A knowledge-driven partial discharge pattern recognition method

Abstract

The present invention discloses a knowledge-driven partial discharge pattern recognition method, which includes the following steps: (1) constructing a convolutional neural network to extract features from partial discharge (PRPS) data of power equipment to obtain a data feature vector; (2) constructing a partial discharge knowledge graph of the power equipment and performing semantic modeling and structural modeling on the partial discharge knowledge graph; (3) constructing a graph convolutional neural network based on the semantically modeled and structurally modeled partial discharge knowledge graph to extract knowledge features of various partial discharge patterns of the power equipment to obtain a knowledge feature vector; (4) comparing the similarity between the data feature vector and the knowledge feature vector, and taking the partial discharge pattern with the greatest similarity as the recognition result.

Description

Technical Field

[0001] The present invention relates to a technology for diagnosing insulation defect faults of electric power equipment, and in particular to a method for identifying a partial discharge mode. Background Art

[0002] As we all know, in the existing energy transmission system, the use of power equipment with high reliability and excellent insulation performance is very important to ensure the safe operation of the system.

[0003] In current traditional power equipment, dielectric insulation degradation is one of the main causes of equipment failure. In severe cases, it may also lead to large-scale power outages, causing significant losses to people's property safety.

[0004] The inventors have found that different partial discharge modes have different effects on the reduction of the insulation level of power equipment, and the severity of faults caused by different partial discharge modes also varies greatly. Therefore, pattern recognition of partial discharge has become an important measure to reduce the risk of insulation defects and faults, which is crucial for the operation and maintenance of power equipment.

[0005] Currently, traditional partial discharge pattern recognition relies heavily on manual feature extraction, while data-driven algorithms are sensitive to the quantity and quality of data sets. Therefore, existing partial discharge pattern recognition is affected by interference signals in actual partial discharge data.

[0006] With the development of knowledge engineering, applying prior knowledge to multi-class object detection in insufficient and unbalanced datasets can significantly improve image recognition performance. Under the supervision of prior knowledge, knowledge-driven algorithms are easier to explain and less likely to violate common sense. Knowledge graphs are multi-relational graphs composed of entities (nodes) and relationships (edges). They represent prior knowledge in the same way that humans understand the world. Due to their unique characteristics, they are widely used in knowledge-driven algorithms and are very effective in representing structured data.

[0007] Based on the above background, in order to compensate for the negative impact of problems such as noise in partial discharge datasets and achieve accurate recognition of partial discharge patterns, the present invention designs and obtains a new knowledge-driven partial discharge pattern recognition method. Summary of the Invention

[0008] The purpose of the present invention is to provide a knowledge-driven partial discharge pattern recognition method. The partial discharge pattern recognition method can automatically extract feature information of partial discharge data using a convolutional neural network and construct a partial discharge knowledge graph of power equipment. The method uses a graph convolutional network to extract knowledge features of various partial discharge patterns in the partial discharge knowledge graph, and finally realizes partial discharge pattern recognition by comparing data features with knowledge features.

[0009] Based on the above objectives, the present invention proposes a knowledge-driven partial discharge pattern recognition method, which includes the following steps:

[0010] (1) Construct a convolutional neural network to extract features from the partial discharge (PRPS) data of power equipment to obtain data feature vectors;

[0011] (2) constructing a partial discharge knowledge graph of power equipment and performing semantic modeling and structural modeling on it; the partial discharge knowledge graph has a plurality of nodes, and the nodes include a plurality of partial discharge mode nodes and a plurality of feature nodes for describing the characteristics of the partial discharge mode; when performing semantic modeling, the Chinese characters of the semantics of each node in the partial discharge knowledge graph are converted into word vectors to form a word vector matrix, and then the word vector matrix corresponding to each node is converted into a node vector to form an embedding matrix; when performing structural modeling, the structural characteristics of the partial discharge knowledge graph are represented as an adjacency matrix A, and the adjacency matrix is ​​a symmetric matrix of size n×n, where n is the number of nodes in the partial discharge knowledge graph;

[0012] (3) Based on the partial discharge knowledge graph that has undergone semantic modeling and structural modeling, a graph convolutional neural network is constructed to extract the knowledge features of various partial discharge modes of power equipment to obtain knowledge feature vectors;

[0013] (4) Comparing the similarity between the data feature vector and the knowledge feature vector, and taking the partial discharge pattern with the greatest similarity as the recognition result.

[0014] In the above technical solution of the present invention, the inventors designed a new knowledge-driven partial discharge pattern recognition method. Compared with the existing technology, the partial discharge pattern recognition method designed by the present invention can effectively identify partial discharge patterns on partial discharge PRPS data with high noise and small data volume. It can compensate for the negative impact caused by problems such as noise in the partial discharge data set and realize accurate recognition of partial discharge patterns.

[0015] Furthermore, in the knowledge-driven partial discharge pattern recognition method described in the present invention, in step (1), the convolutional neural network has 3 convolution layers, 2 pooling layers and 1 stretching layer.

[0016] Furthermore, in the knowledge-driven partial discharge pattern recognition method described in the present invention, in step (1), each convolutional layer has multiple convolution kernels.

[0017] Furthermore, in the knowledge-driven partial discharge pattern recognition method described in the present invention, in step (1), the activation function of the convolutional neural network is a linear rectification function.

[0018] Furthermore, in the knowledge-driven partial discharge pattern recognition method described in the present invention, in step (2), when performing semantic modeling, the BERT Chinese character embedding representation model is used to convert the semantic Chinese characters of each node in the partial discharge knowledge graph into word vectors to form a word vector matrix; and the maximum pooling method is used to convert the word vector matrix corresponding to each node into a node vector.

[0019] Furthermore, in the knowledge-driven partial discharge pattern recognition method described in the present invention, in step (3), the graph convolutional neural network includes two graph convolution layers.

[0020] Furthermore, in the knowledge-driven partial discharge pattern recognition method of the present invention, in step (3), the graph convolution layer is represented by the following formula:

[0021] Y=σ(D -1 (A+I)XΘ)

[0022] Wherein, X represents the embedding matrix, I is the identity matrix, Θ is the trainable convolution kernel of the graph convolution layer, Y is the output matrix of the graph convolution layer, D is the degree matrix of the partial discharge knowledge graph, and A represents the adjacency matrix.

[0023] Furthermore, in the knowledge-driven partial discharge pattern recognition method of the present invention, in step (4), the similarity is calculated using Euler distance.

[0024] Furthermore, in the knowledge-driven partial discharge pattern recognition method of the present invention, the partial discharge patterns include: particle discharge, suspension discharge, tip discharge, surface discharge, and air gap discharge.

[0025] Compared with the prior art, the knowledge-driven partial discharge pattern recognition method of the present invention has the following advantages and beneficial effects:

[0026] In order to compensate for the negative impact of problems such as noise in partial discharge datasets and achieve accurate recognition of partial discharge patterns, the present invention designs and obtains a new knowledge-driven partial discharge pattern recognition method.

[0027] In this invention, the inventors utilize a convolutional neural network to extract partial discharge (PD) feature vectors from PD (Partial Discharge) (PRPS) data detected by power equipment, achieving automatic extraction of PD data feature vectors. Furthermore, the inventors construct a partial discharge (PD) knowledge graph for power equipment to store knowledge about various PD patterns. A graph convolutional network is then used to extract PD knowledge feature vectors for each PD pattern within this PD knowledge graph to obtain knowledge feature vectors. Finally, the PD data feature vectors are compared with the knowledge feature vectors for each PD pattern, and the PD pattern with the most similar knowledge feature vector to the data feature vector is identified as the final result.

[0028] Compared with the existing technology, the partial discharge pattern recognition method designed in the present invention can effectively identify partial discharge patterns on partial discharge PRPS data with high noise and small data volume. It can compensate for the negative impact caused by problems such as noise in the partial discharge data set and realize accurate recognition of partial discharge patterns. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 The present invention provides a flowchart of the steps of constructing a convolutional neural network and obtaining a data feature vector in one embodiment of the partial discharge pattern recognition method.

[0030] Figure 2 A schematic diagram of a partial discharge knowledge graph in one embodiment is schematically shown.

[0031] Figure 3 Schematic diagram showing the conversion of word vector matrix into node vectors.

[0032] Figure 4 This is a flowchart of the steps of obtaining knowledge feature vectors of various types of partial discharge patterns based on a constructed partial discharge knowledge graph in one embodiment of the partial discharge pattern recognition method of the present invention.

[0033] Figure 5 The figure schematically shows a comparison diagram of the similarity between a data feature vector and a knowledge feature vector in one embodiment of the partial discharge pattern recognition method of the present invention. DETAILED DESCRIPTION

[0034] The knowledge-driven partial discharge pattern recognition method of the present invention will be further explained and illustrated below in conjunction with the accompanying drawings and specific embodiments. However, such explanation and illustration do not constitute an undue limitation to the technical solution of the present invention.

[0035] In order to specifically describe the partial discharge pattern recognition method of the present invention, the inventors Figure 1-Figure 5The accompanying drawings shown in FIG. 1 illustrate in detail the step flow of the partial discharge pattern recognition method in one embodiment.

[0036] In this embodiment, the knowledge-driven partial discharge pattern recognition method designed by the present invention may include the following steps (1)-(4):

[0037] (1) Construct a convolutional neural network to extract features from the partial discharge (PRPS) data of power equipment to obtain data feature vectors.

[0038] In the present invention, the specific process of constructing a convolutional neural network and obtaining a data feature vector as shown in step (1) above can be found in Figure 1 , Figure 1 The present invention provides a flowchart of the steps of constructing a convolutional neural network and obtaining a data feature vector in one embodiment of the partial discharge pattern recognition method.

[0039] like Figure 1 As shown, in this embodiment, a convolutional neural network with 3 convolutional layers, 2 pooling layers, and 1 stretching layer is constructed for feature extraction of partial discharge PRPS data.

[0040] Using the convolutional layer, the convolutional neural network can extract various feature information from the original partial discharge (PRPS) data. Assuming that the trainable convolution kernel Ω in the convolutional layer is a release matrix of size s×s, the convolutional layer can be given by the following formula:

[0041]

[0042] In the above formula, x is the input partial discharge PRPS data, y is the matrix output under the convolution kernel Ω convolution, m represents the number of rows of the convolution kernel, n represents the number of columns of the convolution kernel, y(i,j) is the value of the i-th row and j-th column of the output matrix, σ(.) represents the nonlinear activation function, and the activation function of the convolutional neural network adopted in the present invention is a linear rectification function.

[0043] It should be noted that in the convolution layer, since each convolution kernel can learn limited data features, there are multiple convolution kernels in each convolution layer, and the final convolution layer outputs a tensor matrix with multiple channels.

[0044] Accordingly, to prevent overfitting during training, a max pooling layer is performed after each convolution operation (except the final one). This layer has a 2×2 kernel. This pooling kernel continuously scans each channel of the output tensor and retains only the largest value within the kernel, thereby retaining only the most important features learned by the convolutional layer and reducing the number of features. After the final convolution operation, the output tensor is stretched, and all values ​​in the tensor are rearranged in order to form the final partial discharge data feature vector.

[0045] (2) Construct a knowledge graph of partial discharge of power equipment and perform semantic modeling and structural modeling on it.

[0046] In the technical solution designed by the present invention, in order to enable the knowledge-driven algorithm to truly learn the characteristic knowledge of partial discharge, thereby constraining the data feature extraction in step (1), the inventors further constructed Figure 2 The partial discharge knowledge graph shown.

[0047] Figure 2 A schematic diagram of a partial discharge knowledge graph in one embodiment is schematically shown.

[0048] like Figure 2 As shown in this embodiment, the partial discharge knowledge graph constructed by the present invention defines five common partial discharge modes in power equipment: particle discharge, suspended discharge, tip discharge, surface discharge, and air-gap discharge. Furthermore, each discharge mode has its own associated partial discharge characteristic knowledge. For example, air-gap discharge has a weak polarity characteristic. This knowledge graph enables the algorithm to truly understand the various partial discharge modes and learn the relevant characteristic knowledge about these modes.

[0049] Therefore, in Figure 5 In the designed partial discharge knowledge graph, 5 partial discharge mode nodes and 17 feature nodes for describing the characteristics of the partial discharge mode are specifically designed, with a total of 22 nodes designed.

[0050] It should be noted that in step (2), in order to enable the algorithm to understand the graph, the present invention performs semantic modeling and structural modeling on the graph respectively.

[0051] Among them, semantic modeling is to express the semantics represented by each node in the graph (i.e. all 22 nodes) in the form of word embedding vectors. This paper adopts the BERT Chinese word embedding representation model to convert each Chinese character of the semantics of each node in the graph into a word vector, and finally uses the maximum pooling technology to convert the word vector matrix corresponding to each node into a node vector to form an embedding matrix, such as Figure 3shown. Figure 3 Schematic diagram showing the conversion of word vector matrix into node vectors. Figure 3 The “V” in represents the feature vector after pooling, and “max(.)” represents the maximum value operation.

[0052] Accordingly, when performing structural modeling, the structural features of the knowledge graph are represented as an adjacency matrix A. This adjacency matrix A is a symmetric matrix of size n×n, where n is the number of nodes in the partial discharge knowledge graph. The elements of each row and category in this matrix represent the number of connections between one node and another in the knowledge graph. For example, A(i, j) = 1 means that there is an edge between node i and node j.

[0053] (3) Based on the partial discharge knowledge graph that has undergone semantic modeling and structural modeling, a graph convolutional neural network is constructed to extract the knowledge features of various partial discharge modes of power equipment to obtain knowledge feature vectors.

[0054] Similar to the convolution operation on PD data, in order to perform similar convolution operations on the knowledge graph, we adopted and constructed a graph convolutional neural network with two layers of graph convolution layers to extract the pattern features of various PD patterns, such as Figure 4 shown.

[0055] Figure 4 This is a flowchart of the steps of obtaining knowledge feature vectors of various types of partial discharge patterns based on a constructed partial discharge knowledge graph in one embodiment of the partial discharge pattern recognition method of the present invention.

[0056] like Figure 4 As shown, in this embodiment, by using the graph convolution layer of the graph convolutional neural network, the knowledge information of each node on the partial discharge knowledge graph after semantic modeling and structural modeling is transmitted to each other through edges, and ultimately the knowledge features of the five partial discharge pattern nodes in the graph contain the attribute knowledge information of the feature. The graph convolution layer can be expressed by the following formula:

[0057] Y=σ(D -1 (A+I)XΘ)

[0058] Wherein, X represents the embedding matrix, each row of which is a node vector of a knowledge graph node; I is the identity matrix, Θ is the trainable convolution kernel of the graph convolution layer, Y is the output matrix of the graph convolution layer, D is the degree matrix of the partial discharge knowledge graph, and A represents the adjacency matrix.

[0059] Correspondingly, the degree matrix D of the partial discharge knowledge graph is a diagonal matrix, where all elements on the off-diagonal line are 0, and the elements on the diagonal line can be obtained as follows:

[0060]

[0061] In this formula, m represents the row of the adjacency matrix, n represents the number of columns of the adjacency matrix, i represents the i-th row of the adjacency matrix; A(i, m) represents the element in the i-th row and m-th column of the adjacency matrix.

[0062] Finally, the graph convolution layer combines the knowledge feature vectors of all nodes into a matrix as output. In order to realize the pattern recognition of partial discharge, only the knowledge features of the partial discharge pattern nodes are needed. Therefore, the knowledge feature vectors of the five patterns need to be extracted from the matrix.

[0063] (4) Comparing the similarity between the data feature vector and the knowledge feature vector, and taking the partial discharge pattern with the greatest similarity as the recognition result.

[0064] In the present invention, after obtaining the data feature vector and the knowledge feature vector, it is necessary to compare the similarity between the data feature vector and the knowledge vector of each partial discharge pattern in turn, and take the partial discharge pattern with the greatest similarity as the final recognition result, such as Figure 5 shown.

[0065] Figure 5 The figure schematically shows a comparison diagram of the similarity between a data feature vector and a knowledge feature vector in one embodiment of the partial discharge pattern recognition method of the present invention.

[0066] like Figure 5 As shown in FIG, by comparing the data feature vector with the knowledge feature vector, it can be found that the similarity between the data feature vector and the local discharge pattern of the tip discharge is the highest, which is 0.95. Therefore, the local discharge pattern of the tip discharge is output as the final recognition result.

[0067] It should be noted that when evaluating similarity, the present invention uses the Euler distance between vectors as the similarity evaluation method. The Euler distance S c It can be given by the following formula:

[0068]

[0069] Where, f d is the characteristic diagnostic vector of partial discharge data obtained by convolutional neural network, f k is the knowledge feature vector of a certain type of partial discharge pattern obtained by the graph convolutional neural network.

[0070] It should be noted that the prior art within the scope of protection of the present invention is not limited to the embodiments given in this application document. All prior art that does not contradict the solutions of the present invention, including but not limited to prior patent documents, prior publications, prior public uses, etc., can be included in the scope of protection of the present invention.

[0071] In addition, the combination of the various technical features in this case is not limited to the combination described in the claims of this case or the combination described in the specific embodiments. All technical features recorded in this case can be freely combined or combined in any way unless there is a contradiction between them.

[0072] It should also be noted that the embodiments listed above are merely specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and similar changes or modifications made therefrom that can be directly derived from or easily conceived by those skilled in the art based on the disclosure of the present invention are intended to fall within the scope of protection of the present invention.

Claims

1. A knowledge-driven partial discharge pattern recognition method, characterized in that: Including steps: (1) Construct a convolutional neural network to extract features from the partial discharge (PRPS) data of power equipment to obtain data feature vectors; (2) constructing a partial discharge knowledge graph of power equipment and performing semantic modeling and structural modeling on it; the partial discharge knowledge graph has a plurality of nodes, and the nodes include a plurality of partial discharge mode nodes and a plurality of feature nodes for describing the characteristics of the partial discharge mode; when performing semantic modeling, the Chinese characters of the semantics of each node in the partial discharge knowledge graph are converted into word vectors to form a word vector matrix, and then the word vector matrix corresponding to each node is converted into a node vector to form an embedding matrix; when performing structural modeling, the structural characteristics of the partial discharge knowledge graph are represented as an adjacency matrix A, and the adjacency matrix is ​​a symmetric matrix of size n×n, where n is the number of nodes in the partial discharge knowledge graph; (3) Based on the partial discharge knowledge graph that has undergone semantic modeling and structural modeling, a graph convolutional neural network consisting of two graph convolutional layers is constructed to extract the knowledge features of various partial discharge modes of power equipment to obtain knowledge feature vectors; (4) Comparing the similarity between the data feature vector and the knowledge feature vector, and taking the partial discharge pattern with the greatest similarity as the recognition result.

2. The knowledge-driven partial discharge pattern recognition method according to claim 1, characterized in that: In step (1), the convolutional neural network has 3 convolutional layers, 2 pooling layers and 1 stretching layer.

3. The knowledge-driven partial discharge pattern recognition method according to claim 2, characterized in that: In step (1), each of the convolutional layers has multiple convolution kernels.

4. The knowledge-driven partial discharge pattern recognition method according to claim 1, wherein: In step (1), the activation function of the convolutional neural network is a linear rectification function.

5. The knowledge-driven partial discharge pattern recognition method according to claim 1, wherein: In step (2), when performing semantic modeling, the BERT Chinese character embedding representation model is used to convert the semantic Chinese characters of each node in the partial discharge knowledge graph into word vectors to form a word vector matrix; the maximum pooling method is used to convert the word vector matrix corresponding to each node into a node vector.

6. The knowledge-driven partial discharge pattern recognition method according to claim 1, characterized in that: In step (3), the graph convolution layer is expressed as follows: Y=σ(D -1 (A+I)X Θ ) Wherein, X represents the embedding matrix, I is the identity matrix, Θ is the trainable convolution kernel of the graph convolution layer, Y is the output matrix of the graph convolution layer, D is the degree matrix of the partial discharge knowledge graph, and a represents the adjacency matrix.

7. The knowledge-driven partial discharge pattern recognition method according to claim 1, wherein: In step (4), the similarity is calculated using Euler distance.

8. The knowledge-driven partial discharge pattern recognition method according to claim 1, wherein: The local discharge modes include particle discharge, suspension discharge, tip discharge, surface discharge, and air gap discharge.

Images