A power transmission equipment remote diagnosis method

Abstract

The present application belongs to the technical field of electrical measurement, and relates to a kind of power transmission equipment remote diagnosis method.The method collects voltage current phasor data and partial discharge pulse signal in each monitoring node, and establishes space-time correlation matrix according to physical topology connection relationship by edge computing gateway;Extract the time difference and amplitude attenuation ratio of pulse arriving at each node, and perform time alignment under topological constraint in the matrix combined with phase mutation characteristics;Calculate the propagation time delay gradient and energy attenuation gradient of discharge pulse between adjacent nodes, input them into the graph neural network model, the model takes physical topology path as edge feature, and multi-source electrical quantity as node feature, and outputs the forward propagation path of fault signal and discharge source node coordinate.The present application excludes false pulse interference, restores the propagation path of discharge signal in real physical topology, and solves the positioning failure problem caused by long-distance transmission line cumulative time error and topological deviation.

Description

Technical Field

[0001] This invention belongs to the field of electrical measurement technology and relates to a method for remote diagnosis of power transmission equipment. Background Technology

[0002] In the field of remote diagnostics for power transmission equipment, conventional technical solutions typically involve independently deploying partial discharge sensors at each monitoring node for data acquisition. After receiving the discharge pulse signals uploaded by each node, the master station sorts and compares them based on their absolute arrival timestamps. When it is determined that two or more nodes have captured discharge pulses from the same source, the location of the discharge source is directly calculated based on the straight-line geographical distance between these isolated nodes and the signal arrival time difference. This implementation relies on local threshold judgments at each independent measuring point. The processing logic at the master station is limited to isolated comparisons of discrete node data and simple geometric cross-location; the data interaction between nodes is not related to the underlying physical topology of the power transmission network.

[0003] The core flaw of the aforementioned conventional technical solutions lies in the fact that the partial discharge signals collected by isolated nodes are highly susceptible to interference from the complex electromagnetic environment at the site, generating false pulses. The master station cannot identify such false pulses by sorting and comparing them solely based on absolute timestamps. In long-distance transmission lines, the straight-line distance between isolated nodes deviates from the actual line path, and the absolute time synchronization has accumulated errors during long-distance transmission. This causes the calculation results based on simple geometric positioning of discrete nodes to deviate from the actual discharge source location, making it impossible to track and accurately locate the true propagation path of the discharge pulse under complex transmission topology structures. Summary of the Invention

[0004] The purpose of this invention is to provide a remote diagnostic method for power transmission equipment, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A remote diagnostic method for power transmission equipment, the method comprising: deploying phasor measurement units and partial discharge sensors at each monitoring node of the power transmission line to collect voltage and current phasor data and partial discharge pulse signals; The collected multi-source electrical quantity data is injected into the edge computing gateway, and the edge computing gateway establishes a spatiotemporal correlation matrix based on the physical topology connection relationship of the power transmission network; For the partial discharge pulse signal, the time difference and amplitude attenuation ratio of the pulse arriving at each monitoring node are extracted. Combined with the voltage and current phase change characteristics provided by the phasor measurement unit, timing alignment under topological constraints is performed in the spatiotemporal correlation matrix. The propagation delay gradient and energy attenuation gradient of the discharge pulse between adjacent nodes are calculated by the time alignment. The propagation delay gradient and energy attenuation gradient are input into a pre-constructed graph neural network model. The graph neural network model uses the physical topology path as the edge feature and the multi-source electrical quantity as the node feature. It outputs the forward propagation path of the fault signal in the topology network and the coordinates of the discharge source node, thus completing the remote location and type identification of the power transmission equipment fault.

[0006] Preferably, the edge computing gateway establishes a spatiotemporal correlation matrix based on the physical topology of the power transmission network, including: the edge computing gateway obtains the geographical coordinates and line connection relationships of each monitoring node in the power transmission network, and maps each monitoring node to a row vector and a column vector in the spatiotemporal correlation matrix; For two monitoring nodes that have a direct physical line connection, the corresponding matrix element in the spatiotemporal correlation matrix is ​​set to the physical length of the line; for two monitoring nodes that do not have a direct physical line connection, the corresponding matrix element is set to infinity. Phasor measurement units of each monitoring node are injected into the diagonal position of the spatiotemporal correlation matrix to collect timestamps, forming an initial spatiotemporal correlation matrix containing spatial topological distance and time reference.

[0007] Preferably, the extraction of the time difference and amplitude attenuation ratio of the pulses arriving at each monitoring node includes: performing high-pass filtering on the original pulse signal collected by the partial discharge sensor to filter out the power frequency carrier signal, and performing Hilbert transform on the filtered signal to extract the envelope. A dynamic floating threshold is set on the envelope, and pulse peak points exceeding the dynamic floating threshold are captured. The timestamp corresponding to the pulse peak point is used as the pulse arrival time. The amplitude of the pulse peak point is read, the time difference is obtained by subtracting the pulse arrival times captured by different monitoring nodes pairwise, and the amplitude attenuation ratio is obtained by calculating the ratio of the amplitudes captured by different monitoring nodes.

[0008] Preferably, performing temporal alignment under topological constraints in the spatiotemporal correlation matrix includes: extracting the shortest path topological node sequence between any two monitoring nodes from the spatiotemporal correlation matrix; The time difference is matched and verified with the physical length of the line corresponding to the shortest path topology node sequence. When the equivalent wave speed corresponding to the time difference is within the set wave speed range, the time difference is determined to be valid, and the valid time difference is aligned to the shortest path topology node sequence. When the equivalent wave velocity is outside the wave velocity range, the time difference is determined to be an invalid time difference caused by noise interference, and the invalid time difference is removed from the spatiotemporal correlation matrix.

[0009] Preferably, the calculation of the propagation delay gradient and energy attenuation gradient of the discharge pulse between adjacent nodes includes: on the shortest path topology node sequence with completed time alignment, along the propagation direction of the discharge pulse, calculating the local propagation delay between the i-th node and the (i+1)-th node, and dividing the local propagation delay by the physical distance between the two nodes to obtain the propagation delay gradient; Calculate the local amplitude attenuation between the i-th node and the (i+1)-th node, and divide the local amplitude attenuation by the physical distance to obtain the energy attenuation gradient; The propagation delay gradients and energy decay gradients on the shortest path topology node sequence are combined into a gradient feature vector.

[0010] Preferably, the graph neural network model uses physical topology paths as edge features and the multi-source electrical quantities as node features, including: concatenating the voltage and current phasor data of the monitoring nodes with the frequency domain features of the partial discharge pulse signal to form the node feature vector; The line impedance parameters, line type identifier, propagation delay gradient, and energy attenuation gradient of the physical topology path are concatenated to form the edge feature vector; In the graph convolutional layer of the graph neural network model, attention weight coefficients for information transmission between adjacent nodes are calculated based on the edge feature vectors. The node feature vectors are then weighted and aggregated using these attention weight coefficients to generate higher-order node representation vectors that contain topological spatial relationships.

[0011] Preferably, after forming the initial spatiotemporal correlation matrix containing spatial topological distance and time reference, the method further includes: the edge computing gateway monitoring the status change signals of the power transmission network circuit breakers and disconnectors in real time; When the state change signal is received, the branch with the topology change is identified, the matrix element of the corresponding branch in the initial spatiotemporal correlation matrix is ​​updated from the physical length of the line to infinity, and the timestamp synchronization association of the corresponding branch is disconnected. When a closing signal indicating the end of maintenance is received, the matrix elements of the corresponding branch in the initial spatiotemporal correlation matrix are restored from infinity to the updated physical length of the line, thereby generating a dynamic spatiotemporal correlation matrix.

[0012] Preferably, after capturing the pulse peak point exceeding the dynamic floating threshold, the method further includes: extracting the voltage and current phase change characteristics provided by the phasor measurement unit within the preset time window where the pulse peak point is located; Calculate the change energy value of the voltage and current phase change characteristics, and determine whether the change energy value reaches the set phase change threshold value; If the mutation energy value does not reach the phase mutation threshold value, the pulse peak point is determined to be an interference pulse generated by spatial electromagnetic radiation, and the time difference and amplitude attenuation ratio corresponding to the interference pulse are marked as abnormal data and deleted. If the mutation energy value reaches the phase mutation threshold value, the corresponding feature data is retained.

[0013] Preferably, combining all the propagation delay gradients and the energy decay gradients on the shortest path topology node sequence into a gradient feature vector includes: detecting whether there are branch nodes in the shortest path topology node sequence; When the branch node exists, identify the multiple discharge pulse branch paths flowing through the branch node, and calculate the branch propagation delay gradient and branch energy attenuation gradient corresponding to each discharge pulse branch path. A reflection attenuation coefficient is introduced at the point of impedance discontinuity to correct the distortion of the branch propagation delay gradient and the branch energy attenuation gradient at the point of reflection. The corrected gradient data is then concatenated into the gradient feature vector.

[0014] Preferably, after generating the higher-order node representation vector containing topological spatial association, the method further includes: inputting the higher-order node representation vector into the fully connected output layer of the graph neural network model, wherein the fully connected output layer outputs the discharge probability matrix and coordinate offset. Obtain the geographic coordinates of each node in the shortest path topology node sequence, and superimpose the coordinate offsets onto the starting node with the highest geographic coordinates to generate the initial discharge source coordinates; Determine whether the initial discharge source coordinates fall on the geometric line segment formed by the physical topology path. If not, orthogonally project the initial discharge source coordinates onto the nearest geometric line segment, and use the coordinates of the projection point as the final discharge source node coordinates.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention establishes a spatiotemporal correlation matrix based on physical topology connections using an edge computing gateway. It aligns the time difference and amplitude attenuation ratio of partial discharge pulse signals within the topology path for time sequence adjustment, calculates the propagation delay gradient and energy attenuation gradient between adjacent nodes, and combines this with a graph neural network model, using the physical topology path as edge features and multi-source electrical quantities as node features for joint solution. This scheme transforms the electrical quantity measurement of isolated nodes into spatiotemporal evolution feature extraction on the entire network topology path, eliminating interference from spurious pulses, restoring the forward propagation path of the discharge signal in the real physical topology, and solving the positioning failure problem caused by accumulated time errors and topology path deviations in long-distance transmission lines.

[0016] 2. This invention improves the reliability of input data by introducing a cross-validation mechanism of dynamic floating threshold and phase change energy value to eliminate interference pulses generated by spatial electromagnetic radiation; it eliminates invalid time differences through equivalent wave velocity verification and introduces the reflection and attenuation coefficient of wave impedance discontinuity points to correct gradient characteristics when branch nodes exist, thus eliminating the influence of network branches on signal propagation characteristics; it updates the spatiotemporal correlation matrix in real time by monitoring the circuit breaker status change signal and corrects the discharge source coordinates by using orthogonal projection, making the diagnostic method adaptable to the dynamic changes of the transmission network topology and ensuring the close fit between the fault location results and the actual physical line. Attached Figure Description

[0017] Figure 1 This is a flowchart of the overall method for remote diagnosis of power transmission equipment according to the present invention; Figure 2 This is a flowchart illustrating the spatiotemporal correlation matrix establishment and dynamic update process of the present invention. Figure 3 This is a flowchart of the pulse feature extraction and interference removal process of the present invention; Figure 4 This is a flowchart of the timing alignment under topological constraints of the present invention; Figure 5 This is a flowchart of the gradient calculation and distortion correction process of the present invention; Figure 6 This is a flowchart of the graph neural network model processing and coordinate projection of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Please refer to Figure 1This embodiment provides a remote diagnostic method for power transmission equipment. Phasor measurement units and partial discharge sensors are deployed at various monitoring nodes of the transmission line to collect voltage and current phasor data and partial discharge pulse signals. The acquisition process of the phasor measurement units and partial discharge sensors is time-synchronized, with the synchronization reference coming from a unified satellite navigation system time source. The acquired multi-source electrical quantity data is marked with a unified format timestamp. The multi-source electrical quantity data includes voltage and current phasor data output by the phasor measurement units, and time-domain pulse waveform data output by the partial discharge sensors. The collected multi-source electrical quantity data is injected into an edge computing gateway. The edge computing gateway is deployed at the substation bay layer or key nodes of the transmission network, receiving multi-source electrical quantity data uploaded by all monitoring nodes within its jurisdiction. The received data is cached and preprocessed, including data format unification, outlier removal, and timestamp alignment verification. The edge computing gateway establishes a spatiotemporal correlation matrix based on the physical topology of the power transmission network. The dimension of the spatiotemporal correlation matrix is ​​consistent with the number of monitoring nodes in the power transmission network. Each monitoring node corresponds to a set of row and column vectors in the spatiotemporal correlation matrix. The values ​​of the matrix elements are related to the physical connection status between the corresponding two monitoring nodes. The diagonal elements of the matrix carry the data acquisition timestamp information of the corresponding monitoring node. The mathematical expression of the spatiotemporal correlation matrix is: , Where N represents the total number of monitoring nodes deployed in the transmission network, and the row and column indices of the matrix correspond to the unique numbers of the monitoring nodes, thus mapping the monitoring nodes to the matrix vector space. For the partial discharge pulse signals collected by the partial discharge sensor, the time difference and amplitude attenuation ratio of the pulses arriving at each monitoring node are extracted. The time difference is the difference in timestamps of the same-source discharge pulses arriving at different monitoring nodes, and the amplitude attenuation ratio is the ratio of the peak amplitudes of the same-source discharge pulses at different monitoring nodes. Simultaneously, the voltage and current phase abrupt change features provided by the phasor measurement unit are extracted. These features represent the change in phase angle between the voltage and current phasors before and after the discharge pulse occurs. Combining the extracted time difference, amplitude attenuation ratio, and phase abrupt change features, time-series alignment under topological constraints is performed in the spatiotemporal correlation matrix. The time-series alignment process uses the physical topology path of the transmission network as a constraint, excluding time difference data that does not match the topology path, and mapping valid time difference data to the corresponding topology path node sequence. During the time-series alignment process, data validity is determined through equivalent wave velocity verification. The equivalent wave velocity is calculated as follows: , in, This represents the equivalent propagation velocity of the discharge pulse between the two monitoring nodes. This represents the total length of the shortest topological path between the two monitoring nodes. The arrival time difference of discharge pulses from the same source captured by two monitoring nodes is used to verify the equivalent wave velocity within a set wave velocity range, and effective time difference data that conforms to the electromagnetic wave propagation law of transmission lines are selected. Using the time-aligned effective data, the propagation delay gradient and energy attenuation gradient of the discharge pulse between adjacent nodes are calculated. The propagation delay gradient is the ratio of the propagation delay of the discharge pulse between adjacent nodes to the corresponding physical length of the line, characterizing the change in propagation delay of the discharge pulse per unit length of line. The calculation method for the propagation delay gradient is as follows: , in, Let be the propagation delay gradient between the i-th node and the (i+1)-th node. Let be the local propagation delay of the discharge pulse between the i-th node and the (i+1)-th node. Let be the physical length of the line between the i-th node and the (i+1)-th node. The energy attenuation gradient is the ratio of the amplitude attenuation of the discharge pulse between adjacent nodes to the corresponding physical length of the line, characterizing the change in energy attenuation of the discharge pulse per unit length of line. The energy attenuation gradient is calculated as follows: , in, Let be the energy decay gradient between the i-th node and the (i+1)-th node. This represents the local amplitude attenuation of the discharge pulse between the i-th and (i+1)-th nodes. The propagation delay gradients and energy attenuation gradients along the shortest path topology node sequence are combined into a gradient feature vector, which is input into a pre-constructed graph neural network model. This model uses the physical topology path of the transmission network as edge features and the multi-source electrical quantity data collected by monitoring nodes as node features. It performs joint learning and inference on the input gradient data, topology features, and node features. The graph neural network model adopts a graph convolutional architecture, using the physical topology of the transmission network as the graph structure. The node feature matrix is ​​generated by concatenating the multi-source electrical quantity features of all monitoring nodes, and the edge feature matrix is ​​generated by concatenating the physical parameters of the topology path and gradient features. In the graph convolutional layer, the features of adjacent nodes are aggregated to generate node representations containing topological spatial correlation information. Finally, the model outputs the forward propagation path of the fault signal in the topology network and the coordinates of the discharge source node, while simultaneously identifying the type of fault in the transmission equipment.

[0020] In this embodiment, to clarify the mapping relationship between monitoring nodes and the spatiotemporal correlation matrix, as well as the basic characteristics of multi-source electrical quantity data, the following table is set up: Table 1. Mapping of Transmission Line Monitoring Nodes and Composition of Multi-Source Electrical Quantity Characteristics 001 0 16 2 1 002 1 16 3 1 003 2 16 2 1 004 3 16 2 1 005 4 16 3 1 In Table 1, the unique node number is consistent with the resource numbering system of the transmission network. The matrix row / column index realizes a one-to-one mapping between the monitoring node and the spatiotemporal correlation matrix. The node feature composition dimension corresponds to the feature dimension of the multi-source electrical quantity data. The number of associated topology branches corresponds to the connectivity of the monitoring node in the physical topology. The timestamp synchronization accuracy level is used to identify the time synchronization reliability of the node's data collection. This table provides a basic mapping basis for the construction of the spatiotemporal correlation matrix, ensuring that the correspondence between matrix elements and physical topology nodes is accurate.

[0021] The technical solution provided in this embodiment constructs a spatiotemporal correlation matrix that integrates physical topology and time reference through an edge computing gateway. It aligns the timing and amplitude characteristics of partial discharge pulses under topological constraints, extracts the propagation gradient characteristics of discharge pulses on the topological path, and combines them with a graph neural network model based on physical topology for inference. This transforms the electrical quantity measurement of isolated nodes into spatiotemporal evolution feature extraction on the entire network topology path, restores the forward propagation path of the discharge signal in the real physical topology, and realizes remote location and type identification of power transmission equipment faults.

[0022] In a preferred embodiment, reference Figure 2 The edge computing gateway acquires the geographical coordinates and line connection relationships of each monitoring node in the power transmission network. The geographical coordinates adopt a unified Cartesian coordinate system, and the line connection relationships are derived from the public information model data of the power transmission network, including the conduction status and physical length parameters of the branches. Each monitoring node is mapped to a row vector and a column vector in the spatiotemporal correlation matrix. Both the row index and column index are encoded using the unique identifier of the monitoring node to ensure that the row vector and column vector corresponding to the same monitoring node have a one-to-one correspondence in the matrix. For two monitoring nodes with a direct physical line connection, the corresponding matrix element in the spatiotemporal correlation matrix is ​​set to the physical length of the line. For two monitoring nodes without a direct physical line connection, the corresponding matrix element is set to infinity. The phasor measurement units of each monitoring node are injected into the diagonal position of the spatiotemporal correlation matrix to collect timestamps, forming an initial spatiotemporal correlation matrix containing spatial topological distance and time reference. The complete definition of the initial spatiotemporal correlation matrix is: , in, The initial spatiotemporal correlation matrix, The physical length of the line between node i and node j, in meters. The timestamp, in microseconds, is used to collect data from the phasor measurement unit of node i. The timestamp is synchronized using a unified satellite navigation system second pulse, and the values ​​of elements at symmetrical positions in the matrix remain consistent, conforming to the undirected graph characteristics of the transmission line topology. After the initial spatiotemporal correlation matrix is ​​constructed, the edge computing gateway monitors the status change signals of circuit breakers and disconnectors in the transmission network in real time. These status change signals include the unique identifier of the switching equipment, its branch number, the time of the change, and its operating status after the change. The edge computing gateway parses and verifies the received status change signals to confirm their validity and the topology region to which they belong. When a status change signal is received and the post-change status is open, the branch with the topology change is identified, the monitoring node numbers at both ends of the branch are located, and the matrix elements of the corresponding branch in the initial spatiotemporal correlation matrix are updated from the physical line length to infinity. Simultaneously, the timestamp synchronization association of the corresponding branch is disconnected. In subsequent timing alignment and path calculation processes, this branch is marked as an invalid path and excluded from the shortest path search. When a closing signal indicating the end of maintenance is received, the updated physical length parameter of the line for that branch is obtained. The matrix elements of the corresponding branch in the initial spatiotemporal correlation matrix are restored from infinity to the updated physical length of the line. Simultaneously, the timestamp synchronization association of the corresponding branch is restored, generating a dynamic spatiotemporal correlation matrix. The definition of the dynamic spatiotemporal correlation matrix is ​​as follows: ,

[0023] in, It is a dynamic spatiotemporal correlation matrix. The updated physical length of branch ij is given. The matrix elements are updated in real time as the switching state of the transmission network changes. The updated matrix is ​​synchronously applied to subsequent time alignment, path calculation and feature extraction processes.

[0024] In this embodiment, to clarify the correspondence between the state of the transmission network topology branches and the elements of the spatiotemporal correlation matrix, and to standardize the execution logic of matrix updates, the following table is set: Table 2 Mapping Rules for Elements of the Spatiotemporal Correlation Matrix of Transmission Network Topology Branch States

[0025] In Table 2, the branch numbering is consistent with the primary equipment numbering system of the transmission network. The node indices at both ends of the branch correspond to the row / column indices of the spatiotemporal correlation matrix. The switch operation status comes from the real-time acquired change signals. The matrix element values ​​are strictly executed according to the mapping rules. The timestamp synchronization correlation status corresponds to the timestamp alignment logic of the nodes at both ends of the branch. The path validity flag is used to identify whether the branch is included in the shortest path search range. This table provides a clear execution specification for the dynamic updating of the spatiotemporal correlation matrix, ensuring that the matrix matches the actual operating topology of the transmission network in real time.

[0026] The technical solution provided in this embodiment clarifies the mapping rules between monitoring nodes and the spatiotemporal correlation matrix, constructs an initial spatiotemporal correlation matrix that integrates spatial topological distance and time reference, and updates matrix elements and path validity in real time based on the state change signals of switching equipment. This enables the diagnostic method to adapt to the dynamic changes in the transmission network topology, eliminates the impact of topological changes on timing alignment and path calculation, and ensures the accuracy of basic data for subsequent feature extraction and model inference processes.

[0027] In another preferred embodiment, reference Figures 3 to 5 The raw pulse signal acquired by the partial discharge sensor is subjected to high-pass filtering to remove the power frequency carrier signal and low-frequency interference components, retaining the high-frequency signal components corresponding to the partial discharge pulse. The envelope of the filtered signal is then extracted using a Hilbert transform. The calculation method for the Hilbert transform is as follows: , in, This is the original pulse signal after high-pass filtering. For Hilbert transform operators, The Hilbert transform of the original pulse signal is given. Based on the transform result, an analytic signal is constructed. The expression of the analytic signal is: , in, To analyze the signal, The imaginary unit is used; the modulus of the analytic signal is the envelope of the original pulse signal. The envelope is calculated as follows: , in, The envelope of the pulse signal completely characterizes the amplitude variation of the pulse signal over time, eliminating the influence of carrier phase on peak detection. A dynamic floating threshold is set on the envelope, which is calculated in real time based on the background noise statistical characteristics of the envelope. The calculation method is as follows: , in, Let be the dynamic floating threshold at time t. Let be the mean of the envelope within the preset time window at time t. To correspond to the standard deviation of the envelope within the time window, the preset time window is updated using a sliding window method. The window length and sliding step size are set according to the frequency characteristics of the pulse signal. Pulse peaks exceeding a dynamic floating threshold on the envelope are captured. The timestamp corresponding to the pulse peak is used as the pulse arrival time. The amplitude corresponding to the pulse peak is read. The arrival times of the same source pulses captured by different monitoring nodes are subtracted pairwise to obtain the time difference. The amplitude attenuation ratio is obtained by calculating the ratio of the amplitudes of the same source pulses captured by different monitoring nodes.

[0028] After capturing pulse peaks exceeding the dynamic floating threshold, the voltage and current phase abrupt change characteristics provided by the phasor measurement unit within the preset time window containing the pulse peak are extracted. These characteristics represent the changes in voltage and current phase angles within a preset time period before and after the pulse peak. The abrupt change energy value of these characteristics is calculated; it is the sum of squares of the phase abrupt change characteristics within the preset time window. The calculation method is as follows: , in, This represents the energy value of the phase transition. for The voltage and current phase abrupt change characteristics at any given time, where K is the number of phasor data points within a preset time window. It is determined whether the abrupt change energy value reaches the set phase abrupt change threshold. If the abrupt change energy value does not reach the phase abrupt change threshold, the pulse peak point is determined to be an interference pulse generated by spatial electromagnetic radiation, and the time difference and amplitude attenuation ratio corresponding to the interference pulse are marked as abnormal data and deleted. If the abrupt change energy value reaches the phase abrupt change threshold, the corresponding characteristic data is retained, and the process proceeds to the subsequent timing alignment process.

[0029] When performing time-series alignment under topological constraints in the spatiotemporal correlation matrix, the shortest path topological node sequence between any two monitoring nodes is extracted from the spatiotemporal correlation matrix. The shortest path topological node sequence is obtained using a path search algorithm based on the line length in the spatiotemporal correlation matrix as the weight. The time difference between the two monitoring nodes is matched and verified with the total physical length of the line corresponding to the shortest path topological node sequence. The equivalent wave velocity corresponding to the time difference is calculated. When the equivalent wave velocity is within the set wave velocity range, the time difference is determined to be valid, and the valid time difference is aligned to the shortest path topological node sequence. When the equivalent wave velocity is outside the wave velocity range, the time difference is determined to be an invalid time difference caused by noise interference, and the invalid time difference is removed from the spatiotemporal correlation matrix and not included in the subsequent gradient calculation process.

[0030] On the shortest path topology node sequence with completed timing alignment, along the propagation direction of the discharge pulse, the local propagation delay between the i-th node and the (i+1)-th node is calculated. Dividing the local propagation delay by the physical distance between the two nodes yields the propagation delay gradient. The local amplitude attenuation between the i-th node and the (i+1)-th node is also calculated, and divided by the physical distance, yields the energy attenuation gradient. All propagation delay gradients and energy attenuation gradients on the shortest path topology node sequence are combined into a gradient feature vector. During the combination of the gradient feature vector, it is detected whether there are branch nodes in the shortest path topology node sequence. Branch nodes are nodes with a connectivity greater than 2 in the topology network, corresponding to T-junction nodes in the transmission network. When branch nodes exist, the branch paths of multiple discharge pulses flowing through the branch nodes are identified, and the branch propagation delay gradient and branch energy attenuation gradient corresponding to each discharge pulse branch path are calculated respectively. A catadioptric attenuation coefficient is introduced at the impedance discontinuity point to correct the distortion of the branch propagation delay gradient and branch energy attenuation gradient where catadioptric reflection occurs. The catadioptric attenuation coefficient is calculated based on the impedance matching relationship at the branch node. The calculation methods for the refractive index and reflection coefficient are as follows: , Where T is the refractive index and R is the reflection index. The line impedance is the incident wave path. The equivalent load wave impedance at the branch node is the parallel value of the wave impedances of all branch paths. The catadioptric attenuation coefficient is the magnitude of the refractive index. The amplitude attenuation of the branch path is corrected based on the catadioptric attenuation coefficient. The corrected local amplitude attenuation is the original amplitude attenuation divided by the catadioptric attenuation coefficient. The corresponding energy attenuation gradient is corrected synchronously. The propagation delay gradient is not affected by catadioptric reflection and does not need to be corrected. The corrected gradient data is concatenated into the gradient feature vector to complete the complete construction of the gradient feature.

[0031] In this embodiment, to standardize the feature extraction process and data validity verification logic of partial discharge pulse signals and clarify the processing rules for different features, the following table is set: Table 3. Rules for Feature Extraction and Validity Verification of Partial Discharge Pulse Signals

[0032] Table 3 shows the feature types corresponding to the core feature dimensions extracted during pulse signal processing, the extraction methods clarify the specific calculation and extraction process for each feature, the validity verification conditions specify the screening criteria for feature data, and the data processing methods clarify the subsequent processing logic for valid and invalid data. This table provides a standardized execution process for pulse feature extraction and data screening, ensuring that the feature data input to subsequent processing steps has high reliability and high signal-to-noise ratio.

[0033] The technical solution provided in this embodiment extracts the envelope of the pulse signal through Hilbert transform, accurately captures the pulse peak point by combining a dynamic floating threshold, eliminates interference pulses generated by spatial electromagnetic radiation by cross-validation of phase change energy value and pulse characteristics, achieves time alignment under topological constraints by equivalent wave velocity verification, and filters out effective time difference data that conforms to the electromagnetic wave propagation law. At the same time, to address the wave impedance discontinuity problem at branch nodes, a reflection and attenuation coefficient is introduced to correct the distortion of gradient features, eliminating the influence of network branches and electromagnetic interference on feature extraction and ensuring the accuracy of subsequent model input data.

[0034] In yet another preferred embodiment, reference is made to Figure 6 The voltage and current phasor data of the monitoring nodes are concatenated with the frequency domain features of the partial discharge pulse signal to form a node feature vector. The voltage and current phasor data includes six basic features: voltage amplitude, voltage phase angle, current amplitude, current phase angle, active power, and reactive power. The frequency domain features of the partial discharge pulse signal are extracted through Fourier transform. Multiple characteristic frequency points within a preset frequency band are selected to form a frequency domain feature vector. The basic phasor features and the frequency domain features are concatenated dimensionally to generate the node feature vector corresponding to each monitoring node. The node feature vectors of all monitoring nodes are concatenated row by row to generate a node feature matrix. The line impedance parameters, line type identifiers, propagation delay gradients, and energy attenuation gradients of the physical topology path are concatenated into an edge feature vector. The line impedance parameters include two-dimensional features: positive-sequence impedance and zero-sequence impedance. The line type identifier is represented by one-hot encoding and includes three-dimensional features: overhead line, cable line, and mixed line. The propagation delay gradient and energy attenuation gradient are each one-dimensional features. The above features are concatenated dimensionally to generate the edge feature vector corresponding to each topology path. The edge feature vectors of all topology paths are concatenated row by row to generate an edge feature matrix.

[0035] In the graph convolutional layer of the graph neural network model, attention weight coefficients for information transfer between adjacent nodes are calculated based on edge feature vectors. These attention weight coefficients are then used to weight and aggregate node feature vectors, generating higher-order node representation vectors that contain topological spatial relationships. The graph neural network model employs a graph attention convolutional architecture. The calculation process for the attention weight coefficients involves first performing a linear transformation on the node features and edge features, then concatenating the transformed features, and finally calculating the correlation between nodes through an attention mechanism. The correlation is calculated as follows: , in, Let represent the attentional association between node i and node j, and LeakyReLU be the non-linear activation function. This is the weight vector for the attention mechanism. The linear transformation weight matrix for node features. Let be the linear transformation weight matrix of the edge features. Let be the node feature vector of node i. Let be the node feature vector of node j. Let be the feature vector of the edge between node i and node j. This is a feature concatenation operation. The calculated attention correlation is then normalized to obtain the final attention weight coefficients. The normalization method is as follows: , in, These are the normalized attention weight coefficients. Given the set of neighboring nodes of node i, normalization ensures that the sum of the attention weight coefficients of all neighboring nodes of node i is 1, thus quantifying the information contribution of different neighboring nodes. Based on the attention weight coefficients, the features of neighboring nodes are weighted and aggregated to generate the higher-order node representation vector of node i. The aggregation calculation method is as follows: , in, Let i be the higher-order node representation vector. As a nonlinear activation function, after stacking multiple graph attention convolutional layers, the final output is a high-order node representation matrix. The high-order node representation vector integrates the multi-source electrical quantity characteristics of the node itself, the feature information of adjacent nodes, and the physical characteristics of the topological path, thus fully representing the spatiotemporal evolution characteristics of the discharge pulse in the topological network.

[0036] After generating high-order node representation vectors containing topological spatial associations, these vectors are input into the fully connected output layer of the graph neural network model. The fully connected output layer outputs a discharge probability matrix and coordinate offsets. The dimension of the discharge probability matrix is ​​consistent with the number of monitoring nodes, and each element corresponds to the probability that the monitoring node is the node where the discharge source is located. The coordinate offsets include horizontal and vertical offsets in a Cartesian coordinate system, used to correct the specific location of the discharge source. The geographic coordinates of each node in the shortest path topological node sequence are obtained, and the coordinate offsets are superimposed on the geographic coordinates of the starting node with the highest discharge probability to generate the initial discharge source coordinates. It is determined whether the initial discharge source coordinates fall on the geometric line segment formed by the physical topological path. The geometric line segment is formed by connecting the geographic coordinates of two adjacent monitoring nodes in the topological path. If the initial discharge source coordinates fall on the geometric line segment, they are used as the final discharge source node coordinates; otherwise, they are orthogonally projected onto the nearest geometric line segment, and the coordinates of the projected point are used as the final discharge source node coordinates. The orthogonal projection is calculated as follows: , in, Let be the coordinate vector of the projection point. Let be the coordinate vector of the initial discharge source. and Let be the coordinate vectors of the two endpoints of the geometric line segment closest to the initial discharge source. For vector dot product operation, This involves performing a magnitude operation on the vector while simultaneously constraining the position of the projection point. If the projection parameter is less than 0, the projection point is taken as the starting point of the line segment. If the projection parameter is greater than 1, the projection point is taken as the endpoint of the line segment. This ensures that the final discharge node coordinates fall on the actual physical path of the transmission line. The fully connected output layer simultaneously outputs the probability distribution of fault types, including corona discharge, surface discharge, internal air gap discharge, and metallic grounding faults. The fault type with the highest probability is taken as the final identification result, completing the remote diagnosis of transmission equipment faults.

[0037] In this embodiment, to clarify the layer structure, feature input and output parameters, and core computational logic of the graph neural network model, and to ensure the reproducibility of the model's construction and inference process, the following table is set up: Table 4 shows the layer structure and feature parameters of the neural network model.

[0038] In Table 4, N represents the total number of monitoring nodes, corresponding to the dimension of the discharge probability matrix; 2 represents the dimension of the coordinate offset; M represents the total number of fault types, corresponding to the dimension of the fault type probability distribution; the network layer name corresponds to the hierarchical structure of the model; the input and output feature dimensions clarify the data flow specifications of each layer; the core operation specifies the execution logic of each layer; and the activation function clarifies the nonlinear transformation method of each layer. This table provides clear parameter specifications for the construction and training of graph neural network models, ensuring that those skilled in the art can completely reproduce the inference process of the model based on this table.

[0039] The technical solution provided in this embodiment constructs a graph attention network model that integrates the multi-source electrical quantity characteristics of nodes and the topological path characteristics of edges. It uses the attention mechanism to achieve weighted aggregation of node features under topological constraints, generating a high-order node representation containing spatial correlation information. This fully captures the propagation characteristics of discharge pulses in the topological network. At the same time, it corrects the initial discharge source coordinates through orthogonal projection to ensure that the positioning result falls on the actual physical path of the transmission line. This ensures the close fit between the fault positioning result and the actual physical line, and achieves accurate positioning of the discharge source and accurate identification of the fault type.

Claims

1. A method for remote diagnostics of power transmission equipment, characterized in that, The method includes: deploying phasor measurement units and partial discharge sensors at each monitoring node of the transmission line to collect voltage and current phasor data and partial discharge pulse signals; The collected multi-source electrical quantity data is injected into the edge computing gateway, which establishes a spatiotemporal correlation matrix based on the physical topology connection relationship of the power transmission network. For the partial discharge pulse signal, the time difference and amplitude attenuation ratio of the pulse arriving at each monitoring node are extracted. Combined with the voltage and current phase change characteristics provided by the phasor measurement unit, time alignment under topological constraints is performed in the spatiotemporal correlation matrix. The time alignment process is constrained by the physical topology path of the transmission network, excluding time difference data that does not match the topology path, and mapping the effective time difference data to the corresponding topology path node sequence. The propagation delay gradient and energy attenuation gradient of the discharge pulse between adjacent nodes are calculated by the time alignment. The propagation delay gradient and energy attenuation gradient are input into a pre-constructed graph neural network model. The graph neural network model uses the physical topology path as the edge feature and the multi-source electrical quantity as the node feature. It outputs the forward propagation path of the fault signal in the topology network and the coordinates of the discharge source node, thus completing the remote location and type identification of the power transmission equipment fault.

2. The remote diagnostic method for power transmission equipment according to claim 1, characterized in that, The edge computing gateway establishes a spatiotemporal correlation matrix based on the physical topology of the power transmission network, including: the edge computing gateway obtains the geographical coordinates and line connection relationships of each monitoring node in the power transmission network, and maps each monitoring node to a row vector and a column vector in the spatiotemporal correlation matrix; For two monitoring nodes that have a direct physical line connection, the corresponding matrix element in the spatiotemporal correlation matrix is ​​set to the physical length of the line; for two monitoring nodes that do not have a direct physical line connection, the corresponding matrix element is set to infinity. Phasor measurement units of each monitoring node are injected into the diagonal position of the spatiotemporal correlation matrix to collect timestamps, forming an initial spatiotemporal correlation matrix containing spatial topological distance and time reference.

3. The remote diagnostic method for power transmission equipment according to claim 2, characterized in that, The extraction of the time difference and amplitude attenuation ratio of the pulses arriving at each monitoring node includes: performing high-pass filtering on the original pulse signal collected by the partial discharge sensor to filter out the power frequency carrier signal, and performing Hilbert transform on the filtered signal to extract the envelope. A dynamic floating threshold is set on the envelope, and pulse peak points exceeding the dynamic floating threshold are captured. The timestamp corresponding to the pulse peak point is used as the pulse arrival time. The amplitude of the pulse peak point is read, the time difference is obtained by subtracting the pulse arrival times captured by different monitoring nodes pairwise, and the amplitude attenuation ratio is obtained by calculating the ratio of the amplitudes captured by different monitoring nodes.

4. The remote diagnostic method for power transmission equipment according to claim 3, characterized in that, The step of performing time-series alignment under topological constraints in the spatiotemporal correlation matrix includes: extracting the shortest path topological node sequence between any two monitoring nodes from the spatiotemporal correlation matrix; The time difference is matched and verified with the physical length of the line corresponding to the shortest path topology node sequence. When the equivalent wave speed corresponding to the time difference is within the set wave speed range, the time difference is determined to be valid, and the time difference is aligned to the shortest path topology node sequence. When the equivalent wave velocity is outside the wave velocity range, the time difference is determined to be an invalid time difference caused by noise interference, and the invalid time difference is removed from the spatiotemporal correlation matrix.

5. The remote diagnostic method for power transmission equipment according to claim 4, characterized in that, The calculation of the propagation delay gradient and energy attenuation gradient of the discharge pulse between adjacent nodes includes: on the shortest path topology node sequence with completed time alignment, along the propagation direction of the discharge pulse, calculating the local propagation delay between the i-th node and the (i+1)-th node, and dividing the local propagation delay by the physical distance between the two nodes to obtain the propagation delay gradient; Calculate the local amplitude attenuation between the i-th node and the (i+1)-th node, and divide the local amplitude attenuation by the physical distance to obtain the energy attenuation gradient; The propagation delay gradients and energy decay gradients on the shortest path topology node sequence are combined into a gradient feature vector.

6. The remote diagnostic method for power transmission equipment according to claim 5, characterized in that, The graph neural network model uses physical topology paths as edge features and the multi-source electrical quantities as node features, including: concatenating the voltage and current phasor data of the monitoring nodes with the frequency domain features of the partial discharge pulse signal to form a node feature vector; The line impedance parameters, line type identifier, propagation delay gradient, and energy attenuation gradient of the physical topology path are concatenated into an edge feature vector. In the graph convolutional layer of the graph neural network model, attention weight coefficients for information transmission between adjacent nodes are calculated based on the edge feature vectors. The node feature vectors are then weighted and aggregated using these attention weight coefficients to generate higher-order node representation vectors that contain topological spatial relationships.

7. The remote diagnostic method for power transmission equipment according to claim 2, characterized in that, After forming the initial spatiotemporal correlation matrix containing spatial topological distance and time reference, the method further includes: the edge computing gateway monitoring the status change signals of power transmission network circuit breakers and disconnect switches in real time; When the state change signal is received, the branch with the topology change is identified, the matrix element of the corresponding branch in the initial spatiotemporal correlation matrix is ​​updated from the physical length of the line to infinity, and the timestamp synchronization association of the corresponding branch is disconnected. When a closing signal indicating the end of maintenance is received, the matrix elements of the corresponding branch in the initial spatiotemporal correlation matrix are restored from infinity to the updated physical length of the line, thereby generating a dynamic spatiotemporal correlation matrix.

8. The remote diagnostic method for power transmission equipment according to claim 3, characterized in that, After capturing the pulse peak point that exceeds the dynamic floating threshold, the method further includes: extracting the voltage and current phase change characteristics provided by the phasor measurement unit within the preset time window where the pulse peak point is located; Calculate the change energy value of the voltage and current phase change characteristics, and determine whether the change energy value reaches the set phase change threshold value; If the mutation energy value does not reach the phase mutation threshold value, the pulse peak point is determined to be an interference pulse generated by spatial electromagnetic radiation, and the time difference and amplitude attenuation ratio corresponding to the interference pulse are marked as abnormal data and deleted. If the mutation energy value reaches the phase mutation threshold value, the corresponding feature data is retained.

9. The remote diagnostic method for power transmission equipment according to claim 5, characterized in that, The step of combining all the propagation delay gradients and the energy decay gradients on the shortest path topology node sequence into a gradient feature vector includes: detecting whether there are branch nodes in the shortest path topology node sequence; When the branch node exists, identify the multiple discharge pulse branch paths flowing through the branch node, and calculate the branch propagation delay gradient and branch energy attenuation gradient corresponding to each discharge pulse branch path. A reflection attenuation coefficient is introduced at the point of impedance discontinuity to correct the distortion of the branch propagation delay gradient and the branch energy attenuation gradient at the point of reflection. The corrected gradient data is then concatenated into the gradient feature vector.

10. The remote diagnostic method for power transmission equipment according to claim 6, characterized in that, After generating the high-order node representation vector containing topological spatial association, the method further includes: inputting the high-order node representation vector into the fully connected output layer of the graph neural network model, wherein the fully connected output layer outputs the discharge probability matrix and coordinate offset. Obtain the geographic coordinates of each node in the shortest path topology node sequence, and superimpose the coordinate offset onto the geographic coordinates of the starting node with the highest discharge probability to generate the initial discharge source coordinates. Determine whether the initial discharge source coordinates fall on the geometric line segment formed by the physical topology path. If not, orthogonally project the initial discharge source coordinates onto the nearest geometric line segment, and use the coordinates of the projection point as the final discharge source node coordinates.

Images