Partial discharge classification and identification method

Abstract

The invention relates to the technical field of electrical equipment detection, in particular to a partial discharge classification and identification method, which comprises the following steps: acquiring a partial discharge signal, and performing graph structure feature extraction on the partial discharge signal to obtain a node feature of the partial discharge signal; determining a time dynamic threshold interval based on the KL divergence between the positive sample and each candidate negative sample, wherein the time dynamic threshold interval is used for carrying out negative sample sampling; according to the positive sample and the negative sample, constructing and training a graph encoder for acquiring a high-dimensional feature vector corresponding to the partial discharge signal; the loss function of the graph encoder comprises a bidirectional weighted likelihood loss function and a feature comparison loss function; and carrying out partial discharge type classification on the partial discharge signal based on BK-Means according to the high-dimensional feature vector. Through graph structure feature extraction and high-dimensional feature vector construction, dynamic threshold negative sample sampling and a ResGCN-based graph encoder are combined, so that the classification accuracy and calculation efficiency of partial discharge signals are improved.

Description

Technical Field

[0001] The present application relates to the technical field of electrical equipment detection, and more specifically, to a partial discharge classification and identification method. Background Art

[0002] Accurate identification of partial discharge (PD) is an important guarantee for the safe operation of high-voltage equipment and is of great significance to the stability and reliability of power systems. In recent years, the use of supervised learning to identify partial discharge has made significant progress. By building a model and training it with sample data with known labels, this method enables the model to learn and identify the characteristics of different types of partial discharges. However, in the process of modern electrical equipment monitoring, the amount of data generated is huge. To determine the label of each set of data, a detailed analysis is required, which is time-consuming and increases high labor costs. Given these limitations, the unsupervised learning paradigm provides a viable alternative, offering the potential to bypass the laborious data labeling process.

[0003] In the study of partial discharge, pulse signals have significant time series characteristics, and the temporal relationship between pulses is the key to studying discharge characteristics. Traditional clustering algorithms, such as k-means and DBSCAN, usually rely only on the feature information of data points and ignore the structural relationship between samples, so the clustering effect is not good when processing complex data. In recent years, deep graph clustering (Graph Neural Networks, GNNs) has made significant progress in many fields. GNN is able to process graph structured data, where nodes represent data points and edges represent the relationship between data points. Thanks to its powerful graph information processing capabilities, graph convolutional networks perform well in graph clustering tasks such as social networks and recommendation systems. By combining node features and graph structural information, GNN can achieve more accurate node representation learning and clustering.

[0004] Recently, deep graph node clustering is mainly divided into two categories: generative methods and contrastive methods. Generative methods usually use autoencoders to reconstruct the structure or node features of the graph, but constrain the latent representation through reconstruction loss, which easily leads to fluctuations and instability during training. In contrast, the goal of contrastive methods is to learn representations that can distinguish between positive and negative samples, where positive samples are pairs of nodes belonging to the same cluster and negative samples are from different clusters. By pulling positive samples closer and negative samples farther away, the representation ability of the model is improved.

[0005] In contrastive learning, data augmentation techniques (such as mask features, node inactivation, edge perturbations, subgraph partitioning, etc.) have been shown to significantly improve the effect of representation learning. However, for time series spike data, the construction of negative samples is particularly important. These methods usually only focus on the static features of the data and cannot effectively capture the temporal relationship. Therefore, in the classification of time series spike data, the sampling and construction of negative samples requires more targeted strategies to improve the model's sensitivity to temporal dependencies and classification accuracy. Summary of the invention

[0006] In view of this, the present application provides a local discharge classification and recognition method to solve the problem that when the prior art classifies time series pulse data based on contrastive learning, the sampling and construction of negative samples only focus on the static characteristics of the data and cannot effectively capture the timing relationship, resulting in the model's sensitivity to timing dependence and low classification accuracy.

[0007] The technical solutions provided by this application are as follows:

[0008] A partial discharge classification and identification method, comprising:

[0009] Acquire a partial discharge signal, and extract graph structure features of the partial discharge signal to obtain node features of the partial discharge signal;

[0010] Taking the node features as positive samples, and extracting a negative sample candidate set from the local discharge signal;

[0011] Determine a temporal dynamic threshold interval based on the KL divergence between the positive sample and each candidate negative sample in the negative sample candidate set, for performing negative sample sampling; the KL divergence between the sampled negative sample and the positive sample is within the temporal dynamic threshold interval;

[0012] According to the positive samples and negative samples, two ResGCN-based graph encoders are constructed and trained to obtain high-dimensional feature vectors corresponding to partial discharge signals; the loss function of the graph encoder includes a bidirectional weighted likelihood loss function and a feature contrast loss function;

[0013] According to the high-dimensional feature vector, the partial discharge signal is classified into partial discharge types based on BK-Means.

[0014] In one possible implementation, extracting graph structure features of the partial discharge signal to obtain node features corresponding to the partial discharge signal includes:

[0015] According to the partial discharge signal, construct an adjacency matrix for characterizing a graph structure, and determine a degree matrix D of the adjacency matrix A;

[0016] Based on the degree matrix D, the adjacency matrix A is standardized to obtain a standardized adjacency matrix It is expressed as: Where I is the identity matrix;

[0017] Based on the standardized adjacency matrix Enhanced graph structure representation to obtain improved adjacency matrix It is expressed as: Among them, α is used to represent the transition probability from one node to another.

[0018] In one possible implementation, the KL divergence between the positive sample and each candidate negative sample in the negative sample candidate set is expressed as:

[0019]

[0020] Among them, x represents the positive sample, represents a candidate negative sample in the negative sample candidate set.

[0021] In one possible implementation, determining a temporal dynamic threshold interval based on the KL divergence between the positive sample and each candidate negative sample in the negative sample candidate set, and performing negative sample sampling includes:

[0022] Determine the minimum KL divergence value KL between the positive sample and each of the candidate negative samples min and the maximum KL divergence value KL max ;

[0023] According to the KL min and the KL max , determine the dynamic threshold increment θ, expressed as:

[0024] θ=φ(KL max -KL min ), where φ is the adjustment factor;

[0025] According to the KL min , the KL max and the dynamic threshold increment θ, determine the time dynamic threshold interval for negative sample sampling, and the negative sample obtained by sampling The KL divergence between the positive sample and the time dynamic threshold is within the time dynamic threshold interval, expressed as:

[0026]

[0027] in, is a candidate negative sample in the negative sample candidate set, w t is the time window, t is the time corresponding to the collection of the positive sample, t iis the time corresponding to the collection of the negative sample.

[0028] In one possible implementation, two ResGCN-based graph encoders are constructed and trained according to the positive samples and the negative samples, including:

[0029] The positive samples and the negative samples are encoded based on two independent graph convolutional networks GCNs, which are expressed as:

[0030] In the formula, is the graph convolution operation of the i-th encoder, is the normalized adjacency matrix, X is the node feature matrix composed of positive samples ×, is the parameter of the i-th GCN, δ is the PRelu activation function;

[0031] According to the graph convolution operation, the output z of the positive sample is obtained 1 、z 2 And the output of the negative sample

[0032] Among them, the output of the positive sample is expressed as:

[0033]

[0034] The output of the negative sample is expressed as:

[0035] In the formula, Negative samples The negative sample matrix is ​​formed. represents residual connection;

[0036] According to the output of the positive sample and the negative sample, the feature mean of all samples is calculated to obtain the global feature vector h i , expressed as:

[0037]

[0038] In the formula, j represents a single sample, σ is the activation function, and n is the number of samples. is the output of the graph encoder;

[0039] The similarity between features is calculated based on a bilinear interaction network, expressed as:

[0040]

[0041] Among them, F k represents the output of the bilinear interaction network, W k are the parameters of the bilinear interaction network, is the transposed matrix of the output of the graph encoder.

[0042] In one possible implementation, the bidirectional weighted likelihood loss function L c , expressed as:

[0043]

[0044] In the formula, N is the number of positive samples, M is the number of negative samples; is the feature of the positive sample, is the feature of the negative sample, y is the positive sample label, h 2 is the global vector obtained by the second graph encoder.

[0045] In one possible implementation, the feature contrast loss function is expressed as:

[0046]

[0047] Where index(j) is the negative example index sampled from a uniform distribution from 0 to μ. e is the current training round, E is the total training round, is a predetermined constant, R is the index range; N is the number of positive samples, M is the number of negative samples; S ii is the feature-level similarity matrix S i，j The diagonal elements of the feature-level similarity matrix S i，j Used to quantify the cosine similarity between feature i and feature j, S index(j) are the non-diagonal elements corresponding to negative samples; γ is used to control the sensitivity of similarity.

[0048] Compared with the prior art, the technical solution of this application has the following beneficial effects:

[0049] Based on the technical solution of this application, by extracting graph structure features and constructing high-dimensional feature vectors, the complex features of local discharge signals can be effectively captured, and the accuracy of classification can be improved. The negative sample sampling strategy within the dynamic threshold interval avoids the blindness of negative sample selection, so that the model can better learn the characteristic differences of local discharge signals during training. At the same time, the use of ResGCN-based graph encoder and BK-Means clustering algorithm can not only process large-scale data sets, but also achieve accurate classification of local discharge types while maintaining computational efficiency, with high practicality and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A flowchart of a partial discharge classification and identification method provided in Example 1 of the present application.

[0051] Figure 2 A framework diagram of graph comparison unsupervised clustering provided in Example 1 of the present application. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0053] Embodiment 1

[0054] See also Figure 1 , is a flow chart of a partial discharge classification and identification method provided in Example 1 of the present application. Figure 1 As shown in , the specific implementation steps of the above method include:

[0055] Step 101: Collect partial discharge signals at preset time intervals.

[0056] Optionally, the above-mentioned preset time interval can be set according to actual conditions, and the embodiment of the present application does not make any specific limitation.

[0057] Step 102: extracting graph structure features of the partial discharge signal to obtain node features corresponding to the partial discharge signal.

[0058] The graph structure is used to describe the complex relationship between data elements. It is a nonlinear structure composed of vertex sets and edge sets. Each signal point in the partial discharge signal is regarded as a node in the graph. A graph structure is represented as: G = {V, E, X}, where V and E are vertex sets and edge sets, that is, V = {v1, v2, v3...}, E = {v1v2, v2v3...}, v i represents a node. X is the corresponding node feature matrix.

[0059] According to the collected partial discharge signals, the adjacency matrix A is constructed to represent the topological structure of the graph G. The adjacency matrix A is expressed as That is, if two sets of partial discharge signals are adjacent, the edge between them is a i，j =1, otherwise a ij is 0.

[0060] Based on the above adjacency matrix A, determine the degree matrix D of the adjacency matrix A. The degree matrix D is a diagonal matrix, and the elements on the diagonal are the degrees of each node. i The degree represents the number of edges associated with the node.

[0061] Standardize A where I is the identity matrix.

[0062] Specifically, the input of each layer of GCN is the node feature matrix X and the adjacency matrix A. The inner product of these two is directly taken, and then multiplied by a parameter matrix W, and activated with an activation function to form a neural network layer.

[0063] However, because the diagonal of the adjacency matrix is ​​0, the inner product with the node feature matrix is ​​equivalent to the weighted sum of the adjacency matrix. The value of the node feature becomes the weight of the adjacency matrix, and its own features are ignored. To avoid this situation, add a unit matrix I to the adjacency matrix A so that the diagonal elements of the adjacency matrix become 1.

[0064] Furthermore, since the adjacency matrix A has not been normalized, it is difficult to limit the data to the required range. Through normalization, the data can be made comparable, while the relationship between the data is relatively maintained. Therefore, in order to avoid the inner product multiplication of the adjacency matrix and the feature matrix changing the original distribution of the features, this embodiment performs a normalization process on the adjacency matrix A, which is expressed as:

[0065] Based on the above standardized adjacency matrix, the above graph structure representation is enhanced to obtain an improved adjacency matrix, which is expressed as

[0066] Specifically, enhancing the graph structure is a method to improve the original graph structure by adding additional information or connections. This enhancement can be reflected in many aspects. Figure 2 As shown, this embodiment enhances the structure of the graph by adding new nodes or edges. Where α is the transition probability, which is used to describe the probability of going from one node to another. In the field of deep learning, graph convolutional networks (GCNs) can be used to learn node representations on graphs and calculate transition probabilities based on this. GCN updates the representation of nodes by aggregating information from neighboring nodes, thereby capturing the structural characteristics of the graph. Optionally, different methods for enhancing the graph structure can be set according to actual conditions, and the embodiments of this application do not make specific limitations.

[0067] Step 103: Take the node features as positive samples (x), and extract a negative sample candidate set (x1, x2, ..., x n ). Where n is a positive integer.

[0068] In this embodiment, the positive samples and negative samples are derived from the same data set, that is, a data set consisting of collected partial discharge signals. For example, assuming that there are 100 sets of data in the data set, a corresponding negative sample is found for each set of data in order. Specifically, the method for screening negative samples provided in this application is as follows:

[0069] Step 104: quantify the probability distribution between each candidate negative sample in the positive sample and negative sample candidate set using KL divergence.

[0070] In this embodiment, a more effective data representation is learned by minimizing the similarity between positive and negative samples and maximizing the similarity between samples of the same type. Specifically, for each candidate negative sample Calculate the KL divergence between it and the positive sample ×, expressed as:

[0071]

[0072] Step 105: According to each candidate negative sample The KL divergence between x and the positive sample x determines the time dynamic threshold interval and selects more suitable negative samples It is expressed as:

[0073]

[0074] Specifically, the embodiment of the present application is based on all candidate negative samples The KL divergence between the positive sample x and the positive sample x determines the range of a temporal dynamic threshold, which is determined by the minimum KL divergence KL min , maximum KL divergence KL max and the dynamic threshold increment θ. Where θ=φ(KL max -KL min ). φ is the adjustment factor, which is used to adjust the threshold range so that the selected negative samples More suitable. t is the time window, t is the time corresponding to the collection of positive samples, t i The time corresponding to the collection of negative samples. Set the time window w t It is used to ensure that there is a certain time interval between the selected negative samples and the positive samples, and avoid selecting samples that are too close in time to the positive samples, thereby improving the negative samples. diversity.

[0075] Through the above method, this embodiment can more effectively select negative samples that have a moderate difference in probability distribution from positive samples and a certain time interval. It not only takes into account the independence of time series, but also ensures the diversity of negative samples. The negative samples obtained by screening are used to train the model, so as to improve the model's ability to recognize the difference between positive and negative samples.

[0076] Step 106: Based on the above positive samples and the corresponding negative samples, two ResGCN-based graph encoders are constructed and trained.

[0077] In this embodiment, a graph convolutional network with residual connections (ResGCN) is used as a graph encoder, and the positive samples and the corresponding negative samples obtained in steps 102 to 105 are used together in the training process of the graph encoder, so that the signal features of each node can be mapped to a high-dimensional space for encoding. Specifically, Figure 2 As shown in , two independent graph convolutional networks (GCNs) are used to encode positive and negative samples. Each GCN encodes the adjacency matrix And node feature x as input, perform the following transformation to extract features:

[0078]

[0079] in, is the graph convolution operation of the i-th encoder, is the normalized adjacency matrix, X is the node feature matrix composed of positive samples x, is the parameter of the i-th GCN, and δ is the PRelu activation function.

[0080] After being processed by the graph encoder, the output z of the positive sample i The feature representation is enhanced by adding the node feature matrix X, which is expressed as:

[0081]

[0082] In the formula, +X represents the residual connection. Based on the above residual connection, the original feature information can be retained while allowing the model to learn more complex feature interactions.

[0083] The negative sample output is and It is expressed as:

[0084] In the formula, Negative samples The negative sample matrix is ​​formed. represents a residual connection.

[0085] Calculate the feature mean of all samples to obtain the global feature vector hi, whose calculation formula is expressed as:

[0086]

[0087] Among them, j represents a single sample, σ is the activation function, and n is the number of samples. is the output of the graph encoder.

[0088] A bilinear interaction network is used to calculate the similarity between features, expressed as:

[0089]

[0090] Among them, F k represents the output of the bilinear interaction network, W k are the parameters of the bilinear interaction network, is the transposed matrix of the output of the graph encoder.

[0091] The method provided in the embodiment of the present application learns feature representation through graph structure information and node features, and allows the model to discover the intrinsic structure and pattern of data without labels. Through the above steps, the model can map the original features to a high-dimensional space, where the similarity of the features is jointly determined by the graph structure and node features, which is helpful for subsequent tasks such as clustering, classification or dimensionality reduction.

[0092] Step 107: Set a loss function to optimize the two graph encoders to complete the training of the two graph encoders. The loss function includes a bidirectional weighted likelihood loss function L c And feature contrast loss function L f .

[0093] Specifically, define the loss function To optimize the classification performance of the model, the loss function is trained by maximizing the likelihood of positive examples and minimizing the likelihood of negative samples, expressed as:

[0094]

[0095] Among them, N and M are the number of positive samples and negative samples respectively. The loss functions in the two directions are weighted averaged to obtain the final bidirectional weighted likelihood loss function L c , expressed as:

[0096]

[0097] This embodiment uses the joint loss function in the model training process, comprehensively considers the losses in two directions, and optimizes the performance of the model in the classification task by balancing the maximization likelihood of positive samples and the minimization likelihood of negative samples, thereby improving its ability to distinguish between positive and negative samples.

[0098] Furthermore, in order to enhance the independence of features in the latent space, this embodiment provides a feature contrast loss function L f Specifically, define the feature matrix Where D is the feature dimension and N is the number of positive samples. Its transposed matrix

[0099] By calculating the transpose of the feature matrix Z The dot product with itself is then normalized to obtain the feature-level similarity matrix S i，j, which is used to quantify the similarity between each feature and is expressed as:

[0100]

[0101] Among them, S i，j Represents the cosine similarity between feature i and feature j, matrix S i，j The value range of is [-1, 1].

[0102] Feature contrast loss function L f , by transforming the matrix S i，j The diagonal elements of are treated as positive samples, and the off-diagonal elements are treated as negative samples. Feature contrast loss function L f The expression is as follows:

[0103]

[0104] Among them, index(j) is the negative example index sampled from a uniform distribution from 0 to μ, e is the current training round, E is the total training round, is a predetermined constant (set to 0.5), R is the index range. N is the number of positive samples, and M is the number of negative samples. S ii is the diagonal element of the feature-level similarity matrix S, S index(j) are the off-diagonal elements corresponding to negative examples. γ controls the sensitivity of the similarity.

[0105] The final loss function is:

[0106] Loss = L c +L f

[0107] Step 108: Based on the two trained graph encoders, a high-dimensional feature vector corresponding to the partial discharge signal is output.

[0108] Step 109: input the high-dimensional feature vector corresponding to the partial discharge signal into the skeleton initialization K-Means clustering (Baackbone initializtion K-means; hereinafter referred to as BK-means) to perform partial discharge (Partial Discharge, hereinafter referred to as PD) type classification.

[0109] This embodiment applies BK-Means to cluster PD signals based on the obtained high-dimensional feature vector. BK-means clustering uses the K-means algorithm to obtain a set of local optimal solutions (clustering results), and obtains skeleton clusters by intersecting the local optimal solutions. The skeleton clusters are used to construct the skeleton initial solution and the new search space, and finally the skeleton initial solution is used to guide the K-means algorithm to search for clustering results in the new search space.

[0110] Compared with the prior art, the technical solution provided in the first embodiment of the present application has the following beneficial effects:

[0111] The technical solution provided in the embodiment of the present application can effectively capture the complex features of local discharge signals and improve the accuracy of classification by extracting graph structure features and constructing high-dimensional feature vectors. The negative sample sampling strategy within the dynamic threshold interval avoids the blindness of negative sample selection, so that the model can better learn the characteristic differences of local discharge signals during training. At the same time, the use of the ResGCN-based graph encoder and BK-Means clustering algorithm can not only process large-scale data sets, but also achieve accurate classification of local discharge types while maintaining computational efficiency, with high practicality and reliability.

[0112] Embodiment 2

[0113] In Example 2 of this application, 11 representative clustering techniques are compared and verified. The experimental results are shown in Table 1. BK-means has the ability to obtain highly cohesive and highly separated clustering results. Based on the new graph data clustering method provided in this embodiment, the clustering accuracy is significantly improved, and it exceeds the existing traditional and deep learning methods in multiple performance indicators.

[0114] See Table 1 for a comparison of the performance of different models.

[0115] Method ACC NMI ARI F1 K-Means 63.89 63.17 56.71 55.94 Spectral Clustering 54.83 39.6 34.38 50.82 Ward Hierarchical 62.81 62.69 55.36 54.74 BK-Means 77.28 66.23 69.93 66.08 GAE 68.74 63.06 61.08 65.95 DFCN 74.65 69.61 66.35 63.18 GCSEE 76.27 53.49 54.62 60.85 DCRN 83.16 79.50 73.76 68.49 DCLN 88.34 81.51 84.35 81.27 EGAE 92.36 85.75 87.62 89.16 SGCC 95.54 85.92 89.81 94.61

[0116] As shown in Table 1, the method provided in the embodiment of the present application shows a significant improvement in the four indicators of accuracy (ACC), normalized mutual information (NMI), adjusted Rand index (ARI) and F1 value. Compared with the closest existing method EGAE, this method improves ACC by 3.18%, NMI by 0.17%, ARI by 2.19%, and F1 value by 5.45%. This improvement is attributed to the innovative introduction of the positive and negative sample contrast learning mechanism in this application, which effectively emphasizes the differences between classes, thereby enhancing the clustering ability during the encoding process.

[0117] In comparison with deep graph clustering methods (such as GAE, DFCN, GCSEE, etc.), the SGCC method provided in the embodiment of the present application also showed obvious performance improvement, especially under the guidance of contrastive learning, achieving a higher clustering effect.

[0118] The technical effect of the embodiment of the present application has been verified in the clustering task of partial discharge signals, proving that the method can effectively improve the clustering ability of graph data, especially in the processing of unlabeled signal data, and can significantly improve the accuracy and robustness of clustering.

[0119] Although the embodiments of the present application have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present application, and that the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A partial discharge classification and identification method, characterized in that: include: Acquire a partial discharge signal, and extract graph structure features of the partial discharge signal to obtain node features of the partial discharge signal; Taking the node features as positive samples, and extracting a negative sample candidate set from the local discharge signal; Determine a temporal dynamic threshold interval based on the KL divergence between the positive sample and each candidate negative sample in the negative sample candidate set, for performing negative sample sampling; The KL divergence between the sampled negative sample and the positive sample is within the time dynamic threshold interval; According to the positive samples and negative samples, two ResGCN-based graph encoders are constructed and trained to obtain high-dimensional feature vectors corresponding to partial discharge signals; the loss function of the graph encoder includes a bidirectional weighted likelihood loss function and a feature contrast loss function; According to the high-dimensional feature vector, the partial discharge signal is classified into partial discharge types based on BK-Means.

2. A partial discharge classification and identification method according to claim 1, characterized in that: Extracting graph structure features of the partial discharge signal to obtain node features corresponding to the partial discharge signal includes: According to the partial discharge signal, construct an adjacency matrix for characterizing a graph structure, and determine a degree matrix D of the adjacency matrix A; Based on the degree matrix D, the adjacency matrix A is standardized to obtain a standardized adjacency matrix It is expressed as: Where I is the identity matrix; Based on the standardized adjacency matrix Enhanced graph structure representation to obtain improved adjacency matrix It is expressed as: Among them, α is the transition probability.

3. A partial discharge classification and identification method according to claim 1, characterized in that: The KL divergence between the positive sample and each candidate negative sample in the negative sample candidate set is expressed as: Among them, x represents the positive sample, represents a candidate negative sample in the negative sample candidate set.

4. A partial discharge classification and identification method according to claim 1, characterized in that: Determining a temporal dynamic threshold interval based on the KL divergence between the positive sample and each candidate negative sample in the negative sample candidate set, and performing negative sample sampling, including: Determine the minimum KL divergence value KL between the positive sample and each of the candidate negative samples min and the maximum KL divergence value KL max ; According to the KL min and the KL max , determine the dynamic threshold increment θ, expressed as: θ=φ(KL max -KL min ), where φ is the adjustment factor; According to the KL min , the KL max and the dynamic threshold increment θ, determine the time dynamic threshold interval for negative sample sampling, and the negative sample obtained by sampling The KL divergence between the positive sample and the time dynamic threshold is within the time dynamic threshold interval, expressed as: in, is a candidate negative sample in the negative sample candidate set, w t is the time window, t is the time corresponding to the collection of the positive sample, t i is the time corresponding to the collection of the negative sample.

5. A partial discharge classification and identification method according to claim 1, characterized in that: According to the positive samples and negative samples, two ResGCN-based graph encoders are constructed and trained, including: The positive samples and the negative samples are encoded based on two independent graph convolutional networks GCNs, which are expressed as: In the formula, is the graph convolution operation of the i-th encoder, is the adjacency matrix, X is the node feature matrix composed of positive samples x, is the parameter of the i-th GCN, δ is the PRelu activation function; According to the graph convolution operation, the output z of the positive sample is obtained 1 、z 2 And the output of the negative sample Among them, the output of the positive sample is expressed as: The output of the negative sample is expressed as: In the formula, Negative samples The negative sample matrix is ​​formed. represents residual connection; According to the output of the positive sample and the negative sample, the feature mean of all samples is calculated to obtain the global feature vector h i , expressed as: In the formula, j represents a single sample, σ is the activation function, and n is the number of samples. is the output of the graph encoder; The similarity between features is calculated based on a bilinear interaction network, expressed as: Among them, F k represents the output of the bilinear interaction network, W k are the parameters of the bilinear interaction network, is the transposed matrix of the output of the graph encoder.

6. A partial discharge classification and identification method according to claim 1, characterized in that: The bidirectional weighted likelihood loss function L c , expressed as: In the formula, N is the number of positive samples, M is the number of negative samples; is the feature of the positive sample, is the feature of the negative sample, y is the positive sample label, h 2 is the global vector obtained by the second graph encoder.

7. A partial discharge classification and identification method according to claim 1, characterized in that: The feature contrast loss function is expressed as: Where index(j) is the negative example index sampled from a uniform distribution from 0 to μ. e is the current training round, E is the total training round, is a predetermined constant, R is the index range; N is the number of positive samples, M is the number of negative samples; S ii is the feature-level similarity matrix S i，j The diagonal elements of the feature-level similarity matrix S i，j Used to quantify the cosine similarity between feature i and feature j, S index(j) are the non-diagonal elements corresponding to negative samples; γ is used to control the sensitivity of similarity.

Images