A high-voltage cable fault diagnosis method based on space-time attention

Abstract

The application discloses a high-voltage cable fault diagnosis method based on space-time attention, relates to the field of high-voltage cable fault diagnosis, and comprises the following steps: acquiring reference operation data of a high-voltage cable, and constructing a one-dimensional node path coordinate, a node space correlation matrix and a propagation time delay reference matrix; dividing a time window of a reference sampling interval, extracting partial discharge, sheath current and temperature characteristics, forming a reference state matrix, a fluctuation boundary matrix and a characteristic effectiveness mask; acquiring real-time monitoring data and constructing a real-time state matrix, calculating a boundary deviation matrix, extracting spatial deviation characteristics through CNN and extracting fault time sequence characteristics through LSTM; combining space attention and time attention to generate comprehensive fault diagnosis characteristics, outputting a fault type, a fault position and a probability, and realizing real-time diagnosis and explainable positioning of high-voltage cable faults.

Description

Technical Field

[0001] This invention relates to the field of high-voltage cable fault diagnosis technology, and in particular to a high-voltage cable fault diagnosis method based on spatiotemporal attention. Background Technology

[0002] In recent years, with the rapid development of smart grids, Internet of Things sensing and high-voltage cable online monitoring systems, a large-scale, multi-source heterogeneous data system has been generated for high-voltage cable operation, partial discharge, sheath current, temperature and load status. This system includes both structured data such as continuous current, voltage and temperature rise collected by sensors, and unstructured data such as monitoring videos, maintenance records and text logs. Meanwhile, the development of edge computing and deep learning technologies has enabled high-voltage cable fault diagnosis to evolve from traditional periodic inspections and post-event analysis to a closed-loop model of "real-time monitoring - anomaly alarm - time series analysis - fault location". To support this model, the industry is conducting time alignment, missing data repair, feature cleaning, and effectiveness evaluation at the data level, and exploring CNN-LSTM and spatiotemporal attention mechanisms at the model level to improve adaptability and interpretability to multiple nodes, multiple features, and multiple operating conditions. However, existing solutions mostly focus on static feature analysis or single-source signal processing, lacking spatiotemporal correlation and anomaly causal expression across nodes and time windows. This makes it difficult to fully associate fault time series features with spatial nodes, temporal evolution, and multi-source sensor information, and there is still room for improvement in interpretability and location accuracy. Summary of the Invention

[0003] In view of the aforementioned existing problems, the present invention is proposed.

[0004] Therefore, this invention provides a high-voltage cable fault diagnosis method based on spatiotemporal attention to solve the problem in the prior art of performing spatiotemporal correlation analysis, real-time anomaly detection, and interpretable fault location on multi-node, multi-source heterogeneous sensor data.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a high-voltage cable fault diagnosis method based on spatiotemporal attention, which includes: Obtain the baseline operating data of the target high-voltage cable, select the seven-day continuous normal operating history interval of the target high-voltage cable as the baseline sampling interval, calculate the cumulative path length of each node along the cable to form a one-dimensional path coordinate and construct a node set; sort by path coordinate, calculate the path distance between adjacent nodes, further construct the node spatial correlation matrix and calculate the propagation delay between any nodes to form a propagation delay baseline matrix; The benchmark sampling interval is divided into time windows, and the characteristics of partial discharge, sheath current and temperature are statistically analyzed and combined to form a node window state vector. Invalid features are removed based on the validity mask. The mean, fluctuation and upper and lower boundaries of the state are calculated and a benchmark matrix is ​​formed. Real-time monitoring data is acquired and a state matrix is ​​constructed according to time window and node order. A validity mask is applied and standardized deviation values ​​and boundary deviation matrices are calculated. The boundary deviation matrix of the continuous window is input into a CNN to extract spatial deviation features, which are then corrected using a spatial correlation matrix. Temporal features are extracted using an LSTM to form fault temporal features. The node anomaly intensity is calculated based on the boundary deviation matrix and spatial attention weights are generated by combining spatial correlation. Spatial attention enhancement features are obtained by weighting spatial deviation features. Temporal attention weights are calculated and continuous window features are weighted and fused to form spatiotemporal attention features. These features are combined with fault time sequence features to generate comprehensive fault diagnosis features. Finally, the fault type, node location and probability are determined and the fault diagnosis results are output.

[0006] As a preferred embodiment of the high-voltage cable fault diagnosis method based on spatiotemporal attention described in this invention, the following steps are included: acquiring the baseline operating data of the target high-voltage cable, selecting a seven-day continuous normal operating history interval of the target high-voltage cable as the baseline sampling interval, calculating the cumulative path length along the cable for each node to form a one-dimensional path coordinate, and constructing a node set. ; Using the cable's starting and ending points as the origin of the coordinate system, for any node on the target high-voltage cable... Calculate the distance from the starting point to this node along the actual cable laying direction. The cumulative path length is calculated and used as the one-dimensional path coordinate of the node. ; By unifying the cable's starting and ending points, intermediate joints, sheath grounding points, cross-connection boxes, and sensor installation points as nodes, a node set for the target high-voltage cable is generated. .

[0007] As a preferred embodiment of the high-voltage cable fault diagnosis method based on spatiotemporal attention described in this invention, the following steps are taken: sorting by path coordinates, calculating the path distance between adjacent nodes, further constructing a node spatial correlation matrix, and calculating the propagation delay between any nodes to form a propagation delay reference matrix based on the node set. Sort the nodes in ascending order of their one-dimensional path coordinates, and calculate the path distance between adjacent nodes. ; Calculate the first using adjacent priority space association. The node and the first Spatial correlation value between nodes ; Based on spatial correlation value Constructing the node space association matrix ; Calculate the propagation delay between any two nodes. And form a propagation delay reference matrix. .

[0008] As a preferred embodiment of the high-voltage cable fault diagnosis method based on spatiotemporal attention described in this invention, the step of dividing the reference sampling interval into time windows, statistically analyzing partial discharge, sheath current, and temperature characteristics, and combining them to form a node window state vector refers to dividing the reference sampling interval into multiple windows of equal length. ; In each time window Internally, a partial discharge sensor is used to perform pulse detection and feature statistics on the original partial discharge waveform to obtain the number of partial discharge pulses. and the maximum amplitude of partial discharge pulse And calculate the total energy of partial discharge. and the proportion of high-frequency energy and combine to form the first The node at the th Partial discharge reference feature vector within a time window ; In each time window Inside, obtain the first Time window The sheath current sampling sequence is obtained by the sheath current sensor. Calculate the effective value of the sheath current. Sheath current change rate Harmonic distortion characteristics And combine them into a base characteristic vector of sheath current. ; In each time window Inside, the temperature value of the connector is acquired by the temperature sensor. Ambient temperature value and operating load current value Calculate the relative ambient temperature rise Further, the load-corrected temperature rise is calculated based on the relative ambient temperature rise and the operating load current. Then, based on the load-corrected temperature rise in the current time window and the load-corrected temperature rise in the previous time window, calculate the load-corrected temperature rise change rate. The calculated load-corrected temperature rise and the load-corrected temperature rise rate of change are combined into a temperature reference eigenvector. ; Based on partial discharge reference feature vector Sheath current reference eigenvector and temperature reference eigenvector All features within the same node and the same time window are combined in a fixed order to form a node window state vector. .

[0009] As a preferred embodiment of the high-voltage cable fault diagnosis method based on spatiotemporal attention described in this invention, the step of eliminating invalid features based on validity masks, calculating the state mean, fluctuation, and upper and lower boundaries, and forming a reference matrix refers to establishing a node feature validity mask matrix. ; Calculate the first The node State mean of each feature and standard deviation of fluctuation Forming a baseline mean matrix of operating states and state benchmark fluctuation matrix ; Based on state mean and standard deviation of fluctuation Calculate the upper and lower boundaries of the fluctuation and And combined into a wave boundary matrix .

[0010] As a preferred embodiment of the high-voltage cable fault diagnosis method based on spatiotemporal attention described in this invention, wherein: acquiring real-time monitoring data and constructing a state matrix according to time windows and node order refers to acquiring real-time monitoring data of the target high-voltage cable during its current operation; The real-time monitoring data is processed to obtain the first... The node at the th Real-time node window state vector within a real-time diagnostic time window ; Based on the one-dimensional path coordinates of each node in the node set, the real-time node window state vectors of all nodes within the same real-time diagnostic time window are combined into a real-time state matrix. .

[0011] As a preferred embodiment of the high-voltage cable fault diagnosis method based on spatiotemporal attention described in this invention, the step of applying a validity mask and calculating the standardized deviation value and boundary deviation matrix refers to calling the node feature validity mask matrix. For the real-time state matrix By applying validity constraints, the real-time valid state matrix is ​​obtained. ; Based on the benchmark mean matrix of operating status and state benchmark fluctuation matrix Calculate the first The node The feature in the first Standardized state deviation within a real-time diagnostic time window and form the first State deviation matrix for each real-time diagnostic time window ; According to the wave boundary matrix The system judges whether the real-time feature value exceeds the normal fluctuation range, and obtains the result. The node The feature in the first Boundary deviation value within a real-time diagnostic time window and form the first Boundary deviation matrix corresponding to each real-time diagnostic time window .

[0012] As a preferred embodiment of the high-voltage cable fault diagnosis method based on spatiotemporal attention described in this invention, the method involves: inputting the boundary deviation matrix of a continuous window into a CNN to extract spatial deviation features, correcting it with a spatial correlation matrix, and extracting temporal features using an LSTM to form fault temporal features. Input the CNN module, perform convolution calculations on the boundary deviation features of each node along the node arrangement direction, and obtain the first... Spatial deviation characteristics corresponding to each real-time diagnostic time window ; Using the node space correlation matrix The spatial deviation features are corrected to obtain the corrected spatial deviation features. ; Spatial deviation features corresponding to multiple consecutive real-time diagnostic time windows are input into the LSTM module in chronological order to extract the evolution of boundary deviation features over time. ; The first The LSTM hidden states corresponding to each real-time diagnosis time window are used as the fault time-series features extracted by CNN-LSTM. .

[0013] As a preferred embodiment of the high-voltage cable fault diagnosis method based on spatiotemporal attention described in this invention, wherein: the node anomaly intensity is calculated based on the boundary deviation matrix and spatial attention weights are generated by combining spatial correlation; the spatial deviation features are weighted to obtain the spatial attention enhancement feature index according to the first... Boundary deviation matrix corresponding to each real-time diagnostic time window Calculate the first The node anomaly strength within this time window ; Combining the node space correlation matrix Calculate the first Spatial attention score of each node ; Normalize the spatial attention scores of each node to obtain the first... The node at the th Spatial attention weights within a real-time diagnostic time window ; Determine the node number corresponding to the fault location based on spatial attention weights. ; Based on spatial attention weights, for the th Corrected spatial deviation features corresponding to each real-time diagnostic time window Weighting is performed to obtain spatial attention enhancement features. .

[0014] As a preferred embodiment of the high-voltage cable fault diagnosis method based on spatiotemporal attention described in this invention, the method involves: calculating temporal attention weights and weighting and fusing continuous window features to form spatiotemporal attention features, combining these features with fault time-series features to generate comprehensive fault diagnosis features, and finally determining the fault type, node location, and probability, and outputting the fault diagnosis result based on the hidden state. Calculate attention score over time ; Normalize the temporal attention score of the continuous real-time diagnostic time window to obtain the first... Time attention weights corresponding to each real-time diagnostic time window ; Based on temporal attention weights, the spatial attention enhancement features corresponding to continuous real-time diagnostic time windows are weighted and fused to obtain spatiotemporal attention fusion features. ; Fusing spatiotemporal attention features With fault timing characteristics By fusion, comprehensive fault diagnosis features are obtained. ; Based on comprehensive fault diagnosis features Calculate the fault type judgment result ; Based on the fault type probability vector Determine the fault type of the target high-voltage cable. ; Output the first Fault diagnosis results corresponding to each real-time diagnostic time window .

[0015] The beneficial effects of this invention are as follows: By acquiring benchmark operating data of high-voltage cables, this invention constructs one-dimensional node path coordinates, node spatial correlation matrices, and propagation delay benchmark matrices along the cable laying direction; further, it divides the benchmark sampling interval into time windows, extracts multi-source state features such as partial discharge, sheath current, and temperature, and forms a benchmark state matrix, a fluctuation boundary matrix, and a feature validity mask; in the real-time diagnosis stage, it constructs a real-time state matrix according to the same node order and feature arrangement, calculates the standardized deviation value and boundary deviation matrix, and combines CNN to extract local spatial anomaly features, corrects them using the spatial correlation matrix, and then extracts fault temporal evolution features through LSTM; finally, it performs weighted fusion of abnormal nodes and key time windows based on spatial attention and temporal attention to generate comprehensive fault diagnosis features, outputting fault type, fault node, fault location, fault probability, and attention weight, thereby realizing spatiotemporal collaborative diagnosis and interpretable localization of high-voltage cable faults. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a high-voltage cable fault diagnosis method based on spatiotemporal attention.

[0018] Figure 2 This is a flowchart for calculating fault timing characteristics based on real-time detection data of high-voltage cables. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0022] Reference Figures 1-2 As one embodiment of the present invention, this embodiment provides a high-voltage cable fault diagnosis method based on spatiotemporal attention, comprising the following steps: S1. Obtain the target high-voltage cable's baseline operating data, select the target high-voltage cable's seven-day continuous normal operating history interval as the baseline sampling interval, calculate the cumulative path length of each node along the cable to form a one-dimensional path coordinate and construct a node set; sort by path coordinate, calculate the path distance between adjacent nodes, further construct the node spatial correlation matrix and calculate the propagation delay between any nodes to form a propagation delay baseline matrix; First, the baseline operating data of the target high-voltage cable is obtained, and the historical range of normal operation of the target high-voltage cable for 7 consecutive days is selected as the baseline sampling range. This includes the cable's starting and ending points, intermediate joint locations, sheath grounding points, cross-connection box locations, sensor installation locations, total cable length, cable type, rated voltage level, and cable signal propagation speed. Using the cable's starting and ending points as the origin of the coordinate system, for any node on the target high-voltage cable... Calculate the distance from the starting point to this node along the actual cable laying direction. The cumulative path length is calculated and used as the one-dimensional path coordinate of the node. : ; in, Indicates the first Nodes One-dimensional path coordinates; Indicates from the cable start point to the end point The number of cable segments that a node passes through; Indicates the first The actual length of the cable segment; if the first Each node is recorded directly along the line from the starting point to the terminal point. It equals the distance along the line; By unifying the cable's starting and ending points, intermediate joints, sheath grounding points, cross-connection boxes, and sensor installation points as nodes, a node set for the target high-voltage cable is generated. ,in, Indicates the first 1 node Represents the total number of nodes; each node is represented as ,in, Indicates the node number. Represents the one-dimensional path coordinates of the node. The code indicates the node category (e.g., the starting terminal code is 1, the ending terminal code is 2, the intermediate joint code is 3, the sheath grounding point code is 4, the cross-connection box code is 5, and the ordinary sensor node code is 6). The type of sensor contained in the node is encoded using binary encoding. , This indicates whether the node has a partial discharge sensor. This indicates whether the node has a sheath current sensor. This indicates whether the node has a temperature sensor. If a corresponding sensor exists, the corresponding bit is set to 1; otherwise, it is set to 0. Based on node set Sort the nodes in ascending order of their one-dimensional path coordinates, and calculate the path distance between adjacent nodes. ,in Indicates the first The node and the first The cable path distance between nodes; Further calculations using adjacent priority space associations are performed. The node and the first Spatial correlation value between nodes : ; in, and They represent the first The node and the first One-dimensional path coordinates of each node; This represents the spatial attenuation coefficient, with a preferred value of 500m, which enables nearby nodes to maintain a high correlation. This represents the threshold for truncating long-distance associations, with a preferred value of 1500m, which can prevent irrelevant long-distance nodes from interfering with subsequent spatiotemporal attention diagnosis. Based on spatial correlation value Constructing the node space association matrix ; Calculate the propagation delay between any two nodes. And form a propagation delay reference matrix. : ; in, The propagation speed of the fault traveling wave in the cable is preferably taken as... It can better reflect the engineering propagation speed of traveling wave signals in high-voltage cables; S2. Divide the reference sampling interval into time windows, statistically analyze the characteristics of partial discharge, sheath current and temperature, combine them to form a node window state vector, and remove invalid features based on the validity mask; calculate the state mean, fluctuation and upper and lower boundaries and form a reference matrix; The reference sampling interval is divided into multiple windows of equal time duration. ， Indicates the first A time window, Indicates the total number of windows. , This can prevent the original high-frequency data from being directly entered into the baseline matrix, which would result in excessively high dimensionality. In each time window Internally, a partial discharge sensor is used to perform pulse detection and feature statistics on the original partial discharge waveform to obtain the number of partial discharge pulses. and the maximum amplitude of partial discharge pulse And calculate the total energy of partial discharge. and the proportion of high-frequency energy and combine to form the first The node at the th Partial discharge reference feature vector within a time window ; Among them, the total energy of partial discharge , Indicates the first The amplitude of each detected partial discharge pulse; the proportion of high-frequency energy. , Indicates the first The node at the th Frequency within a window Power spectral density at; This represents the high-frequency boundary frequency, with a preferred value of 5MHz, which can effectively distinguish between low-frequency background interference and high-frequency discharge pulse energy. and These represent the lowest and highest frequencies in the partial discharge spectrum analysis, respectively. This represents a constant to prevent the denominator from being zero; In each time window Inside, obtain the first Time window The sheath current sampling sequence is obtained by the sheath current sensor. Calculate the effective value of the sheath current. Sheath current change rate Harmonic distortion characteristics And combine them into a base characteristic vector of sheath current. ; Among them, the effective value of the protective layer current , Indicates the first The node at the th The first window The sheath current value at each sampling point This indicates the number of sheath current sampling points within the window; the sheath current change rate. , This indicates the current effective value of the sheath current in the window. This represents the RMS value of the sheath current in the previous window. For the first window, the RMS value of the sheath current can be set as the average value of the RMS value of the sheath current at that node within the reference sampling interval; harmonic distortion characteristics. , The fundamental current component is obtained by spectral decomposition of the current sampling sequence within the window. Indicates the first The subharmonic current components are obtained by performing harmonic analysis on the current sampling sequence within the window. This indicates the highest harmonic order involved in the calculation, preferably 5, which balances anomaly identification capability and computational load. This represents a constant to prevent the denominator from being zero; In each time window Inside, the temperature value of the connector is acquired by the temperature sensor. Ambient temperature value and operating load current value Calculate the relative ambient temperature rise Further, the load-corrected temperature rise is calculated based on the relative ambient temperature rise and the operating load current. Then, based on the load-corrected temperature rise in the current time window and the load-corrected temperature rise in the previous time window, calculate the load-corrected temperature rise change rate. The calculated load-corrected temperature rise and the load-corrected temperature rise rate of change are combined into a temperature reference eigenvector. ; Among them, relative ambient temperature rise , Indicates the first Temperature of each node joint, Indicates the first Ambient temperature within a window; load-corrected temperature rise , Indicates the first Operating load current within each window Indicates the rated current of the cable. This represents the temperature load correction stability coefficient, with a preferred value of [value missing]. This avoids excessive amplification of the temperature rise correction value under low load conditions, while not significantly weakening the temperature correction effect under medium and high load conditions; load correction temperature rise change rate , This indicates the current window load-corrected temperature rise. This indicates the load-corrected temperature rise of the previous window. The initial value can be set to the average value of the load-corrected temperature rise in the reference sampling interval. Based on partial discharge reference feature vector Sheath current reference eigenvector and temperature reference eigenvector All features within the same node and the same time window are combined in a fixed order to form a node window state vector. =[ , ]; In practical high-voltage cable online monitoring systems, the sensor configurations at different nodes are usually different. Therefore, a node feature validity mask matrix needs to be established. ; ; in, Indicates the first Feature validity mask vector corresponding to each node for Partial discharge sensor corresponding The first four digits correspond to the temperature sensor. The last two bits correspond to the remaining three middle bits for the sheath current sensor. For example, for a node with only a temperature sensor installed, the first seven features in its 9-dimensional state vector are invalid, while the last two features are valid. Therefore, the mask is... ; Calculate the first The node State mean of each feature and standard deviation of fluctuation Forming a baseline mean matrix of operating states and state benchmark fluctuation matrix ; in, ; ; Indicates the first The node at the th Node window state vector Inner One eigenvalue; Based on state mean and standard deviation of fluctuation Calculate the upper and lower boundaries of the fluctuation and And combined into a wave boundary matrix Among them, the lower boundary of the fluctuation upper boundary of fluctuation , This represents the normal fluctuation boundary coefficient, with a preferred value of [value missing]. It can achieve a good balance between false alarm rate and sensitivity; Finally, the reference state matrix of the high-voltage cable is output. ; S3. Acquire real-time monitoring data and construct a state matrix according to time window and node order. Apply validity mask and calculate standardized deviation value and boundary deviation matrix. Input the boundary deviation matrix of continuous window into CNN to extract spatial deviation features, correct it with spatial correlation matrix, and then extract temporal features through LSTM to form fault temporal features. Acquire real-time monitoring data of the target high-voltage cable during its current operation, and process the real-time monitoring data according to the same time window division method, the same node sorting method, and the same feature arrangement order to obtain the [number of data]. The node at the th Real-time node window state vector within a real-time diagnostic time window: ;in, Indicates the first The node at the th Real-time node window state vector within a real-time diagnostic time window; Indicates the number of real-time partial discharge pulses; This indicates the maximum amplitude of the real-time partial discharge pulse; This represents the total energy of real-time partial discharge; This indicates the proportion of high-frequency energy in real-time partial discharge; This indicates the real-time effective value of the sheath current; This indicates the real-time rate of change of sheath current; This indicates the real-time harmonic distortion characteristics of the sheath current; This indicates the real-time load-corrected temperature rise. This indicates the real-time load-corrected rate of temperature rise. Based on the one-dimensional path coordinates of each node in the node set, the real-time node window state vectors of all nodes within the same real-time diagnostic time window are combined into a real-time state matrix. : ; in, Indicates the first The node at the th The real-time node window state vector within a real-time diagnostic time window; since each node is already arranged according to the one-dimensional path coordinate order, therefore At the same time, the spatial order of the nodes along the high-voltage cable was preserved; Call the node feature validity mask matrix For the real-time state matrix By applying validity constraints, the real-time valid state matrix is ​​obtained. : ; in, This indicates that corresponding elements are multiplied. Through this process, the features of nodes without corresponding sensors are masked. Based on the benchmark mean matrix of operating status and state benchmark fluctuation matrix Calculate the first The node The feature in the first Standardized state deviation within a real-time diagnostic time window and form the first State deviation matrix for each real-time diagnostic time window : ; According to the wave boundary matrix The system judges whether the real-time feature value exceeds the normal fluctuation range, and obtains the result. The node The feature in the first Boundary deviation value within a real-time diagnostic time window and form the first Boundary deviation matrix corresponding to each real-time diagnostic time window : ; Specifically, when the real-time feature value is within the corresponding normal fluctuation boundary, When the real-time characteristic value is below the lower boundary of normal fluctuation Or higher than the upper limit of normal fluctuations At that time, retain its standardized state deviation value. ; The boundary deviation matrices corresponding to multiple consecutive real-time diagnostic time windows are arranged in chronological order to form the input sequence of the CNN-LSTM. , deviate the boundary from the matrix Input the CNN module, perform convolution calculations on the boundary deviation features of each node along the node arrangement direction, and obtain the first... Spatial deviation characteristics corresponding to each real-time diagnostic time window : ; in, This represents the kernel parameters, and the kernel size is preferably... , Represents the boundary deviation matrix The number of features corresponding to each node, i.e., the node window state vector. The nine state features can extract the local spatial anomaly combination relationship between adjacent nodes near intermediate joints, sheath grounding points or cross-interconnection boxes. The number of convolution kernels is preferably 32, which is suitable for real-time fault diagnosis in online monitoring scenarios. This represents the convolution bias, and the initial value is preferably selected. And during model training, the parameters of the convolution kernel change. The simultaneous update can adaptively correct the overall bias of the convolution output, improving the ability to represent spatial deviation features; This represents the convolution operation; The nonlinear activation function is preferred, and the ReLU activation function can suppress the negative response to 0 and retain the positive anomalous response, so that the CNN module can better highlight the positive anomalous enhancement trend of features such as partial discharge, sheath current and temperature relative to the baseline state. Indicates the first The node at the th The spatial deviation feature vector within the first real-time diagnostic time window is used to characterize the first... Spatial combination relationship of anomalous features such as partial discharge, sheath current and temperature in local areas near each node; Using the node space correlation matrix The spatial deviation features are corrected to obtain the corrected spatial deviation features. : ; The spatial deviation features corresponding to multiple consecutive real-time diagnostic time windows are input into the LSTM module in chronological order to extract the evolution of boundary deviation features over time: ; in, Indicates the first The LSTM hidden state corresponding to each real-time diagnostic time window; Indicates the first The LSTM hidden states corresponding to each real-time diagnostic time window are initially set as zero vectors. Finally, the first The LSTM hidden states corresponding to each real-time diagnostic time window are used as the fault time-series features extracted by CNN-LSTM. It characterizes the degree of abnormality of the current operating state of the target high-voltage cable relative to the baseline normal state and its temporal variation pattern.

[0023] S4. Calculate the node anomaly intensity based on the boundary deviation matrix and generate spatial attention weights by combining spatial correlation. Weight the spatial deviation features to obtain spatial attention enhancement features. Further calculate the temporal attention weights and weight and fuse the continuous window features to form spatiotemporal attention features. Combine them with the fault time sequence features to generate comprehensive fault diagnosis features. Finally, determine the fault type, node location and probability and output the fault diagnosis results. According to the Boundary deviation matrix corresponding to each real-time diagnostic time window Calculate the first The node anomaly strength within this time window : ; in, This indicates the number of features in the node window state vector, including nine state features related to partial discharge, sheath current, and temperature. Combining the node space correlation matrix Calculate the first Spatial attention score of each node : ; in, Indicates the first The node at the th The intensity of node anomalies within a real-time diagnostic time window; Indicates the total number of nodes; Normalize the spatial attention scores of each node to obtain the first... The node at the th Spatial attention weights within a real-time diagnostic time window : ; in, Indicates the first Spatial attention score for each node; Indicates the first The spatial attention score corresponding to each node; the higher the degree of spatial anomaly and the stronger the anomaly correlation with surrounding nodes, the greater the spatial attention weight of the node. Determine the node number corresponding to the fault location based on spatial attention weights. : ; in, Indicates the first The spatial attention weights correspond to each node; the one-dimensional path has been sorted by coordinates, and each node corresponds to one-dimensional path coordinates. Therefore, based on the fault node number The corresponding one-dimensional path coordinates of the fault can be obtained. ,in, Indicates the first The fault location coordinates corresponding to each real-time diagnostic time window. Indicates the faulty node One-dimensional path coordinates; Based on spatial attention weights, for the th Corrected spatial deviation features corresponding to each real-time diagnostic time window Weighting is performed to obtain spatial attention enhancement features. : ; Based on hidden state Calculate attention score over time : ; in, Indicates the first The time attention score corresponding to each real-time diagnostic time window; Indicates the first The LSTM hidden state corresponding to each real-time diagnostic time window; The time attention weight parameter is preferably initialized using the Xavier initialization method. ,in, Indicates a uniform distribution; Represents the hidden state of LSTM The dimension enables different real-time diagnostic time windows to have a relatively balanced initial attention response in the early stages of training; This represents the temporal attention bias parameter, with an initial value of [value to be filled in]. And it is updated through backpropagation during model training, the update method is as follows: ,in, Cross-entropy loss function can be used. This represents the learning rate, with a preferred value of [value missing]. This can avoid artificially altering attention scores at different time windows during the early stages of training; Normalize the temporal attention score of the continuous real-time diagnostic time window to obtain the first... Time attention weights corresponding to each real-time diagnostic time window : ; in, Indicates the first The time attention score corresponding to each real-time diagnostic time window; Indicates the first The time attention score corresponding to each real-time diagnostic time window; Based on temporal attention weights, the spatial attention enhancement features corresponding to continuous real-time diagnostic time windows are weighted and fused to obtain spatiotemporal attention fusion features. : ; in, Indicates the first The time attention weights corresponding to each real-time diagnostic time window; Indicates the first Spatial attention enhancement features corresponding to each real-time diagnostic time window; Fusing spatiotemporal attention features With fault timing characteristics By fusion, comprehensive fault diagnosis features are obtained. : ; in, This represents the fault time-series features extracted by CNN-LSTM; Indicates feature concatenation operation; Predefined fault types ， in, Indicates a normal state. This indicates an abnormal partial discharge. This indicates an abnormal sheath current. This indicates an abnormal temperature. This indicates a composite anomaly (at least two of the following anomalies: partial discharge, sheath current, and temperature occur simultaneously). Based on comprehensive fault diagnosis features Calculate the fault type judgment result : ; in, Indicates the first A probability vector of fault types corresponding to each real-time diagnostic time window. The fault type judgment weight parameter is also initialized using the Xavier method. This represents the bias parameter for fault type determination, with an initial value of [value missing]. And update it based on the backpropagation results of the loss function during model training; This represents the normalized exponential function, used to map comprehensive fault diagnosis features to probabilities of different fault types; Based on the fault type probability vector Determine the fault type of the target high-voltage cable. : ; in, Indicates the first Fault type judgment results corresponding to each real-time diagnostic time window; This indicates the operation that takes the category with the highest probability. Finally, output the first... Fault diagnosis results corresponding to each real-time diagnostic time window : ; in, Indicates the result of the fault type determination; Indicates the fault node number; Represents the one-dimensional path coordinates corresponding to the fault location; Represents the probability vector of fault types; Indicates spatial attention weights; This represents the time attention weight.

[0024] In summary, this invention acquires benchmark operating data of high-voltage cables and constructs one-dimensional node path coordinates, node spatial correlation matrices, and propagation delay benchmark matrices along the cable laying direction. It further divides the benchmark sampling interval into time windows, extracting multi-source state features such as partial discharge, sheath current, and temperature to form a benchmark state matrix, a fluctuation boundary matrix, and a feature validity mask. In the real-time diagnosis stage, a real-time state matrix is ​​constructed according to the same node order and feature arrangement. Standardized deviation values ​​and boundary deviation matrices are calculated, and local spatial anomaly features are extracted using CNN and corrected using the spatial correlation matrix. Then, fault temporal evolution features are extracted using LSTM. Finally, based on spatial and temporal attention, anomaly nodes and key time windows are weighted and fused to generate comprehensive fault diagnosis features, outputting fault type, fault node, fault location, fault probability, and attention weights, thus achieving spatiotemporal collaborative diagnosis and interpretable localization of high-voltage cable faults. It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A high-voltage cable fault diagnosis method based on spatiotemporal attention, characterized in that, include: Obtain the baseline operating data of the target high-voltage cable, select the seven-day continuous normal operating history interval of the target high-voltage cable as the baseline sampling interval, calculate the cumulative path length of each node along the cable to form a one-dimensional path coordinate and construct a node set; sort by path coordinate, calculate the path distance between adjacent nodes, further construct the node spatial correlation matrix and calculate the propagation delay between any nodes to form a propagation delay baseline matrix; The reference sampling interval is divided into time windows, and the characteristics of partial discharge, sheath current and temperature are statistically analyzed and combined to form a node window state vector. Invalid features are then removed based on the validity mask. Calculate the state mean, fluctuation, and upper and lower boundaries, and form a baseline matrix; Acquire real-time monitoring data and construct a state matrix according to time window and node order, apply validity mask and calculate standardized deviation value and boundary deviation matrix; The boundary deviation matrix of the continuous window is input into the CNN to extract spatial deviation features, and then corrected with the spatial correlation matrix. Temporal features are extracted through LSTM to form fault temporal features. The node anomaly strength is calculated based on the boundary deviation matrix and spatial attention weights are generated by combining spatial correlation. Spatial attention enhancement features are obtained by weighting the spatial deviation features. The temporal attention weights are calculated and the continuous window features are weighted and fused to form spatiotemporal attention features. These features are then combined with the fault time sequence features to generate comprehensive fault diagnosis features. Finally, the fault type, node location and probability are determined and the fault diagnosis results are output.

2. The high-voltage cable fault diagnosis method based on spatiotemporal attention as described in claim 1, characterized in that: The process of acquiring the target high-voltage cable's baseline operating data involves selecting a seven-day continuous operating history interval of the target high-voltage cable as the baseline sampling interval, calculating the cumulative path length along the cable for each node to form one-dimensional path coordinates, and constructing a node set. ; Using the cable's starting and ending points as the origin of the coordinate system, for any node on the target high-voltage cable... Calculate the distance from the starting point to this node along the actual cable laying direction. The cumulative path length is calculated and used as the one-dimensional path coordinate of the node. ; By unifying the cable's starting and ending points, intermediate joints, sheath grounding points, cross-connection boxes, and sensor installation points as nodes, a node set for the target high-voltage cable is generated. .

3. The high-voltage cable fault diagnosis method based on spatiotemporal attention as described in claim 2, characterized in that: The process involves sorting nodes by path coordinates, calculating the path distance between adjacent nodes, constructing a node spatial correlation matrix, and calculating the propagation delay between any two nodes to form a propagation delay baseline matrix based on the node set. Sort the nodes in ascending order of their one-dimensional path coordinates, and calculate the path distance between adjacent nodes. ; Calculate the first using adjacent priority space association. The node and the first Spatial correlation value between nodes ; Based on spatial correlation value Constructing the node space association matrix ; Calculate the propagation delay between any two nodes. And form a propagation delay reference matrix. .

4. The high-voltage cable fault diagnosis method based on spatiotemporal attention as described in claim 3, characterized in that: The process of dividing the reference sampling interval into time windows, statistically analyzing partial discharge, sheath current, and temperature characteristics, and combining them to form a node window state vector refers to dividing the reference sampling interval into multiple equal-length windows. ; In each time window Internally, a partial discharge sensor is used to perform pulse detection and feature statistics on the original partial discharge waveform to obtain the number of partial discharge pulses. and the maximum amplitude of partial discharge pulse And calculate the total energy of partial discharge. and the proportion of high-frequency energy , and combine to form the first The node at the th Partial discharge reference feature vector within a time window ; In each time window Inside, obtain the first Time window The sheath current sampling sequence is obtained by the sheath current sensor. Calculate the effective value of the sheath current. Sheath current change rate Harmonic distortion characteristics And combine them into a base characteristic vector of sheath current. ; In each time window Inside, the temperature value of the connector is acquired by the temperature sensor. Ambient temperature value and operating load current value Calculate the relative ambient temperature rise Further, the load-corrected temperature rise is calculated based on the relative ambient temperature rise and the operating load current. Then, based on the load-corrected temperature rise in the current time window and the load-corrected temperature rise in the previous time window, calculate the load-corrected temperature rise change rate. The calculated load-corrected temperature rise and the load-corrected temperature rise rate of change are combined into a temperature reference eigenvector. ; Based on partial discharge reference feature vector Sheath current reference eigenvector and temperature reference eigenvector All features within the same node and the same time window are combined in a fixed order to form a node window state vector. .

5. The high-voltage cable fault diagnosis method based on spatiotemporal attention as described in claim 4, characterized in that: The process of eliminating invalid features based on validity masks, calculating the state mean, fluctuation, and upper and lower boundaries, and forming a baseline matrix refers to establishing a node feature validity mask matrix. ; Calculate the first The node State mean of each feature and standard deviation of fluctuation Forming a baseline mean matrix of operating states and state benchmark fluctuation matrix ; Based on state mean and standard deviation of fluctuation Calculate the upper and lower boundaries of the fluctuation and And combined into a wave boundary matrix .

6. The high-voltage cable fault diagnosis method based on spatiotemporal attention as described in claim 5, characterized in that: The acquisition of real-time monitoring data and the construction of a state matrix according to time windows and node order refer to the acquisition of real-time monitoring data of the target high-voltage cable during its current operation. The real-time monitoring data is processed to obtain the first... The node at the th Real-time node window state vector within a real-time diagnostic time window ; Based on the one-dimensional path coordinates of each node in the node set, the real-time node window state vectors of all nodes within the same real-time diagnostic time window are combined into a real-time state matrix. .

7. The high-voltage cable fault diagnosis method based on spatiotemporal attention as described in claim 6, characterized in that: The application of validity mask and calculation of standardized deviation value and boundary deviation matrix refers to calling the node feature validity mask matrix. For the real-time state matrix By applying validity constraints, the real-time valid state matrix is ​​obtained. ; Based on the benchmark mean matrix of operating status and state benchmark fluctuation matrix Calculate the first The node The feature in the first Standardized state deviation within a real-time diagnostic time window and form the first State deviation matrix for each real-time diagnostic time window ; According to the wave boundary matrix The system judges whether the real-time feature value exceeds the normal fluctuation range, and obtains the result. The node The feature in the first Boundary deviation value within a real-time diagnostic time window and form the first Boundary deviation matrix corresponding to each real-time diagnostic time window .

8. The high-voltage cable fault diagnosis method based on spatiotemporal attention as described in claim 7, characterized in that: The process involves inputting the boundary deviation matrix of the continuous window into a CNN to extract spatial deviation features, correcting them with a spatial correlation matrix, and then extracting temporal features using an LSTM to form fault temporal features. This process involves using the boundary deviation matrix... Input the CNN module, perform convolution calculations on the boundary deviation features of each node along the node arrangement direction, and obtain the first... Spatial deviation characteristics corresponding to each real-time diagnostic time window ; Using the node space correlation matrix The spatial deviation features are corrected to obtain the corrected spatial deviation features. ; Spatial deviation features corresponding to multiple consecutive real-time diagnostic time windows are input into the LSTM module in chronological order to extract the evolution of boundary deviation features over time. ; The first The LSTM hidden states corresponding to each real-time diagnosis time window are used as the fault time-series features extracted by CNN-LSTM. .

9. The high-voltage cable fault diagnosis method based on spatiotemporal attention as described in claim 8, characterized in that: The method involves calculating node anomaly strength based on the boundary deviation matrix and generating spatial attention weights by combining spatial correlation. This weighted spatial deviation features are then used to obtain spatial attention enhancement features. Boundary deviation matrix corresponding to each real-time diagnostic time window Calculate the first The node anomaly strength within this time window ; Combined with the node space correlation matrix Calculate the first Spatial attention score of each node ; Normalize the spatial attention scores of each node to obtain the first... The node at the th Spatial attention weights within a real-time diagnostic time window ; Determine the node number corresponding to the fault location based on spatial attention weights. ; Based on spatial attention weights, for the th Corrected spatial deviation features corresponding to each real-time diagnostic time window Weighting is performed to obtain spatial attention enhancement features. .

10. The high-voltage cable fault diagnosis method based on spatiotemporal attention as described in claim 9, characterized in that: The calculation of temporal attention weights and the weighted fusion of continuous window features form spatiotemporal attention features, which are then combined with fault temporal features to generate comprehensive fault diagnosis features. Finally, the fault type, node location, and probability are determined, and the fault diagnosis result is output based on the hidden state. Calculate attention score over time ; Normalize the temporal attention score of the continuous real-time diagnostic time window to obtain the first... Time attention weights corresponding to each real-time diagnostic time window ; Based on temporal attention weights, the spatial attention enhancement features corresponding to continuous real-time diagnostic time windows are weighted and fused to obtain spatiotemporal attention fusion features. ; Fusing spatiotemporal attention features With fault timing characteristics By fusion, comprehensive fault diagnosis features are obtained. ; Based on comprehensive fault diagnosis features Calculate the fault type judgment result ; Based on the fault type probability vector Determine the fault type of the target high-voltage cable. ; Output the first Fault diagnosis results corresponding to each real-time diagnostic time window .

Images