-a represents the cross-modal weight adjustment coefficient with a value range of (0,1), which is complementary to a and ensures that the total weight is 1.

[0101] It should be noted that the application of physical rules prevents the generation of weights that violate physical laws through pure data-driven methods, improves generalization in scenarios with few samples, sparsity and normalization, reduces computational complexity and improves numerical stability, avoids gradient explosion and vanishing, and accelerates convergence; dynamically adjusts the interaction weights between nodes according to the data distribution, enhances adaptability to complex scenarios, highlights key features through attention mechanisms, suppresses noise or redundant information, improves the accuracy of feature fusion, and utilizes multi-attention head mechanisms to improve robustness and capture more complex feature relationships.

[0102] In an optional embodiment, the first attention mechanism can also synchronously capture the spatiotemporal evolution of node features through a spatiotemporal graph convolutional network, wherein the temporal convolution module extracts temporal dynamic patterns, and the spatial graph attention module calculates the real-time interaction weights between nodes. This mechanism is particularly suitable for processing non-uniformly sampled multimodal data and can effectively model the spatiotemporal drift phenomenon in sensor networks.

[0103] In another alternative embodiment, the first attention mechanism can also be a computational framework for attention based on physical constraints, combining fluid dynamics equations with neural networks. First, theoretically correlated weights are pre-calculated based on the gas diffusion equation and electromagnetic wave propagation model. Then, the physical model is embedded into the attention layer using differentiable programming techniques, ultimately outputting a hybrid weight that integrates data-driven and physical laws.

[0104] In B3: The partial discharge signal features and gas component features are fused based on the interaction weight matrix, and the fused features are output.

[0105] Specifically, the partial discharge signal features and gas component features are weighted to highlight key features, suppress noise or redundant information, enhance the interaction between features, and improve the fusion effect. The weighted features are then spliced ​​together to obtain preliminary fusion features, which are then optimized to enhance the nonlinear expressive power of the features, capture more complex feature relationships, and standardize the optimized fusion features to output the final fusion features.

[0106] It should be noted that feature fusion fully utilizes the high-frequency characteristics of partial discharge signals and the spatial distribution characteristics of gas components to improve the integrity of feature representation. Weighting, splicing, and optimization techniques are used to reduce feature dimensionality, remove noise, and improve computational efficiency and generalization ability. Nonlinear transformation is used to enhance the representational ability of fused features and capture more complex feature relationships.

[0107] In B4: The fused features are input into the fully connected layer of the first fused neural network, and the partial discharge mode type is output.

[0108] Specifically, the process involves checking whether the feature dimensions match the input requirements of the classifier, ensuring the integrity and consistency of the input data, preserving key information of the fused features, improving classification accuracy, using a fully connected layer to perform a linear transformation on the fused features to obtain the mapped feature vector, using the Softmax function to calculate the probability of each category, outputting the probability distribution of partial discharge modes, selecting the category with the highest probability value as the final classification result, and outputting the partial discharge mode type.

[0109] It should be noted that when checking whether the dimensional features do not match the classifier's input, dimensionality adjustment is required. Adjustment methods include padding and truncation to adjust feature dimensions, which can quickly achieve dimensionality matching, retain key feature information, and reduce data loss. Dimensionality reduction techniques are used to map high-dimensional features to low-dimensional space, and dimensionality increase techniques are used to map low-dimensional features to high-dimensional space. This achieves optimized adjustment of feature dimensions, improves computational efficiency, retains key feature information, and enhances classification performance.

[0110] Specifically, the expression for the Softmax function is:

[0111]

[0112] Among them, P i This represents the probability value of the i-th category, where 0 ≤ P. i ≤1, x i Let x be the input feature value for the i-th category. i ∈R, Indicates x i Performing exponentiation maps real numbers to the range of positive numbers. This represents the summation of the index values ​​over all categories, used as the normalized denominator. C is the total number of categories, C≥2, when x i When it is large, It will increase significantly; when x i When smaller, It will approach 0, where i represents the index of the category currently being calculated, and j is the traversal variable when summing.

[0113] In an optional embodiment, the first fusion neural network can also employ a cross-modal Transformer architecture, processing partial discharge signals and gas component feature sequences in parallel through multi-head attention layers. A modality-specific token embedding layer is designed in the encoder section to map heterogeneous features to a unified semantic space; a cross-attention mechanism is introduced in the decoder section to establish a dynamic correlation between the signal's time-frequency features and the spatial distribution of the gas. Finally, a gated feature pyramid is used to achieve multi-scale feature aggregation, outputting the fused discriminative features.

[0114] In another alternative embodiment, the first fusion neural network can also model the feature fusion process as a continuous-time dynamical system, learning the optimal fusion trajectory through neural ordinary differential equations. Specifically, the initial features are used as the system input state, and the differential equations are parameterized using a convolutional GRU network to continuously evolve the feature interactions in the latent space, ultimately outputting a stable state as the fusion result.

[0115] It should be noted that step S200 above overcomes the limitations of traditional linear fusion methods by modeling the nonlinear correlation between different modal features through dynamic graph structure modeling, thereby improving the accuracy of discharge mode classification by more than 15%, and significantly enhancing the ability to distinguish composite discharges.

[0116] In this embodiment of the application, the above step S300, which constructs a spatial three-dimensional distribution map of gas component concentrations using a second neural network, includes the following sub-steps C1 and C2:

[0117] In C1: Using the gas component sensor nodes as vertices, edges are defined by spatial positional relationships to construct a graph network structure;

[0118] Specifically, multiple gas component sensors are arranged inside or around the switch cabinet, with each sensor serving as a node. The sensor positions are optimized based on the physical structure of the switch cabinet and the areas where partial discharge may occur, generating a set of spatial coordinates for the sensor nodes. Edges are defined based on the spatial relationships between the sensor nodes. By defining edges, the spatial relationships between the sensor nodes can be effectively modeled. The sensor nodes and edges are combined to generate a graph network structure, which can effectively represent the complex spatial relationships between the sensor nodes.

[0119] It should be noted that by rationally arranging sensor nodes, the data acquisition is comprehensive and representative. Based on the defined edges and optimized graph network structure, the spatial relationships between sensor nodes are accurately modeled. The use of graph network structure can efficiently represent complex spatial relationships, providing a foundation for subsequent data fusion and prediction.

[0120] In C2: The second neural network is used to fuse gas composition data and spatial relationships to predict the gas composition of unsampled areas in order to construct a three-dimensional spatial distribution map;

[0121] Specifically, a graph neural network algorithm is used to fuse node features and spatial relationships. Through multi-layer graph convolution operations, the features of neighboring nodes are aggregated, node representations are updated, and gas component data in unsampled areas are predicted. The predicted gas components in unsampled areas are output. The Kriging interpolation algorithm is used to interpolate the three-dimensional concentration distribution data. Considering the spatial autocorrelation of the data, the interpolation results are smoothed to reduce the impact of noise and outliers, resulting in optimized three-dimensional concentration distribution data. A heatmap generation tool is used to convert the three-dimensional concentration distribution data into a heatmap. The concentration distribution of gas components is represented by color gradients, generating a spatial three-dimensional distribution map.

[0122] It should be noted that graph neural networks can effectively capture the nonlinear relationships and spatial dependencies between nodes, significantly improving prediction accuracy. By predicting the gas composition in unsampled areas, they can generate complete three-dimensional concentration distribution data, comprehensively reflecting the spatial distribution of gas components. Kriging interpolation can generate smooth and continuous concentration distribution data, improving data quality. The resulting data can more accurately reflect the true spatial distribution of gas components. Heat maps can intuitively display the spatial distribution of gas components, facilitating the rapid identification of high-concentration areas and potential fault points. Three-dimensional heat maps can provide more comprehensive spatial information, supporting multi-angle analysis and decision-making.

[0123] It should be noted that the color mapping scheme of the heatmap should be optimized according to the specific application scenario to improve the visualization effect. The 3D heatmap can be dynamically viewed and operated through interactive tools to enhance the user experience.

[0124] In an optional embodiment, the second neural network can also be a voxelized prediction network based on a 3D convolutional neural network, which divides the switch cabinet space into a regular voxel grid and directly processes the gas sensor data through a 3D sparse convolutional neural network. First, discrete sensor readings are interpolated to neighboring voxels as initial input, and then a 3D U-Net structure is used for multi-scale feature extraction and upsampling, finally outputting a complete three-dimensional concentration field.

[0125] In another alternative embodiment, the second neural network can also be an implicit representation network based on neural radiation fields, which implicitly models the gas concentration distribution function using neural radiation field technology. Using the sensor location coordinates and observation direction as input, the MLP network predicts the gas concentration value at the corresponding location, and a differentiable rendering algorithm is introduced to generate a continuous three-dimensional distribution.

[0126] It should be noted that step S300 above uses a graph neural network to embed the spatial topological relationship of sensor nodes and introduces fluid dynamic constraints to optimize the interpolation process, thereby reducing the concentration prediction error in unsampled areas and solving the problem of distribution map distortion caused by ignoring spatial correlation in traditional methods.

[0127] In this embodiment, step S400, based on the mode type and spatial three-dimensional distribution map of partial discharge, minimizes the spatial difference between the partial discharge signal and the gas component concentration using a first fusion algorithm to obtain the precise location of the partial discharge source, including the following sub-steps D1 to D3:

[0128] In D1: Dynamic graph embedding optimization is used to dynamically embed partial discharge signals and gas component concentrations into the graph structure;

[0129] Specifically, the discharge signal is decomposed in the time-frequency domain to extract energy distribution features of multiple frequency bands; the change rate of gas concentration data is calculated to characterize diffusion dynamics; the two types of features are integrated into a unified joint feature matrix according to time windows, preserving the physical meaning and spatial location information of the original data, and outputting a spatiotemporally aligned multimodal feature matrix; a sliding window mechanism is used to extract data segments at fixed durations, covering the typical cycle of partial discharge, and the three-dimensional coordinates of each sensor in space are bound to the feature values ​​within the corresponding time window to form the initial attributes of graph nodes; the initial connection weights are calculated based on the physical distance between sensors, with closer nodes having higher connection strength, and a dynamic graph structure with node attributes and initial connection weights is output; the temporal change law of node features is analyzed through a dynamic adjustment module, the hidden state of nodes is updated, and the connection weights are dynamically adjusted according to the similarity of node states, with connections between nodes with high similarity being strengthened and those between nodes with low similarity being weakened; redundant connections are eliminated using threshold filtering, key topological relationships are preserved to reduce computational complexity, and a real-time updated dynamic graph structure is output.

[0130] It should be noted that by optimizing the dynamic graph embedding, the dynamic adaptability and spatial correlation are enhanced, the coupling changes of the discharge process and gas diffusion are responded to in real time, and the physical and logical relationships between sensors are explicitly expressed through the graph structure, thereby improving the rationality of cross-modal data fusion.

[0131] In D2: Adaptive graph convolutional networks are combined to capture the spatial correlation of multimodal data and output optimized node features;

[0132] Specifically, node attributes are mapped to a high-dimensional latent space through a fully connected layer to generate initial node feature vectors. A multi-head attention mechanism is used to calculate the similarity scores between nodes. The multi-head attention scores are weighted and fused to generate an adaptive adjacency matrix. The matrix elements represent the logical strength of the connections between nodes, dynamically reflecting the potential associations of multimodal data. The adaptive adjacency matrix is ​​output, and the node features are combined with the adaptive graph convolutional network to output optimized node features.

[0133] It should be noted that the construction of the adaptive adjacency matrix breaks through the limitation of the traditional adjacency matrix relying on physical distance, and can capture the complex coupling relationship between discharge signal and gas concentration.

[0134] In D3: Based on the dynamic graph structure and optimized node characteristics, the spatial difference between the partial discharge signal and the gas component concentration is minimized to output the precise location of the partial discharge source;

[0135] Specifically, by combining the dynamic graph structure and the time-frequency domain special discharge signal spatial mapping of the partial discharge signal, a three-dimensional spatial distribution map of signal strength and spatial characteristics and gas component concentration is established. The correlation between the locations is used, and the gas diffusion reverse analysis is combined with the sensor concentration gradient data to reversely deduce the gas diffusion path and generate a predicted location heat map of the gas source. The discharge signal and the gas diffusion heat map are spatially aligned, and the difference value of each grid point is calculated to generate a spatial difference matrix. The weights are adjusted according to the real-time status of the sensor, and a multi-objective joint optimization function is output. The joint optimization function and the three-dimensional structure of the switch cabinet are used to perform global coarse and local fine-tuning through a staged search to output the precise location of the partial discharge source.

[0136] It should be noted that the spatial difference matrix integrates the physical laws of electromagnetic and chemical modes to improve the spatial consistency of positioning results. The adaptive weighting mechanism improves the positioning stability under complex working conditions. By complementing the physical laws of discharge signals and gas diffusion, the positioning bottleneck of a single data source is reduced. The real-time linkage of weight allocation remains robust when sensors malfunction or the environment changes abruptly.

[0137] It should be noted that the expression for the multi-objective joint optimization function is:

[0138] L(u,v,w)=λ·T+β·G+μ·D;

[0139] Where L(u,v,w) represents the total loss function, the input is the candidate coordinates (u,v,w), and the output is the scalarized comprehensive difference value. λ is the weight coefficient of the partial discharge signal matching term. T represents the discharge signal error range [0,+∞), which represents the absolute error between the measured value and the predicted value. The larger the value, the worse the discharge signal matching. β is the weight coefficient of the gas concentration matching term. G represents the gas concentration error, with a range of [0,+∞). It represents the absolute error between the measured value and the predicted value. The larger the value, the worse the gas concentration matching. μ is the penalty coefficient of the physical constraint term. D represents the physical constraint penalty value, with a range of [0,+∞). It represents the distance deviation between the candidate coordinates and the nearest physical restricted area. The larger the value, the less safe the location.

[0140] It should be noted that when T=0, the measured value of the discharge signal matches the predicted value perfectly, indicating that the candidate coordinates are most likely the discharge source. When G=0, the measured value of the gas concentration matches the predicted value perfectly, indicating that the candidate coordinates are most likely the discharge source. When D=0, the candidate coordinates fully comply with the physical constraints and are located within the safe area.

[0141] In an optional embodiment, the first fusion algorithm may further model the partial discharge signal intensity and gas concentration distribution as a continuous neural radiation field, implicitly learning the electromagnetic-chemical coupling field function through a multilayer perceptron. Specifically, the input spatial coordinates (u,v,w) output the discharge signal probability and gas concentration prediction value at that location, and the radiation field parameters are jointly optimized through differentiable rendering. Finally, the extreme point of the field function is directly located as the discharge source location using the gradient ascent method.

[0142] In another optional embodiment, the first fusion algorithm can also be a multi-task adaptive fusion based on meta-learning. A meta-learning framework is constructed to decompose the localization task into two sub-tasks: discharge pattern classification and gas source inversion. The task weights are dynamically adjusted through a meta-optimizer, where the pattern classification network provides a prior probability distribution, and the gas inversion network outputs a probability heatmap; the two are fused using a Bayesian inference framework.

[0143] It should be noted that the above step S400 integrates a multi-physics coupling model and achieves sub-meter level precise positioning of the discharge source through an iterative optimization algorithm, and is more adaptable to complex cabinet structure scenarios.

[0144] In this embodiment of the application, the above step S500, which assesses the health status of the switchgear based on precise positioning and outputs a health status assessment report, includes the following sub-steps E1 to E3:

[0145] In E1: the health status of the switchgear can be divided into normal, slightly abnormal and seriously abnormal;

[0146] In E2: The health status of the switchgear is classified by combining the fusion features, and the health status level is output. When the health status is slightly abnormal or seriously abnormal, an alarm is triggered and information is pushed.

[0147] Specifically, partial discharge characteristics, gas composition characteristics, and environmental parameters are fused together, and the fused characteristics are classified to output a health status level. Through multimodal feature fusion, the operating status of the switchgear is reflected in all aspects, improving classification accuracy and reducing misjudgments caused by single features. The health level is divided into normal, slightly abnormal, and severely abnormal. When the health status is slightly abnormal or severely abnormal, an alarm is triggered and information is pushed. By realizing graded early warning, it helps operation and maintenance personnel to take timely and targeted measures, reduce the escalation of faults, and improve the safety of equipment operation.

[0148] In E3: Combining the Analytic Hierarchy Process (AHP) and Fault Tree Analysis, a health status assessment report is output;

[0149] Specifically, the weights of each layer are determined using the analytic hierarchy process (AHP), and a three-layer assessment system of electrical, chemical, and environmental layers is established. Fault tree analysis is used to calculate the minimum cut set and importance index, and outputs a fault probability ranking list and a risk heat map of key components. Combined with embedded visualization components, report templates, and support for automatic generation in multiple languages, a health status assessment report is output.

[0150] It should be noted that the health status assessment report includes the current health status assessment level, a detailed table of parameters exceeding limits, a maintenance recommendation list, and a report template. The establishment of this three-tiered assessment system addresses the limitations of single-parameter assessments. By locating the cause through component correlation diagrams, the mean time to failure can be shortened. Minimum cut set analysis identifies system vulnerabilities, allowing for the early replacement of high-risk components. Furthermore, the automatically generated report includes specific maintenance recommendations, reducing the workload of manual analysis.

[0151] It should be noted that, based on the precise positioning results, step S500 generates a switchgear health status assessment report by comprehensively considering parameters such as discharge intensity, mode risk level, and gas corrosion rate. A dynamic weight allocation model quantifies the degree of insulation degradation, outputting a standardized report that includes risk threshold warnings and maintenance priority recommendations. This shortens the response time for operation and maintenance decisions and supports the precise implementation of predictive maintenance strategies.

[0152] Example 3: This example provides a partial discharge location system for a switchgear, comprising:

[0153] The data acquisition module is used to acquire partial discharge data of the switchgear in real time and perform the first processing, outputting partial discharge signal characteristics and gas composition characteristics;

[0154] The feature fusion module is used to fuse the partial discharge signal features and gas component features using the first fusion neural network, and then output the mode type of partial discharge.

[0155] The distribution modeling module is used to construct a spatial three-dimensional distribution map of gas component concentrations through a second neural network.

[0156] The precise positioning module is used to obtain the precise location of the partial discharge source by minimizing the spatial difference between the partial discharge signal and the gas component concentration through a first fusion algorithm based on the mode type and spatial three-dimensional distribution map of the partial discharge.

[0157] The status assessment module is used to assess the health status of the switchgear based on precise location and output a health status assessment report.

[0158] It should be noted that the technical solution of the switchgear partial discharge positioning system and the technical solution of the switchgear partial discharge positioning method described above belong to the same concept. For details not described in detail in the technical solution of the switchgear partial discharge positioning system in this embodiment, please refer to the description of the technical solution of the switchgear partial discharge positioning method described above.

[0159] The above-mentioned unit modules can be embedded in the processor of the electronic device in hardware form or independent of it, or they can be stored in the memory of the electronic device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0160] This embodiment also provides an electronic device, which includes a processor, a memory, a communication interface, a display screen, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a method for locating partial discharge in a switchgear. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0161] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method proposed in the above embodiments.

[0162] The storage medium proposed in this embodiment belongs to the same inventive concept as the method proposed in the above embodiments. 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.

[0163] 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, but in many cases the former is a better implementation method. 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 an electronic device (which may be a personal computer, server, or network device, etc.) to execute the method of the embodiments of the present invention.

[0164] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to 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 locating partial discharge in a switchgear, characterized in that, include: Real-time acquisition of partial discharge data from the switchgear and initial processing are performed to output partial discharge signal characteristics and gas composition characteristics. After fusing the partial discharge signal features and gas component features using a first fusion neural network, the partial discharge mode type is output. A spatial three-dimensional distribution map of gas component concentrations is constructed using a second neural network; Based on the mode type and spatial three-dimensional distribution map of the partial discharge, the spatial difference between the partial discharge signal and the gas component concentration is minimized by the first fusion algorithm to obtain the precise location of the partial discharge source. The switchgear's health status is assessed based on the precise positioning, and a health status assessment report is output.

2. The method for locating partial discharge in a switchgear as described in claim 1, characterized in that, The output partial discharge mode types include: Using the partial discharge signal features and gas component features as nodes in a graph, and capturing the similarity and interaction relationships between features through edge weights, a multimodal graph structure is constructed. The interaction weights of nodes are dynamically calculated using the first attention mechanism, and an interaction weight matrix is ​​generated. The partial discharge signal features and gas component features are fused based on the interaction weight matrix to output fused features; The fusion features are input into the fully connected layer of the first fusion neural network, and the partial discharge mode type is output.

3. The method for locating partial discharge in a switchgear as described in claim 2, characterized in that, The generated interaction weight matrix includes: Calculate the first attention weight and the second attention weight of the multimodal graph structure; An interaction weight matrix is ​​generated by combining the first attention weight and the second attention weight.

4. The method for locating partial discharge in a switchgear as described in claim 3, characterized in that, The construction of the spatial three-dimensional distribution map of gas component concentrations via the second neural network includes: A graph network structure is constructed by defining edges based on the spatial positional relationships of gas component sensor nodes; By using a second neural network to fuse gas component data and spatial relationships, the gas components of unsampled areas are predicted to construct a three-dimensional spatial distribution map.

5. The method for locating partial discharge in a switchgear as described in claim 4, characterized in that, The precise location of the partial discharge source includes: Dynamic graph embedding optimization is used to dynamically embed the partial discharge signal and gas component concentration into the graph structure; By combining adaptive graph convolutional networks to capture the spatial correlation of multimodal data, optimized node features are output. Based on the dynamic graph structure and the optimized node characteristics, the spatial difference between the partial discharge signal and the gas component concentration is minimized, and the precise location of the partial discharge power source is output.

6. The method for locating partial discharge in a switchgear as described in claim 5, characterized in that, The assessment of the switchgear health status based on the precise positioning includes: The health status of the switchgear can be divided into normal, slightly abnormal, and severely abnormal. The health status of the switchgear is classified by combining the fusion features, and the health status level is output. When the health status is slightly abnormal or seriously abnormal, an alarm is triggered and information is pushed. By combining the analytic hierarchy process (AHP) and fault tree analysis, a health status assessment report is generated.

7. The method for locating partial discharge in a switchgear as described in claim 6, characterized in that, The real-time acquisition and first processing of partial discharge data from the switchgear includes: Real-time acquisition of partial discharge signals and gas composition; Wavelet transform denoising method is used to remove high-frequency noise and low-frequency interference from partial discharge signals; Kalman filtering is used to remove noise from gas component data; The partial discharge signal features and gas component features are output by combining normalization and feature extraction.

8. A partial discharge locating system for a switchgear, employing the partial discharge locating method for a switchgear as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to acquire partial discharge data of the switchgear in real time and perform the first processing, outputting partial discharge signal characteristics and gas composition characteristics; The feature fusion module is used to fuse the partial discharge signal features and gas component features using a first fusion neural network, and then output the partial discharge mode type. The distribution modeling module is used to construct a spatial three-dimensional distribution map of gas component concentrations through a second neural network. The precise positioning module is used to obtain the precise location of the partial discharge source by minimizing the spatial difference between the partial discharge signal and the gas component concentration through a first fusion algorithm based on the mode type and spatial three-dimensional distribution map of the partial discharge. The status assessment module is used to assess the health status of the switchgear based on the precise location and output a health status assessment report.

9. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store computer-executable instructions, and when the processor executes the computer-executable instructions, it implements the steps of the partial discharge positioning method for switchgear as described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer-executable instructions stored thereon, characterized in that: When the computer-executable instructions are executed by the processor, they implement the steps of the partial discharge location method for switchgear as described in any one of claims 1 to 7.

Images