Substation primary equipment partial discharge intelligent identification method and system

Abstract

The application provides a substation primary equipment partial discharge intelligent identification method and system, relates to the technical field of power equipment monitoring, and comprises the following steps: synchronously collecting discharge electromagnetic signals through a sensor array and extracting features, and constructing a multi-modal feature set; obtaining a discharge source feature representation vector through multi-scale time-frequency transformation and graph neural network processing; matching the discharge source feature representation vector with a discharge mode knowledge graph, determining an optimal discharge type through variational inference and Bayesian quantization; performing consistency verification in combination with spatial positioning and matching historical data, evaluating the severity of the discharge, and outputting a report. The application realizes accurate identification of the type of partial discharge and quantitative evaluation of the severity, and improves the intelligent level and reliability of substation equipment state monitoring.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, and in particular to a method and system for intelligent identification of partial discharge in primary equipment of substations. Background Technology

[0002] In the field of substation equipment condition monitoring, partial discharge detection is a key technology for assessing insulation degradation and predicting potential faults. Existing technologies typically rely on various sensors deployed at critical locations on the equipment, such as ultra-high frequency sensors, ultrasonic sensors, or transient ground voltage sensors, to collect the physical signals generated by the discharge. After preprocessing, the physical signals are used to extract features in the time or frequency domain. Pattern recognition algorithms are then used to compare and classify the extracted features with a database of known discharge types, thereby achieving preliminary identification of the discharge type.

[0003] However, existing technologies still suffer from insufficient exploration of the deep correlations and coupling relationships between multimodal signals (such as electromagnetic and acoustic signals) captured by different sensors. This results in limited feature representation capabilities in complex electromagnetic interference substation environments, making it difficult to effectively distinguish different types of discharge modes and lacking quantitative assessment of the reliability of identification results. It is also unable to provide a measure of the uncertainty of classification decisions. When encountering novel discharge modes not fully covered by training data or strong noise interference, it may produce high-confidence misclassifications. Summary of the Invention

[0004] This invention provides a method and system for intelligent identification of partial discharge in primary equipment of a substation, which can at least solve some of the problems existing in the prior art.

[0005] A first aspect of this invention provides a method for intelligent identification of partial discharge in primary equipment of a substation, comprising: The discharge electromagnetic signals of the primary equipment in the substation are synchronously acquired by a sensor array. The discharge electromagnetic signals are synchronously calibrated and the pulse arrival time difference and signal amplitude features are extracted. Spatial positioning coordinates are constructed based on the pulse arrival time difference and multimodal feature set is generated by combining the signal amplitude features. The multimodal feature set is decomposed by multi-scale time-frequency transformation to obtain localized feature matrices for different frequency bands. The localized feature matrices are correlated to obtain an intermodal coupling relationship graph. The intermodal coupling relationship graph is encoded by a graph neural network to obtain a discharge source feature representation vector. The discharge source feature representation vector is semantically matched with a preset discharge mode knowledge graph to obtain a set of candidate discharge types. The posterior probability distribution of the discharge type corresponding to the discharge source feature representation vector is calculated based on the variational inference algorithm. The confidence interval of the posterior probability distribution is estimated by combining Bayesian uncertainty quantization to obtain the probability ranking result and determine the optimal discharge type. Based on the optimal discharge type and the spatial positioning coordinates, the consistency of the equipment structural parameters is verified to obtain the verification result. The verified discharge data is then matched with the pre-acquired historical discharge data to obtain a similarity score. Based on the similarity score, the severity rating of the partial discharge is calculated and an identification report is output.

[0006] In one alternative implementation, The discharge electromagnetic signals of primary equipment in a substation are synchronously acquired using a sensor array. These signals are then synchronously calibrated, and pulse arrival time difference (TAD) and signal amplitude features are extracted. Spatial positioning coordinates are constructed based on the TAD, and a multimodal feature set is generated by combining the signal amplitude features. The spatial position information and clock parameters of each sensor are obtained from the sensor array. Based on the clock parameters, the discharge electromagnetic signals collected by each sensor are time-aligned to obtain a synchronous electromagnetic signal. The synchronous electromagnetic signal is pulse-detected to identify the pulse start time and the pulse arrival time difference between adjacent sensors is calculated. Based on the pulse arrival time difference and the spatial position information, triangulation is performed to obtain the spatial positioning coordinates. The peak amplitude and effective amplitude of each frequency band are extracted from the synchronous electromagnetic signal to form the signal amplitude feature. The statistical features of the signal amplitude feature in the time domain are calculated and the power spectral density distribution in the frequency domain is extracted. The statistical features and the power spectral density distribution are vector-concatenated to obtain the single-mode feature vector. The single-mode feature vectors corresponding to different sensors are normalized and arranged according to the sensor number to form a feature matrix. The spatial distance between the power source and each sensor is calculated based on the spatial positioning coordinates, and a distance weight matrix is ​​constructed. The distance weight matrix and the feature matrix are used to solve for the distance weight matrix to obtain the distance weighted feature matrix. Principal component analysis is performed on the distance weighted feature matrix to reduce its dimensionality and retain the principal components to obtain the multimodal feature set.

[0007] In one alternative implementation, The multimodal feature set is decomposed by multi-scale time-frequency transformation to obtain localized feature matrices for different frequency bands. Correlation measurement is then performed on these localized feature matrices to obtain an intermodal coupling relationship graph, including: The multimodal feature set is grouped according to sensor modes to obtain feature subsets corresponding to each sensor. For each feature subset, the corresponding optimal wavelet basis is determined and wavelet transform is performed to obtain multi-scale time-frequency coefficients. The multi-scale time-frequency coefficients are subjected to singular value decomposition and the principal singular vectors are extracted to form frequency band feature vectors. The frequency band feature vectors at different scales are arranged in ascending order of frequency to obtain the localized feature matrix. The mutual information between the localized feature matrices corresponding to different sensors is calculated. Based on the mutual information, the nonlinear correlation strength is calculated to obtain the information correlation matrix. The information correlation matrix is ​​divided into modal clusters by spectral clustering algorithm and the inter-cluster connection relationship is identified to obtain the sparse correlation matrix. The non-zero elements in the sparse correlation matrix are used as edge weights to construct the sensor connection graph to obtain the initial coupling relationship graph. For each sensor node in the initial coupling graph, the node access frequency is calculated using a random walk sampling strategy to obtain a node centrality index. Based on the node centrality index, key sensor nodes are identified and the weights of the connection edges of the key sensor nodes are adjusted. The initial coupling graph after weight adjustment is then normalized using a Laplace matrix to obtain the intermodal coupling graph.

[0008] In one alternative implementation, The graph neural network encoding operation performed on the intermodal coupling relationship graph yields the following feature representation vectors of the discharge source: The connectivity distribution and edge weight distribution of each sensor node are extracted from the intermodal coupling relationship graph to obtain graph topology statistics. Based on the graph topology statistics, the aggregation range of the node neighborhood is calculated and the multi-order neighborhood boundary is determined. The localized feature matrix corresponding to each sensor node is used as the initial feature of the node and combined with the multi-order neighborhood boundary to construct the node feature propagation matrix. The node feature propagation matrix is ​​subjected to tensor decomposition and low-rank approximation reconstruction to obtain the node embedding feature. The node embedding features are arranged into a graph feature matrix according to the sensor number and multi-layer convolution operation is performed with the adjacency matrix of the intermodal coupling relationship graph. In each layer convolution operation, the spectral correlation of the neighborhood features is calculated based on the spectral domain frequency decomposition and the node representation is updated by combining the residual connection. The updated node representation is batch normalized and multi-layer convolution output is accumulated to obtain a multi-scale node representation. The global graph representation vector is calculated based on the multi-scale node representation. The global graph representation vector is adaptively sampled and aggregated, and the node importance weights are calculated by combining the multi-scale node representations. The multi-scale node representations are then weighted and fused based on the node importance weights to obtain fused node features. The fused node features are then mapped and transformed through a fully connected layer and combined with linear activation to obtain the discharge source feature representation vector.

[0009] In one alternative implementation, The discharge source feature representation vector is semantically matched with a preset discharge mode knowledge graph to obtain a candidate discharge type set. The posterior probability distribution of the discharge type corresponding to the discharge source feature representation vector is calculated based on a variational inference algorithm, including: Feature template vectors and hierarchical relationships between discharge types are extracted from a pre-defined discharge mode knowledge graph. A discharge type feature library is constructed based on the feature template vectors, and the Riemann manifold distance between the discharge source feature representation vector and each feature template vector in the discharge type feature library is calculated to obtain a manifold similarity matrix. Based on the manifold similarity matrix, high-similarity discharge types are determined, and related parent types are expanded by combining the hierarchical relationships between types to obtain the candidate discharge type set. Gaussian mixture distribution parameter estimation is performed on the discharge source feature representation vector to calculate the feature space probability density distribution. Based on the probability density distribution, multiple sets of feature variables are sampled and generated, and the similarity with the feature template vectors of each type in the candidate discharge type set is calculated to obtain the type similarity sequence. The occurrence frequency of each discharge type in the type similarity sequence is counted and normalized to obtain the type frequency distribution. Based on each discharge type in the candidate discharge type set, a multinomial distribution prior is constructed. The likelihood probability is obtained by performing expectation maximization iterative calculation on the type frequency distribution and the multinomial distribution prior. The likelihood probability is then randomly sampled to obtain a posterior sample set, and the posterior probability distribution of the discharge type corresponding to the discharge source feature representation vector is calculated.

[0010] In one alternative implementation, By combining Bayesian uncertainty quantification to estimate the confidence interval of the posterior probability distribution, the probability ranking results are obtained, and the optimal discharge type is determined, including: The probability distribution statistics of each discharge type are extracted from the posterior probability distribution and the confidence ellipse boundary is calculated to obtain a set of confidence intervals. The intersection overlap rate of each discharge type in the set of confidence intervals is calculated based on the interval volume. High confusion type groups are marked based on the intersection overlap rate and the prediction uncertainty is quantified. Monte Carlo uncertainty propagation is performed on the probability distribution statistics based on the prediction uncertainty to obtain a corrected probability distribution. A probability matrix is ​​constructed based on the modified probability distribution, and the conditional entropy is calculated. The inter-type discrimination is evaluated based on the conditional entropy, and an adaptive weight coefficient is set. The modified probability distribution is subjected to a nonlinear transformation by the adaptive weight coefficient, and the type confidence is calculated in combination with the corresponding interval volume. The discharge types are sorted based on the type confidence to obtain the probability ranking result. The discharge type with the highest confidence level is selected from the probability ranking results as the initial optimal type. The statistical significance difference between the initial optimal type and the discharge type with the second highest confidence level is calculated. It is determined whether the statistical significance difference is greater than a preset discrimination threshold. If so, the initial optimal type is the optimal discharge type. Otherwise, the type with the largest information gain is selected from the probability ranking results according to the information gain maximization criterion, and the optimal discharge type is determined by combining the information gain maximization criterion with Nash equilibrium.

[0011] In one alternative implementation, Based on the optimal discharge type and the spatial positioning coordinates, the consistency of the device structural parameters is verified to obtain the verification result. The verified discharge data is then matched with pre-acquired historical discharge data to obtain a similarity score. Based on the similarity score, the severity rating of the partial discharge is calculated, and an identification report is output, including: The standard structural parameters corresponding to the optimal discharge type are obtained from the preset equipment structural parameter database. The geometric parameters of the equipment structure are calculated based on the spatial positioning coordinates and the difference is calculated with the standard structural parameters. When the difference is less than the preset consistency threshold, the verification is marked as passed and the verification result is obtained. Extract feature parameters from the verified discharge data to construct the current discharge feature vector, extract historical feature vectors of the same type of discharge from the historical discharge data, calculate the time-normalized distance between the current discharge feature vector and the historical feature vector and normalize it to obtain the similarity score; Adaptive fuzzy clustering analysis is performed based on the similarity score to obtain the severity classification boundary. The similarity score is then mapped to a severity rating based on the severity classification boundary. Finally, an identification report is generated by combining the optimal discharge type and the spatial positioning coordinates.

[0012] A second aspect of the present invention provides an intelligent identification system for partial discharge of primary equipment in a substation, comprising: The signal feature extraction unit is used to synchronously acquire the discharge electromagnetic signals of the primary equipment of the substation through the sensor array, synchronously calibrate the discharge electromagnetic signals and extract the pulse arrival time difference and signal amplitude features, construct spatial positioning coordinates based on the pulse arrival time difference and generate a multimodal feature set in combination with the signal amplitude features; The feature representation encoding unit is used to decompose the multimodal feature set through multi-scale time-frequency transformation to obtain localized feature matrices of different frequency bands, perform correlation measurement on the localized feature matrices to obtain an intermodal coupling relationship graph, and perform graph neural network encoding operation on the intermodal coupling relationship graph to obtain a discharge source feature representation vector; The discharge type identification unit is used to semantically match the discharge source feature representation vector with a preset discharge mode knowledge graph to obtain a candidate discharge type set, calculate the posterior probability distribution of the discharge type corresponding to the discharge source feature representation vector based on the variational inference algorithm, and estimate the confidence interval of the posterior probability distribution by combining Bayesian uncertainty quantization to obtain the probability ranking result and determine the optimal discharge type. The severity assessment unit is used to verify the consistency of the device structural parameters based on the optimal discharge type and the spatial positioning coordinates to obtain the verification result, perform similarity matching operation between the verified discharge data and the pre-acquired historical discharge data to obtain a similarity score, calculate the severity rating of partial discharge based on the similarity score, and output an identification report.

[0013] A third aspect of the present invention provides an electronic device, comprising: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] In this invention, a sensor array is used to synchronously acquire discharge electromagnetic signals. Synchronous calibration ensures the time consistency of multi-channel data, effectively eliminating errors caused by signal transmission delays. Pulse arrival time difference and signal amplitude features are extracted and combined with constructed spatial positioning coordinates to form a multimodal feature set. This provides a comprehensive and complementary information foundation for subsequent analysis, enhancing the robustness of feature representation. Multi-scale time-frequency transformation decomposition of the multimodal feature set captures localized discharge characteristics at different frequency bands, adapting to the non-stationary characteristics of discharge signals. A coupling relationship diagram between modes is constructed through correlation metrics, revealing the intrinsic connections between different feature dimensions. By encoding the relationship graph using a graph neural network, the generated discharge power source feature representation vector deeply integrates spatial, temporal, and amplitude information, significantly improving the feature representation ability and discriminative power. Semantic matching of the feature vector with the discharge mode knowledge graph can quickly screen out candidate discharge types, improving recognition efficiency. Based on the variational inference algorithm, the posterior probability distribution is calculated, and confidence interval estimation is performed by combining Bayesian uncertainty quantification. This not only provides the probability ranking of the optimal discharge type, but also quantifies the credibility of the recognition results, enhancing the reliability and interpretability of the decision-making, and effectively addressing the ambiguity and complexity of discharge modes in real-world scenarios. Attached Figure Description

[0016] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating the intelligent identification method for partial discharge of primary equipment in substations according to an embodiment of the present invention. Figure 2This is a flowchart illustrating the confidence assessment process for the discharge type in the intelligent identification method for partial discharge of primary equipment in substations, as described in this embodiment of the invention. Figure 3 This is a structural block diagram of the intelligent identification system for partial discharge of primary equipment in a substation, according to an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0019] Figure 1 This is a flowchart illustrating the intelligent identification method for partial discharge of primary equipment in substations according to an embodiment of the present invention. Figure 1 As shown, the method includes: The discharge electromagnetic signals of the primary equipment in the substation are synchronously acquired by a sensor array. The discharge electromagnetic signals are synchronously calibrated and the pulse arrival time difference and signal amplitude features are extracted. Spatial positioning coordinates are constructed based on the pulse arrival time difference and multimodal feature set is generated by combining the signal amplitude features. The multimodal feature set is decomposed by multi-scale time-frequency transformation to obtain localized feature matrices for different frequency bands. The localized feature matrices are correlated to obtain an intermodal coupling relationship graph. The intermodal coupling relationship graph is encoded by a graph neural network to obtain a discharge source feature representation vector. The discharge source feature representation vector is semantically matched with a preset discharge mode knowledge graph to obtain a set of candidate discharge types. The posterior probability distribution of the discharge type corresponding to the discharge source feature representation vector is calculated based on the variational inference algorithm. The confidence interval of the posterior probability distribution is estimated by combining Bayesian uncertainty quantization to obtain the probability ranking result and determine the optimal discharge type. Based on the optimal discharge type and the spatial positioning coordinates, the consistency of the equipment structural parameters is verified to obtain the verification result. The verified discharge data is then matched with the pre-acquired historical discharge data to obtain a similarity score. Based on the similarity score, the severity rating of the partial discharge is calculated and an identification report is output.

[0020] In one optional implementation, the discharge electromagnetic signals of primary equipment in a substation are synchronously acquired using a sensor array. The discharge electromagnetic signals are then synchronously calibrated, and pulse arrival time difference and signal amplitude features are extracted. Spatial positioning coordinates are constructed based on the pulse arrival time difference, and a multimodal feature set is generated by combining the signal amplitude features. The spatial position information and clock parameters of each sensor are obtained from the sensor array. Based on the clock parameters, the discharge electromagnetic signals collected by each sensor are time-aligned to obtain a synchronous electromagnetic signal. The synchronous electromagnetic signal is pulse-detected to identify the pulse start time and the pulse arrival time difference between adjacent sensors is calculated. Based on the pulse arrival time difference and the spatial position information, triangulation is performed to obtain the spatial positioning coordinates. The peak amplitude and effective amplitude of each frequency band are extracted from the synchronous electromagnetic signal to form the signal amplitude feature. The statistical features of the signal amplitude feature in the time domain are calculated and the power spectral density distribution in the frequency domain is extracted. The statistical features and the power spectral density distribution are vector-concatenated to obtain the single-mode feature vector. The single-mode feature vectors corresponding to different sensors are normalized and arranged according to the sensor number to form a feature matrix. The spatial distance between the power source and each sensor is calculated based on the spatial positioning coordinates, and a distance weight matrix is ​​constructed. The distance weight matrix and the feature matrix are used to solve for the distance weight matrix to obtain the distance weighted feature matrix. Principal component analysis is performed on the distance weighted feature matrix to reduce its dimensionality and retain the principal components to obtain the multimodal feature set.

[0021] The discharge electromagnetic signals of primary equipment in the substation are synchronously acquired using a sensor array. In this implementation, eight high-sensitivity broadband electromagnetic sensors are arranged in an array and deployed around the primary equipment to be monitored within the substation in a three-dimensional spatial structure. Each sensor has a unique identifier and its spatial coordinates are recorded. The sensor sampling frequency is 100 MHz, the acquisition bandwidth is 3 MHz to 3 GHz, and the sensitivity is not less than -110 dB / mW. All sensors are connected to the same data acquisition system and synchronized via a global timer with a synchronization accuracy better than 10 nanoseconds.

[0022] The collected discharge electromagnetic signals are synchronously calibrated to obtain the spatial position information of each sensor, including three-dimensional coordinates, with an accuracy down to the centimeter level. Simultaneously, the clock parameters of each sensor are acquired, including clock deviation and drift rate. Based on these parameters, the signals collected by each sensor are time-aligned. Specifically, a reference sensor is selected as the time base, and the time deviation of other sensors relative to the reference sensor is calculated. The time axes of each signal are then shifted to compensate for this shift. For example, if the clock deviation between reference sensor 1 and sensor 2 is 15 nanoseconds, the entire time axis of sensor 2 is shifted forward by 15 nanoseconds to achieve time alignment and obtain a synchronous electromagnetic signal.

[0023] Pulse detection is performed on the synchronous electromagnetic signal, and time difference features are extracted. A dynamic threshold detection method is used to identify the pulse start time. Specifically, the mean of the signal within a 25-microsecond sliding window plus three times the standard deviation is calculated as the dynamic threshold. The pulse start time is marked when the signal amplitude first exceeds this threshold. This operation is performed on the signal of each sensor to obtain a series of pulse detection points. The pulse arrival time difference between adjacent sensors is calculated to form a time difference matrix. For example, in a practical case, the pulse start time detected by sensor 1 is 25.436 microseconds, and that of sensor 2 is 25.489 microseconds, so the arrival time difference is 53 nanoseconds.

[0024] Triangulation is performed based on the pulse arrival time difference and sensor spatial location information to determine the spatial location of the discharge source. A system of time difference equations is constructed using the hyperbolic positioning principle. Assuming the discharge source location is unknown, and based on the electromagnetic wave propagation speed being the speed of light, combined with the pulse arrival time differences detected by each sensor and the known sensor spatial coordinates, the spatial coordinates of the discharge source are obtained by iteratively solving the system of equations. In practical applications, the Taylor series expansion method is used to solve the nonlinear equations, with the iterative convergence condition set to a position change of less than 1 millimeter, determining the three-dimensional coordinates of the discharge source with centimeter-level positioning accuracy. For example, in one test, the determined coordinates of the discharge source were insulators at 0.76 meters, 1.53 meters, and 2.14 meters within the substation.

[0025] Multimodal features are constructed by extracting signal amplitude features from synchronous electromagnetic signals. The signal is divided into three bands: low frequency (3-30 MHz), mid frequency (30-300 MHz), and high frequency (300 MHz-3 GHz), respectively. Peak amplitude and RMS amplitude are extracted for each band to form a six-dimensional amplitude feature. For example, in a certain measurement, sensor 1's low-frequency peak amplitude is 86 mV, and its RMS amplitude is 25 mV; the mid-frequency peak amplitude is 65 mV, and its RMS amplitude is 18 mV; the high-frequency peak amplitude is 42 mV, and its RMS amplitude is 12 mV.

[0026] The statistical characteristics of the amplitude features in the time domain are calculated, including mean, standard deviation, skewness, kurtosis, and variance, resulting in a 30-dimensional statistical feature vector. A Fast Fourier Transform is performed on the original signal to extract the power spectral density distribution. The spectrum is divided into 50 equally divided frequency bands, and the power proportion within each band is calculated, resulting in a 50-dimensional frequency domain feature vector. The statistical features are concatenated with the frequency domain feature vector to form an 80-dimensional single-mode feature vector. This process is performed on each of the eight sensors. The resulting single-mode feature vectors are then normalized so that the feature values ​​of each dimension are distributed between 0 and 1. These vectors are then arranged in order of sensor number to form an 8x80 feature matrix.

[0027] A distance weight matrix is ​​constructed based on the spatial distances between the discharge source and each sensor, calculated using spatial positioning coordinates. For a given location of the discharge source and the spatial coordinates of each sensor, the Euclidean distance between them is calculated. For example, the distance between the discharge source and sensor 1 is 1.25 meters, and the distance to sensor 2 is 1.47 meters. The distance weight matrix is ​​constructed using an inverse distance function, with closer sensors receiving greater weights. In the specific implementation, a negative exponential function of distance is used to generate the weights, ensuring that the weight values ​​are between 0 and 1, and that the weights decrease as the distance increases. For example, the weight of sensor 1 is 0.9, and the weight of sensor 2 is 0.86.

[0028] Multiplying the distance weight matrix by the feature matrix yields the distance-weighted feature matrix, which reflects the influence of spatial location information on feature extraction. Principal component analysis is then performed on the distance-weighted feature matrix, selecting principal components with a cumulative contribution rate reaching 95%, typically reducing the number of dimensions to 15 to 20, thus forming a multimodal feature set.

[0029] In this embodiment, the discharge electromagnetic signals of the primary equipment in the substation are synchronously acquired by a sensor array. Combined with a unified clock parameter, time alignment of multi-sensor signals and calculation of pulse arrival time difference are achieved. This allows for accurate comparison of discharge information acquired by different sensors under a unified time reference, improving the time consistency and positioning accuracy of discharge signal location calculation. By introducing sensor spatial location information during the positioning process and using pulse arrival time difference for triangulation, accurate inversion of the spatial coordinates of the discharge source is achieved. This transforms discharge source positioning from traditional experience-based judgment or single-dimensional signal analysis to multi-point joint positioning under spatial geometric constraints, significantly improving the reliability and spatial resolution of discharge source positioning. By extracting multi-band peak amplitude and effective amplitude from the synchronous electromagnetic signal and constructing multi-dimensional signal features by combining time-domain statistical characteristics and frequency-domain power spectral density distribution, and then uniformly normalizing and matrixing the features of different sensors, the time-domain and frequency-domain information of the discharge signal can be systematically integrated. This allows for a more comprehensive characterization of the dynamic characteristics of the discharge electromagnetic signal, improving the completeness and discriminative ability of discharge feature expression.

[0030] In one optional implementation, the multimodal feature set is decomposed using multi-scale time-frequency transformation to obtain localized feature matrices for different frequency bands, and the intermodal coupling relationship graph is obtained by performing correlation measurement on the localized feature matrices, including: The multimodal feature set is grouped according to sensor modes to obtain feature subsets corresponding to each sensor. For each feature subset, the corresponding optimal wavelet basis is determined and wavelet transform is performed to obtain multi-scale time-frequency coefficients. The multi-scale time-frequency coefficients are subjected to singular value decomposition and the principal singular vectors are extracted to form frequency band feature vectors. The frequency band feature vectors at different scales are arranged in ascending order of frequency to obtain the localized feature matrix. The mutual information between the localized feature matrices corresponding to different sensors is calculated. Based on the mutual information, the nonlinear correlation strength is calculated to obtain the information correlation matrix. The information correlation matrix is ​​divided into modal clusters by spectral clustering algorithm and the inter-cluster connection relationship is identified to obtain the sparse correlation matrix. The non-zero elements in the sparse correlation matrix are used as edge weights to construct the sensor connection graph to obtain the initial coupling relationship graph. For each sensor node in the initial coupling graph, the node access frequency is calculated using a random walk sampling strategy to obtain a node centrality index. Based on the node centrality index, key sensor nodes are identified and the weights of the connection edges of the key sensor nodes are adjusted. The initial coupling graph after weight adjustment is then normalized using a Laplace matrix to obtain the intermodal coupling graph.

[0031] The multimodal feature set was divided into eight feature subsets, each corresponding to the feature representation of a single sensor. Each feature subset contains 15- to 20-dimensional feature vectors obtained after dimensionality reduction using principal component analysis. For each feature subset, Debord wavelet, Haar wavelet, Sinica wavelet, and Kuiper wavelet were used for testing. The performance of the wavelet basis was evaluated by calculating three metrics: information entropy, energy concentration, and correlation coefficient. The wavelet basis with the lowest information entropy, highest energy concentration, and lowest correlation coefficient was selected as the optimal wavelet basis. In practical applications, the 4th order Debord wavelet typically performs best for transformer partial discharge signals, while the 5th order Sinica wavelet performs better for switchgear partial discharge signals.

[0032] A five-level wavelet transform is performed on the feature subset using the selected optimal wavelet basis to obtain multi-scale time-frequency coefficients. Taking transformer partial discharge as an example, a fourth-order Debord wavelet transform is applied to the feature subset of sensor 1 to obtain five scale coefficients from low frequency to high frequency, each scale corresponding to the time-frequency characteristics of different frequency bands. Singular value decomposition is performed on these multi-scale time-frequency coefficients to extract the first principal singular vector as the frequency band feature vector. In specific implementation, a matrix is ​​constructed for the time-frequency coefficients of each scale, and the corresponding singular values ​​and singular vectors are calculated. The singular vector corresponding to the largest singular value is selected as the feature representation of that frequency band. In a certain test case, 12-dimensional, 15-dimensional, 18-dimensional, 20-dimensional, and 22-dimensional frequency band feature vectors are extracted from the five scales of sensor 1, respectively.

[0033] The frequency band feature vectors at different scales are arranged in ascending order of frequency to form a localized feature matrix. The frequency band feature vectors corresponding to the low-frequency scale are located at the front of the matrix, and the frequency band feature vectors corresponding to the high-frequency scale are located at the back of the matrix, forming a complete spectral representation. Taking sensor 1 as an example, the localized feature matrix has a size of 87 rows and 16 columns, containing a complete signal feature representation from low frequency to high frequency. The same operation is performed on all 8 sensors to obtain 8 localized feature matrices.

[0034] Calculating the mutual information between the localized feature matrices of different sensors is fundamental to constructing the information correlation matrix. For any two sensors' localized feature matrices, the mutual information is calculated by estimating the joint probability distribution and marginal probability distribution. The probability distribution is calculated using kernel density estimation with a Gaussian kernel function and a bandwidth parameter of 0.2. The mutual information calculation yields an 8×8 mutual information matrix, where element values ​​range from 0 to 1; larger values ​​indicate stronger information correlation between the two sensors. For example, in a transformer partial discharge detection case, the mutual information between sensor 1 and sensor 2 is 0.72, while the mutual information between sensor 1 and sensor 8 is only 0.15.

[0035] The information correlation matrix is ​​obtained by calculating the nonlinear correlation strength based on mutual information. An exponential function is used to convert the mutual information into nonlinear correlation strength; sensor pairs with high mutual information receive higher correlation strength values. A threshold of 0.4 is set, and correlation strengths below this threshold are set to 0, thus sparsifying the correlation matrix. A spectral clustering algorithm is applied to the information correlation matrix to divide the sensor nodes into modal clusters. In practice, the eigenvalues ​​and eigenvectors of the information correlation matrix are calculated, and positions with significant intervals in the eigenvalue sorting are selected as the basis for determining the number of clusters; typically, 8 sensors are divided into 2 to 3 modal clusters.

[0036] A sparse correlation matrix is ​​constructed by identifying the connections between modal clusters. Strong intra-cluster connections and representative inter-cluster connections are retained, while other weak connections are deleted. For example, in the partial discharge detection of a circuit breaker in a substation, eight sensors are divided into three modal clusters: sensors 1, 2, and 3 form the first cluster; sensors 4 and 5 form the second cluster; and sensors 6, 7, and 8 form the third cluster. The correlation strength of all sensor pairs within a cluster is retained, while only the correlation strength between sensors 3 and 4, and sensors 5 and 6, is retained between clusters; other correlation strengths are set to 0. The non-zero elements in the sparse correlation matrix are used as edge weights to construct a connection graph between sensors, resulting in an initial coupling graph.

[0037] For each sensor node in the initial coupling graph, a random walk sampling strategy is applied to calculate the node access frequency. The restart probability is set to 0.15, and the number of walks is 10,000. The number of times each node is visited is recorded and normalized to obtain the node centrality index. Nodes with high access frequencies are identified as critical sensor nodes; typically, the top three nodes with the highest centrality index are selected. For example, in a transformer partial discharge detection case, the centrality indices of sensor 2, sensor 3, and sensor 5 are 0.22, 0.19, and 0.17, respectively, and they are identified as critical sensor nodes.

[0038] The edge weights connecting key sensor nodes are adjusted to enhance their influence in the network. This is achieved by multiplying the original weights of the edges connecting key nodes by an amplification factor of 1.5, thus strengthening the connections between key nodes and other nodes. The Laplacian matrix is ​​then calculated and normalized on the initial coupling graph after weight adjustment to obtain the intermodal coupling graph.

[0039] In this embodiment, by grouping the multimodal feature set according to sensor modes and adaptively selecting the optimal wavelet basis for multi-scale time-frequency analysis for different sensor feature subsets, the features of the discharge signal at different time and frequency scales can be fully unfolded. This allows for a more accurate characterization of the transient change characteristics and spectral distribution of the discharge electromagnetic signal, thereby improving the refinement and adaptability of signal feature extraction. By performing singular value decomposition on the multi-scale time-frequency coefficients and extracting the principal singular vector to construct the frequency band feature vector, complex time-frequency information can be effectively compressed and structured while maintaining the main energy characteristics, thereby reducing redundant information and noise interference. This significantly improves the stability and discriminative ability of feature expression. By calculating the mutual information of the localized feature matrices between different sensors and further constructing the information correlation matrix, a quantitative description of the nonlinear correlation between signals from different sensors is achieved, which can more accurately reflect the coupling characteristics of complex discharge signals among multiple sensors.

[0040] In one optional implementation, performing graph neural network encoding on the intermodal coupling graph to obtain the discharge source feature representation vector includes: The connectivity distribution and edge weight distribution of each sensor node are extracted from the intermodal coupling relationship graph to obtain graph topology statistics. Based on the graph topology statistics, the aggregation range of the node neighborhood is calculated and the multi-order neighborhood boundary is determined. The localized feature matrix corresponding to each sensor node is used as the initial feature of the node and combined with the multi-order neighborhood boundary to construct the node feature propagation matrix. The node feature propagation matrix is ​​subjected to tensor decomposition and low-rank approximation reconstruction to obtain the node embedding feature. The node embedding features are arranged into a graph feature matrix according to the sensor number and multi-layer convolution operation is performed with the adjacency matrix of the intermodal coupling relationship graph. In each layer convolution operation, the spectral correlation of the neighborhood features is calculated based on the spectral domain frequency decomposition and the node representation is updated by combining the residual connection. The updated node representation is batch normalized and multi-layer convolution output is accumulated to obtain a multi-scale node representation. The global graph representation vector is calculated based on the multi-scale node representation. The global graph representation vector is adaptively sampled and aggregated, and the node importance weights are calculated by combining the multi-scale node representations. The multi-scale node representations are then weighted and fused based on the node importance weights to obtain fused node features. The fused node features are then mapped and transformed through a fully connected layer and combined with linear activation to obtain the discharge source feature representation vector.

[0041] Graph topology statistics are extracted from the established intermodal coupling relationship graph. For each sensor node, the corresponding connectivity degree, i.e., the number of edges directly connected to it, is calculated. For example, in a transformer partial discharge detection case, sensor 2 has a connectivity degree of 4, and sensor 7 has a connectivity degree of 2. Simultaneously, the weight distribution of the edges connecting each node is extracted, including statistics such as maximum weight, minimum weight, average weight, and weight variance. Statistics are obtained by statistically analyzing the connectivity degree and edge weight distribution of all 8 nodes, including the mean node connectivity degree of 3.25, the standard deviation of connectivity degree of 1.28, the mean edge weight of 0.54, and the standard deviation of edge weight of 0.17.

[0042] The aggregation range of node neighborhoods is calculated based on graph topological statistics to determine the multi-order neighborhood boundary. An adaptive threshold method is used to determine the range of neighborhood hop counts based on the connectivity and weight distributions. In the specific implementation, the upper limit of the neighborhood hop count is set to 1.5 times the mean connectivity and rounded up; in the aforementioned example, this is 5 hops. For each node, the set of neighboring nodes is calculated sequentially from 1-hop neighborhood to 5-hop neighborhood. When the number of newly added nodes in the neighborhood is less than the threshold or the neighborhood hop count reaches the upper limit, the multi-order neighborhood boundary is determined. Taking sensor 3 as an example, the 1-hop neighborhood includes sensors 1, 2, and 4; the 2-hop neighborhood adds sensors 5 and 6; the 3-hop neighborhood adds sensor 7; the 4-hop neighborhood adds sensor 8; and the 5-hop neighborhood has no newly added nodes. Therefore, its multi-order neighborhood boundary is 4 hops.

[0043] The localized feature matrix corresponding to each sensor node is used as the initial feature of the node, and a node feature propagation matrix is ​​constructed by combining it with multi-order neighborhood boundaries. The initial feature of each node is the aforementioned generated localized feature matrix, with a matrix size of approximately 87×16. When constructing the feature propagation matrix, a weighted combination of the node's own features and the features of neighboring nodes of different orders is considered. The propagation weights decrease as the number of hops in the neighborhood increases, with a decay coefficient set to 0.6. For example, in the feature propagation matrix of sensor 3, the weight of its own feature is 1, the weight of the 1-hop neighbor feature is 0.6, the weight of the 2-hop neighbor feature is 0.36, the weight of the 3-hop neighbor feature is 0.216, and the weight of the 4-hop neighbor feature is 0.13. The resulting feature propagation matrix has a dimension of 8×8×87×16, containing the complete feature propagation relationship between nodes.

[0044] Tensor decomposition and low-rank approximation reconstruction are performed on the node feature propagation matrix to obtain node embedding features. A higher-order singular value decomposition method is used to decompose the fourth-order tensor into the product of a core tensor and factor matrices of each pattern. A rank of 32 is chosen for the core tensor, preserving over 95% of the information. The low-rank approximation tensor is reconstructed using the factor matrices to obtain compressed node embedding features with a dimension of 8×32. For example, the embedding features of sensor 1 are a 32-dimensional vector with values ​​ranging from -0.8 to 0.8, capturing the key feature patterns of this node and its neighboring nodes.

[0045] The node embedding features are arranged into a graph feature matrix of size 8×32 according to the sensor number, representing the 32-dimensional embedding features of 8 sensor nodes. An adjacency matrix of size 8×8 is extracted from the inter-modal coupling graph, representing the connection relationships and weights between nodes. Multi-layer convolution operations are performed on the graph feature matrix and the adjacency matrix to construct a deep graph convolutional network. The network is set to a 3-layer structure, with 64, 128, and 256 kernels in each layer, respectively. In each convolution operation, the spectral correlation of neighborhood features is calculated based on spectral domain frequency decomposition. Specifically, the adjacency matrix is ​​eigenvalued and its eigenvectors are obtained, and a frequency domain filter is constructed to filter the node features in the frequency domain. The filtering parameters are approximated using Chebyshev polynomials, with the polynomial order set to 3.

[0046] In each convolutional layer, residual connections are used to update the node representations. The input features are added to the convolutional output features to form a residual structure, mitigating the gradient vanishing problem in deep networks. The updated node representations are then batch-normalized using parameters including mean and variance, with a momentum coefficient set to 0.9. The outputs of the three convolutional layers are accumulated to obtain multi-scale node representations. The first convolutional layer captures local structural features with an output dimension of 8×64; the second convolutional layer captures medium-range structural features with an output dimension of 8×128; and the third convolutional layer captures global structural features with an output dimension of 8×256. These three outputs are concatenated to form an 8×448 multi-scale node representation matrix.

[0047] The global graph representation vector is calculated based on multi-scale node representations. An attention pooling mechanism is employed to calculate an attention score for each node, followed by a weighted summation. The attention score is calculated using a single-layer neural network, with the multi-scale representation of the node as input and a scalar score as output. After weighted summation, a 448-dimensional global graph representation vector is obtained. For example, in a partial discharge detection case of a circuit breaker in a substation, the attention scores of sensors 2, 3, and 5 are 0.22, 0.19, and 0.17, respectively, while the attention scores of the remaining sensors are all below 0.15.

[0048] An adaptive sampling and aggregation process is performed on the global graph representation vector, selecting key feature dimensions through an importance sampling strategy. With a sampling ratio of 0.75, 336 important dimensions are selected from 448 features. Node importance weights are calculated by combining multi-scale node representations; these weights reflect the contribution of each node to the discharge source feature representation. The calculation method is the cosine similarity between the node representation and the global graph representation, followed by soft maximum normalization. Taking the partial discharge detection of a transformer as an example, the node importance weights of sensors 2, 3, and 5 are 0.24, 0.21, and 0.18, respectively, significantly higher than other sensor nodes.

[0049] The multi-scale node representations are weighted and fused based on node importance weights to obtain fused node features. The multi-scale representation of each node is multiplied by its corresponding importance weight, and then summed to obtain a 448-dimensional fused feature vector. The fused node features are then transformed through a fully connected layer and combined with linear activation to obtain the discharge source feature representation vector. The fully connected layer has an input dimension of 448 and an output dimension of 128. The weight matrix is ​​optimized using the backpropagation algorithm to obtain a 128-dimensional discharge source feature representation vector.

[0050] In this embodiment, by extracting the connectivity distribution and edge weight distribution of sensor nodes from the intermodal coupling graph, and determining the multi-order neighborhood range of nodes based on graph topological statistics, each sensor node can fully consider neighborhood structural information during feature propagation. This enables a systematic modeling of the spatial relationships between sensors, more comprehensively reflecting the structural dependencies between multiple sensors, and improving the ability of feature representation to depict the propagation laws of complex discharge signals. By using the localized feature matrix as the initial feature of the nodes and constructing the node feature propagation matrix, and then reconstructing it through tensor decomposition and low-rank approximation pairs, the compactness and stability of the feature representation can be improved, and the computational efficiency of the subsequent feature learning process can be enhanced. By performing multi-layer convolution operations on node embedding features and graph structure adjacency matrices, and introducing spectral frequency decomposition during the convolution process to calculate the spectral correlation between neighborhood features, the node feature update not only considers spatial adjacency relationships but also characterizes the correlation characteristics between different frequency components. This enhances the ability to express complex changing patterns of discharge electromagnetic signals and improves the discriminative ability of feature representation. By accumulating the multi-layer convolution outputs to form multi-scale node representations and further calculating the global graph representation vector, local node features and overall network structure information can be obtained simultaneously, thereby constructing a more complete sensor network feature representation that can more accurately reflect the overall feature distribution of the discharge source in the multi-sensor network.

[0051] In one optional implementation, semantic matching is performed between the discharge source feature representation vector and a preset discharge mode knowledge graph to obtain a candidate discharge type set. The posterior probability distribution of the discharge type corresponding to the discharge source feature representation vector is calculated based on a variational inference algorithm, including: Feature template vectors and hierarchical relationships between discharge types are extracted from a pre-defined discharge mode knowledge graph. A discharge type feature library is constructed based on the feature template vectors, and the Riemann manifold distance between the discharge source feature representation vector and each feature template vector in the discharge type feature library is calculated to obtain a manifold similarity matrix. Based on the manifold similarity matrix, high-similarity discharge types are determined, and related parent types are expanded by combining the hierarchical relationships between types to obtain the candidate discharge type set. Gaussian mixture distribution parameter estimation is performed on the discharge source feature representation vector to calculate the feature space probability density distribution. Based on the probability density distribution, multiple sets of feature variables are sampled and generated, and the similarity with the feature template vectors of each type in the candidate discharge type set is calculated to obtain the type similarity sequence. The occurrence frequency of each discharge type in the type similarity sequence is counted and normalized to obtain the type frequency distribution. Based on each discharge type in the candidate discharge type set, a multinomial distribution prior is constructed. The likelihood probability is obtained by performing expectation maximization iterative calculation on the type frequency distribution and the multinomial distribution prior. The likelihood probability is then randomly sampled to obtain a posterior sample set, and the posterior probability distribution of the discharge type corresponding to the discharge source feature representation vector is calculated.

[0052] Discharge type features are extracted from a pre-defined discharge mode knowledge graph, which includes 10 common discharge types: levitation discharge, corona discharge, insulation failure discharge, air gap discharge, surface discharge, internal void discharge, dendritic discharge, particle discharge, levitation potential discharge, and tip discharge. Each discharge type has a specific feature template vector with the same 128 dimensions as the aforementioned discharge source feature representation vector. For example, the feature template vector for levitation discharge exhibits a high proportion of high-frequency components in the frequency domain, with values ​​ranging from 0.15 to 0.82; while the feature template vector for corona discharge displays a regular pulse sequence characteristic in the time domain, with values ​​ranging from -0.45 to 0.68.

[0053] The discharge types exhibit a hierarchical relationship, forming a tree-like structure. Within this tree structure, the first-level types include surface discharge, internal discharge, and external discharge; the second-level types include the aforementioned 10 specific discharge types: surface discharge includes surface discharge and dendritic discharge; internal discharge includes internal void discharge and particle discharge; and external discharge includes suspended discharge, corona discharge, insulation failure discharge, air gap discharge, suspended potential discharge, and tip discharge. A discharge type feature library is constructed based on the aforementioned feature template vectors, storing the standard feature representations of each type.

[0054] The Riemannian manifold distance between the discharge source feature representation vector and each feature template vector in the discharge type feature library is calculated to obtain the manifold similarity matrix. The Riemannian manifold distance calculation process includes: estimating the local curvature of the manifold containing the feature vector, constructing the geodesic equation, and calculating the geodesic distance between vectors. Specifically, the Riemannian metric tensor approximation method is used, with a local neighborhood size of 15 and a manifold dimension of 8. The calculated distance values ​​are converted into similarity scores, ranging from 0 to 1, with higher values ​​indicating greater similarity. For example, in the partial discharge identification of a transformer in a substation, the similarity between the discharge source feature representation vector and the internal void discharge feature template vector is 0.85, and the similarity with the dendritic discharge is 0.24.

[0055] High-similarity discharge types are determined based on the manifold similarity matrix. A similarity threshold of 0.7 is set, and discharge types with similarities greater than this threshold are selected as candidate types. The hierarchical relationship between types is then considered to expand related parent types. In the previous example, the internal cavity discharge has a similarity of 0.85, exceeding the threshold, and is therefore included in the candidate discharge type set; its parent type, internal discharge, is also included in the candidate discharge type set. The final candidate discharge type set includes three types: internal cavity discharge, particle discharge, and internal discharge.

[0056] Gaussian mixture distribution parameters are estimated for the feature representation vector of the discharge source, and the probability density distribution in the feature space is calculated. The number of Gaussian mixture components is set to 5, and the mean vector, covariance matrix, and mixture weights of each component are estimated using the expectation-maximization algorithm. The iterative convergence condition is that the log-likelihood change is less than 0.001 or the maximum number of iterations (100) is reached. The obtained probability density distribution can describe the distribution characteristics of the feature vector in high-dimensional space, providing a probabilistic model for subsequent feature sampling.

[0057] Multiple sets of feature variables were generated based on probability density distribution sampling. The sampling quantity was set to 1000, and the Markov chain Monte Carlo method was used for sampling. The generated feature variables have similar distribution characteristics to the original discharge source feature representation vector. The similarity between the aforementioned feature variables and the feature template vectors of each type in the candidate discharge type set was calculated to obtain a type similarity sequence. The similarity calculation method is the same as the Riemannian distance method mentioned above. In the aforementioned example, among the 1000 feature variables, 628 have the highest similarity to internal void discharge, 245 have the highest similarity to particle discharge, and 127 have the highest similarity to internal discharge.

[0058] The frequency of each discharge type in the statistical type similarity sequence was calculated and normalized to obtain the type frequency distribution. In the above case, the frequency proportion of internal void discharge was 0.628, particle discharge was 0.245, and internal discharge was 0.127. The type frequency distribution reflects the degree of matching between the discharge source characteristics and each candidate discharge type in the sampling space, providing a data foundation for subsequent probabilistic inference.

[0059] A multinomial distribution prior is constructed based on each discharge type in the candidate discharge type set. The prior probabilities are set based on historical statistical data, reflecting the probability of each type occurring in practical applications. For example, the prior probability of internal cavity discharge is set to 0.35, particle discharge to 0.25, and internal discharge to 0.4. The likelihood probability is obtained by iteratively maximizing the expectation of the type frequency distribution and the multinomial distribution prior. During the iteration process, the parameters are updated until convergence. The convergence condition is that the parameter change is less than 0.0001 or the maximum number of iterations (50) is reached.

[0060] Random sampling is performed on the likelihood probabilities to obtain the posterior sample set. The Gibbs sampling method is used, with 10,000 samples taken and a combustion period of 1,000. Based on the posterior sample set, the posterior probability distribution of the discharge type corresponding to the discharge source feature representation vector is calculated. In the aforementioned transformer partial discharge example, the posterior probability of internal void discharge is 0.72, particle discharge is 0.21, and internal discharge is 0.07. The posterior probability distribution comprehensively considers feature similarity, type frequency, and prior knowledge, providing a probabilistic framework for discharge type identification.

[0061] In this embodiment, a manifold similarity matrix is ​​constructed by calculating the Riemannian manifold distance between the discharge source feature representation vector and the feature template vectors of each discharge type. This allows the similarity measurement between features to be described in a non-Euclidean space, more accurately reflecting the true geometric relationship between complex feature distributions. This effectively improves the accuracy of discharge type matching. By expanding the relevant parent types based on the hierarchical relationship between types to form a candidate discharge type set on the basis of high-similarity discharge types, the relationship between the specific type and its upper-level category can be considered simultaneously when performing discharge pattern recognition. This avoids misjudgment caused by a single feature matching error and improves the accuracy of discharge type recognition. To improve the fault tolerance and stability of the identification process, Gaussian mixture distribution modeling is performed on the characteristic representation vector of the discharge source, and the probability density distribution of the feature space is calculated. This allows the discharge characteristics to be described probabilistically in a statistical sense, which can characterize the uncertainty and multi-peak characteristics of the feature distribution. It can more realistically reflect the distribution law of the actual discharge signal in the feature space. By generating multiple sets of feature variables based on probability density distribution sampling and statistically analyzing the occurrence frequency of different discharge types to form a type frequency distribution, the discharge type identification results can obtain statistically stable judgment basis through multiple samplings, reducing the randomness of single matching results and improving the stability and reliability of the identification results.

[0062] In one optional implementation, combining Bayesian uncertainty quantification to estimate the confidence interval of the posterior probability distribution to obtain the probability ranking result and determine the optimal discharge type includes: The probability distribution statistics of each discharge type are extracted from the posterior probability distribution and the confidence ellipse boundary is calculated to obtain a set of confidence intervals. The intersection overlap rate of each discharge type in the set of confidence intervals is calculated based on the interval volume. High confusion type groups are marked based on the intersection overlap rate and the prediction uncertainty is quantified. Monte Carlo uncertainty propagation is performed on the probability distribution statistics based on the prediction uncertainty to obtain a corrected probability distribution. A probability matrix is ​​constructed based on the modified probability distribution, and the conditional entropy is calculated. The inter-type discrimination is evaluated based on the conditional entropy, and an adaptive weight coefficient is set. The modified probability distribution is subjected to a nonlinear transformation by the adaptive weight coefficient, and the type confidence is calculated in combination with the corresponding interval volume. The discharge types are sorted based on the type confidence to obtain the probability ranking result. The discharge type with the highest confidence level is selected from the probability ranking results as the initial optimal type. The statistical significance difference between the initial optimal type and the discharge type with the second highest confidence level is calculated. It is determined whether the statistical significance difference is greater than a preset discrimination threshold. If so, the initial optimal type is the optimal discharge type. Otherwise, the type with the largest information gain is selected from the probability ranking results according to the information gain maximization criterion, and the optimal discharge type is determined by combining the information gain maximization criterion with Nash equilibrium.

[0063] For precise identification, probability distribution statistics for each discharge type are extracted from the posterior probability distribution. For each discharge type, the mean vector, covariance matrix, skewness vector, and kurtosis value of the corresponding probability distribution are calculated. Taking the partial discharge detection of a transformer in a substation as an example, the posterior probability of internal void discharge has a mean of 0.72, a standard deviation of 0.08, a skewness of -0.35, and a kurtosis of 2.86; the posterior probability of particle discharge has a mean of 0.21, a standard deviation of 0.06, a skewness of 0.42, and a kurtosis of 3.12; and the posterior probability of internal discharge has a mean of 0.07, a standard deviation of 0.03, a skewness of 0.56, and a kurtosis of 3.45. Based on the aforementioned statistics, the confidence ellipse boundaries for each discharge type are calculated, and the set of 95% confidence intervals is determined. The covariance matrix eigenvalue decomposition method is used to calculate the principal axis direction and length, constructing a multidimensional ellipsoid. In the two-dimensional projection space, the center coordinates of the confidence ellipse for internal void discharge are (0.72, 0.68), the major axis length is 0.16, and the minor axis length is 0.12; the center coordinates of the confidence ellipse for particle discharge are (0.21, 0.25), the major axis length is 0.12, and the minor axis length is 0.08.

[0064] The overlap rate between different discharge types is calculated based on the volume of the confidence intervals. The volume calculation uses the Monte Carlo integration method, uniformly sampling 10,000 points in a high-dimensional space and counting the proportion of points falling within each confidence interval. The overlap rate is obtained by calculating the proportion of points simultaneously falling within two confidence intervals out of the total number of points. In the aforementioned example, the overlap rate between internal void discharge and particle discharge is 0.12, between internal void discharge and internal discharge is 0.06, and between particle discharge and internal discharge is 0.15. High-confusion type groups are labeled based on the overlap rate, with a threshold of 0.10. When the overlap rate of two types exceeds this threshold, they are labeled as high-confusion type pairs. This identifies internal void discharge and particle discharge, and particle discharge and internal discharge, as high-confusion type groups.

[0065] The uncertainty in predictions is quantified based on highly confused type groups, calculated using the weighted entropy of the posterior probabilities of each type within the highly confused type pair. In the aforementioned example, the prediction uncertainty for internal cavity discharge and particle discharge is 0.28, and the prediction uncertainty for particle discharge and internal discharge is 0.22. Monte Carlo uncertainty propagation is performed on the probability distribution statistics based on the prediction uncertainty, generating 5000 randomly perturbed samples. The perturbed probability distribution is then calculated to obtain the corrected probability distribution. The corrected posterior probability has a mean of 0.68 and a standard deviation of 0.09 for internal cavity discharge; a mean of 0.24 and a standard deviation of 0.07 for particle discharge; and a mean of 0.08 and a standard deviation of 0.04 for internal discharge.

[0066] A probability matrix is ​​constructed based on a modified probability distribution, with matrix elements representing the joint and marginal probabilities of each discharge type. The conditional entropy of the probability matrix is ​​calculated to quantify the distinguishability between types. The conditional entropy is calculated using the Shannon entropy formula, with the natural logarithm as the logarithm base. In the previous example, the conditional entropy of internal void discharge relative to particle discharge is 0.45, the conditional entropy of internal void discharge relative to internal discharge is 0.28, and the conditional entropy of particle discharge relative to internal discharge is 0.62. A smaller conditional entropy value indicates higher distinguishability. The distinguishability between types is evaluated based on the conditional entropy, and adaptive weighting coefficients are set. The weighting coefficients are inversely proportional to the conditional entropy; types with higher distinguishability receive greater weights. The weighting coefficient for internal void discharge is 0.58, for particle discharge it is 0.26, and for internal discharge it is 0.16.

[0067] An adaptive weighting coefficient was applied to the corrected probability distribution for nonlinear transformation using a power function transformation method, with the exponent value as the weighting coefficient. After the transformation, the corrected probability for internal void discharge was 0.74, for particle discharge it was 0.18, and for internal discharge it was 0.08. The type confidence score was calculated based on the corresponding confidence interval volume, using the formula: weighted corrected probability divided by the square root of the confidence interval volume. The confidence scores for internal void discharge were 0.85, for particle discharge it was 0.26, and for internal discharge it was 0.15. Based on the type confidence scores, the discharge types were ranked, resulting in the following probability ranking: internal void discharge (0.85), particle discharge (0.26), and internal discharge (0.15).

[0068] The discharge type with the highest confidence level from the probability ranking results is selected as the initial optimal type, i.e., internal void discharge. The statistical significance difference between the initial optimal type and the discharge type with the second highest confidence level is calculated. A modified Welch t-test is used to calculate the significance difference, considering unequal sample sizes and variances. In the previous example, the t-test statistic for internal void discharge and particle discharge is 18.65, corresponding to a p-value of 0.0000023, and the significance difference is 0.59. It is determined whether the statistical significance difference is greater than a preset discrimination threshold of 0.45. If it is greater than the threshold, the initial optimal type is determined as the optimal discharge type. In this embodiment, the significance difference of 0.59 is greater than the threshold of 0.45, thus internal void discharge is determined as the optimal discharge type.

[0069] In some challenging cases of partial discharge identification in substation circuit breakers, the significant difference between the initial optimal and suboptimal discharge types may not be substantial enough. For example, the confidence level for air gap discharge is 0.62, and the confidence level for insulation failure discharge is 0.58, with a statistically significant difference of 0.32, which is less than the preset discrimination threshold of 0.45. In such cases, the type with the highest information gain needs to be selected from the probability ranking results based on the information gain maximization criterion. Information gain calculation is based on the reduction of cross-entropy. In a certain circuit breaker case, the information gain for air gap discharge is 0.38, and the information gain for insulation failure discharge is 0.45. Simultaneously, Nash equilibrium theory is introduced to construct a decision matrix and calculate the expected returns for each type under different decision conditions. The return value of insulation failure discharge at the Nash equilibrium point is 0.64, higher than that of air gap discharge (0.56). Therefore, insulation failure discharge is ultimately determined as the optimal discharge type.

[0070] In this embodiment, by extracting probability statistics from the posterior probability distribution and constructing confidence ellipse boundaries, a corresponding set of confidence intervals is formed. This allows the discharge type prediction result to not only provide the probability magnitude but also reflect the spatial uncertainty range of the probability distribution, thereby achieving a quantitative description of the credibility of the recognition result. This provides a more comprehensive characterization of the stability and reliability of the prediction result. By calculating the intersection overlap rate between confidence intervals of different discharge types and identifying highly confused type groups, the recognition reliability in complex scenarios can be effectively improved. By introducing Monte Carlo uncertainty propagation based on the quantified prediction uncertainty, the original probability distribution is subjected to multiple random sampling and propagation calculations. This allows the probability estimation to comprehensively consider feature perturbations and model uncertainties, significantly reducing the impact of random errors and noise on the recognition result. By combining interval volume information to calculate the type confidence and sorting the discharge types, the discharge type evaluation considers not only the probability magnitude but also the stability of the probability distribution, improving the robustness of the recognition decision.

[0071] Figure 2 This is a flowchart illustrating the confidence assessment of discharge type in the intelligent identification method for partial discharge of primary equipment in substations, as described in this embodiment of the invention.

[0072] In one optional implementation, the consistency of device structural parameters is verified based on the optimal discharge type and the spatial positioning coordinates to obtain a verification result. The verified discharge data is then compared with pre-acquired historical discharge data to obtain a similarity score. Based on the similarity score, the severity rating of the partial discharge is calculated, and an identification report is output, including: The standard structural parameters corresponding to the optimal discharge type are obtained from the preset equipment structural parameter database. The geometric parameters of the equipment structure are calculated based on the spatial positioning coordinates and the difference is calculated with the standard structural parameters. When the difference is less than the preset consistency threshold, the verification is marked as passed and the verification result is obtained. Extract feature parameters from the verified discharge data to construct the current discharge feature vector, extract historical feature vectors of the same type of discharge from the historical discharge data, calculate the time-normalized distance between the current discharge feature vector and the historical feature vector and normalize it to obtain the similarity score; Adaptive fuzzy clustering analysis is performed based on the similarity score to obtain the severity classification boundary. The similarity score is then mapped to a severity rating based on the severity classification boundary. Finally, an identification report is generated by combining the optimal discharge type and the spatial positioning coordinates.

[0073] Verify the identification results and assess the severity of the discharge. Retrieve the standard structural parameters corresponding to the optimal discharge type from a pre-defined equipment structural parameter database. This database covers standard structural parameters for various primary equipment in substations, including geometric dimensions, material properties, and design gap values. For example, for a 220 kV oil-immersed transformer, the standard structural parameters for internal void discharge include: bushing outer diameter 85 mm, thickness 12 mm, insulation distance 85 mm, and dielectric constant 3.2. For a 330 kV circuit breaker, the standard structural parameters for air gap discharge include: contact gap 4.5 mm, insulator length 2250 mm, and gas pressure 0.6 MPa.

[0074] Calculating the geometric parameters of an equipment structure based on spatial positioning coordinates involves three-dimensional spatial measurement and geometric calculation. Spatial positioning coordinates are typically obtained using ultrasonic arrays or electromagnetic interference positioning systems, with an accuracy of ±5 cm. In a 220 kV transformer case, the spatial positioning coordinates of the power supply are (125 cm, 78 cm, 210 cm), corresponding to the upper insulating bushing area inside the transformer. Based on these coordinates, relevant geometric parameters are calculated: distance from the top cover 45 cm, distance from the main tank wall 72 cm, and distance from the core column 83 cm. The calculated geometric parameters are then compared with standard structural parameters using Euclidean distance normalization. The difference calculation considers the importance weight of each parameter, assigning higher weights to critical parameters. In this transformer case, the calculated difference value is 0.18, which is less than the preset consistency threshold of 0.25; therefore, the verification is passed, and the verification result is obtained.

[0075] Feature parameters are extracted from the verified discharge data to construct the current discharge feature vector. These features include two categories: frequency domain features and time domain features. Frequency domain features include: discharge signal main frequency of 4.8 MHz, bandwidth of 0.7 MHz, high-frequency component proportion of 36%, and spectral attenuation rate of 0.82. Time domain features include: pulse amplitude of 65 mV, rise time of 0.4 μs, duration of 2.2 μs, repetition rate of 72 Hz, and phase angle distribution concentrated around 25 degrees and 205 degrees. These features are combined into a 32-dimensional current discharge feature vector.

[0076] Historical feature vectors of similar discharge types were extracted from historical discharge data. The historical database stores discharge records of substation equipment over the past 5 years, containing more than 5,000 labeled discharge data entries. For the aforementioned case of internal void discharge in the transformer, 358 historical internal void discharge records were retrieved. Each record contains a 32-dimensional feature vector and a corresponding severity label. The time warping distance between the current discharge feature vector and historical feature vectors was calculated. The time warping distance calculation adopted a dynamic time warping algorithm, setting the time window width to 15% of the sample length and the step factor to 0.05, resulting in 358 time warping distance values. The distance values ​​were then subjected to max-min normalization and converted into similarity scores, ranging from 0 to 1, with higher values ​​indicating greater similarity. In the aforementioned transformer case, the obtained similarity scores ranged from 0.23 to 0.92, with an average of 0.58 and a median of 0.62.

[0077] Adaptive fuzzy clustering analysis based on similarity scores was used to obtain severity grading boundaries. A fuzzy C-means clustering algorithm was employed, with a fuzzy coefficient of 1.8, a maximum of 200 iterations, and a convergence threshold of 0.001. The initial number of cluster centers was set to 4, corresponding to four severity levels: mild, moderate, severe, and critical. Through iterative optimization, the membership degree of each sample to each cluster center was calculated, and the cluster centers were updated. After the clustering analysis was completed, the boundary values ​​between adjacent cluster centers were calculated to obtain the severity grading boundaries. In the aforementioned case of internal void discharge in a transformer, the calculated grading boundary values ​​were 0.35, 0.58, and 0.75. These boundary values ​​divided the similarity score space into four intervals, corresponding to the four severity levels.

[0078] The similarity score is mapped to a severity rating based on the severity grading boundary. The current case has a similarity score of 0.82, which is greater than the boundary value of 0.75, therefore it is classified as critical. To further quantify the severity, the relative position of the current score within the corresponding interval is calculated and mapped as a percentage. The relative position is calculated by subtracting the lower bound of the interval from the current score and then dividing by the interval width. In the example above, the critical level interval is 0.75 to 1.00, with an interval width of 0.25. The current score of 0.82 has a relative position of 28% within this interval, indicating it is at the 28th percentile of the critical level, and can be described as "Critical Level (28%)".

[0079] Generate an identification report by combining the optimal discharge type and spatial positioning coordinates. The identification report contains the following key information: equipment type (220 kV oil-immersed transformer), discharge type (internal cavity discharge), spatial positioning (upper bushing area, coordinates 125 cm, 78 cm, 210 cm), severity level (critical level, 28%), discovery time (June 15, 2023, 10:35:42), feature description (pulse amplitude 65 mV, main frequency 4.8 MHz, phase distribution concentrated at 25 degrees and 205 degrees), development trend (signal amplitude growth rate within the past 72 hours is 42%), recommended measures (recommend immediate outage for maintenance, with focus on inspecting the insulating cardboard in the area where the upper bushing is connected to the oil tank). The identification report is automatically pushed to the operation and maintenance personnel through the substation monitoring system, and different-level alarm mechanisms are triggered according to the severity level. For critical-level discharges, text messages and phone calls are simultaneously sent to the on-duty supervisor and technical person in charge.

[0080] In this embodiment, by obtaining the standard structure parameters corresponding to the optimal discharge type from the equipment structure parameter database and calculating the geometric parameters of the actual equipment structure in combination with the spatial positioning coordinates for consistency difference analysis, the discharge type recognition result can be cross-validated with the actual equipment structure conditions, improving the consistency and credibility between the discharge type recognition result and the actual equipment state. By extracting the discharge data that passes the verification to construct the current discharge feature vector and calculating the temporal alignment distance with the historical feature vectors in the historical discharge data of the same type, the discharge feature changes at different time scales can be effectively aligned and compared. By performing adaptive fuzzy clustering analysis on the normalized similarity scores, the classification boundaries of the discharge severity are automatically determined, enabling the severity division to be dynamically adjusted according to the actual data distribution without relying on fixed threshold rules, thereby improving the adaptability of the severity assessment to different equipment states and operating environments. By mapping the similarity scores to the corresponding severity classifications and generating a comprehensive identification report in combination with the optimal discharge type and spatial positioning results, more comprehensive, intuitive, and decision-making reference diagnostic information can be provided, thereby improving the identification efficiency and disposal accuracy of equipment operation and maintenance personnel for discharge faults.

[0081] In a preferred implementation manner, to further improve the reliability of partial discharge type recognition and severity rating in the complex substation site environment, this embodiment further introduces spatial positioning reliability, type structure adaptation score, and comprehensive severity score on the basis of the foregoing sensor array synchronous acquisition, multi-scale time-frequency transformation, graph neural network encoding, knowledge graph matching, variational inference, Bayesian uncertainty quantification, equipment structure parameter consistency verification, and historical discharge data similarity matching to form a closed-loop verification of the partial discharge recognition result.

[0082] Specifically, after obtaining the spatial positioning coordinates through triangulation based on the pulse arrival time difference and sensor spatial location information, the spatial positioning reliability is further determined based on the residuals from solving the positioning equations. Let the spatial positioning coordinates of the power supply be P, and the spatial positions of each sensor be... The pulse arrival time difference between the i-th sensor and the reference sensor is If the propagation speed of electromagnetic waves is c, then the residual of the positioning equation system can be expressed as:

[0083] in, Indicates the spatial position of the reference sensor. Let represent the positioning residual corresponding to the i-th sensor. The positioning residuals corresponding to all sensors involved in the positioning process are normalized to obtain the mean positioning residual r. avg Based on the mean positioning residual, the number of sensors involved in positioning, and the spatial distribution dispersion of the sensors, the spatial positioning reliability R is calculated. p Spatial positioning reliability R p The value ranges from 0 to 1, and the larger the value, the more reliable the spatial positioning coordinates.

[0084] In one specific implementation, the spatial positioning reliability R p It can be determined in the following way:

[0085] Where η represents the residual scale parameter, and N represents the number of effective sensors involved in the positioning. This indicates the preset baseline value for the number of sensors. This represents the normalized value of the spatial distribution dispersion of the sensors. Using this calculation method, spatial positioning reliability is high when the positioning residual is small, the number of effective sensors is sufficient, and the spatial distribution of sensors is relatively balanced; conversely, spatial positioning reliability decreases when the positioning residual is large, the number of effective sensors is insufficient, or the concentrated distribution of sensors leads to poor positioning geometry.

[0086] Furthermore, based on spatial positioning reliability The distance weight matrix is ​​corrected. Specifically, it is based on the spatial positioning coordinates P of the discharge source and the spatial position of the i-th sensor. Spatial distance between Calculate the base distance weight :

[0087] Where λ represents the distance attenuation coefficient. Subsequently, the spatial positioning reliability... By introducing a distance weight correction process, the corrected spatial weighting coefficients are obtained. :

[0088] in, This represents the balanced weight of each sensor. Through this processing method, when the spatial positioning reliability is high, the features of sensors that are closer together have a higher contribution to the distance weight matrix; when the spatial positioning reliability is low, the system reduces the influence of a single spatial positioning result on feature weighting, avoiding the multimodal feature set from being biased towards the wrong spatial region due to positioning errors.

[0089] After forming the corrected distance weight matrix, it is applied to the feature matrix composed of the amplitude characteristics, time-domain statistical characteristics, and frequency-domain power spectral density distribution of each sensor signal to obtain the corrected distance-weighted feature matrix. Principal component analysis is then performed to reduce the dimensionality of this matrix, resulting in a multimodal feature set. Thus, the spatial positioning coordinates are not only used for subsequent consistency verification of equipment structural parameters, but also participate in adjusting the feature contribution of each sensor during the feature construction stage.

[0090] Furthermore, when constructing the intermodal coupling relationship diagram, spatial positioning reliability can also be considered. The edge weighting selection strategy between sensor nodes is adjusted. Specifically, when the spatial positioning reliability... When the reliability exceeds a preset reliability threshold, strong correlation edges between key sensor nodes adjacent to the spatial positioning coordinates of the discharge source are retained, and the screening threshold for edge weights of sensor nodes far from the spatial positioning coordinates of the discharge source is appropriately increased; when the spatial positioning reliability... When the reliability threshold is less than the preset threshold, the influence of spatial distance factors on edge weight selection is reduced, so that the intermodal coupling relationship graph is determined more based on the mutual information between localized feature matrices, the nonlinear correlation strength, and the random walk node centrality index.

[0091] After completing the graph neural network encoding and obtaining the discharge source feature representation vector, the discharge source feature representation vector is semantically matched with a preset discharge mode knowledge graph to obtain a set of candidate discharge types. Then, variational inference and Bayesian uncertainty quantification are used to obtain the type confidence level corresponding to each candidate discharge type. To avoid relying solely on signal features and probability ranking to determine the discharge type, this embodiment further calculates a structural adaptation score for the candidate discharge types based on the device structural parameter database and the discharge mode knowledge graph.

[0092] Specifically, for each candidate discharge type, the corresponding potential equipment structure region and type structure constraint parameters are obtained from the discharge mode knowledge graph. The potential equipment structure region is used to characterize the equipment space region where this discharge type usually occurs, such as the internal region of oil-paper insulation, the surface region of insulator, the gas gap region, the conductor tip region, or the bushing connection region; the type structure constraint parameters are used to characterize the typical structural conditions corresponding to this discharge type, such as insulation gap, dielectric type, conductor spacing, bushing outer diameter, insulation layer thickness, or gas pressure range.

[0093] The equipment structure region where the current discharge source is located is determined based on the spatial positioning coordinates, and the positional deviation between the equipment structure region where the current discharge source is located and the equipment structure region where the candidate discharge type can occur is calculated. Simultaneously, based on spatial positioning coordinates, the geometric parameters of the equipment components corresponding to the current discharge source are calculated, and the difference between these geometric parameters and the type structural constraint parameters corresponding to the candidate discharge type is calculated to obtain the structural difference. Type structure fit score corresponding to candidate discharge types It can be represented as:

[0094] in, This indicates the position deviation from the scale parameter. Parameters representing the scale of structural differences. Indicates spatial positioning reliability. Type structure adaptation score. The higher the value, the more consistent the candidate discharge type is with the spatial location, equipment structure area, and structural parameters of the current discharge source.

[0095] After obtaining the type structure fit score, the type confidence obtained from Bayesian uncertainty quantification is corrected based on the type structure fit score. Let the type confidence corresponding to candidate discharge type k be... Corrected type confidence It can be represented as:

[0096] in, Indicates candidate discharge type Corrected type confidence; Indicates candidate discharge type The confidence level of the original type obtained after Bayesian uncertainty quantization; Indicates candidate discharge type Corresponding type structure adaptation score; This represents the total number of candidate discharge types.

[0097] The candidate discharge types are reordered based on the corrected type confidence, and the discharge type with the highest corrected type confidence is determined as the optimal discharge type for verifying the consistency of equipment structural parameters.

[0098] Through the above processing, the optimal discharge type is not simply determined by the matching result between the discharge source feature representation vector and the feature template, but is simultaneously verified by spatial positioning reliability, equipment structure region, and type structure constraint parameters. For example, in the process of partial discharge identification of a transformer, the discharge source feature representation vector has high similarity with both internal void discharge and surface discharge, and their confidence intervals overlap to some extent in the Bayesian posterior probability distribution. In this case, if the spatial positioning coordinates are located in the internal region of the oil-paper insulation, and its corresponding geometric parameters match the type structure constraint parameters of the internal void discharge better, then the type structure fit score of the internal void discharge is higher, and the type confidence increases after correction. Conversely, although surface discharge has high signal feature similarity, its type confidence decreases after correction because the equipment structure region where it may occur is inconsistent with the current spatial positioning coordinates. Thus, the risk of misjudgment between highly confusing discharge types can be reduced.

[0099] After verifying the consistency of equipment structural parameters, this embodiment further modifies the severity rating process. Specifically, the current discharge data is only included in the historical discharge data similarity matching process when the consistency verification of the equipment structural parameters corresponding to the optimal discharge type passes. When screening historical discharge data, not only is it required that the historical discharge data be the same as the current optimal discharge type, but the equipment structural region corresponding to the historical discharge data also matches the equipment structural region corresponding to the current spatial positioning coordinates. This screening method avoids introducing historical samples of the same discharge type but with significant differences in structural regions into the severity rating process.

[0100] For the filtered historical discharge data, historical feature vectors of the same type and matching structural regions are extracted, and the time-warped distance between the current discharge feature vector and the historical feature vector is calculated. The historical similarity score is obtained after normalizing the time-warped distance. Furthermore, historical similarity scoring... Structural consistency score And the prediction uncertainty obtained during the Bayesian uncertainty quantification process Together they are used to calculate the overall severity score. :

[0101] Where α, β, and γ represent the weight coefficients of the historical similarity score, structural consistency score, and prediction uncertainty correction term, respectively, and α + β + γ = 1. Prediction uncertainty A higher score indicates greater confusion among candidate discharge types and a lower certainty contribution to the overall severity score; structural consistency score The higher the value, the more consistent the current discharge type is with the spatial structure constraints, and the higher the credibility of the severity rating result.

[0102] After obtaining the overall severity score Subsequently, adaptive fuzzy clustering analysis was performed based on the comprehensive severity score to obtain the severity classification boundary. Compared with severity classification based solely on similarity scores, the comprehensive severity score in this embodiment considers historical similarity, spatial structure consistency, and identification uncertainty simultaneously, making the severity rating results more applicable to substation field scenarios with complex electromagnetic interference, similar discharge types, and uneven distribution of historical samples.

[0103] Finally, an identification report is generated by combining the optimal discharge type, spatial location coordinates, type-structure adaptation score, structural consistency score, historical similarity score, prediction uncertainty, and severity rating. The identification report can further record: changes in confidence level before and after discharge type correction, spatial location reliability, structural consistency verification results, the number of historical samples participating in similarity matching, and the basis for severity rating. In this way, maintenance personnel can not only obtain the partial discharge type and severity results, but also understand the spatial location basis, structural verification basis, and historical data matching basis for the identification conclusions, thereby improving the interpretability of partial discharge diagnosis results and the reliability of maintenance decisions.

[0104] like Figure 3 As shown, in a second aspect of the present invention, a partial discharge intelligent identification system for primary equipment in a substation is provided, comprising: The signal feature extraction unit is used to synchronously acquire the discharge electromagnetic signals of the primary equipment of the substation through the sensor array, synchronously calibrate the discharge electromagnetic signals and extract the pulse arrival time difference and signal amplitude features, construct spatial positioning coordinates based on the pulse arrival time difference and generate a multimodal feature set in combination with the signal amplitude features; The feature representation encoding unit is used to decompose the multimodal feature set through multi-scale time-frequency transformation to obtain localized feature matrices of different frequency bands, perform correlation measurement on the localized feature matrices to obtain an intermodal coupling relationship graph, and perform graph neural network encoding operation on the intermodal coupling relationship graph to obtain a discharge source feature representation vector; The discharge type identification unit is used to semantically match the discharge source feature representation vector with a preset discharge mode knowledge graph to obtain a candidate discharge type set, calculate the posterior probability distribution of the discharge type corresponding to the discharge source feature representation vector based on the variational inference algorithm, and estimate the confidence interval of the posterior probability distribution by combining Bayesian uncertainty quantization to obtain the probability ranking result and determine the optimal discharge type. The severity assessment unit is used to verify the consistency of the device structural parameters based on the optimal discharge type and the spatial positioning coordinates to obtain the verification result, perform similarity matching operation between the verified discharge data and the pre-acquired historical discharge data to obtain a similarity score, calculate the severity rating of partial discharge based on the similarity score, and output an identification report.

[0105] A third aspect of the present invention provides an electronic device, comprising: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0106] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0107] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0108] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0109] 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 method for intelligent identification of partial discharge in primary equipment of a substation, characterized in that, include: The discharge electromagnetic signals of the primary equipment in the substation are synchronously acquired by a sensor array. The discharge electromagnetic signals are synchronously calibrated and the pulse arrival time difference and signal amplitude features are extracted. Spatial positioning coordinates are constructed based on the pulse arrival time difference and multimodal feature set is generated by combining the signal amplitude features. The multimodal feature set is decomposed by multi-scale time-frequency transformation to obtain localized feature matrices for different frequency bands. The localized feature matrices are correlated to obtain an intermodal coupling relationship graph. The intermodal coupling relationship graph is encoded by a graph neural network to obtain a discharge source feature representation vector. The discharge source feature representation vector is semantically matched with a preset discharge mode knowledge graph to obtain a set of candidate discharge types. The posterior probability distribution of the discharge type corresponding to the discharge source feature representation vector is calculated based on the variational inference algorithm. The confidence interval of the posterior probability distribution is estimated by combining Bayesian uncertainty quantization to obtain the probability ranking result and determine the optimal discharge type. Based on the optimal discharge type and the spatial positioning coordinates, the consistency of the equipment structural parameters is verified to obtain the verification result. The verified discharge data is then matched with the pre-acquired historical discharge data to obtain a similarity score. Based on the similarity score, the severity rating of the partial discharge is calculated and an identification report is output.

2. The method according to claim 1, characterized in that, The discharge electromagnetic signals of primary equipment in a substation are synchronously acquired using a sensor array. These signals are then synchronously calibrated, and pulse arrival time difference (TAD) and signal amplitude features are extracted. Spatial positioning coordinates are constructed based on the TAD, and a multimodal feature set is generated by combining the signal amplitude features. The spatial position information and clock parameters of each sensor are obtained from the sensor array. Based on the clock parameters, the discharge electromagnetic signals collected by each sensor are time-aligned to obtain a synchronous electromagnetic signal. The synchronous electromagnetic signal is pulse-detected to identify the pulse start time and the pulse arrival time difference between adjacent sensors is calculated. Based on the pulse arrival time difference and the spatial position information, triangulation is performed to obtain the spatial positioning coordinates. The peak amplitude and effective amplitude of each frequency band are extracted from the synchronous electromagnetic signal to form the signal amplitude feature. The statistical features of the signal amplitude feature in the time domain are calculated and the power spectral density distribution in the frequency domain is extracted. The statistical features and the power spectral density distribution are vector-concatenated to obtain the single-mode feature vector. The single-mode feature vectors corresponding to different sensors are normalized and arranged according to the sensor number to form a feature matrix. The spatial distance between the power source and each sensor is calculated based on the spatial positioning coordinates, and a distance weight matrix is ​​constructed. The distance weight matrix and the feature matrix are used to solve for the distance weight matrix to obtain the distance weighted feature matrix. Principal component analysis is performed on the distance weighted feature matrix to reduce its dimensionality and retain the principal components to obtain the multimodal feature set.

3. The method according to claim 1, characterized in that, The multimodal feature set is decomposed by multi-scale time-frequency transformation to obtain localized feature matrices for different frequency bands. Correlation measurement is then performed on these localized feature matrices to obtain an intermodal coupling relationship graph, including: The multimodal feature set is grouped according to sensor modes to obtain feature subsets corresponding to each sensor. For each feature subset, the corresponding optimal wavelet basis is determined and wavelet transform is performed to obtain multi-scale time-frequency coefficients. The multi-scale time-frequency coefficients are subjected to singular value decomposition and the principal singular vectors are extracted to form frequency band feature vectors. The frequency band feature vectors at different scales are arranged in ascending order of frequency to obtain the localized feature matrix. The mutual information between the localized feature matrices corresponding to different sensors is calculated. Based on the mutual information, the nonlinear correlation strength is calculated to obtain the information correlation matrix. The information correlation matrix is ​​divided into modal clusters by spectral clustering algorithm and the inter-cluster connection relationship is identified to obtain the sparse correlation matrix. The non-zero elements in the sparse correlation matrix are used as edge weights to construct the sensor connection graph to obtain the initial coupling relationship graph. For each sensor node in the initial coupling graph, the node access frequency is calculated using a random walk sampling strategy to obtain a node centrality index. Based on the node centrality index, key sensor nodes are identified and the weights of the connection edges of the key sensor nodes are adjusted. The initial coupling graph after weight adjustment is then normalized using a Laplace matrix to obtain the intermodal coupling graph.

4. The method according to claim 1, characterized in that, The discharge source feature representation vector obtained by performing graph neural network encoding operation on the intermodal coupling relationship graph includes: The connectivity distribution and edge weight distribution of each sensor node are extracted from the intermodal coupling relationship graph to obtain graph topology statistics. Based on the graph topology statistics, the aggregation range of the node neighborhood is calculated and the multi-order neighborhood boundary is determined. The localized feature matrix corresponding to each sensor node is used as the initial feature of the node and combined with the multi-order neighborhood boundary to construct the node feature propagation matrix. The node feature propagation matrix is ​​subjected to tensor decomposition and low-rank approximation reconstruction to obtain the node embedding feature. The node embedding features are arranged into a graph feature matrix according to the sensor number and multi-layer convolution operation is performed with the adjacency matrix of the intermodal coupling relationship graph. In each layer convolution operation, the spectral correlation of the neighborhood features is calculated based on the spectral domain frequency decomposition and the node representation is updated by combining the residual connection. The updated node representation is batch normalized and multi-layer convolution output is accumulated to obtain a multi-scale node representation. The global graph representation vector is calculated based on the multi-scale node representation. The global graph representation vector is adaptively sampled and aggregated, and the node importance weights are calculated by combining the multi-scale node representations. The multi-scale node representations are then weighted and fused based on the node importance weights to obtain fused node features. The fused node features are then mapped and transformed through a fully connected layer and combined with linear activation to obtain the discharge source feature representation vector.

5. The method according to claim 1, characterized in that, The discharge source feature representation vector is semantically matched with a preset discharge mode knowledge graph to obtain a candidate discharge type set. The posterior probability distribution of the discharge type corresponding to the discharge source feature representation vector is calculated based on a variational inference algorithm, including: Feature template vectors and hierarchical relationships between discharge types are extracted from a pre-defined discharge mode knowledge graph. A discharge type feature library is constructed based on the feature template vectors, and the Riemann manifold distance between the discharge source feature representation vector and each feature template vector in the discharge type feature library is calculated to obtain a manifold similarity matrix. Based on the manifold similarity matrix, high-similarity discharge types are determined, and related parent types are expanded by combining the hierarchical relationships between types to obtain the candidate discharge type set. Gaussian mixture distribution parameter estimation is performed on the discharge source feature representation vector to calculate the feature space probability density distribution. Based on the probability density distribution, multiple sets of feature variables are sampled and generated, and the similarity with the feature template vectors of each type in the candidate discharge type set is calculated to obtain the type similarity sequence. The occurrence frequency of each discharge type in the type similarity sequence is counted and normalized to obtain the type frequency distribution. Based on each discharge type in the candidate discharge type set, a multinomial distribution prior is constructed. The likelihood probability is obtained by performing expectation maximization iterative calculation on the type frequency distribution and the multinomial distribution prior. The likelihood probability is then randomly sampled to obtain a posterior sample set, and the posterior probability distribution of the discharge type corresponding to the discharge source feature representation vector is calculated.

6. The method according to claim 1, characterized in that, By combining Bayesian uncertainty quantification to estimate the confidence interval of the posterior probability distribution, the probability ranking results are obtained, and the optimal discharge type is determined, including: The probability distribution statistics of each discharge type are extracted from the posterior probability distribution and the confidence ellipse boundary is calculated to obtain a set of confidence intervals. The intersection overlap rate of each discharge type in the set of confidence intervals is calculated based on the interval volume. High confusion type groups are marked based on the intersection overlap rate and the prediction uncertainty is quantified. Monte Carlo uncertainty propagation is performed on the probability distribution statistics based on the prediction uncertainty to obtain a corrected probability distribution. A probability matrix is ​​constructed based on the modified probability distribution, and the conditional entropy is calculated. The inter-type discrimination is evaluated based on the conditional entropy, and an adaptive weight coefficient is set. The modified probability distribution is subjected to a nonlinear transformation by the adaptive weight coefficient, and the type confidence is calculated in combination with the corresponding interval volume. The discharge types are sorted based on the type confidence to obtain the probability ranking result. The discharge type with the highest confidence level is selected from the probability ranking results as the initial optimal type. The statistical significance difference between the initial optimal type and the discharge type with the second highest confidence level is calculated. It is determined whether the statistical significance difference is greater than a preset discrimination threshold. If so, the initial optimal type is the optimal discharge type. Otherwise, the type with the largest information gain is selected from the probability ranking results according to the information gain maximization criterion, and the optimal discharge type is determined by combining the information gain maximization criterion with Nash equilibrium.

7. The method according to claim 1, characterized in that, Based on the optimal discharge type and the spatial positioning coordinates, the consistency of the device structural parameters is verified to obtain the verification result. The verified discharge data is then matched with pre-acquired historical discharge data to obtain a similarity score. Based on the similarity score, the severity rating of the partial discharge is calculated, and an identification report is output, including: The standard structural parameters corresponding to the optimal discharge type are obtained from the preset equipment structural parameter database. The geometric parameters of the equipment structure are calculated based on the spatial positioning coordinates and the difference is calculated with the standard structural parameters. When the difference is less than the preset consistency threshold, the verification is marked as passed and the verification result is obtained. Extract feature parameters from the verified discharge data to construct the current discharge feature vector, extract historical feature vectors of the same type of discharge from the historical discharge data, calculate the time-normalized distance between the current discharge feature vector and the historical feature vector and normalize it to obtain the similarity score; Adaptive fuzzy clustering analysis is performed based on the similarity score to obtain the severity classification boundary. The similarity score is then mapped to a severity rating based on the severity classification boundary. Finally, an identification report is generated by combining the optimal discharge type and the spatial positioning coordinates.

8. A partial discharge intelligent identification system for primary equipment in a substation, used to implement the method described in any one of claims 1-7, characterized in that, include: The signal feature extraction unit is used to synchronously acquire the discharge electromagnetic signals of the primary equipment of the substation through the sensor array, synchronously calibrate the discharge electromagnetic signals and extract the pulse arrival time difference and signal amplitude features, construct spatial positioning coordinates based on the pulse arrival time difference and generate a multimodal feature set in combination with the signal amplitude features; The feature representation encoding unit is used to decompose the multimodal feature set through multi-scale time-frequency transformation to obtain localized feature matrices of different frequency bands, perform correlation measurement on the localized feature matrices to obtain an intermodal coupling relationship graph, and perform graph neural network encoding operation on the intermodal coupling relationship graph to obtain a discharge source feature representation vector; The discharge type identification unit is used to semantically match the discharge source feature representation vector with a preset discharge mode knowledge graph to obtain a candidate discharge type set, calculate the posterior probability distribution of the discharge type corresponding to the discharge source feature representation vector based on the variational inference algorithm, and estimate the confidence interval of the posterior probability distribution by combining Bayesian uncertainty quantization to obtain the probability ranking result and determine the optimal discharge type. The severity assessment unit is used to verify the consistency of the device structural parameters based on the optimal discharge type and the spatial positioning coordinates to obtain the verification result, perform similarity matching operation between the verified discharge data and the pre-acquired historical discharge data to obtain a similarity score, calculate the severity rating of partial discharge based on the similarity score, and output an identification report.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images