High-voltage GIS partial discharge real-time diagnosis method and system

Abstract

The invention relates to the field of power systems, in particular to a high-voltage GIS partial discharge real-time diagnosis method and system. The method comprises the following steps: preprocessing a video stream of the optical fiber endoscope, and extracting center coordinates of a discharge light spot; constructing a dynamic space-time diagram, and establishing a space edge, a time edge and a physical edge associated with the electric field intensity by taking a light spot coordinate as a vertex; inputting the dynamic space-time diagram into a double-path space-time diagram convolution network for feature extraction, wherein the network is jointly processed through a main path containing space-time diagram convolution and an auxiliary path containing a bottleneck residual unit; weighting the features by using an attention mechanism, and focusing a key feature region; and finally, synchronously outputting defect classification, severity evaluation and a risk distribution diagram through a multi-task output layer. According to the method, the problem of insufficient identification precision caused by dynamic modeling deficiency, physical field disjunction and noise interference in the prior art is solved, and the accurate analysis capability of the discharge dynamic behavior is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of power systems, and specifically to a method and system for real-time diagnosis of partial discharge in high-voltage GIS. Background Technology

[0002] In the field of partial discharge detection in high-voltage gas-insulated switchgear (GIS), existing technologies mainly rely on non-optical methods such as ultra-high frequency (UHF) and ultrasonic methods. For example, existing technology CN120370147A proposes a "partial discharge localization method, device, electronic device and storage medium for GIS". It establishes an electromagnetic attenuation model of GIS and an UHF partial discharge signal source localization model, and uses the signal amplitude of adjacent measuring points to obtain the centroid similarity coefficient of the spectrum for localization, aiming to improve the localization accuracy and compensate for the high equipment requirements of time-of-flight localization. However, the detection effect of this type of detection method based on UHF electromagnetic signals is significantly affected by sensor characteristics, defects and paths between sensors, and the typical structural morphology of GIS, making it difficult to effectively perceive the visual morphological characteristics of discharge and its dynamic evolution process.

[0003] On the other hand, the application of optical detection methods, especially fiber optic endoscopy technology, has made it possible to directly observe the light spots generated by discharges inside GIS equipment, thus compensating for the limitations of non-optical methods to some extent. However, corresponding image analysis methods still have significant shortcomings. Existing models based on static image recognition or conventional convolutional neural networks (CNN) / recurrent neural networks (RNN) typically focus on the analysis of single-frame images, lacking the ability to dynamically model the spatiotemporal propagation trajectory of discharge light spots in continuous multi-frame videos. This results in the system's inability to effectively correlate the motion law of the light spot in the time dimension with the physical process of discharge, leading to limited accuracy in identifying defects with typical dynamic characteristics such as creepage along the insulator surface, and a high false positive rate. In addition, noise such as metallic reflection in strong electric field environments also seriously interferes with the accurate extraction of effective discharge light spot features at the micron level. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time diagnostic method and system for partial discharge in high-voltage GIS, so as to solve the problem of insufficient accuracy in identifying dynamic defects in partial discharge caused by the lack of spatiotemporal dynamic modeling capability of discharge spot, failure to effectively correlate physical field distribution, and severe noise interference under strong electric field environment.

[0005] To achieve the above objectives, the following technical solution is adopted.

[0006] A real-time diagnostic method for partial discharge in high-voltage GIS includes the following steps: The video stream captured by the fiber optic endoscope is preprocessed, and the center coordinates of the discharge spot in the video frames are extracted through multispectral filtering and adaptive threshold segmentation. A dynamic spatiotemporal graph is constructed based on these center coordinates, where the vertices are the center coordinates of the spot, and the edges include spatial, temporal, and physical edges. Spatial edges are established based on the spatial proximity of the spot, temporal edges are established based on the movement of the spot between adjacent frames, and physical edges are established by associating finite element electric field intensity data. The dynamic spatiotemporal graph is then input into a dual-path spatiotemporal graph convolutional network for feature extraction. This network processes data through a main path containing spatiotemporal graph convolutions and an auxiliary path containing bottleneck residual units, outputting deep spatiotemporal features. These deep spatiotemporal features are then input into an attention mechanism module for feature weighting. This module focuses on key feature regions through channel attention and spatial attention. The weighted features are then input into a multi-task output layer, simultaneously outputting defect classification results, severity assessment values, and discharge risk distribution maps.

[0007] Optionally, the preprocessing step specifically includes: separating the discharge spot signal of the first band and the metal reflection noise signal of the second band from the high-resolution video stream captured by the fiber optic endoscope at a frame rate that meets the requirements of real-time processing through frequency domain filtering; performing an adaptive threshold segmentation algorithm on the separated spot signals to calculate the pixel-level center coordinates of all spots in each frame image; the first band corresponds to the emission wavelength range of the discharge spot, and the second band corresponds to the dominant wavelength range of the metal reflection light.

[0008] Optionally, in the construction of the dynamic spatiotemporal graph, the connection rule of the spatial edge is: when the Euclidean distance between the center coordinates of any two light spots in the same frame is less than or equal to a predetermined spatial threshold, a spatial edge is established between the vertices corresponding to these two light spots to characterize the local spatial aggregation characteristics of the discharge region. The construction of the time edge involves tracing the center coordinates of the same light spot between adjacent frames, calculating its displacement vector (Δx, Δy), and deriving the motion velocity v based on this displacement vector. The calculation formula is v = √(Δx / Δy). 2 +Δy 2 ), used to describe the propagation dynamics of the discharge spot in the time dimension; The physical edge is constructed by querying a pre-stored electric field intensity distribution mapping table obtained based on finite element simulation, obtaining the electric field intensity values ​​E(vi) and E(vj) at graph vertices vi and vj, and calculating their absolute difference ΔE=|E(vi)-E(vj)| as the attribute of the edge, thereby associating the dynamic behavior of the light spot with the underlying physical field distribution.

[0009] Optionally, in the dual-path spatiotemporal graph convolutional network, the processing flow of the bottleneck residual unit includes: The number of channels in the input feature map is compressed to a predetermined ratio of the original number of channels using a 1×1 convolutional layer; The compressed features are then aggregated using a spatiotemporal graph convolutional layer. Then, a 1×1 convolutional layer is used to restore the number of feature channels to the original dimension; The recovered features are residually concatenated with the input features of the unit; the calculation process of the bottleneck residual unit is as follows: In the formula This represents the input features of the l-th layer. This represents the weight parameters of the trainable spatiotemporal graph convolution kernel. This represents a non-linear activation function.

[0010] Optionally, in the convolutional block attention module, the calculation process of the channel attention is as follows: Global average pooling and global max pooling are performed on the input feature map F respectively to obtain two different channel descriptors; Input these two descriptors into a shared multilayer perceptron; After summing the outputs of the multilayer perceptron element by element, a channel attention weight map is generated using the sigmoid activation function. Its expression is Where MLP stands for Multilayer Perceptron, AvgPool and MaxPool represent global average pooling and global max pooling, respectively, and σ represents the Sigmoid function. The spatial attention calculation process is as follows: average pooling and max pooling are performed on the input feature map F along the channel dimension to obtain two spatial feature maps; these two spatial feature maps are concatenated along the channel dimension; a standard convolutional layer is used to perform a convolution operation on the concatenated feature map; finally, a spatial attention weight map is generated through the Sigmoid activation function. .

[0011] Optionally, the multi-task output layer includes a defect classification submodule, a severity regression submodule, and a heatmap generation submodule; the defect classification submodule performs a fully connected layer transformation and Softmax activation on the attention-weighted feature vector after global average pooling compression, and outputs classification results representing the probability distribution of different defect types. Its calculation formula is ,in Indicates attention-weighted features Perform global average pooling. and The regression weights and biases are trainable, and Softmax is a normalized exponential function. The severity regression submodule performs a fully connected layer and linear activation on the same pooled feature vector, outputting a continuous scalar value representing the severity of the discharge. Its calculation formula is The heatmap generation submodule upsamples and normalizes the spatial attention map to generate a discharge risk distribution heatmap that matches the spatial size of the original input video frame.

[0012] Optionally, the defect type probability distribution output by the defect classification submodule covers categories including normal discharge, floating potential discharge, tip corona discharge, and insulator surface discharge; the continuous scalar value output by the severity regression submodule is used to quantitatively assess the severity of discharge activity, with higher values ​​representing greater potential risks; the heatmap generation submodule achieves upsampling through deconvolution and uses the Sigmoid function to normalize the output value to the [0,1] interval to generate the discharge risk distribution heatmap.

[0013] Optionally, the main path of the dual-path spatiotemporal graph convolutional network is composed of multiple spatiotemporal graph convolutional layers connected sequentially, and the number of output channels of each layer increases in a pre-set sequence; the auxiliary path contains multiple bottleneck residual units, and the output features of a specific stage in the auxiliary path are stacked and fused with the output features of the corresponding stage in the main path, and the fused result is used as the common input of subsequent network layers.

[0014] A real-time diagnostic system for partial discharge in high-voltage GIS includes: Fiber optic endoscope module, used to acquire high-resolution video data inside GIS equipment in a high-pressure SF6 gas environment; The video preprocessing module is signal-connected to the fiber optic endoscope module and is used to receive video data and perform multispectral filtering and adaptive threshold segmentation to extract the center coordinates of the light spot. The dynamic graph construction module is data-connected to the video preprocessing module and is used to construct a dynamic spatiotemporal graph containing spatial edges, temporal edges, and physical edges based on the sequence of light spot center coordinates. The feature extraction network module is data-connected to the dynamic graph construction module. It is a dual-path spatiotemporal graph convolutional network, which includes a main path and an auxiliary path with bottleneck residual units, and is used to extract deep spatiotemporal features from the dynamic spatiotemporal graph. The attention mechanism module, which is data-connected to the feature extraction network module, includes a channel attention submodule and a spatial attention submodule, and is used to adaptively weight the deep spatiotemporal features. The multi-task diagnostic output module, which is data-connected to the attention mechanism module, includes a defect classification unit, a severity regression unit, and a heatmap generation unit, and is used to output the final diagnostic results.

[0015] Optionally, the dynamic graph construction module integrates a finite element electric field intensity mapping database to provide electric field intensity difference data for the physical edges; the main path of the feature extraction network module consists of multiple spatiotemporal graph convolutional layers with increasing output channel number, and the bottleneck residual unit output of the auxiliary path is stacked with the specific layer output of the main path; the defect classification unit in the multi-task diagnostic output module adopts a fully connected layer and a Softmax activation function, the severity regression unit adopts a fully connected layer and a linear activation function, and the heatmap generation unit includes a deconvolution layer and a Sigmoid normalization layer.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This application fundamentally changes the limitations of traditional analysis methods by constructing a dynamic spatiotemporal graph that integrates multiple physical quantities and employing a dual-path deep network for feature extraction. Discrete light spot sequences are reconstructed into graph-structured data with clear physical meaning, and through the collaborative processing of a dual-path network and an attention mechanism, accurate analysis of discharge dynamics is achieved.

[0017] By constructing a dynamic spatiotemporal graph that includes spatial, temporal, and physical edges, the problem of missing dynamic modeling and disconnection from the physical field in traditional methods is effectively solved. It can simultaneously capture the spatial distribution characteristics and temporal evolution of the discharge spot, and establish a physical correlation between the spot motion and the underlying electric field distribution, significantly improving the accuracy of identifying dynamic discharge behavior. It also solves the problem of insufficient accuracy in identifying dynamic defects in partial discharge caused by the lack of spatiotemporal dynamic modeling capability for discharge spots, failure to effectively correlate with the physical field distribution, and severe noise interference in strong electric field environments.

[0018] Multispectral filtering effectively separates discharge spots from metal reflection noise, and the introduction of electric field intensity difference as a side attribute further enhances the physical interpretability and environmental adaptability of the model. The dual-path network structure, through bottleneck residual design and feature stacking fusion, ensures training stability while maintaining network depth, and improves the efficiency and robustness of feature extraction. The attention mechanism, through dual channel and spatial weighting, enables the model to adaptively focus on the feature regions most relevant to the fault and suppress irrelevant noise interference. Finally, the constructed multi-task output layer realizes accurate classification of defect types, quantitative assessment of discharge severity, and visual location of risk points, providing comprehensive technical support for operation and maintenance decisions. Attached Figure Description

[0019] Figure 1This is a flowchart illustrating the steps of an embodiment of a real-time diagnostic method for partial discharge in high-voltage GIS according to the present invention.

[0020] Figure 2 This is a schematic diagram of the BiST-GCN network structure according to an embodiment of a real-time diagnostic method for partial discharge in high-voltage GIS based on the present invention.

[0021] Figure 3 This is a schematic diagram of the bottleneck residual unit according to an embodiment of a real-time diagnostic method for partial discharge in high-voltage GIS according to the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0023] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0024] Example 1 like Figure 1 As shown, this embodiment provides a real-time diagnostic method for partial discharge in high-voltage GIS equipment. It achieves accurate diagnosis of internal partial discharge in GIS equipment through multi-stage collaborative processing. The specific process is as follows: First, the video stream captured by the fiber optic endoscope is preprocessed. This video stream is in high-resolution format, and the resolution must meet the requirements for detailed observation inside the GIS equipment, clearly displaying millimeter-level discharge spots. The preprocessing stage begins with multispectral filtering. A Fast Fourier Transform (FFT) is used to convert the video frames from the spatial domain to the frequency domain. Within the frequency domain, a filtering window is set to separate the discharge spot signal in the first band from the metallic reflection noise signal in the second band. The first band corresponds to the characteristic spectral wavelength range generated by SF6 gas ionization during partial discharge, while the second band corresponds to the reflectance spectral range of metal components (such as aluminum alloy or steel components) inside the GIS equipment under illumination. After filtering, an Inverse Fast Fourier Transform (IFFT) is used to convert the signal back to the spatial domain, achieving band separation.

[0025] Next, an adaptive threshold segmentation algorithm is applied to the separated spot signals. This can be achieved using either the maximum inter-class variance method or the adaptive local thresholding method. The segmentation threshold is automatically adjusted based on the grayscale difference between the spot and the background in each frame, accurately distinguishing the spot region from the background region. After segmentation, the pixel-level center coordinates of all spots in each frame are calculated using the centroid method. With the top-left corner of the video frame as the origin, the horizontal axis to the right as the positive x-axis, and the vertical axis downwards as the positive y-axis, the coordinates of all pixels within each spot region are weighted by their grayscale values ​​and averaged to obtain the center coordinates of each spot.

[0026] A dynamic spatiotemporal graph is constructed based on the center coordinates of the light spot, with the vertices of the graph being the previously obtained center coordinates of the light spot. When constructing spatial edges, the Euclidean distance between the center coordinates of any two light spots in the same frame is calculated. This distance is calculated based on the coordinate values ​​in the pixel coordinate system. Spatial edges are established when the distance is less than or equal to a predetermined spatial threshold. The predetermined spatial threshold is determined based on actual structural parameters such as the cavity diameter of the GIS equipment and the spacing of the insulating components, in order to accurately characterize the local spatial aggregation characteristics of the discharge area.

[0027] When constructing the time edge, the Kalman filter algorithm or the Hungarian matching algorithm is used to track the center coordinates of the same spot between adjacent frames, determine the position change of the same spot between adjacent frames, obtain the displacement vector, and then calculate the motion speed according to the preset formula. The speed value is the square root of the sum of the squares of the horizontal and vertical coordinates of the displacement vector. The propagation dynamics of the discharge spot in the time dimension are described by the time edge and the corresponding speed, such as stationary, uniform movement or accelerated movement.

[0028] When constructing physical edges, a pre-stored finite element electric field intensity distribution mapping table is relied upon. This mapping table is obtained as follows: First, a 3D model consistent with the actual GIS equipment structure is constructed using 3D modeling software, including the dimensions and material parameters of cavities, conductors, and insulators. Then, the model is imported into finite element simulation software, and parameters such as voltage level and SF6 gas dielectric constant consistent with actual operating conditions are set. The electric field intensity values ​​at different spatial locations inside the equipment are simulated, and the position coordinates are associated with the electric field intensity values ​​to form a mapping table. When constructing physical edges, this mapping table is queried to obtain the electric field intensity values ​​at the corresponding positions of the two vertices to be connected. The absolute difference between the two values ​​is calculated as an attribute of the physical edge, realizing the correlation between the dynamic behavior of the light spot and the underlying physical field distribution.

[0029] After constructing the dynamic spatiotemporal graph, it is input into a dual-path spatiotemporal graph convolutional network for feature extraction. This network includes a main path and an auxiliary path. The main path consists of 3-4 spatiotemporal graph convolutional layers connected sequentially. The number of output channels in each layer increases according to a preset sequence, for example, from 64 to 128 and then to 256. Each layer aggregates the vertex and edge features of the dynamic spatiotemporal graph through graph convolution operations, gradually improving the feature representation capability.

[0030] like Figure 3As shown, the auxiliary path contains 2-3 bottleneck residual units, matching the number of spatiotemporal graph convolutional layers in the main path. Each bottleneck residual unit first compresses the number of channels in the input feature map to 1 / 4 of the original number of channels through a 1×1 convolutional layer, reducing computation and enhancing nonlinear representation; then, it performs spatiotemporal feature aggregation through spatiotemporal graph convolutional layers consistent with the main path structure; subsequently, it restores the number of channels to the original dimension through a 1×1 convolutional layer; finally, it performs residual connections between the restored features and the unit input features to alleviate the gradient vanishing problem and improve training stability. The nonlinear activation function can be either ReLU or LeakyReLU. ReLU is computationally simple and introduces nonlinearity, while LeakyReLU can solve the problem of dead neurons in the negative interval; the choice depends on the training effect.

[0031] In dual-path processing, the output features of a specific stage of the auxiliary path and the corresponding output features of the main path are stacked and fused in the channel dimension. For example, the number of output channels of the first bottleneck residual unit of the auxiliary path is 64, and the number of output channels of the first spatiotemporal graph convolutional layer of the main path is 64. After stacking, the number of fused feature channels is 128. The fusion result is used as the common input of subsequent network layers, integrating dual-path features to improve richness and accuracy.

[0032] After the dual-path spatiotemporal graph convolutional network outputs deep spatiotemporal features, these features are input into an attention mechanism module for feature weighting. This module includes a channel attention submodule and a spatial attention submodule. The channel attention submodule first performs global average pooling and global max pooling on the input feature map. Global average pooling calculates the average value of all pixels in each channel to obtain the average response channel descriptor, while global max pooling calculates the maximum value of all pixels in each channel to obtain the maximum response channel descriptor. The two descriptors are then input into a shared multilayer perceptron, which contains an input layer, hidden layers (the number of neurons is 1 / 4 of the number of channels), and an output layer, to perform nonlinear transformations on the descriptors. Finally, the perceptron outputs are summed element-wise, and a channel attention weight map is generated through a sigmoid activation function. The weight values ​​are mapped to the 0-1 range. The weight map is multiplied with the input feature map channel by channel to enhance important channel features and suppress interference.

[0033] The spatial attention submodule is based on the channel attention-weighted feature map. First, it performs max pooling and average pooling in the channel dimension to obtain two single-channel feature maps. Then, it concatenates the two in the channel dimension to obtain a dual-channel feature map. After fusing information through a 3×3 convolutional layer, it generates a spatial attention weight map through a sigmoid activation function. The weight map and the channel weighted feature map are multiplied pixel by pixel to focus on the discharge spot related area and suppress background interference.

[0034] The weighted features are input to the multi-task output layer, which contains three sub-modules. The defect classification sub-module first performs global average pooling on the weighted features to compress the two-dimensional feature map into a one-dimensional feature vector; then the vector is input into the first and second fully connected layers. The number of neurons in the first layer is set to 256 or 128, and the number of neurons in the second layer is equal to the number of defect categories (normal discharge, floating potential discharge, tip corona discharge, and insulator surface discharge, a total of 4 categories); finally, the output defect category probability distribution is normalized through the Softmax activation function, and the sum of the probabilities is 1, which determines the most likely defect type.

[0035] The severity regression submodule is based on the same one-dimensional feature vector after global average pooling. It is input into a single fully connected layer (with 1 neuron) and outputs a continuous scalar value as the severity assessment value through a linear activation function. The higher the value, the more severe the discharge and the greater the risk of failure. This submodule is trained with a large amount of labeled data to make the output value accurately correspond to physical parameters such as discharge amount and discharge frequency.

[0036] The heatmap generation submodule first obtains the spatial attention weight map, and then uses a deconvolution operation (with 3 layers of 3×3 deconvolution kernels) to match the size of the weight map with the original video frame. Then, it normalizes the weight values ​​to the 0-1 range using the Sigmoid function to generate a heatmap of discharge risk distribution. Areas with values ​​close to 1 have high risk, and areas close to 0 have low risk, which intuitively shows the spatial distribution of discharge risk and provides location guidance for maintenance.

[0037] Example 2 This embodiment is based on the "spatiotemporal graph collaborative perception under physical constraints" mechanism. By integrating the dynamic features of discharge spot and the physical properties of electric field, and combining a dual-path spatiotemporal graph convolutional network with an attention mechanism, it achieves accurate diagnosis of partial discharge in high-voltage GIS. The specific process is as follows: First, video stream data captured by a fiber optic endoscope was acquired. This endoscope is adaptable to high-pressure SF6 gas environments, withstanding temperatures up to 105℃ and pressures up to 0.5MPa, ensuring stable operation under typical GIS equipment operating conditions. The captured video stream has a 5-megapixel resolution and a frame rate set to 25 frames per second, which clearly presents the millimeter-level discharge spot details inside the device while meeting the data processing speed requirements of real-time diagnosis.

[0038] Spot extraction is performed on the aforementioned video stream. First, multispectral filtering separates the discharge spot signal from the metallic reflection noise signal. The multispectral filtering employs a band division strategy. The discharge spot signal corresponds to the 300-500nm band, which is the characteristic emission band generated by SF6 gas ionization during partial discharge and contains core information about the discharge activity. The metallic reflection noise signal corresponds to the 600-800nm ​​band, which mainly originates from the reflected light from metal components such as aluminum alloy conductors and steel cavities inside GIS equipment under illumination, and is considered interference. During the filtering process, a band-specific filtering window is set to separate the signals of the two bands in the frequency domain, avoiding interference from noise signals in subsequent spot detection.

[0039] After filtering, an adaptive threshold segmentation algorithm is applied to the separated discharge spot signals. Based on the grayscale difference between the spot region and the background region in each frame, the segmentation threshold is automatically adjusted to accurately locate the pixel range of the spot. After segmentation, the centroid calculation method is used to determine the center coordinates (x, y) of all spots in each frame. With the top left corner of the video frame as the origin, the positive x-axis is horizontally to the right, and the positive y-axis is vertically downward. The x and y coordinates of all pixels within each spot region are weighted by their pixel grayscale values ​​and then averaged to obtain the precise center coordinates of each spot, providing fundamental data for subsequent construction of a dynamic spatiotemporal map.

[0040] Using the center coordinates of the light spot obtained in the preprocessing stage as the vertices of the dynamic spatiotemporal graph, the edge structure of the graph is constructed according to the three dimensions of space, time, and physics to achieve the fusion of geometric information and physical properties.

[0041] When constructing spatial edges, the Euclidean distance between the center coordinates of any two light spots in the same video frame is calculated. This distance is directly calculated based on the coordinate values ​​in the pixel coordinate system. When the Euclidean distance between the centers of two light spots is less than or equal to 5 pixels, a spatial edge is automatically established between the corresponding vertices of these two light spots. This 5-pixel threshold is determined based on the typical clustering characteristics of the discharge area inside the GIS equipment, which can accurately characterize the local spatial clustering state of discharge activity at the same time. The spatial edge has no additional attributes and is only used to reflect the spatial relationship between light spots.

[0042] When constructing the temporal boundary, a target tracking algorithm (such as the Hungarian matching algorithm) is used to identify the same spot in adjacent video frames, determine the positional change of the same spot between adjacent frames, and then calculate the displacement vector (Δx, Δy), where Δx is the difference in the x-coordinate of the spot center between adjacent frames, and Δy is the difference in the y-coordinate of the spot center between adjacent frames. Based on the displacement vector, the motion velocity v of the spot is calculated using a preset formula: v = √(Δx / Δy). 2 +Δy 2The velocity unit matches the pixel coordinate system and frame rate. The attributes of the time side include the displacement vector (Δx, Δy) and the motion velocity v, which are used to describe the propagation dynamics of the discharge spot in the time dimension, such as the creepage direction and movement speed of the spot along the insulator surface.

[0043] When constructing physical edges, a pre-stored finite element electric field intensity mapping table is relied upon. This mapping table is generated through finite element simulation: First, a 3D model completely identical to the structure of the GIS equipment to be diagnosed is constructed, including key parameters such as cavity dimensions, conductor arrangement, and insulation material. Then, the model is imported into finite element simulation software, and parameters such as voltage level and SF6 gas dielectric constant are set to match the actual operating conditions of the equipment. The electric field intensity values ​​E(x,y) at all spatial locations inside the equipment are simulated. Finally, the spatial coordinates (x,y) and the corresponding electric field intensity values ​​E(x,y) are associated and stored to form the finite element electric field intensity mapping table. When constructing physical edges, the electric field intensity values ​​E(vi) and E(vj) at the corresponding spatial locations of any two spot vertices vi and vj are retrieved from this mapping table. The absolute difference between the two is calculated as ΔE=|E(vi)-E(vj)|, and ΔE is used as an attribute of the physical edge. The construction of physical edges realizes the correlation between the dynamic behavior of the discharge spot and the underlying electric field physical distribution, giving the dynamic spatiotemporal map physical constraint characteristics.

[0044] like Figure 2 As shown, the constructed dynamic spatiotemporal graph is input into the dual-path spatiotemporal graph convolutional network (BiST-GCN), which consists of a main path and an auxiliary path. Deep spatiotemporal feature extraction is achieved through the collaboration of the two paths.

[0045] The main path consists of nine sequentially connected spatiotemporal graph convolutional layers (ST-GCN layers), divided into three stages: the first three layers constitute the first stage, each with 64 output channels; the middle three layers constitute the second stage, each with 128 output channels; and the last three layers constitute the third stage, each with 256 output channels. The main path aggregates vertex and edge features from the dynamic spatiotemporal graph through spatiotemporal graph convolution operations (STGCN), with the specific calculation process following the formula... in This represents the input features of the l-th layer (i.e., the dynamic spatiotemporal graph features output by the previous layer). The weights of the trainable spatiotemporal graph convolution kernel in layer l are represented by `stgcnl`, which represents the spatiotemporal graph convolution operation in layer l. Its calculation formula is: σ uses the ReLU activation function. The main path gradually improves the feature representation capability through the design of increasing channel number, which can effectively extract the spatiotemporal propagation characteristics of the discharge spot, such as the creepage path characteristics of the spot along the insulator surface and the cooperative motion characteristics among multiple spots.

[0046] like Figure 3 As shown, the auxiliary path consists of three bottleneck residual units connected sequentially, with each bottleneck residual unit having the same processing flow and computational logic. A single bottleneck residual unit first compresses the number of channels in the input feature map to 1 / 4 of the original number of channels using a 1×1 convolutional layer, reducing computation while enhancing the non-linear expressive power of the features. Then, the compressed features are input into a spatiotemporal map convolutional layer (with the same structure as the main path ST-GCN layer) to perform spatiotemporal feature aggregation, capturing the dynamics of the light spot and the features associated with the physical field. Subsequently, a 1×1 convolutional layer restores the number of feature channels to the original dimension, ensuring dimensionality matching for subsequent feature fusion. Finally, the restored features are residually connected to the unit's input features, with the specific calculation following the formula... Where F( , The ) represents the characteristic transformation process of the bottleneck residual unit, and the calculation formula is: , For unit input features, Here, σ represents the trainable weight parameters within the unit, and σ is the ReLU activation function. The design of residual connections can solve the gradient vanishing problem during the training process of deep networks, ensuring the stability of feature propagation.

[0047] In the dual-path collaboration process, the feature outputs of the auxiliary path and the main path are stacked and fused in stages: the output features of the first bottleneck residual unit of the auxiliary path are stacked and fused with the output features of the third layer (end of the first stage) of the main path in the channel dimension; the output features of the second bottleneck residual unit of the auxiliary path are stacked and fused with the output features of the sixth layer (end of the second stage) of the main path in the channel dimension; the fused features are used as input features for the fourth and seventh layers of the main path, respectively, and the output features of the third bottleneck residual unit of the auxiliary path are directly used as supplementary input for subsequent layers of the main path. By stacking and fusing features, the deep features of the main path and the residual features of the auxiliary path are integrated, significantly improving the model's efficiency and robustness in extracting discharge features.

[0048] Based on the BiST-GCN network processing, a Convolutional Block Attention (CBAM) module is introduced to enhance features, and diagnostic results are generated through a multi-task output layer. The CBAM module is integrated in two locations: The first integration location is after each bottleneck residual unit, used to enhance local features. After each bottleneck residual unit completes the residual connection, its output features are input into the CBAM module. First, channel weights are generated through the channel attention submodule, and then spatial weights are generated through the spatial attention submodule. This adaptively weights the local features, highlighting key discharge-related features in the local region and suppressing noise interference.

[0049] The second integration point is at the end of the BiST-GCN network, used for global attention enhancement. After all feature extraction and fusion are completed in the main and auxiliary paths, the final deep spatiotemporal features output by the network are input into the CBAM module to perform global channel attention and global spatial attention calculations. This focuses on the key global features related to strong electric field regions and defect-sensitive channels in the entire feature map, providing a high-quality feature foundation for subsequent diagnostic outputs.

[0050] The calculation of channel attention in the CBAM module follows the formula Where F is the input feature map (i.e., the deep spatiotemporal features output by BiST-GCN), AvgPool(F) represents performing a global average pooling operation on F to extract the channel-level average response; MaxPool(F) represents performing a global max pooling operation on F to extract the channel-level salient response; MLP is a shared multilayer perceptron (composed of fully connected layers) used to learn the nonlinear relationships between channels; σ is the sigmoid activation function, which maps the output to the [0,1] interval to generate the channel attention weight map Mc(F). The channel attention weight map is multiplied with the input feature map F channel by channel to enhance the feature channel weights corresponding to the strong electric field region and weaken the influence of irrelevant channels.

[0051] The calculation of spatial attention follows the formula , Where [AvgPool(F);MaxPool(F)] represents the concatenation of two single-channel feature maps after performing average pooling and max pooling on the feature map F after channel attention weighting in the spatial dimension; f7×7 represents a 7×7 convolutional layer used to fuse feature information in the spatial dimension; σ is the Sigmoid activation function, which generates a spatial attention weight map Ms(F). The spatial attention weight map is multiplied pixel by pixel with the feature map to accurately track the creeping path of the discharge spot along the insulator and enhance the feature weights on the path.

[0052] The weighted features from the CBAM module are input to the multi-task output layer, which consists of three functionally independent but parameter-sharing sub-modules: Defect classification submodule (qualitative identification): The network structure consists of a fully connected layer and a softmax activation function concatenated. The input is the attention-weighted feature vector GAP(F) compressed by global average pooling (GAP). att ) (F att The features (weighted by CBAM) are transformed through a fully connected layer, and then a probability distribution vector is generated through the Softmax activation function. The calculation follows the formula , in This is the classification weight matrix. This is the classification bias vector. This submodule outputs four probabilities, corresponding to four defect types: Category 0 is normal or background discharge, Category 1 is floating potential discharge, Category 2 is tip corona discharge, and Category 3 is insulator surface discharge. The category with the highest probability value is the currently diagnosed defect type.

[0053] Severity Regression Submodule (Quantitative Assessment): The network structure consists of a fully connected layer and a linear activation function. The input is the same as that of the defect classification submodule, which is the feature vector GAP(F). att ) Outputs continuous scalar values ​​through fully connected layers and linear activation functions. The calculation follows the formula in For the regression weight matrix, This is the regression bias vector. This scalar value can be mapped to a severity level system of 1-5, where level 1 indicates weak discharge activity and extremely low risk, and level 5 indicates severe discharge activity and serious equipment failure risk, providing quantitative data support for predictive maintenance.

[0054] The heatmap generation submodule (visualized positioning) consists of a deconvolutional layer and a subsequent normalization unit. The input is the spatial attention map Ms(F) output by the CBAM module. First, the spatial attention map is upsampled by the deconvolutional layer to restore its size to be identical to the original video frame, resulting in an initial heatmap. Then, the initial heatmap is normalized using the Sigmoid activation function, mapping the value of each pixel in the heatmap to the [0,1] interval, generating the final discharge risk distribution heatmap. Areas with values ​​close to 1 in the heatmap indicate extremely high discharge risk, while areas with values ​​close to 0 indicate extremely low discharge risk. This visually displays the spatial distribution of discharge risk within GIS equipment, providing maintenance personnel with precise fault location guidance.

[0055] Example 3 This embodiment provides a real-time diagnostic system for partial discharge in high-voltage GIS, which is constructed based on the method described in Embodiment 1. It includes a fiber optic endoscope module, a video preprocessing module, a dynamic image construction module, a feature extraction network module, an attention mechanism module, and a multi-task diagnostic output module. Each module works collaboratively through signal or data connections.

[0056] The fiber optic endoscope module is used to acquire high-resolution video data from inside GIS equipment in a high-pressure SF6 gas environment. Its core components include a fiber optic probe, image sensor, illumination unit, and sealed housing. The fiber optic probe uses flexible fiber optic material to adapt to the complex internal space of the equipment, facilitating access to critical areas such as conductor connections and the surface of insulating components. The image sensor uses a CMOS or CCD sensor with a resolution of at least 1920×1080. CMOS offers low power consumption and high integration, while CCD offers high sensitivity and low noise; the choice depends on the acquisition requirements. The illumination unit uses high-brightness LEDs with wavelength-matched discharge spot characteristic spectra to reduce metal reflection noise. The sealed housing is made of SF6-resistant stainless steel and is sealed with a sealing ring to prevent gas leakage and module damage. During operation, the probe position and angle are adjusted via an external control unit, and the image sensor captures video data and outputs it as a digital signal.

[0057] The video preprocessing module is signal-connected to the fiber optic endoscope module. The hardware utilizes FPGA or DSP chips (FPGA offers strong parallel computing capabilities and high real-time performance, while DSP provides efficient digital signal processing). After receiving video data, it is first buffered to ensure stable transmission. Then, multispectral filtering is achieved using the Fast Fourier Transform IP core in the FPGA / DSP and preset filtering coefficients to separate the light spot from noise signals. Subsequently, the maximum inter-class variance method is run to perform adaptive threshold segmentation, separating the light spot region. Finally, the centroid method is used to calculate the center coordinates of the light spot and output to the dynamic image construction module. The processing frame rate is matched to the acquisition frame rate (e.g., 25fps) to avoid data accumulation.

[0058] The dynamic graph construction module is data-connected to the video preprocessing module, and includes an ARM processor and a storage unit (solid-state drive storing a finite element electric field intensity mapping database). The processor receives the sequence of light spot center coordinates and performs timestamp synchronization to ensure correct frame order; it calculates the Euclidean distance between light spot coordinates in the same frame to establish spatial edges, using a floating-point arithmetic unit for distance calculation; it uses a Kalman filter algorithm to track light spots in adjacent frames, calculating displacement vectors and velocities to establish temporal edges; it queries the electric field intensity mapping database on the solid-state drive (keys are spatial coordinates, values ​​are electric field intensity, and queries are performed quickly using an index) to obtain the electric field intensity value corresponding to the light spot, and calculates the difference to establish physical edges. The constructed dynamic spatiotemporal graph is stored in a temporary region and output to the feature extraction network module. The module also supports updating the database via an external interface to adapt to different device models.

[0059] The feature extraction network module and the dynamic graph construction module are connected via data link. Hardware utilizes GPUs or dedicated AI chips (GPUs offer strong parallel computing capabilities, while dedicated AI chips offer low power consumption). A dual-path spatiotemporal graph convolutional network is constructed based on the TensorFlow or PyTorch framework. The network receives dynamic spatiotemporal graph data, converts it to tensor format, and inputs it into the network. In the main path, 3-4 spatiotemporal graph convolutional layers sequentially extract features, with the number of output channels increasing from 64 to 128 to 256. In the auxiliary path, 2-3 bottleneck residual units process the data in parallel, achieving feature transformation through 1×1 convolutions, spatiotemporal graph convolutions, and residual connections. The outputs of specific stages in the auxiliary path are stacked and fused with the corresponding outputs of the main path in the channel dimension, ultimately outputting a deep spatiotemporal feature tensor to the attention mechanism module. The module improves inference speed to meet real-time requirements through model quantization and pruning optimization.

[0060] The attention mechanism module and the feature extraction network module share hardware and implement their functions through framework layer operations. After receiving the deep spatiotemporal feature tensor, the channel attention submodule generates two channel descriptors through pooling layers, inputs them into a multilayer perceptron transformation composed of fully connected layers, adds them together, generates a weight map through a sigmoid function, and multiplies it with the feature tensor channel by channel; the spatial attention submodule pools and concatenates the channel-weighted feature tensors, generates a weight map through 3×3 convolution and a sigmoid function, multiplies it with the feature tensor pixel by pixel, and outputs the weighted features to subsequent modules.

[0061] The multi-task diagnostic output module and attention mechanism module are data-connected, and include a processor, storage unit, and display unit. The defect classification unit performs global average pooling on weighted features to obtain a one-dimensional vector, which is input to a fully connected layer. After passing through Softmax, the output is a probability distribution of defect categories, and the results are stored and displayed numerically / graphically in the display unit. The severity regression unit inputs the same vector into a fully connected layer, outputs a severity assessment value through linear activation, displays it in the display unit, and sets a threshold alarm. The heatmap generation unit deconvolves and normalizes the spatial attention weight map to generate a heatmap that matches the original video frames. This heatmap is then overlaid on the original video and displayed in the display unit. The module also has a data interface to transmit diagnostic results to the monitoring center server for remote monitoring and archiving.

[0062] The system modules achieve high-speed data transmission via PCIe bus or Ethernet to ensure real-time stability; it has a self-calibration function to periodically calibrate the endoscope imaging effect and the accuracy of the preprocessing algorithm; the software supports online updates and can optimize the network structure and algorithm parameters to adapt to the diagnostic needs of different GIS equipment.

[0063] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative in all respects and are not the only ones. All modifications within the scope of this invention or its equivalents are included in this invention.

Claims

1. A real-time diagnostic method for partial discharge in high-voltage GIS, characterized in that, Includes the following steps: The video stream captured by the fiber optic endoscope is preprocessed, and the center coordinates of the discharge spot in the video frames are extracted through multispectral filtering and adaptive threshold segmentation. A dynamic spatiotemporal graph is constructed based on these center coordinates, where the vertices are the center coordinates of the spot, and the edges include spatial, temporal, and physical edges. Spatial edges are established based on the spatial proximity of the spot, temporal edges are established based on the movement of the spot between adjacent frames, and physical edges are established by associating finite element electric field intensity data. The dynamic spatiotemporal graph is then input into a dual-path spatiotemporal graph convolutional network for feature extraction. This network processes data through a main path containing spatiotemporal graph convolutions and an auxiliary path containing bottleneck residual units, outputting deep spatiotemporal features. These deep spatiotemporal features are then input into an attention mechanism module for feature weighting. This module focuses on key feature regions through channel attention and spatial attention. The weighted features are then input into a multi-task output layer, simultaneously outputting defect classification results, severity assessment values, and discharge risk distribution maps.

2. The real-time diagnostic method for partial discharge in high-voltage GIS according to claim 1, characterized in that, The preprocessing steps specifically include: separating the discharge spot signal of the first band and the metal reflection noise signal of the second band from the high-resolution video stream captured by the fiber optic endoscope at a frame rate that meets the requirements for real-time processing through frequency domain filtering; performing an adaptive threshold segmentation algorithm on the separated spot signals to calculate the pixel-level center coordinates of all spots in each frame image; the first band corresponds to the emission wavelength range of the discharge spot, and the second band corresponds to the dominant wavelength range of the metal reflection light.

3. The real-time diagnostic method for partial discharge in high-voltage GIS according to claim 1, characterized in that, In the construction of the dynamic spatiotemporal graph, the connection rule of the spatial edge is as follows: when the Euclidean distance between the center coordinates of any two light spots in the same frame is less than or equal to a predetermined spatial threshold, a spatial edge is established between the vertices corresponding to these two light spots to characterize the local spatial aggregation characteristics of the discharge region. The construction of the time edge involves tracing the center coordinates of the same light spot between adjacent frames, calculating its displacement vector (Δx, Δy), and deriving the motion velocity v based on this displacement vector. The calculation formula is v = √(Δx / Δy). 2 +Δy 2 ), used to describe the propagation dynamics of the discharge spot in the time dimension; The physical edge is constructed by querying a pre-stored electric field intensity distribution mapping table obtained based on finite element simulation, obtaining the electric field intensity values ​​E(vi) and E(vj) at graph vertices vi and vj, and calculating their absolute difference ΔE=|E(vi)-E(vj)| as the attribute of the edge, thereby associating the dynamic behavior of the light spot with the underlying physical field distribution.

4. The real-time diagnostic method for partial discharge in high-voltage GIS according to claim 1, characterized in that, In the dual-path spatiotemporal graph convolutional network, the processing flow of the bottleneck residual unit includes: The number of channels in the input feature map is compressed to a predetermined ratio of the original number of channels using a 1×1 convolutional layer; The compressed features are then aggregated using a spatiotemporal graph convolutional layer. Then, a 1×1 convolutional layer is used to restore the number of feature channels to the original dimension; The recovered features are residually concatenated with the input features of the unit; the calculation process of the bottleneck residual unit is as follows: In the formula This represents the input features of the l-th layer. This represents the weight parameters of the trainable spatiotemporal graph convolution kernel. This represents a non-linear activation function.

5. The real-time diagnostic method for partial discharge in high-voltage GIS according to claim 1, characterized in that, In the convolutional block attention module, the calculation process of the channel attention is as follows: Global average pooling and global max pooling are performed on the input feature map F respectively to obtain two different channel descriptors; Input these two descriptors into a shared multilayer perceptron; After summing the outputs of the multilayer perceptron element by element, a channel attention weight map is generated using the sigmoid activation function. Its expression is Where MLP represents a multilayer perceptron, AvgPool and MaxPool represent global average pooling and global max pooling, respectively, and σ represents the sigmoid function. The spatial attention calculation process is as follows: Average pooling and max pooling are performed on the input feature map F along the channel dimension to obtain two spatial feature maps; these two spatial feature maps are concatenated along the channel dimension; a standard convolutional layer is used to perform a convolution operation on the concatenated feature map; finally, a spatial attention weight map is generated using the Sigmoid activation function. .

6. The real-time diagnostic method for partial discharge in high-voltage GIS according to claim 1, characterized in that, The multi-task output layer includes a defect classification submodule, a severity regression submodule, and a heatmap generation submodule. The defect classification submodule performs a fully connected layer transformation and Softmax activation on the attention-weighted feature vector after global average pooling compression, and outputs classification results representing the probability distribution of different defect types. Its calculation formula is ,in Indicates attention-weighted features Perform global average pooling. and The regression weights and biases are trainable, and Softmax is a normalized exponential function. The severity regression submodule performs a fully connected layer and linear activation on the same pooled feature vector, outputting a continuous scalar value representing the severity of the discharge. Its calculation formula is The heatmap generation submodule upsamples and normalizes the spatial attention map to generate a discharge risk distribution heatmap that matches the spatial size of the original input video frame.

7. The real-time diagnostic method for partial discharge in high-voltage GIS according to claim 6, characterized in that, The defect type probability distribution output by the defect classification submodule covers categories including normal discharge, floating potential discharge, tip corona discharge, and insulator surface discharge; the continuous scalar value output by the severity regression submodule is used to quantitatively assess the severity of discharge activity, with higher values ​​indicating greater potential risk. The heatmap generation submodule performs upsampling through deconvolution and uses the Sigmoid function to normalize the output value to the [0,1] interval to generate the discharge risk distribution heatmap.

8. The real-time diagnostic method for partial discharge in high-voltage GIS according to claim 4, characterized in that, The main path of the dual-path spatiotemporal graph convolutional network is composed of multiple spatiotemporal graph convolutional layers connected sequentially, and the number of output channels of each layer increases in a pre-set sequence. The auxiliary path contains multiple bottleneck residual units, and the output features of a specific stage in the auxiliary path are stacked and fused with the output features of the corresponding stage in the main path. The fused result is used as the common input of subsequent network layers.

9. A real-time diagnostic system for partial discharge in high-voltage GIS, based on the real-time diagnostic method for partial discharge in high-voltage GIS according to any one of claims 1 to 8, characterized in that, include: Fiber optic endoscope module, used to acquire high-resolution video data inside GIS equipment in a high-pressure SF6 gas environment; The video preprocessing module is signal-connected to the fiber optic endoscope module and is used to receive video data and perform multispectral filtering and adaptive threshold segmentation to extract the center coordinates of the light spot. The dynamic graph construction module is data-connected to the video preprocessing module and is used to construct a dynamic spatiotemporal graph containing spatial edges, temporal edges, and physical edges based on the sequence of light spot center coordinates. The feature extraction network module is data-connected to the dynamic graph construction module. It is a dual-path spatiotemporal graph convolutional network, which includes a main path and an auxiliary path with bottleneck residual units, and is used to extract deep spatiotemporal features from the dynamic spatiotemporal graph. The attention mechanism module, which is data-connected to the feature extraction network module, includes a channel attention submodule and a spatial attention submodule, and is used to adaptively weight the deep spatiotemporal features. The multi-task diagnostic output module, which is data-connected to the attention mechanism module, includes a defect classification unit, a severity regression unit, and a heatmap generation unit, and is used to output the final diagnostic results.

10. A real-time diagnostic system for partial discharge in high-voltage GIS according to claim 9, characterized in that, The dynamic graph construction module integrates a finite element electric field intensity mapping database to provide electric field intensity difference data for the physical edges; the main path of the feature extraction network module consists of multiple spatiotemporal graph convolutional layers with an increasing number of output channels, and the bottleneck residual unit output of the auxiliary path is stacked with the specific layer output of the main path; the defect classification unit in the multi-task diagnostic output module uses a fully connected layer and a Softmax activation function, the severity regression unit uses a fully connected layer and a linear activation function, and the heatmap generation unit includes a deconvolution layer and a Sigmoid normalization layer.

Images