An abnormal detection method for partial discharge faults of vehicle-mounted cable terminals

Abstract

The present invention discloses a method for detecting abnormal partial discharge faults in on-vehicle cable terminals, specifically: directly using the time-series data of partial discharge signals as input, and constructing a single sample through sliding window segmentation; dividing the sample and determining the adjacency relationship, edge weight matrix, and constructing a graph signal according to the similarity index and adjacent windows; constructing the node degree centrality in the heat map of the original partial discharge waveform, and determining the window size and step size by judging whether the size of the node degree centrality is in the main position of partial discharge; constructing a graph self-attention convolutional layer as the main body, extracting the global features of the partial discharge graph signal through a graph adaptive pooling layer, and finally using a three-layer fully connected layer as the classification layer to output the probability of the discharge type. The present invention realizes the abnormal detection of partial discharge signals in on-vehicle cable terminals, and is a fast and accurate method for detecting abnormal partial discharge signals in on-vehicle cable terminals, which is beneficial to ensuring the safe and stable operation of cable terminals.

Description

Technical Field

[0001] The present invention belongs to the technical field of on-train cable terminal fault diagnosis, and particularly relates to a method for detecting abnormal partial discharge faults of on-vehicle cable terminals. Background Art

[0002] Partial discharge defects in high-voltage cable terminals in the on-train power supply system of high-speed railways are one of the potential hazards in the high-speed railway power supply system. Its internal insulation structure is complex and is prone to partial discharge due to external factor interference. In recent years, with the increase in the speed and traction power of high-speed railways, the impact pressure on the terminals has increased, which may lead to the expansion of internal insulation defects, and then trigger partial discharge in the cable terminals, deteriorating the defects, and thus resulting in terminal failures or even train operation accidents. Therefore, accurately detecting internal defects in on-vehicle cable terminals is crucial for preventing terminal failures and ensuring the safe operation of high-speed trains.

[0003] Currently, for the detection of partial discharge defects in cable terminals, it is commonly divided into extracting features manually and classifying them through machine learning algorithms, or automatically extracting features based on deep learning such as typical CNN, RNN and other models. After data training, it can be determined which type of partial discharge the defect belongs to. Then, classification through machine learning algorithms requires manual experience, and the classification results are generally average. For networks such as CNN, due to the type of input data, the required computing resources are large and it is not easy to be applied in practice. Summary of the Invention

[0004] In order to effectively detect the partial discharge types of different defects in on-vehicle cable terminals and achieve accurate, efficient, and rapid detection of abnormal detection, the present invention provides a method for detecting abnormal partial discharge faults of on-vehicle cable terminals.

[0005] A method for detecting abnormal partial discharge faults of on-vehicle cable terminals according to the present invention includes the following steps:

[0006] S1. Directly use the time-series data of partial discharge signals as input, and construct a single sample through sliding window segmentation.

[0007] S2. Segment the samples and determine the adjacency matrix and edge weights according to the similarity index and adjacent sub-windows to construct a graph signal.

[0008] S3. Construct a heat map of the node degree centrality in the original partial discharge waveform, and determine the sub-window size and step size by judging whether the size of the node degree centrality is in the main position of the partial discharge.

[0009] S4. Construct a graph self-attention convolutional layer as the main body, extract the global features of the partial discharge graph signal through a graph adaptive pooling layer, and finally use a three-layer fully connected layer as the classification layer to output the probability of the discharge type.

[0010] Further, in step S1, the partial discharge time series data of the cable terminal is obtained by the pulse current detection method, and the time series data of 20 ms is acquired. It is divided into sliding windows with a window size of 1000, where the step size is 900, and each window is used as an input sample.

[0011] Further, step S2 is specifically as follows:

[0012] S21. The input sample is divided into sub-windows by sliding windows. The size of each sub-window is set to window_size, and the step size is step_size. A total of num_windows sub-windows can be generated as graph nodes, where It is calculated that the sub-window contains time series values as node features.

[0013] S22. Using the segmented graph nodes, for the graph nodes according to where x i represents the element in the first node, and y i represents the element in the second node. Calculate the Euclidean distance and cosine similarity, traverse all nodes, and judge whether the distance between different nodes meets the threshold. If it meets, it is set to be connected; otherwise, it is unconnected. And according to the window order of the nodes, corresponding connections are made, and the connection relationship is constructed into an adjacency matrix, where the connection is set to 1 and the unconnection is set to 0. Secondly, the cosine similarity matrix is used as the edge weight matrix, and the graph nodes and connection relationships are constructed into a graph signal.

[0014] Further, in step S3, by calculating the node degree centrality under different sub-windows and mapping it to the corresponding partial discharge waveform window graph after segmentation, the sizes of the sub-window and step size are reasonably determined through a heat map.

[0015] Further, step S4 includes:

[0016] S41. Construct a model based on the SAG-GAT model, which consists of three layers of graph self-attention convolutional layers GAT for feature extraction, two layers of batch normalization layers, three layers of adaptive pooling layers for global feature extraction, and a dropout layer to discard some neurons to prevent overfitting. Finally, the outputs of the three pooling layers are concatenated as the input of the classification layer, and the outputs of the three fully connected layers are the probabilities of the partial discharge signal belonging to the defect.

[0017] S42. Use the constructed model for training and testing. Finally, for the graph signal constructed in step S3 from the time series signal, prediction is performed to achieve end-to-end abnormal detection of the partial discharge time series signal.

[0018] The beneficial technical effects of the present invention are:

[0019] First, using the partial discharge time-series data of various different defects of the cable terminal, individual samples are segmented through a sliding window, and graph signals are constructed for the samples. By constructing a graph network model based on SAG-GAT, through training and testing, it is possible to accurately judge the time-series signals of different defects of the cable terminal. Finally, end-to-end anomaly detection of the partial discharge signals of the cable terminal is completed. The present invention realizes the accurate detection of the partial discharge time-series signals of the in-vehicle cable terminal, and is an efficient and accurate detection method for the time-series signals of the partial discharge faults of the in-vehicle cable terminal, which is conducive to ensuring the safe and stable operation of the cable terminal. Description of the Drawings

[0020] Figure 1 It is a flowchart of the method for detecting anomalies in partial discharge faults of the in-vehicle cable terminal of the present invention.

[0021] Figure 2 It is a flowchart of converting the partial discharge time-series signal of the cable terminal into a graph signal.

[0022] Figure 3 It is a schematic diagram of the structure of the SAG-GAT model of the present invention.

[0023] Figure 4 It is a graph of the model training loss value and accuracy rate for constructing graph signals under different windows of the present invention.

[0024] Figure 5 It is a schematic diagram of the confusion matrix after testing the test set of the model of the present invention. Detailed Implementation Manner

[0025] The following further elaborates on the present invention in detail in conjunction with the drawings and specific implementation methods.

[0026] The process of a method for detecting anomalies in partial discharge faults of an in-vehicle cable terminal of the present invention is as Figure 1 shown, and includes the following steps:

[0027] S1. Directly using the partial discharge signal time-series data as input, individual samples are formed through sliding window segmentation.

[0028] The partial discharge time-series data of the cable terminal is obtained by the pulse current detection method, and 20 ms of time-series data is acquired. It is slid and segmented according to a window of 1000, with a step size of 900, and each window is used as an input sample.

[0029] S2. The samples are segmented and graph signals are constructed based on the Euclidean distance and adjacent sub-windows.

[0030] S21. Split the input sample into sub - windows using a sliding window. The size of each sub - window is set to window_size, and the step size is step_size. num_windows sub - windows can be generated as graph nodes. Among them it is calculated that the sub - window contains time - series values as node features.

[0031] S22. Using the segmented graph nodes, for the graph nodes according to where x i represents the element in the first node, and y i represents the element in the second node. Calculate the Euclidean distance and cosine similarity, traverse all nodes, judge whether the distance between different nodes meets the threshold. If it meets, set it as connected, otherwise as unconnected, and make corresponding connections according to the window order of the nodes. Construct an adjacency matrix for the connection relationship, where the connection is set to 1 and the non - connection is set to 0. Secondly, use the cosine similarity matrix as the edge weight matrix, and construct the graph nodes and connection relationship into a graph signal. The conversion process of the cable terminal partial discharge time - series signal to the graph signal is as Figure 2 shown.

[0032] S3. Construct the node degree centrality and the heat map of the original partial discharge waveform. By calculating the node degree centrality under different sub - windows, map it to the corresponding segmented partial discharge waveform window graph, and reasonably determine the sub - window size and step size through the heat map.

[0033] The formula for calculating the node degree centrality is as follows:

[0034]

[0035] where, k i represents the number of edges connecting the existing node and node i, and N - 1 represents the number of nodes minus 1.

[0036] S4. Construct a graph self - attention convolutional layer as the main body, use a graph adaptive pooling layer to extract the global features of the partial discharge graph signal, and finally use three fully - connected layers as the classification layer to output the probability of the discharge type.

[0037] S41. Construct a model based on the SAG - GAT model. The specific model is shown as Figure 3 shown, where there are three layers of graph self - attention convolutional layers GAT for feature extraction, two layers of batch normalization layers, three layers of adaptive pooling layers for global feature extraction, and a dropout layer to discard some neurons to prevent overfitting. Finally, the outputs of the three pooling layers are concatenated as the input of the classification layer, and the three fully - connected layers output the probability of the defect to which the partial discharge signal belongs.

[0038] S42. Use the constructed model to train and test it. The training results of the graph signals constructed by different windows are asFigure 4 As shown, its predicted confusion matrix is as Figure 5 shown, realizing accurate identification of partial discharge of different defects in cable terminals.

[0039] The present invention uses a sliding window to segment time series signals, constructs a graph signal based on the Euclidean distance of the window and the adjacent situation, constructs a graph network based on SAG-GAT to realize automatic extraction of graph signal features, and realizes anomaly detection of time series signals. This method can effectively classify and detect anomalies in partial discharge signals of in-vehicle cable terminals, so as to take corresponding measures in time to prevent cable terminal failures and improve the operation reliability and safety of cable terminals.

Claims

1. An abnormal detection method for partial discharge faults of in-vehicle cable terminals, characterized in that, it includes the following steps: S1. Directly use the time-series data of partial discharge signals as input, and through sliding window segmentation, form a single sample; S2. Divide the sample by window and determine the adjacency relationship according to the similarity index and adjacent sub-windows, and the edge weight matrix to construct a graph signal; S3. Construct the node degree centrality in the original partial discharge waveform heat map, and determine the sub-window size and step size by judging whether the node degree centrality is in the main position of partial discharge; S4. Construct a graph self-attention convolutional layer as the main body, use a graph adaptive pooling layer to extract the global features of the partial discharge graph signal, and finally use three fully connected layers as the classification layer to output the probability of the discharge type.

2. The abnormal detection method for partial discharge faults of in-vehicle cable terminals according to claim 1, characterized in that, in step S1, the time-series data of partial discharge of the cable terminal is obtained by the pulse current detection method, 20 ms of time-series data is obtained, and it is slid and segmented by a window of 1000, where the step size is 900, and each window is used as an input sample.

3. The abnormal detection method for partial discharge faults of in-vehicle cable terminals according to claim 1, characterized in that, step S2 is specifically: S21. Split the input sample into sub - windows by a sliding window. The size of each sub - window is set to window_size, and the step size is step_size. num_windows sub - windows can be generated as graph nodes, where is calculated, and the sub - window contains time - series values as node features; S22. Using the segmented graph nodes, the graph nodes are arranged according to where x i represents the element in the first node, and y i represents the element in the second node. Calculate the Euclidean distance and cosine similarity, traverse all nodes, and determine whether the distance between different nodes meets the threshold. If it meets the threshold, set it as connected; otherwise, it is unconnected. And make corresponding connections according to the window order of the nodes, construct an adjacency matrix for the connection relationship, where the connection is set to 1 and the unconnection is set to 0. Secondly, use the cosine similarity matrix as the edge weight matrix, and construct the graph nodes and the connection relationship into a graph signal.

4. The abnormal detection method for partial discharge faults of in-vehicle cable terminals according to claim 1, characterized in that, in step S3, the node degree centrality under different sub-windows is calculated, mapped to the partial discharge waveform window graph after corresponding segmentation, and the sub-window size and step size are reasonably determined through the heat map.

5. The abnormal detection method for partial discharge faults of in-vehicle cable terminals according to claim 1, characterized in that, step S4 includes: S41. Construct a model based on the SAG-GAT model, which consists of three layers of graph self-attention convolutional layers GAT for feature extraction, two layers of batch normalization layers, three layers of adaptive pooling layers for global feature extraction, and a dropout layer to discard some neurons to prevent overfitting. Finally, the outputs of the three pooling layers are concatenated as the input of the classification layer, and the probability of the defect to which the partial discharge signal belongs is output by the last three fully connected layers; S42. Use the constructed model to train and test it, and finally realize the prediction of the graph signal constructed in step S3 for the time-series signal, and realize the abnormal detection of the end-to-end partial discharge time-series signal.

Images