Fault diagnosis positioning method and system applied to power equipment operation and maintenance

Abstract

The application provides a fault diagnosis positioning method and system applied to power equipment operation and maintenance, relates to the technical field of power equipment operation and maintenance, and first acquires real-time operation state monitoring data flow of a target power equipment; the real-time operation state monitoring data flow comprises a time sequence collected sensing response signal fragment sequence of a plurality of sensing units; then, a time sequence dependent relationship mining operation is performed on the real-time operation state monitoring data flow to generate a time sequence state transmission chain network; a spatial topology correlation analysis operation is further performed to generate a spatial correlation knowledge graph; then, the two are input into a time-space fault feature tracing model for joint deduction analysis to generate a fault source positioning indication mark; finally, a fault diagnosis positioning report is generated according to the fault source positioning indication mark and the spatial correlation knowledge graph and is pushed to an operation and maintenance management terminal. The application can efficiently and accurately diagnose and position power equipment faults and improve the reliability and stability of a power system.

Description

Technical Field

[0001] This invention relates to the field of power equipment operation and maintenance technology, and more specifically, to a fault diagnosis and location method and system for power equipment operation and maintenance. Background Technology

[0002] With the continuous expansion and increasing complexity of power equipment, traditional fault diagnosis and location methods face numerous challenges. Currently, common power equipment fault diagnosis and location methods mainly rely on manual inspection and monitoring analysis based on single parameters. While manual inspection allows direct observation of the equipment's appearance and operating status, it is inefficient and struggles to detect faults hidden within the equipment. Furthermore, manual inspection is limited by the experience and skill level of the inspectors, making it prone to missed detections and misdiagnosis.

[0003] Single-parameter monitoring and analysis methods, such as relying solely on monitoring parameters like temperature, current, or voltage to determine equipment malfunctions, while capable of detecting faults to some extent, have limitations. Power equipment failures are often caused by a combination of factors, and changes in a single parameter may not fully reflect the actual operating status of the equipment, easily leading to false alarms or missed alarms. Furthermore, the aforementioned methods struggle to pinpoint the exact location of the fault within the equipment, failing to provide accurate maintenance guidance for operators, resulting in low maintenance efficiency and increased equipment downtime and maintenance costs. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a fault diagnosis and location method applied to the operation and maintenance of power equipment, the method comprising: Acquire real-time operating status monitoring data stream of target power equipment, wherein the real-time operating status monitoring data stream includes a sequence of sensor response signal segments continuously collected in chronological order by multiple sensor units deployed on the target power equipment; A time-series dependency mining operation is performed on the real-time operating status monitoring data stream. By analyzing the transmission and evolution of signal amplitude fluctuation patterns along the time axis in the sensor response signal segment sequence, a time-series state transmission chain network reflecting the internal state transition path of the target power equipment is generated. A spatial topology association parsing operation is performed on the real-time operation status monitoring data stream. By analyzing the degree of correlation between the deployment location adjacency relationship of the multiple sensing units on the physical structure of the target power equipment and the synchronous change of their respective sensing response signal segment sequences, a spatial association knowledge graph characterizing the mutual influence relationship between the components of the target power equipment is generated. The temporal state transmission chain network and the spatial association knowledge graph are input into a preset spatiotemporal fault feature tracing model for joint inference and analysis to generate a fault source location indicator for locating the source of the fault inside the target power equipment. Based on the fault source location indication and the spatial association knowledge graph, a fault diagnosis and location report is generated, which includes the unique identifier code of the faulty component and a list of affected components. The fault diagnosis and location report is then pushed to a preset operation and maintenance management terminal.

[0005] Furthermore, embodiments of the present invention also provide a fault diagnosis and location system for power equipment operation and maintenance, comprising: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the above-described fault diagnosis and location method for power equipment operation and maintenance by executing the machine-executable instructions.

[0006] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, a processor of a fault diagnosis and location system for power equipment operation and maintenance reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the fault diagnosis and location system for power equipment operation and maintenance to execute the aforementioned fault diagnosis and location method for power equipment operation and maintenance.

[0007] Based on the above, firstly, by acquiring the real-time operating status monitoring data stream of the target power equipment, a time-series dependency mining operation is performed on the real-time operating status monitoring data stream. By analyzing the transmission and evolution law of signal amplitude fluctuation patterns along the time axis, a time-series state transmission chain network reflecting the internal state migration path of the target power equipment is generated. This allows for in-depth mining of the changing law of equipment operating status in the time dimension, accurately understanding the dynamic migration process of the internal state of the equipment. Secondly, a spatial topology association analysis operation is performed on the real-time operating status monitoring data stream. By analyzing the adjacency relationship of the deployment positions of multiple sensing units in the physical structure of the equipment and the degree of synchronous change correlation between their respective sensing response signal segment sequences, a spatial association knowledge graph representing the mutual influence relationship between the components of the target power equipment is generated. This can reveal the mutual influence relationship between equipment components from a spatial dimension, which helps to understand the propagation path and impact range of faults within the equipment. By inputting a temporal state transmission chain network and a spatial association knowledge graph into a pre-defined spatiotemporal fault feature tracing model for joint inference and analysis, a fault source location indicator is generated to pinpoint the origin of the fault within the target power equipment. This comprehensive utilization of information from both temporal and spatial dimensions enables accurate tracing of the fault's origin, significantly improving fault location accuracy. Finally, based on the fault source location indicator and the spatial association knowledge graph, a fault diagnosis and location report is generated, containing the unique identifier of the faulty component and a list of affected components. This report is then pushed to a pre-defined operation and maintenance management terminal, providing maintenance personnel with fault information to facilitate rapid development of repair plans and timely fault resolution. This effectively reduces equipment downtime and improves the reliability and stability of the power system, thereby achieving efficient and accurate diagnosis and location of power equipment faults. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the execution flow of the fault diagnosis and location method for power equipment operation and maintenance provided in the embodiments of the present invention.

[0009] Figure 2 This is a schematic diagram of exemplary hardware and software components of a fault diagnosis and location system for power equipment operation and maintenance provided in an embodiment of the present invention. Detailed Implementation

[0010] Figure 1 This is a flowchart illustrating a fault diagnosis and location method for power equipment operation and maintenance provided by an embodiment of the present invention, which will be described in detail below.

[0011] The fault diagnosis and location method for power equipment operation and maintenance provided in this application can be applied to online monitoring and fault diagnosis scenarios for core power equipment such as large power transformers, high-voltage circuit breakers, and gas-insulated metal-enclosed switchgear in substations. This method analyzes real-time operating status data collected by various types of sensors deployed on the equipment to uncover the propagation and evolution patterns of fault symptoms in time and space, thereby accurately locating the source of the fault. The following explanation uses an oil-immersed power transformer as an example; those skilled in the art will understand that this method is also applicable to other types of power equipment such as generators, motors, reactors, and surge arresters.

[0012] Step S110: Obtain the real-time operating status monitoring data stream of the target power equipment. The real-time operating status monitoring data stream includes a sequence of sensor response signal segments continuously collected in chronological order by multiple sensing units deployed on the target power equipment.

[0013] Multiple sensing units are deployed on the target power equipment, with different types of units used to sense changes in different physical quantities. Current sensing units collect the amplitude and phase of the current in the equipment's grounding lead; vibration sensing units are installed on the equipment's casing to collect mechanical vibration signals during operation; oil chromatography sensing units extract oil from inside the equipment via a circulating oil circuit and analyze the content of characteristic gases dissolved in the insulating oil; and partial discharge sensing units are installed at the equipment's flange interface or observation window to collect electromagnetic wave signals generated by partial discharge. Each sensing unit continuously collects signals at a fixed sampling frequency, and the data collected by each unit is arranged in time sequence to form a sequence of sensing response signal segments for that unit. The sequences from all sensing units together constitute a real-time operational status monitoring data stream.

[0014] Step S120: Perform time-series dependency mining operation on the real-time operation status monitoring data stream. By analyzing the transmission evolution law of the signal amplitude fluctuation mode along the time axis in the sensor response signal segment sequence, generate a time-series state transmission chain network that reflects the internal state transition path of the target power equipment.

[0015] Step S121: Expand the sequence of sensor response signal segments corresponding to each sensor unit in the real-time operation status monitoring data stream along the time axis into a sequence of signal amplitudes with a chronological order, and perform time window division processing on the signal amplitude sequence to obtain a set of signal amplitude segments composed of signal amplitude segments within continuous and partially overlapping time windows. Perform fluctuation pattern encoding processing on the signal amplitude segments within each time window in the set of signal amplitude segments, extract the arrangement order features of the alternating local peaks and valleys within the signal amplitude segment and the amplitude difference features between adjacent peaks and valleys, and generate a local fluctuation pattern characterization code that uniquely corresponds to each time window.

[0016] For each sensor unit's acquired signal amplitude sequence, a time window length parameter L and a sliding step size parameter S are set. Starting from the beginning of the sequence, signal amplitude segments with window starting indices of 1, 1+S, 1+2S, ... are sequentially extracted. Each signal amplitude segment has a length of L, and there are LS overlapping sampling points between adjacent windows. For each extracted signal amplitude segment, local extremum detection is performed: all sampling points within the segment are traversed. If the amplitude of the current sampling point is greater than the amplitude of its preceding and following sampling points, it is marked as a peak; if the amplitude of the current sampling point is less than the amplitude of its preceding and following sampling points, it is marked as a trough. The peak and trough points are extracted in chronological order, for example, in a sequence pattern of "peak-trough-peak-trough". Simultaneously, the amplitude difference between adjacent peaks and troughs is recorded, forming an amplitude drop sequence. The permutation sequence feature is encoded into a binary sequence (e.g., peaks are encoded as 1 and troughs as 0). The amplitude drop sequence is quantized and then concatenated to the end of the binary sequence to form a local fluctuation pattern representation code for the amplitude segment of the signal.

[0017] Step S122: Construct the wave pattern evolution sequence corresponding to the sensing unit by encoding all local wave patterns belonging to the same sensing unit and arranged in chronological order. Perform state transition probability analysis on the wave pattern evolution sequence, count the frequency distribution of transitions between adjacent local wave pattern encodings, generate the wave state transition probability distribution matrix corresponding to the sensing unit, traverse the wave state transition probability distribution matrices corresponding to all sensing units, calculate the similarity value of the state transition trajectory between any two sensing units according to the preset state transition similarity measurement rules, and mark the sensing unit pairs whose state transition trajectory similarity value exceeds the preset similarity threshold as associated sensing unit pairs with temporal correlation.

[0018] The local fluctuation mode characterization codes corresponding to each time window generated in step S121 are arranged in chronological order to form the fluctuation mode evolution sequence of the sensing unit. For two adjacent codes c_i and c_{i+1} in this sequence, the number of times the transition from c_i to c_{i+1} occurs is counted. After traversing all adjacent code pairs in the sequence, the frequency distribution of all possible transitions is obtained. The transition frequency from each starting code c_i to each target code c_j is divided by the total number of transitions starting from c_i to obtain the transition probability P(c_i→c_j). Using all possible local fluctuation mode characterization code values ​​as the row and column indices of the matrix, the transition probabilities are filled into the corresponding positions in the matrix to obtain the fluctuation state transition probability distribution matrix M_i of the sensing unit. For any two sensing units a and b, calculate the Frobenius norm difference D = sqrt(Σ(M_a[p][q]-M_b[p][q])^2) between their transition probability distribution matrices M_a and M_b, and then calculate the similarity value of the state transition trajectories S_sim = 1 / (1+D). Set a similarity threshold T_sim; if S_sim > T_sim, then the sensing units a and b are marked as a pair of associated sensing units with a temporal correlation.

[0019] Step S123: Use the associated sensing unit pairs as nodes in a temporal network, and use the similarity value of the state transition trajectory as the edge weight of the directed edge connecting the temporal network nodes to construct a preliminary temporal directed association network.

[0020] Create a directed graph G_temporal, where each node represents a sensor unit. For each associated sensor unit pair (a, b) marked in step S122, create two directed edges in opposite directions between nodes a and b: one from a to b, and the other from b to a. Use the state transition trajectory similarity value S_sim calculated in step S122 as the edge weight of these two directed edges. For sensor unit pairs that are not associated, do not create connecting edges.

[0021] Step S124: Perform a network path backtracking search operation on the preliminary temporal directed association network. Starting from each temporal network node, search downstream associated nodes level by level along the direction of the directed edges, record the sequence of temporal network nodes passed through in each search process, and generate a forward influence propagation path corresponding to each temporal network node; and perform a reverse path backtracking search operation on the preliminary temporal directed association network. Starting from each temporal network node, search upstream associated nodes level by level against the direction of the directed edges, record the sequence of temporal network nodes passed through in each reverse search process, and generate a reverse influence propagation path corresponding to each temporal network node.

[0022] For each node v in the initial temporally directed association network, perform a depth-first search: starting from v, recursively visit all reachable nodes along the direction of the directed edges (i.e., from the node to its downstream neighbor), recording the sequence of nodes visited during the visit, to obtain the forward influence propagation path P_forward(v). Simultaneously, starting from v, recursively visit all reachable nodes in the opposite direction of the directed edges (i.e., from the node to its upstream neighbor), recording the sequence of nodes visited during the visit, to obtain the backward influence propagation path P_backward(v).

[0023] Step S125: Combine and process the forward influence propagation path and the reverse influence propagation path corresponding to the same time-series network node to form a complete influence propagation link with the time-series network node as the central node, and deduplicate and integrate the complete influence propagation links corresponding to all time-series network nodes to generate a time-series state transfer chain network that reflects the internal state transition path of the target power equipment.

[0024] For each node v, its backward propagation path P_backward(v) (the path from the upstream node to v) is concatenated with its forward propagation path P_forward(v) (the path from v to the downstream node), resulting in a complete propagation chain L_v = [P_backward(v), v, P_forward(v)] centered at node v. All complete propagation chains corresponding to all nodes are merged, and duplicate chains are removed to obtain the temporal state transfer chain network G_chain.

[0025] Step S126: Obtain the historical maintenance records of the target power equipment, and mark the fault propagation path correlation of the link where the corresponding component is located in the time-series state transmission chain network according to the historical maintenance records. In the time-series state transmission chain network, the complete impact propagation link passing through the component is marked as a verified fault propagation path. The historical maintenance records include the specific component identifiers targeted by each maintenance operation and the health status description text of the component corresponding to the specific component identifier after the maintenance operation.

[0026] Historical maintenance records are retrieved from the equipment management system. Each record includes the maintenance date, the identifier of the component involved (e.g., "Phase A high-voltage bushing", "Phase B winding temperature sensor"), the maintenance operation type (replacement, repair, cleaning), and a text description of the component's health status after maintenance. For each maintenance record, the identifier of the component involved is extracted, the corresponding node in the time-series state propagation chain network G_chain is located, and all complete impact propagation links passing through that node are marked as verified fault propagation paths.

[0027] Step S130: Perform a spatial topology association parsing operation on the real-time operation status monitoring data stream. By analyzing the adjacency relationship of the deployment positions of the multiple sensing units on the physical structure of the target power equipment and the degree of correlation between the synchronous changes of their respective sensing response signal segment sequences, a spatial association knowledge graph characterizing the mutual influence relationship between the components of the target power equipment is generated.

[0028] Step S131: Obtain the physical assembly structure topology diagram of the target power equipment. The physical assembly structure topology diagram includes the component identifiers of each functional component inside the target power equipment and a description of the mechanical connection relationship type between any two functional components.

[0029] Extract the physical assembly structure topology diagram G_assembly from the equipment design drawing file. Each node in this physical assembly structure topology diagram represents a functional component (e.g., iron core, winding, bushing, tap changer, oil tank), and the node attributes include the component identifier. Each edge in the diagram represents a direct physical connection between two functional components, and the edge attributes include a description of the mechanical connection relationship type (e.g., bolted connection, welded connection, insulated support connection, oil circuit connection).

[0030] Step S132: Perform position mapping matching processing on multiple sensing units deployed on the target power equipment and the functional components in the physical assembly structure topology diagram to determine the target functional component identifier directly monitored by each sensing unit, and establish a one-to-one mapping relationship table between sensing unit identifier and target functional component identifier.

[0031] For each sensing unit, based on its installation location description (e.g., "at the iron core grounding lead," "at the top of the A-phase winding oil tank wall," "at the oil chromatograph sampling port"), the functional component corresponding to that installation location is located in the physical assembly structure topology diagram G_assembly. The component identifier of that functional component is determined as the target functional component identifier directly monitored by that sensing unit. A mapping table T_map is established, with the sensing unit identifier as the key and the corresponding functional component identifier as the value.

[0032] Step S133: Based on the description of the mechanical connection relationship type in the physical assembly structure topology diagram, determine whether there is a direct physical connection between any two target functional component identifiers, and mark the target functional component identifier pairs with direct physical connections as spatial adjacent component pairs. For each spatial adjacent component pair with direct physical connections, obtain the sensing response signal segment sequence of the sensing unit corresponding to each of the two target functional component identifiers in the spatial adjacent component pair, and perform synchronization time point alignment processing on the two sensing response signal segment sequences to obtain a synchronization signal segment pair.

[0033] Traverse all edges in the physical assembly structure topology graph G_assembly. For each edge connecting functional components c_i and c_j, mark (c_i, c_j) as a spatially adjacent component pair. According to the mapping table T_map, find the sensing unit s_i corresponding to c_i and the sensing unit s_j corresponding to c_j, and obtain the sensing response signal segment sequence collected by s_i and s_j in the same time period. Since the two sequences have the same time axis, they are directly aligned according to the time index to obtain the synchronization signal segment pair (X_i(t), X_j(t)).

[0034] Step S134: Perform synchronization fluctuation consistency analysis on the synchronization signal segment pair, calculate the frequency of the same-direction change and the frequency of the opposite-direction change of the signal amplitude change trend of the two sensor response signal segment sequences within the same time window, and determine the synchronization response consistency value between the two sensing units based on the ratio of the same-direction change frequency to the opposite-direction change frequency.

[0035] For a pair of synchronized signal segments (X_i(t), X_j(t)), a sliding time window length W is set (e.g., W = 10 sampling points). Within each window, the difference sign of X_i(t) is calculated as sign(ΔX_i) = sign(X_i(t+1) - X_i(t)), and similarly, the difference sign of X_j(t) is calculated as sign(ΔX_j). If both sign(ΔX_i) and sign(ΔX_j) are positive or both are negative, it is counted as a same-direction change; if their signs are opposite, it is counted as a opposite-direction change. After traversing all windows, the total number of same-direction changes N_same and the total number of opposite-direction changes N_opp are counted. The synchronization response consistency value C_sync = (N_same - N_opp) / (N_same + N_opp) is calculated. The value of C_sync ranges from [-1, 1], where the closer to 1, the more consistent the fluctuation direction of the two sequences, and the closer to -1, the more opposite the fluctuation direction.

[0036] Step S135: Traverse all spatial adjacency component pairs with direct physical connections, repeatedly perform synchronization fluctuation consistency analysis processing, generate a synchronization response consistency value uniquely corresponding to each spatial adjacency component pair, and mark spatial adjacency component pairs whose synchronization response consistency value exceeds a preset consistency threshold as target associated component pairs.

[0037] Set a consistency threshold T_consistency (e.g., 0.6). For each spatially adjacent component pair (c_i, c_j), if its corresponding synchronization response consistency value C_sync > T_consistency, then the component pair is marked as a target associated component pair, indicating that the two components have a strong mutual influence relationship during operation.

[0038] Step S136: Take all functional components in the physical assembly structure topology as spatial network nodes, and establish an undirected connection edge between the spatial network nodes corresponding to the two functional components in the target associated component pair. Assign the synchronization response consistency value to the undirected connection edge as the edge weight to construct a preliminary spatial undirected coupled network. Perform a coupling relationship extension operation on the preliminary spatial undirected coupled network. For any two spatial network nodes in the preliminary spatial undirected coupled network that do not have a direct undirected connection edge but have common adjacent spatial network nodes, analyze the product value of the edge weights between the two spatial network nodes and the common adjacent spatial network node. Add an indirect coupling connection edge between the two spatial network nodes whose product value exceeds a preset extension threshold and assign the product value as the indirect edge weight.

[0039] Create an undirected graph G_spatial, using all functional component nodes from the physical assembly topology graph G_assembly as nodes of G_spatial. For each target associated component pair (c_i, c_j), add an undirected edge between nodes c_i and c_j in G_spatial, with the edge weight set to C_sync. Then perform outward extension: traverse all node pairs (c_u, c_v) in G_spatial that do not have direct connecting edges, and search for a common adjacent node c_w such that there are connecting edges between c_u and c_w, and between c_v and c_w. If it exists, calculate the product P = w(c_u, c_w) * w(c_v, c_w). Set an extension threshold T_extend; if P > T_extend, add an indirect coupling connecting edge between c_u and c_v, with the edge weight set to P.

[0040] Step S137: Merge the initial spatial undirected coupled network with the newly added indirect coupled connection edges after the coupling relationship expansion operation to generate a spatial association knowledge graph representing the mutual influence relationship between the target power equipment components.

[0041] The initial spatial undirected coupled network G_spatial is merged with the indirect coupled connection edges added in step S136 to obtain the final spatial association knowledge graph G_knowledge. Nodes in this spatial association knowledge graph represent functional components of a device, edges represent direct or indirect mutual influence relationships between components, and edge weights represent the strength of these mutual influences.

[0042] Step S138: Obtain the actual operating environment deployment information of the target power equipment. The actual operating environment deployment information includes the geographical orientation description of the target power equipment in the deployment site and the distribution description of surrounding interference sources. Based on the actual operating environment deployment information, perform environmental attenuation factor correction processing on the edge weights of the undirected connection edges of the corresponding edge components in the spatial association knowledge graph.

[0043] The actual operating environment deployment information is read from the equipment deployment record, such as the equipment installation direction (south or north), whether there are other devices generating electromagnetic interference around the equipment, and whether there are vibration sources near the equipment. For components located at the edge of the equipment or close to interference sources, the weight of their corresponding connection edges needs to be multiplied by the environmental attenuation factor α (0 < α < 1) to reduce their influence weight in fault propagation.

[0044] Step S140: Input the temporal state transmission chain network and the spatial association knowledge graph into the preset spatiotemporal fault feature tracing model for joint inference and analysis to generate a fault source location indicator for locating the source of the fault inside the target power equipment.

[0045] In this embodiment, the spatiotemporal fault feature tracing model adopts an encoder-decoder architecture based on graph neural networks, specifically including a graph structure encoding layer, a spatiotemporal attention fusion layer, and a fault tracing decoding layer. The graph structure encoding layer is composed of two stacked graph convolutional networks. The update formula for each graph convolutional network is H^{(l+1)}=σ(D^{-1 / 2}AD^{-1 / 2}H^{(l)}W^{(l)}), where A is the adjacency matrix of the spatiotemporal fusion network, D is the degree matrix of A, H^{(l)} is the feature matrix of the l-th layer node, W^{(l)} is the learnable weight matrix, and σ is the ReLU activation function. The first graph convolutional network maps the input node feature dimension from d_in to d_hidden, and the second layer maps d_hidden to d_out. The spatiotemporal attention fusion layer includes a multi-head self-attention mechanism with 8 attention heads. Each head is calculated as Attention(Q, K, V) = softmax(QK^T / sqrt(d_k))V, where Q, K, and V are obtained from the node feature matrix through different linear transformations. This layer performs weighted fusion of node representations at different time steps and outputs the fused node feature matrix.

[0046] The training process of the spatiotemporal fault feature tracing model is as follows. The training dataset comes from a historical fault case library, containing a total of 5000 fault samples. Each sample contains the sensor response signal fragment sequence of all sensor units within 30 seconds before and after the fault occurs, as well as the corresponding actual fault source component identification label. Input data preprocessing: The sensor response signal fragment sequence of each sensor unit is Z-score normalized, and then steps S120 and S130 are executed to generate a temporal state transfer chain network and a spatial association knowledge graph. The two are fused into a spatiotemporal fusion network as the model input. The output is the probability distribution vector of the target node, with a length equal to the total number of device components. The loss function adopts cross-entropy loss L=-Σy_ilog(p_i), where y_i is the one-hot encoding of the actual fault source component identification, and p_i is the probability of the i-th component output by the model. The optimizer uses Adam, with an initial learning rate set to 0.001 and a weight decay parameter set to 0.0001. The batch size is set to 32, and the training epochs are set to 200. An early stopping mechanism is employed, halting training when the validation set loss fails to decrease for 20 consecutive epochs. Top-1 accuracy and Top-3 accuracy are used as evaluation metrics.

[0047] When applying the model, the real-time operational status monitoring data stream is standardized using the same preprocessing method as the training data. Steps S120 and S130 are executed to generate the spatiotemporal fusion network for the current moment, which is then input into the trained spatiotemporal fault feature tracing model. After forward propagation, the model outputs the probability distribution vector of each component as the fault source, and the identifier of the component with the highest probability is taken as the fault source location indicator.

[0048] Step S141: Obtain a temporal state transfer chain network reflecting the internal state transition path of the target power equipment. The temporal state transfer chain network includes multiple temporal network nodes and directed edges connecting the temporal network nodes. Each temporal network node corresponds to a sensing unit within the target power equipment, and each directed edge is accompanied by a temporal dependency transfer weight. Obtain a spatial association knowledge graph characterizing the mutual influence relationships between components of the target power equipment. The spatial association knowledge graph includes multiple spatial network nodes and undirected connecting edges connecting the spatial network nodes. Each spatial network node corresponds to a functional component within the target power equipment, and each undirected connecting edge is accompanied by a spatial coupling strength weight.

[0049] Obtain the temporal state transfer chain network G_chain from step S125, and obtain the spatial association knowledge graph G_knowledge from step S137.

[0050] Step S142: The temporal network nodes in the temporal state transfer chain network and the spatial network nodes in the spatial association knowledge graph are aligned and fused according to the one-to-one mapping relationship table between the sensing unit identifier and the target functional component identifier to generate a spatiotemporal fusion network. Each fusion node in the spatiotemporal fusion network inherits the corresponding temporal network node attributes and spatial network node attributes.

[0051] Based on the mapping table T_map, align each sensor unit node s_i in the temporal state propagation chain network G_chain with its corresponding functional component node c_i in the spatial association knowledge graph G_knowledge. Create a new spatiotemporal fusion network G_fusion, whose node set is the functional component node set. For a directed edge (s_i→s_j) in G_chain, if the functional component corresponding to s_i is c_i and the functional component corresponding to s_j is c_j, then add a directed edge between nodes c_i and c_j in G_fusion. The edge's attributes inherit both the temporal dependency propagation weight and the spatial coupling strength weight (if any).

[0052] Step S143: Perform fault symptom correlation initialization processing on each fusion node in the spatiotemporal fusion network, obtain the latest signal segment of the real-time sensing response signal segment sequence of the corresponding sensing unit of the fusion node, extract the abnormal amplitude offset that exceeds the preset normal operation range in the latest signal segment, and map the abnormal amplitude offset to a normalized fault symptom level value according to the preset abnormal level mapping table of each sensing unit, and use it as the initial fault symptom level value of the fusion node.

[0053] For each fusion node c_i in the spatiotemporal fusion network G_fusion, the latest signal segment of its corresponding sensor unit is obtained from the real-time operational status monitoring data stream. The amplitude of all sampling points in this signal segment is calculated and compared with the preset normal operating range [L_norm, H_norm]. For sampling points that exceed the normal range, the offset offset = max(|value - H_norm|, |value - L_norm|) is calculated. The maximum offset offset_max of all exceeding points is taken, and the initial fault symptom level value L_init(c_i) is obtained according to the preset anomaly level mapping table (which maps the offset range to level values ​​between 0 and 1).

[0054] Step S144: Perform a fault source probability inference operation based on a random walk mechanism on the spatiotemporal fusion network. With each fusion node as a possible fault starting candidate point, perform multiple rounds of random walk propagation process. In each round of random walk propagation process, at each step, select the next fusion node to walk based on the joint weighted probability distribution of the temporal dependency propagation weight indicated by the directed edge of the current fusion node and the spatial coupling strength weight indicated by the undirected connection edge.

[0055] Set the total number of random walk rounds R (e.g., R=10000). For each round, starting from candidate node c_i, perform a walk with a step size of T_step (e.g., T_step=10). At the current node c_cur, consider all directed edges (temporally dependent propagation weight w_t) and all undirected edges (spatial coupling strength weight w_s) originating from c_cur. Calculate the joint weight w_joint=w_t*w_s for each edge (if there is only one type of edge, use that edge's weight). Then calculate the probability P(c_cur→c_next)=w_joint(c_cur, c_next) / Σw_joint(c_cur, *) of moving from c_cur to each neighbor node c_next. Randomly select the next walk node based on this probability distribution. Repeat until the maximum number of steps is reached.

[0056] Step S145: Record the sequence of all fusion nodes traversed from the starting candidate point to the end of the walk in each round of random walk propagation. Calculate the cumulative frequency of each fusion node being traversed in all rounds of random walk propagation. Divide the cumulative frequency of each fusion node being traversed by the product of the total number of walk rounds and the number of network summary points to obtain the fault impact range score of the fusion node.

[0057] For each round of walking, record the sequence of nodes traversed. After the walk is completed, increment the count of each node in the sequence by 1. After completing all R rounds of walking, for each node c_i, its cumulative frequency of being traversed is cnt_i. Calculate the fault ripple impact range score S_range(c_i) = cnt_i / (R*N), where N is the total number of nodes in the spatiotemporal fusion network G_fusion.

[0058] Step S146: For each fusion node that serves as a starting candidate point, summarize the initial fault symptom level values ​​of the fusion nodes that have passed through during all rounds of random walk propagation from that starting candidate point. Calculate the sum of the initial fault symptom level values ​​of the fusion nodes that have passed through the fusion nodes as the initial source explanation capability score of that starting candidate point. Divide this initial source explanation capability score by the largest initial source explanation capability score among all starting candidate points to obtain the normalized source explanation capability score of that starting candidate point.

[0059] For all walks starting from node c_i, collect the initial fault symptom level values ​​L_init of all nodes visited in these walks, and calculate the sum of these level values ​​S_explain(c_i) = ΣL_init(c_j) for all walks starting from c_i in all visited c_j. Find the maximum S_explain value S_max among all candidate starting points. Calculate the normalized source explanation score S_explain_norm(c_i) = S_explain(c_i) / S_max.

[0060] Step S147: Perform a weighted fusion calculation on the fault impact range score value and the normalized source explanation capability score value corresponding to each starting candidate point to generate a comprehensive confidence score value for the starting candidate point as the fault source location. Sort the comprehensive confidence scores corresponding to all starting candidate points in descending order. Determine the target functional component identifier corresponding to the starting candidate point with the highest comprehensive confidence score value as the fault source location indication identifier for locating the source location of the fault inside the target power equipment.

[0061] Set weighting coefficients w_range and w_explain (e.g., w_range=0.4, w_explain=0.6). Calculate the overall confidence score S_conf(c_i) = w_range * S_range(c_i) + w_explain * S_explain_norm(c_i). Sort all nodes' S_conf(c_i) from largest to smallest, and take the functional component identifier corresponding to the node with the highest ranking c_max as the fault source location indicator.

[0062] Step S148: Perform dynamic weight decay operation on the directed edges in the spatiotemporal fusion network. Calculate the time decay factor based on the signal acquisition time interval of the corresponding sensing units of the fusion nodes at both ends of the directed edge. Use the time decay factor to correct the temporal dependency transit weight of the directed edge to obtain the time decay corrected temporal dependency transit weight.

[0063] For a directed edge (c_i→c_j), obtain the sampling frequencies f_i and f_j of the corresponding sensing units for c_i and c_j, and calculate the signal acquisition time interval Δt=max(1 / f_i, 1 / f_j). Calculate the time decay factor γ=exp(-Δt / τ), where τ is a preset time constant. Multiply γ by the original time-dependent transit weight w_t to obtain the corrected weight w'_t=γ*w_t.

[0064] Step S150: Based on the fault source location indication identifier and the spatial association knowledge graph, generate a fault diagnosis location report containing the unique identifier code of the fault component and a list of affected components, and push the fault diagnosis location report to the preset operation and maintenance management terminal.

[0065] Step S151: Locate the target spatial network node in the spatial association knowledge graph that matches the target functional component identifier corresponding to the fault source component, and use the target spatial network node as the search center point. Starting from the search center point, expand the search layer by layer outward along the undirected connection edge in the spatial association knowledge graph, record all spatial network nodes passed during the search process, and record the cumulative spatial coupling strength weight value of the undirected connection edge passed from the search center point to each searched spatial network node.

[0066] In the spatial association knowledge graph G_knowledge, find the node corresponding to the fault source location indicator c_fault. Starting from this node, perform a breadth-first search. For each visited node c_visited, calculate the product (or weighted sum) of the weights of all edges on the path from c_fault to c_visited, and use this as the cumulative spatial coupling strength weight value W_cum(c_visited).

[0067] Step S152: Compare the cumulative spatial coupling strength weight value with the preset influence range determination threshold, mark the searched spatial network nodes whose cumulative spatial coupling strength weight value exceeds the preset influence range determination threshold as affected component nodes, and stop the search for the searched spatial network nodes whose cumulative spatial coupling strength weight value does not exceed the preset influence range determination threshold.

[0068] Set an impact range threshold T_influence. During the breadth-first search, for each visited node c_visited, if its W_cum(c_visited) > T_influence, it is marked as an affected component node, and the search for its neighbors continues. If W_cum(c_visited) ≤ T_influence, further diffusion to that node stops.

[0069] Step S153: Collect the target functional component identifiers corresponding to all spatial network nodes marked as affected component nodes, and construct a list of affected component ranges that are spatially coupled with the fault source component. Each item in the list of affected component ranges contains the affected target functional component identifier and the corresponding cumulative spatial coupling strength weight value.

[0070] Iterate through all marked affected component nodes, extract their component identifiers and cumulative spatial coupling strength weight values, and form a list L_affected=[(comp_id1, W1), (comp_id2, W2), ...].

[0071] Step S154: Obtain the mapping relationship of the target functional component identifier corresponding to the fault source component in the device coding database, query the unique identifier code of the fault component that matches the target functional component identifier corresponding to the fault source component, and construct a fault diagnosis and location report containing the unique identifier code of the fault component and the list of affected components. The fault diagnosis and location report also includes a timestamp of the report generation time.

[0072] The system searches the equipment coding database for the unique code (e.g., asset number or material code) corresponding to the faulty component identifier c_fault. It then combines the faulty component's unique identifier code with the affected component range list L_affected, adding the current timestamp to generate a fault diagnosis and location report.

[0073] Step S155: Push the fault diagnosis and location report to the preset operation and maintenance management terminal through the preset communication protocol interface, triggering the operation and maintenance management terminal to generate an audible and visual alarm signal containing the unique identifier code of the faulty component.

[0074] The report is sent to the operations and maintenance management terminal via MQTT or HTTP protocol. After receiving the report, the terminal highlights the location of the faulty component on the screen and issues an alarm sound.

[0075] Step S156: Perform encrypted transmission processing on the fault diagnosis and location report, and use the public key information preset in the operation and maintenance management terminal to perform asymmetric encryption on the fault diagnosis and location report to generate encrypted fault diagnosis and location report ciphertext.

[0076] Before sending the report, the report is encrypted using the RSA public key of the operations and maintenance management terminal to obtain the ciphertext C_report, which is then transmitted over the network. Upon receiving the report, the terminal decrypts it using its own private key to ensure that the report content is not stolen or tampered with.

[0077] Step S210: Obtain the set of physical failure model parameters for key components inside the target power equipment. The set of physical failure model parameters includes the aging rate constant of insulating materials, the fatigue crack propagation index of metallic conductors, the wear coefficient of mechanical parts, and the demagnetization rate parameter of magnetic materials.

[0078] From the design documents or material databases provided by the equipment manufacturer, retrieve the physical failure model parameters of key components. For insulating components (such as transformer winding insulation paper), obtain their aging rate constant k_aging, which is used in the Arrhenius formula to describe the relationship between temperature and aging rate. For metallic conductor components (such as circuit breaker contacts), obtain their fatigue crack propagation index m, which is used in the Paris formula to describe the relationship between crack propagation rate and stress intensity factor amplitude. For mechanical transmission components (such as the gearbox of a tap changer), obtain the wear coefficient K_wear, which describes the influence of contact pressure and sliding distance on wear. For magnetic material components (such as the core of a current transformer), obtain the demagnetization parameter λ_demag, which describes the influence of temperature and time on permeability decay. These parameters are subsequently used to correct the coupling strength weights in the spatial association knowledge graph, making the fault propagation analysis more consistent with physical reality.

[0079] Step S220: After generating the spatial association knowledge graph, the spatial coupling strength weights of the corresponding components in the spatial association knowledge graph are subjected to physical degradation trend weighting correction processing based on the physical failure model parameter set, thereby generating a physically degraded spatial association knowledge graph.

[0080] For each component node c in the spatial association knowledge graph G_knowledge, the corresponding physical failure model parameters are queried based on its component type. The ratio of the component's current runtime t_current to its design life t_life is calculated to obtain the lifespan depletion factor f_life = t_current / t_life. Combining the physical failure model parameters, the degradation correction coefficient δ_degrade is calculated. For example, for an insulating component, δ_degrade = exp(k_aging * f_life); for a metallic conductor component, δ_degrade = (1 + f_life)^m. The spatial coupling strength weights of all connection edges of the component are multiplied by δ_degrade to obtain the degradation-corrected weights. After completing the correction for all nodes, the physically degradation-corrected spatial association knowledge graph G_knowledge_degrade is generated.

[0081] Step S230: Perform a condition change event detection process on the real-time operation status monitoring data stream. By comparing the slope of the signal amplitude change within the sliding window with a preset change slope threshold, identify the change moment when the signal amplitude changes abruptly within a preset short time. Mark the time window corresponding to the change moment as the condition change event window. Extract the sensor response signal segment sequence corresponding to all sensor units within the condition change event window. Analyze the similarity and difference of the amplitude change direction of each sensor unit signal segment sequence within the condition change event window. Group sensor units with the same amplitude change direction into the same source response group and sensor units with opposite amplitude change directions into the opposite source response group.

[0082] Set the sliding window length W_surge (e.g., 5 sampling points). For the signal sequence of each sensing unit, within each window, calculate the amplitude change slope k = (x_{t+W} - x_t) / W. If |k| exceeds the preset abrupt change slope threshold T_slope, the window is determined to be an abrupt change window, and the time corresponding to the last sampling point within the window is the abrupt change time point. Extract the signal segments of all sensing units within ±ΔT around the abrupt change time point. For each sensing unit, calculate the signal mean μ_before of the window before the abrupt change and the signal mean μ_after of the window after the abrupt change. If μ_after > μ_before, the direction of change is marked as "increasing"; otherwise, it is marked as "decreasing". Group all sensing units with the same direction of change into the same source response group, and sensing units with opposite directions of change into the opposite source response group.

[0083] Step S240: Based on the division results of the same-source response group and the heterogeneous response group, adjust the direction of the directed edges between the corresponding time-series network nodes of the sensing units in the same-source response group and between the time-series network nodes of the heterogeneous response group in the time-series state transmission chain network. Input the time-series state transmission chain network after the direction of the directed edges is adjusted and the spatial association knowledge graph after the physical degradation correction into the spatiotemporal fault feature tracing model for joint inference and analysis to generate the fault source location indication.

[0084] In the temporal state propagation chain network G_chain, for any two sensor unit nodes within the same source response group, a directed edge is added from the node with the earlier response time to the node with the later response time, indicating the direction of fault propagation. For nodes in the opposite source response group, a directed edge is added from the node with the significant change in response amplitude to the node with the insignificant change in response amplitude. Existing directed edges that contradict these directional inferences are deleted. The adjusted G_chain and the spatial association knowledge graph G_knowledge_degrade after physical degradation correction are input into the spatiotemporal fault feature tracing model, and steps S141 to S148 are executed to generate fault source location indication markers.

[0085] Step S250: Obtain partial discharge spectrum data of internal components of the target power equipment collected by a portable testing instrument during each planned shutdown and maintenance period. The partial discharge spectrum data of internal components includes the phase distribution characteristics and amplitude distribution characteristics of the discharge pulse.

[0086] During planned equipment shutdowns for maintenance, portable partial discharge detectors (such as high-frequency current transformers, ultrasonic sensors, and UHF antennas) are used to perform live-line testing on the internal components of the equipment. The collected partial discharge spectrum data includes: a histogram of the phase angle distribution of the discharge pulse relative to the AC voltage power frequency cycle, and statistical characteristics of the amplitude distribution of the discharge pulse in different phase intervals (maximum value, average value, and standard deviation).

[0087] Step S260: Extract the phase distribution features of the discharge pulses from the partial discharge spectrum data of the internal components of the device, convert the phase distribution features of the discharge pulses into phase-resolved partial discharge feature vectors, perform similarity matching processing between the phase-resolved partial discharge feature vectors and the reference discharge feature vectors in the preset typical insulation defect discharge fingerprint database, and determine the list of components with suspected insulation defects.

[0088] Phase-resolved partial discharge feature vectors are extracted from the partial discharge pattern data. These feature vectors contain the following components: the proportion of discharge occurrences within a phase interval to the total number of discharges, the asymmetry of discharges in the positive and negative half-cycles, the discharge initiation phase angle, and the discharge extinction phase angle. The feature vectors are then compared with reference vectors in a typical insulation defect discharge fingerprint database using cosine similarity calculation. Fingerprints with similarity exceeding a preset matching threshold (e.g., 0.8) are identified as suspected insulation defect components, along with their associated components (e.g., "winding inter-turn insulation defect," "bushing capacitor core defect," "core multi-point grounding defect"), forming a suspected insulation defect component list L_suspect.

[0089] Step S270: Compare the component identifier in the list of suspected insulation defect components with the fault source location indicator identifier. If the fault source location indicator identifier exists in the list of suspected insulation defect components, then increase the diagnostic confidence level identifier of the fault component in the fault diagnosis and location report.

[0090] Check if the fault source location indication identifier c_fault generated in step S148 is in L_suspect. If it is, in the fault diagnosis location report, upgrade the diagnostic confidence level of the faulty component from "medium" to "high" and add the note "Partial discharge detection results are consistent".

[0091] Step S280: The fault diagnosis and location report is synchronously pushed to the preset spare parts inventory management system, triggering the spare parts inventory management system to retrieve the spare parts inventory quantity and storage location information that match the unique identifier code of the faulty component.

[0092] The unique identifier of the faulty component in the report is sent to the spare parts inventory management system via REST API. The spare parts inventory management system queries the database, returns information such as the inventory quantity, storage warehouse location, and shelf number of the spare part, and automatically generates a spare parts requisition form so that maintenance personnel can prepare replacement materials in advance.

[0093] Step S310: Obtain the design structure blueprint file of the target power equipment. The design structure blueprint file includes a description of the component assembly hierarchy and a description of the electrical connection lines between components. Determine the parent-child relationship of each component inside the target power equipment based on the description of the component assembly hierarchy and determine the energy flow path between each component inside the target power equipment based on the description of the electrical connection lines between components.

[0094] Read the CAD-format design structure blueprint file from the equipment design document. Parse the component assembly tree in the file to determine the parent and child components of each component; for example, "winding" is a child component of "core column," and "core column" is a child component of "core." Parse the electrical schematic diagram to determine the path of current flowing from the input terminals through each component to the output terminals; for example, "bushing → winding → core → oil tank → ground."

[0095] Step S320: Based on the description of the component assembly hierarchy and the description of the electrical connection lines between the components, construct a physical mechanism causal graph model of the target power equipment. The physical mechanism causal graph model includes component hierarchy nodes and directed edges for energy flow. The direction of the directed edges for energy flow is consistent with the direction of energy flow from the upstream component to the downstream component.

[0096] Create a directed graph G_physics. Treat each component as a node in G_physics. Add directed edges from parent components to child components based on the component assembly hierarchy. Add directed edges from components where current flows in to components where current flows out based on electrical connections. Merge these two types of edges to obtain the physical mechanism causal graph model. This model reflects the physical direction of fault energy propagation naturally within the equipment; for example, overvoltage energy propagates from the high-voltage side to the low-voltage side, and mechanical vibration energy propagates from the vibration source to the connector.

[0097] Step S330: After generating the temporal state transit chain network, align the temporal network nodes in the temporal state transit chain network with the component-level nodes in the physical mechanism causal graph model to establish a mapping association table between data-driven temporal correlation and physical mechanism causal relationship.

[0098] Based on the mapping table T_map, the sensor unit nodes in the time-series state transfer chain network G_chain are aligned with the component nodes in the physical mechanism causal graph model G_physics. A mapping association table M_align is created to record the component node corresponding to each sensor unit node, as well as the upstream and downstream neighbor relationships of that component node in G_physics.

[0099] Step S340: Extract the temporal network node identifiers at both ends of each directed edge in the temporal state transfer chain network, find the corresponding component-level node identifiers based on the mapping association table, and determine whether the direction of the directed edge is consistent with the direction of the energy flow between the corresponding component-level nodes in the physical mechanism causal graph model.

[0100] For each directed edge (s_i→s_j) in G_chain, its corresponding component nodes c_i and c_j are found using M_align. G_physics checks if a directed edge exists from c_i to c_j with the direction of the edge aligned with the energy flow. If it exists, the edges are considered consistent; otherwise, if they do not exist, or if a directed edge exists from c_j to c_i, the edges are considered inconsistent.

[0101] Step S350: Mark directed edges with inconsistent pointing directions as physically contradictory abnormal related edges, and mark directed edges with consistent pointing directions as physically consistent related edges. Retain the physically consistent related edges, and remove the physically contradictory abnormal related edges from the temporal state transmission chain network to obtain the target temporal state transmission chain network after physical mechanism consistency pruning. Replace the original temporal state transmission chain network with the target temporal state transmission chain network, and input it into the spatiotemporal fault feature tracing model for joint inference analysis and processing to generate the fault source location indication identifier.

[0102] Remove all directed edges marked as contradictory to physical mechanisms from G_chain, retaining only those that conform to physical mechanisms. This yields the pruned G_chain_pruned. Input G_chain_pruned and the spatial association knowledge graph into the spatiotemporal fault feature tracing model, and re-execute fault source localization to obtain fault source localization indicators that better conform to physical laws.

[0103] Step S360: Obtain the sequence record of switch status change events inside the target power equipment. The sequence record of switch status change events includes circuit breaker opening and closing status change events, protection device action signal trigger events, and corresponding millisecond-level precise event occurrence time tags.

[0104] Read the sequence record (SOE record) of switch status change events from the substation monitoring system. Each record contains: event type (e.g., "circuit breaker A phase trip", "differential protection action activated", "gas relay action"), and the time of event occurrence (accurate to milliseconds). The above records provide a precise time reference point for equipment status changes.

[0105] Step S370: Extract the protection device action signal triggering events and their millisecond-level precise event occurrence time tags within the preset backtracking time period in the sequence record of the switch state change event. Using the millisecond-level precise event occurrence time tag as the time reference benchmark, trace back to the abnormal fluctuation start time of the corresponding sensing response signal segment sequence of each sensing unit in the time sequence state transmission chain network.

[0106] Extract the time t_protect of the most recent protection device action signal trigger event (such as "differential protection action") from the SOE record. Using t_protect as a reference, backtrack by a time window T_back (e.g., 100 milliseconds) to analyze the signal sequence of each sensing unit and find the time t_deviation(i) when the signal first deviates from the normal operating range.

[0107] Step S380: Calculate the absolute value of the time difference between the start time of the abnormal fluctuation of the sensing response signal segment sequence corresponding to each sensing unit and the millisecond-level precise event occurrence time tag of the protection device action signal trigger event. Mark the timing network node corresponding to the sensing unit with the smallest absolute value of the time difference as the monitoring node most relevant to the protection action event. Superimpose the information of the monitoring node most relevant to the protection action event into the fault diagnosis and location report to form an enhanced fault diagnosis and location report that supplements the protection action-related monitoring point information.

[0108] Calculate Δt_i = |t_protect - t_deviation(i)|. Find the sensing unit s_min with the smallest Δt_i, mark it as the "monitoring point most relevant to protection action" and add it to the fault diagnosis and location report, while recording the time difference Δt_min. Send the enhanced report to the operation and maintenance terminal.

[0109] Step S410: Obtain hourly environmental meteorological element sequences recorded by multiple external environmental meteorological monitoring stations for a continuous preset time period for the target power equipment. The hourly environmental meteorological element sequences include environmental temperature numerical sequences, environmental relative humidity numerical sequences, and atmospheric pressure numerical sequences.

[0110] Hourly meteorological data for the past 72 hours, including ambient temperature T_env(t), ambient relative humidity RH(t), and atmospheric pressure P_atm(t), are obtained from meteorological monitoring stations in the area where the equipment is located. This data is used to analyze the impact of environmental factors on the equipment monitoring signal.

[0111] Step S420: The hourly environmental meteorological element sequence and the sensor response signal segment sequence corresponding to each sensor unit in the real-time operation status monitoring data stream are time-aligned and spliced ​​to form an extended sensor response signal segment sequence with additional meteorological background information. The data item at each time point in the extended sensor response signal segment sequence with additional meteorological background information includes the sensor response signal value and the corresponding ambient temperature value and ambient relative humidity value.

[0112] For each sensing unit's signal sequence X(t), the extended sequence X_ext(t) = [X(t), T_env(t), RH(t), P_atm(t)] is generated by aligning it with the meteorological data T_env(t), RH(t), P_atm(t) according to the timestamp.

[0113] Step S430: Perform meteorological factor and signal drift correlation analysis on the extended sensor response signal segment sequence with additional meteorological background information, calculate the temperature drift correlation coefficient between the sensor response signal value and the ambient temperature value, and the humidity drift correlation coefficient between the sensor response signal value and the ambient relative humidity value, to obtain the temperature drift correlation coefficient and humidity drift correlation coefficient corresponding to each sensor unit. Based on the temperature drift correlation coefficient and humidity drift correlation coefficient corresponding to each sensor unit, when generating the time-series state transmission chain network, perform a meteorological drift component stripping operation on the signal amplitude fluctuation pattern in the real-time operating status monitoring data stream, and subtract the temperature drift component formed by the product of the temperature drift correlation coefficient and the ambient temperature change, and the humidity drift component formed by the product of the humidity drift correlation coefficient and the ambient relative humidity change from the signal amplitude fluctuation pattern.

[0114] For each sensing unit, the sensing response signal value X(t) in the extended sequence is used as the dependent variable, and the ambient temperature T_env(t) and ambient relative humidity RH(t) are used as independent variables to establish a multiple linear regression model X(t) = a*T_env(t) + b*RH(t) + c + ε. The temperature drift correlation coefficient a and the humidity drift correlation coefficient b are obtained by solving the model. The drift component caused by weathering in the signal D_weather(t) = a*(T_env(t) - T_ref) + b*(RH(t) - RH_ref) is calculated, where T_ref and RH_ref are baseline values ​​(e.g., 20℃ and 50%). The weathering drift component is subtracted from the original signal to obtain the stripped signal X_clean(t) = X(t) - D_weather(t). Steps S121 to S125 are re-executed using X_clean(t) to generate the weathering stripped temporal state transfer chain network G_chain_weather.

[0115] Step S440: Replace the original signal amplitude fluctuation pattern with the signal amplitude fluctuation pattern after removing the meteorological drift component, and re-execute the step of analyzing the transmission evolution law of the signal amplitude fluctuation pattern along the time axis in the sequence of sensor response signal segments to generate a meteorological stripping time-series state transmission chain network that reflects the internal state migration path of the target power equipment. Input the meteorological stripping time-series state transmission chain network and the spatial association knowledge graph into the spatiotemporal fault feature tracing model for joint inference and analysis to generate the fault source location indication.

[0116] By inputting G_chain_weather and the spatial association knowledge graph G_knowledge into the spatiotemporal fault feature tracing model, the fault source localization is performed, and the fault localization result is obtained after eliminating environmental interference.

[0117] Step S450: Obtain the system frequency disturbance event record published by the power grid dispatch automation system in the area where the target power equipment is located. The system frequency disturbance event record includes the start time of the frequency deviation from the rated value and the frequency deviation magnitude value.

[0118] Read the system frequency disturbance event record from the power grid dispatch automation system. For example, “2024-05-15 14:23:17.500, system frequency dropped to 49.85Hz, duration 200 milliseconds”. The above record indicates that a disturbance event has occurred in the power grid, which may cause synchronous fluctuations in the equipment monitoring signal.

[0119] Step S460: Based on the start time of the frequency deviation from the rated value in the system frequency disturbance event record, when generating the real-time operation status monitoring data stream, mark the frequency disturbance event for the sensor response signal segment sequence corresponding to all sensor units within the start time of the frequency deviation from the rated value and the subsequent preset time period.

[0120] When generating real-time operational status monitoring data streams, for each signal segment of a sensing unit, if the timestamp of the segment falls within ±ΔT_freq of the start time of the system frequency disturbance event, then the metadata of the segment is labeled "freq_disturbance=1", otherwise it is 0.

[0121] Step S470: Obtain the vibration acceleration sensing response signal segment sequence of the rotating mechanical components in the target power equipment, perform envelope spectrum analysis on the vibration acceleration sensing response signal segment sequence, extract the amplitude sequence of the fault characteristic frequency harmonic component in the envelope spectrum, and use the amplitude sequence of the fault characteristic frequency harmonic component as an additional vibration fault feature vector.

[0122] For vibration acceleration signals of rotating mechanical components (such as transformer cooling fans and on-load tap changer motors), bandpass filtering is first performed (the center frequency is the component fault characteristic frequency, such as the bearing outer ring fault frequency), then Hilbert transform is performed to obtain the envelope signal, and then fast Fourier transform is performed on the envelope signal to obtain the envelope spectrum. The amplitudes of the 1st, 2nd, and 3rd harmonic fault characteristic frequencies in the envelope spectrum are extracted to form the vibration fault feature vector F_vib=[A1, A2, A3].

[0123] Step S480: When inputting the temporal state transfer chain network and the spatial association knowledge graph into the spatiotemporal fault feature tracing model, the vibration fault feature vector is also input as an auxiliary feature into the spatiotemporal fault feature tracing model for joint inference and analysis.

[0124] During the random walk in step S144, when the walk reaches the node corresponding to the rotating mechanical component, the vibration fault feature vector F_vib of that node is used as an additional feature for selecting neighboring nodes: the similarity between the vibration fault feature of the neighboring node c_next and the vibration fault feature of the current node c_cur is calculated, and this similarity is used as an additional multiplier factor for the edge weight to enhance the walk probability between nodes with similar vibration features.

[0125] For example, the method may further include: step S510: obtaining a historical fault case library of the target power equipment, wherein each case record in the historical fault case library includes a unique identifier field of the fault component, a sensor data snapshot field at the time of the fault, and an identifier field of the actual fault source component, wherein the sensor data snapshot field at the time of the fault stores a truncated copy of the sensor response signal segment sequence corresponding to all sensor units within a preset time period before and after the time of the fault.

[0126] Retrieves all fault cases recorded in the past 5 years from the equipment operation and maintenance database. Each case includes: the actual confirmed fault component identifier at the time of the fault (e.g., "B-phase high-voltage bushing"), a snapshot of signal segments from all sensing units within ±10 seconds before and after the fault occurred, and the diagnostic conclusions of the maintenance personnel after the fault occurred.

[0127] Step S520: Extract the actual fault source component identifier field from each case record in the historical fault case library, and in the sensor data snapshot field at the fault time corresponding to the case record, calculate the cross-correlation function between the sensor response signal segment sequence of the other sensor units and the sensor response signal segment sequence of the central sensor unit, with the sensor unit corresponding to the actual fault source component identifier as the center; extract the time delay value corresponding to the maximum peak value of the cross-correlation function, and use the time delay value as the fault symptom propagation time delay parameter of the corresponding sensor unit relative to the fault source component; traverse all case records in the historical fault case library, and count the mode of the time delay values ​​that appear in different cases for each pair of sensor unit combinations as the empirical propagation time delay reference value of the sensor unit combination, and use it as the prior time delay constraint information of the time-series network nodes corresponding to the two sensor units under the direction of the directed edge, and attach it to the directed edge attribute of the time-series state transmission chain network to obtain the time-series state transmission chain network with attached prior time delay constraint information.

[0128] For each fault case, taking the signal of the sensing unit s_root corresponding to the fault source component as a reference, calculate the cross-correlation function R(τ) = ∫X_s_root(t)·X_s_i(t+τ)dt between the signals of other sensing units s_i and the signal of s_root. Find the τ_i corresponding to the maximum peak value of R(τ), which is the propagation delay of s_i relative to s_root. Collect the delay values ​​of the same pair (s_root, s_i) in all cases, and take the delay value that appears most frequently as the empirical propagation delay reference value τ_ref(s_root, s_i). In the temporal state propagation chain network G_chain, attach the attribute delay=τ_ref to the directed edge (s_root→s_i).

[0129] Step S530: When inputting the temporal state transfer chain network with the additional prior time delay constraint information and the spatial association knowledge graph into the spatiotemporal fault feature tracing model for joint inference and analysis, the empirical propagation time delay reference value attached to the directed edge attribute is used as the time delay weighting factor for selecting the next walk fusion node in the random walk propagation process. The time delay weighting factor is inversely correlated with the empirical propagation time delay reference value.

[0130] In the random walk of step S144, when selecting the next node c_next from the current node c_cur, the time delay weighting factor w_delay=exp(-τ_ref / τ_0) for the directed edge (c_cur→c_next) is calculated, where τ_0 is a normalization constant. This time delay weighting factor is inversely proportional to the propagation delay, meaning that edges with shorter delays have a higher probability of being selected. The joint probability distribution P(c_cur→c_next)=w_t*w_s*w_delay / Σ(...).

[0131] Step S540: After joint inference and analysis processing that integrates the time delay weighting factors, the fault source location indication identifier is generated.

[0132] Perform steps S144 to S148 to obtain the fault source location indication.

[0133] Step S550: Obtain real-time time synchronization deviation monitoring data of multiple sensing units deployed on the target power equipment. Based on the real-time time synchronization deviation monitoring data, perform time deviation correction processing on the absolute time tag attached to the sensing response signal segment sequence of each sensing unit in the real-time operation status monitoring data stream. Add or subtract the corresponding current time deviation value to the absolute time tag to obtain the time-synchronized corrected absolute time tag. The real-time time synchronization deviation monitoring data records the current time deviation value between the internal clock of each sensing unit and the unified time reference source.

[0134] Each sensing unit contains an internal clock chip, which synchronizes with a unified time reference source (such as a GPS clock) via the NTP protocol or IRIG-B code. However, slight deviations may exist, and the deviation value δ_i is recorded in real time by the time synchronization monitoring module. When generating the sequence of sensing response signal segments, the original timestamp of each signal point is t_raw, and the corrected timestamp t_corr = t_raw - δ_i (if the sensing unit clock is ahead) or t_corr = t_raw + δ_i (if it is behind).

[0135] Step S560: Obtain the historical load current curve data of the target power equipment. The historical load current curve data records the curve trajectory of the current carrying capacity of the target power equipment changing with time in a continuous operating cycle. Extract the time of occurrence of periodic peaks and valleys in the historical load current curve data to form a time table of periodic load fluctuations.

[0136] Load current data from the past year is read from the equipment monitoring system, with a sampling interval of 1 minute. Peak current peak and valley times are identified daily, weekly, and monthly using a peak detection algorithm, forming a timetable of load periodic fluctuations.

[0137] Step S570: When generating the sensor response signal segment sequence in the real-time operation status monitoring data stream, the sensor response signal segments at the time of periodic peak occurrence are marked with load peak periods according to the load periodic fluctuation time table. In the subsequent time-series dependency mining operation, the signal amplitude fluctuation pattern of the sensor response signal segment sequence with load peak period marking is distinguished from the signal amplitude fluctuation pattern of the unmarked sensor response signal segment sequence.

[0138] When generating a signal segment, check if the segment's timestamp falls within ±Δt_peak of the peak time in the load periodic fluctuation timetable. If so, add the tag "load_peak=1" to the segment's metadata. In the fluctuation pattern encoding of step S121, perform independent pattern statistics on tagged and untagged segments respectively, generating two sets of fluctuation pattern characterization codes to avoid misjudging normal signal changes caused by load fluctuations as fault symptoms.

[0139] Step S610: Obtain the equipment outage and maintenance event log recorded by the target power equipment within a preset historical operating cycle. The equipment outage and maintenance event log includes the maintenance start time tag, maintenance end time tag, replacement component identifier involved in the maintenance operation, and equipment commissioning time tag for each outage and maintenance event.

[0140] Retrieve the equipment shutdown and maintenance logs for the past 3 years from the equipment management system. Each record includes: maintenance start time t_start_repair, maintenance end time t_end_repair, a list of component identifiers replaced during maintenance (e.g., ["B-phase bushing", "tap switch contact"]), and equipment restart time t_commission after maintenance.

[0141] Step S620: Extract the replacement component identifiers involved in each maintenance operation from the equipment shutdown and maintenance event log, associate and locate the replacement component identifiers with the corresponding time-series network nodes in the time-series state transmission chain network, and determine the set of time-series network nodes corresponding to the sensing unit where the component replacement operation occurred.

[0142] For each maintenance, the list of replacement component identifiers is extracted, and the corresponding sensing units of these components are found through the mapping table T_map, thereby locating the corresponding node set N_replaced in the time-series state transfer chain network G_chain.

[0143] Step S630: In the time-series state transmission chain network, the complete influence propagation link starting from any node in the time-series network node set corresponding to the sensing unit where the component replacement operation occurred is segmented and truncated. Taking the post-maintenance equipment commissioning time label as the time boundary point, the complete influence propagation link crossing the time boundary point is divided into a pre-maintenance propagation sub-link and a post-maintenance propagation sub-link.

[0144] For each complete impact propagation link L in G_chain, if L contains nodes from N_replaced and L spans the maintenance and commissioning time t_commission, then L is cut off at t_commission: the part before t_commission is the pre-maintenance propagation sub-link L_before, and the part after t_commission is the post-maintenance propagation sub-link L_after.

[0145] Step S640: Perform link topology comparison analysis on the pre-maintenance propagation sub-link and the post-maintenance propagation sub-link respectively, extract the first link topology fingerprint formed by the connection order of the time-series network nodes and the direction of the directed edges in the pre-maintenance propagation sub-link, and extract the second link topology fingerprint formed by the connection order of the time-series network nodes and the direction of the directed edges in the post-maintenance propagation sub-link.

[0146] For L_before, its node order and edge direction are encoded into a string, such as "A→B→C→D", which serves as the first link topology fingerprint F_before. Similarly, L_after is encoded as F_after.

[0147] Step S650: Calculate the link topology structure difference degree between the first link topology fingerprint and the second link topology fingerprint, mark the sensor unit pairs whose link topology structure difference degree exceeds the preset topology variation threshold as post-maintenance propagation characteristic change association pairs, and record the post-maintenance propagation characteristic change association pairs in the propagation characteristic variation record table.

[0148] Calculate the edit distance between F_before and F_after (the minimum number of edit operations required to convert one string to another). If the edit distance is greater than a preset threshold D_edit (e.g., 2), mark the corresponding sensor unit pair as a post-maintenance propagation characteristic change association pair and record it in the variation record table T_variation.

[0149] Step S660: Construct a deep temporal graph neural network model. The deep temporal graph neural network model includes a graph convolutional feature extraction layer, a temporal attention convergence layer, and a link prediction output layer. The graph convolutional feature extraction layer is used to perform neighbor node feature aggregation on the input network topology. The temporal attention convergence layer is used to perform weighted fusion on the node representation vectors at different time steps. The link prediction output layer is used to generate the probability value of the existence of directed connection edges between nodes.

[0150] A graph neural network model (GNN_model) is constructed. The graph convolutional feature extraction layer uses a graph convolutional network, with the update formula H^{(l+1)}=σ(D^{-1 / 2}AD^{-1 / 2}H^{(l)}W^{(l)}), where A is the adjacency matrix, D is the degree matrix, H^{(l)} represents the node features of the l-th layer, and W^{(l)} is the learnable weight matrix. The temporal attention convergence layer uses a multi-head attention mechanism to weightedly fuse node representations from different time steps. The link prediction output layer uses an inner product decoder to calculate the probability p_ij=sigmoid(h_i^Th_j) that an edge exists between node i and node j.

[0151] Step S670: Use the temporal state transfer chain network segment corresponding to the pre-maintenance propagation sub-link as positive sample training data, and the false propagation link generated by random perturbation as negative sample training data. Input the data into the deep temporal graph neural network model for iterative optimization training of model parameters. By minimizing the cross-entropy loss value between the probability value output by the link prediction output layer and the real link label, adjust the network weight parameters of the graph convolutional feature extraction layer and the temporal attention convergence layer.

[0152] Extract all directed edges from the pre-maintenance propagation sub-links as positive samples. Randomly generate the same number of node pairs that do not exist in the real network as negative samples. Input the samples into GNN_model and output the predicted probability p_ij. Calculate the cross-entropy loss L=-[y_ijlog(p_ij)+(1-y_ij)log(1-p_ij)], where y_ij is the true label (1 indicates an edge exists, 0 indicates no edge exists). Update the network parameters using backpropagation with the Adam optimizer.

[0153] Step S680: Input the post-maintenance propagation sub-link into the trained deep temporal graph neural network model, perform neighbor feature aggregation on each temporal network node in the post-maintenance propagation sub-link through the graph convolutional feature extraction layer to generate a node-level topological representation vector, perform temporal dependency modeling on the node-level topological representation vector of continuous time steps through the temporal attention convergence layer to generate a node-level spatiotemporal fusion representation vector, input the node-level spatiotemporal fusion representation vector into the link prediction output layer to generate a predicted probability value for forming a new directed connection edge between any two temporal network nodes in the post-maintenance propagation sub-link, and mark the node pairs whose predicted probability values ​​exceed a preset new link generation threshold as potential new propagation path candidate pairs.

[0154] The node features of the post-repair propagation sub-link L_after are input into the trained GNN_model to obtain the spatiotemporal fusion representation vector of each node. For any two nodes i and j, p_ij = sigmoid(h_i^Th_j). If p_ij > 0.7, then (i, j) is marked as a potential candidate pair for new propagation paths.

[0155] Step S690: Add the potential new propagation path candidate pairs to the temporal state transit chain network as supplementary directed edges, and assign initial temporal dependency transit weights to the supplementary directed edges according to the predicted probability values ​​to generate an augmented temporal state transit chain network containing potential new propagation paths after component replacement. Input the augmented temporal state transit chain network into the spatiotemporal fault feature tracing model to generate the fault source location indication identifier.

[0156] Add a new directed edge (i→j) to G_chain, with edge weight w_new=p_ij. This yields the augmented network G_chain_aug. Input G_chain_aug into the spatiotemporal fault feature tracing model to perform fault source localization, obtaining localization results that reflect potential new fault propagation paths after component replacement.

[0157] This deep temporal graph neural network model employs a hybrid architecture based on gated recurrent units and graph convolutional networks. The graph convolutional feature extraction layer consists of two graph convolutional network layers: the first layer maps the node input feature dimension from d_node to d_gcn1, and the second layer maps d_gcn1 to d_gcn2. The temporal attention convergence layer uses a Transformer encoder architecture, containing two identical stacked encoder layers. Each encoder layer includes a multi-head self-attention sub-layer (with 4 attention heads) and a feedforward neural network sub-layer (hidden layer dimension d_ff, output dimension d_gcn2). Each sub-layer is followed by layer normalization and residual connections. The link prediction output layer uses an inner product decoder to calculate the probability p_ij = sigmoid(h_i^Th_j) that there is an edge between node i and node j, where h_i and h_j are the final representation vectors of node i and node j after graph convolution and temporal attention processing.

[0158] The training process of this deep temporal graph neural network model is as follows. The training dataset comes from the temporal state transfer chain network segment corresponding to the pre-maintenance propagation sub-link, containing a total of 8000 positive samples (real-existing directed edges) and 8000 negative samples (randomly generated non-existent node pairs). Input format: For each sample, input the feature vectors of two nodes (node ​​features include the historical signal statistical features of the corresponding sensor unit, node degree, clustering coefficient, and other topological features), and output 0 or 1 to indicate whether an edge exists. The loss function adopts the binary classification cross-entropy loss L=-[ylog(p)+(1-y)log(1-p)]. The optimizer uses Adam, with an initial learning rate set to 0.0005, a batch size set to 64, and 150 training epochs. Evaluation metrics include precision, recall, and F1 score.

[0159] When applying the model, the feature vectors of all nodes in the post-maintenance propagation sub-link are input into the trained model. The node-level topological representation vector is obtained through the graph convolution feature extraction layer, and the node-level spatiotemporal fusion representation vector is obtained through the temporal attention convergence layer. Then, the inner product is calculated for each pair of nodes and the prediction probability is obtained through the sigmoid function. Node pairs with a prediction probability greater than 0.7 are marked as potential new propagation path candidate pairs.

[0160] Step S710: Construct an adversarial generative fault sample expansion model for target power equipment. The adversarial generative fault sample expansion model includes a fault signal generator network and a fault signal discriminator network. The fault signal generator network is composed of cascaded layers of multi-layer transposed convolutional neural networks, and the fault signal discriminator network is composed of cascaded layers of multi-layer convolutional neural networks.

[0161] Construct a generative adversarial network (GAN). The generator G takes a random noise vector z (length 100) as input, passes it through 5 transposed convolutional layers, each followed by batch normalization and ReLU activation, progressively upsampling it to output a fake signal segment of the same length as the real sensor signal segment (e.g., 1000 sampling points). The discriminator D takes the signal segment as input, passes it through 5 convolutional layers, each followed by LeakyReLU activation and Dropout, and finally outputs a scalar representing the probability that the input is a real sample through a fully connected layer.

[0162] Step S720: Obtain a set of sensor response signal fragment sequences of the target power equipment within a preset acquisition window before and after each real fault occurrence time, and use the set of sensor response signal fragment sequences as a real fault signal sample library. Each sample in the real fault signal sample library is accompanied by a corresponding real fault source component identification tag.

[0163] Signal segments from all sensing units within ±5 seconds before and after each fault occurrence are extracted from the historical fault case database to form the real fault signal sample database X_real. Each sample is labeled y_real, indicating the identifier of the fault source component corresponding to that sample.

[0164] Step S730: Input the random noise vector into the fault signal generator network, and perform layer-by-layer upsampling and feature mapping operations on the random noise vector through the multi-layer transposed convolutional neural network layer to generate an artificially synthesized fault signal segment sequence with the same dimensions as the real sensor response signal segment sequence. Mix the artificially synthesized fault signal segment sequence with a real sensor response signal segment sequence randomly selected from the real fault signal sample library and input it into the fault signal discriminator network. Perform layer-by-layer convolution and pooling operations on the input signal segment sequence through the multi-layer convolutional neural network layer, and output a discrimination probability value for determining whether the input signal segment sequence comes from the real fault signal sample library or from the fault signal generator network.

[0165] Sample z from the noise distribution p(z) and input it into G to generate a fake signal X_fake=G(z). Randomly select a batch of real samples X_batch_real from X_real. Mix X_fake and X_batch_real and input them into D. D outputs the discrimination probabilities d_real=D(X_batch_real) and d_fake=D(X_fake).

[0166] Step S740: With the goal of maximizing the difference between the discrimination probability value of the fault signal discriminator network for the real sensor response signal segment sequence and the discrimination probability value for the artificially synthesized fault signal segment sequence, update the network weight parameters of the fault signal discriminator network. With the goal of minimizing the discrimination probability value of the fault signal discriminator network for the artificially synthesized fault signal segment sequence, update the network weight parameters of the fault signal generator network.

[0167] The discriminator loss L_D = -[log(D(X_real)) + log(1 - D(G(z)))], and the generator loss L_G = -log(D(G(z))). The weights of D and G are updated using these two loss functions respectively.

[0168] Step S750: Alternately and iteratively execute the weight parameter update operations of the fault signal discriminator network and the fault signal generator network until the artificially synthesized fault signal segment sequence generated by the fault signal generator network can make the discrimination probability value output by the fault signal discriminator network converge to a preset equilibrium interval, thereby obtaining the trained fault signal generator network. Use the trained fault signal generator network to generate a massive number of artificially synthesized fault signal segment sequences. Combine each artificially synthesized fault signal segment sequence with the corresponding fault source component identification tag to form an artificially synthesized fault sample. Merge the artificially synthesized fault samples with the real fault signal sample library to form an augmented fault signal sample library.

[0169] D and G are trained alternately until D(G(z)) converges to around 0.5. At this point, the signal generated by G is difficult to distinguish from the real signal. 100,000 artificially synthesized fault signal segments are generated using G, and each is randomly assigned a fault source component identifier label. These are then merged with the real sample library to obtain the augmented sample library X_aug.

[0170] Step S760: Based on the augmented fault signal sample library, perform secondary reinforcement training on the spatiotemporal fault feature tracing model, replace the corresponding part in the original real-time operation status monitoring data stream with the sensor response signal segment sequence of each sample in the augmented fault signal sample library, re-execute the temporal dependency mining operation and spatial topology association parsing operation, and generate the sample temporal state transmission chain network and the sample spatial coupling association graph.

[0171] For each sample in the augmented sample library, its signal segment sequence is used as the input of the sensing unit. Steps S120 to S130 are executed to generate the temporal state transfer chain network G_sample and the spatial association knowledge graph G_sample_space corresponding to the sample.

[0172] Step S770: Input the sample temporal state transmission chain network and the sample spatial coupling correlation graph into the spatiotemporal fault feature tracing model, obtain the predicted fault source location indicator output by the spatiotemporal fault feature tracing model, and calculate the location error loss value between the predicted fault source location indicator and the actual fault source component identification label attached to the sample.

[0173] Input G_sample and G_sample_space into the spatiotemporal fault feature tracing model to obtain the predicted fault source identifier c_pred. Calculate the localization error loss L_loc = 0 if c_pred equals the true label, otherwise L_loc = 1.

[0174] Step S780: With minimizing the positioning error loss value as the optimization objective, the weight parameters of each network layer inside the spatiotemporal fault feature tracing model are iteratively fine-tuned and updated using the backpropagation algorithm until the positioning error loss value converges to a preset loss threshold, thus obtaining the spatiotemporal fault feature tracing model after adversarial generative sample enhancement training, which is used for subsequent fault source location deduction of the real-time operating status monitoring data stream of the target power equipment.

[0175] The spatiotemporal fault feature tracing model was fine-tuned using an augmented sample library to optimize model parameters, enabling it to recognize rare fault types with high accuracy. After training, the model was deployed to an online monitoring system for real-time fault diagnosis.

[0176] This adversarial generative fault sample augmentation model comprises a fault signal generator network and a fault signal discriminator network. The fault signal generator network adopts a transposed convolutional neural network architecture, containing one fully connected layer and five transposed convolutional layers. The fully connected layer maps the input 100-dimensional random noise vector to 1024 dimensions, reshaping it into a (1, 32, 32) tensor. The transposed convolutional layers are configured as follows: Transposed Convolutional Layer 1 (32 input channels, 256 output channels, kernel size 5x5, stride 2, padding 2, output size 64x64), Transposed Convolutional Layer 2 (256→128, 5x5, stride 2, padding 2, output 128x128), Transposed Convolutional Layer 3 (128→64, 5x5, stride 2, padding 2, output 256x256), Transposed Convolutional Layer 4 (64→32, 5x5, stride 2, padding 2, output 512x512), and Transposed Convolutional Layer 5 (32→1, 5x5, stride 2, padding 2, output 1024x1024). Each transposed convolutional layer is followed by a batch normalization layer and a ReLU activation function (the last layer uses the Tanh activation function). The fault signal discriminator network adopts a convolutional neural network architecture, containing 5 convolutional layers and 2 fully connected layers. The convolutional layers are configured as follows: Convolutional Layer 1 (1 input channel, 32 output channels, kernel size 5x5, stride 2, padding 2), Convolutional Layer 2 (32→64, 5x5, stride 2, padding 2), Convolutional Layer 3 (64→128, 5x5, stride 2, padding 2), Convolutional Layer 4 (128→256, 5x5, stride 2, padding 2), and Convolutional Layer 5 (256→512, 5x5, stride 2, padding 2). Each convolutional layer is followed by a batch normalization layer and a LeakyReLU activation function (slope 0.2). The outputs of the convolutional layers are flattened and fed into two fully connected layers (512→256, 256→1), and finally, the discriminant probability is output through a Sigmoid activation function.

[0177] The training process of this adversarial generative fault sample augmentation model is as follows. The training dataset comes from a real fault signal sample library, containing a total of 5000 real fault signal segment sequences, each with a length of 1024 sampling points. Input format: The generator's input is a 100-dimensional random noise vector, and the output is a 1024-dimensional artificially synthesized fault signal segment sequence; the discriminator's input is a 1024-dimensional signal segment sequence, and the output is a discrimination probability between 0 and 1. Loss function: Generator loss L_G=-log(D(G(z))), discriminator loss L_D=-[log(D(x_real))+log(1-D(G(z)))]. The optimizer uses Adam, with an initial learning rate set to 0.0002, beta1 parameter set to 0.5, and beta2 parameter set to 0.999. The batch size is set to 64, and the training epochs are set to 500. An alternating training strategy is adopted, updating the generator after each update of the discriminator. The evaluation metric used is the Fréchet initial distance, which measures the similarity between the generated samples and the real sample distribution.

[0178] In model application, a 100-dimensional noise vector is randomly sampled from a standard normal distribution and input into the trained fault signal generator network. After forward propagation, it outputs a sequence of artificially synthesized fault signal fragments. This sequence is combined with the corresponding fault source component identifiers to form artificially synthesized fault samples, which are used to expand the fault signal sample library.

[0179] In one exemplary embodiment, a fault diagnosis and location system for power equipment operation and maintenance is provided. This system can be a terminal, server, etc., and its internal structure diagram can be as follows: Figure 2As shown, the fault diagnosis and location system for power equipment operation and maintenance includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements a fault diagnosis and location method for power equipment operation and maintenance. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the casing of a fault diagnosis and positioning system used for power equipment operation and maintenance, or an external keyboard, touchpad, or mouse, etc.

[0180] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A fault diagnosis and location method applied to the operation and maintenance of power equipment, characterized in that, The method includes: Acquire real-time operating status monitoring data stream of target power equipment, wherein the real-time operating status monitoring data stream includes a sequence of sensor response signal segments continuously collected in chronological order by multiple sensor units deployed on the target power equipment; A time-series dependency mining operation is performed on the real-time operating status monitoring data stream. By analyzing the transmission and evolution of signal amplitude fluctuation patterns along the time axis in the sensor response signal segment sequence, a time-series state transmission chain network reflecting the internal state transition path of the target power equipment is generated. A spatial topology association parsing operation is performed on the real-time operation status monitoring data stream. By analyzing the degree of correlation between the deployment location adjacency relationship of the multiple sensing units on the physical structure of the target power equipment and the synchronous change of their respective sensing response signal segment sequences, a spatial association knowledge graph characterizing the mutual influence relationship between the components of the target power equipment is generated. The temporal state transmission chain network and the spatial association knowledge graph are input into a preset spatiotemporal fault feature tracing model for joint inference and analysis to generate a fault source location indicator for locating the source of the fault inside the target power equipment. Based on the fault source location indication and the spatial association knowledge graph, a fault diagnosis and location report is generated, which includes the unique identifier code of the faulty component and a list of affected components. The fault diagnosis and location report is then pushed to a preset operation and maintenance management terminal.

2. The fault diagnosis and location method for power equipment operation and maintenance according to claim 1, characterized in that, The step of performing time-series dependency mining on the real-time operating status monitoring data stream involves analyzing the transmission and evolution of signal amplitude fluctuation patterns along the time axis in the sensor response signal segment sequence to generate a time-series state transmission chain network reflecting the internal state transition path of the target power equipment, including: The sensor response signal segment sequence corresponding to each sensor unit in the real-time operation status monitoring data stream is expanded along the time axis into a signal amplitude sequence with a sequential order. The signal amplitude sequence is then divided into time windows to obtain a set of signal amplitude segments consisting of continuous and partially overlapping signal amplitude segments within the time window. Fluctuation pattern encoding is then performed on the signal amplitude segments within each time window of the signal amplitude segment set to extract the arrangement order features of the alternating local peaks and troughs and the amplitude difference features between adjacent peaks and troughs, generating a local fluctuation pattern characterization code that uniquely corresponds to each time window. All local fluctuation patterns belonging to the same sensing unit and arranged in chronological order are encoded to construct a fluctuation pattern evolution sequence corresponding to that sensing unit. State transition probability analysis is performed on the fluctuation pattern evolution sequence to statistically analyze the frequency distribution of transitions between adjacent local fluctuation pattern encodings, generating a fluctuation state transition probability distribution matrix corresponding to that sensing unit. The fluctuation state transition probability distribution matrices corresponding to all sensing units are traversed. According to the preset state transition similarity measurement rules, the similarity value of the state transition trajectory between the fluctuation state transition probability distribution matrices corresponding to any two sensing units is calculated. Sensing unit pairs whose similarity value of the state transition trajectory exceeds the preset similarity threshold are marked as associated sensing unit pairs with temporal correlation. The associated sensing unit pairs are used as nodes in a temporal network, and the similarity value of the state transition trajectory is used as the edge weight of the directed edge connecting the temporal network nodes to construct a preliminary temporal directed association network. A network path backtracking search operation is performed on the preliminary temporal directed association network. Starting from each temporal network node, the downstream associated nodes are searched level by level along the direction of the directed edges. The sequence of temporal network nodes passed through in each search process is recorded, and a forward influence propagation path corresponding to each temporal network node is generated. A reverse path backtracking search operation is also performed on the preliminary temporal directed association network. Starting from each temporal network node, the upstream associated nodes are searched level by level in the opposite direction of the directed edges. The sequence of temporal network nodes passed through in each reverse search process is recorded, and a reverse influence propagation path corresponding to each temporal network node is generated. The forward influence propagation path and the reverse influence propagation path corresponding to the same time-series network node are spliced ​​and processed to form a complete influence propagation link with the time-series network node as the central node. The complete influence propagation links corresponding to all time-series network nodes are deduplicated and integrated to generate a time-series state transfer chain network that reflects the internal state transition path of the target power equipment. Obtain historical maintenance records of the target power equipment, and mark the fault propagation path correlation of the link where the corresponding component is located in the time-series state transmission chain network according to the historical maintenance records. In the time-series state transmission chain network, mark the complete impact propagation link that passes through the component as a verified fault propagation path. The historical maintenance records include the specific component identifiers targeted by each maintenance operation and the health status description text of the component corresponding to the specific component identifier after the maintenance operation.

3. The fault diagnosis and location method for power equipment operation and maintenance according to claim 1, characterized in that, The spatial topology association parsing operation performed on the real-time operation status monitoring data stream generates a spatial association knowledge graph characterizing the mutual influence relationships between components of the target power equipment by analyzing the adjacency relationship of the deployment positions of the multiple sensing units on the physical structure of the target power equipment and the degree of correlation between the synchronous changes of their respective sensing response signal segment sequences. This includes: Obtain the physical assembly structure topology diagram of the target power equipment. The physical assembly structure topology diagram includes the component identifiers of each functional component inside the target power equipment and a description of the mechanical connection relationship type between any two functional components. Multiple sensing units deployed on the target power equipment are matched with the functional components in the physical assembly structure topology diagram by performing position mapping and matching processing. The target functional component identifier directly monitored by each sensing unit is determined, and a one-to-one mapping relationship table between sensing unit identifier and target functional component identifier is established. Based on the description of the mechanical connection relationship type in the physical assembly structure topology diagram, determine whether there is a direct physical connection between any two target functional component identifiers, and mark the target functional component identifier pairs with direct physical connections as spatial adjacent component pairs. For each spatial adjacent component pair with direct physical connections, obtain the sensing response signal segment sequence of the sensing unit corresponding to each of the two target functional component identifiers in the spatial adjacent component pair, and perform synchronization time point alignment processing on the two sensing response signal segment sequences to obtain a synchronization signal segment pair. The synchronization signal segment pairs are subjected to synchronization fluctuation consistency analysis processing. The frequency of the same-direction change and the frequency of the opposite-direction change of the signal amplitude change trend of the two sensor response signal segment sequences within the same time window are calculated. The synchronization response consistency value between the two sensor units is determined based on the ratio of the same-direction change frequency to the opposite-direction change frequency. Traverse all spatial adjacency component pairs with direct physical connections, repeatedly perform synchronization fluctuation consistency analysis processing, generate a synchronization response consistency value uniquely corresponding to each spatial adjacency component pair, and mark spatial adjacency component pairs whose synchronization response consistency value exceeds a preset consistency threshold as target associated component pairs; All functional components in the physical assembly structure topology diagram are taken as spatial network nodes, and an undirected connection edge is established between the spatial network nodes corresponding to the two functional components in the target associated component pair. The synchronization response consistency value is assigned to the undirected connection edge as the edge weight to construct a preliminary spatial undirected coupled network. The coupling relationship extension operation is performed on the preliminary spatial undirected coupled network. For any two spatial network nodes in the preliminary spatial undirected coupled network that do not have a direct undirected connection edge but have common adjacent spatial network nodes, the product value of the edge weights between each of the two spatial network nodes and the common adjacent spatial network node is analyzed. An indirect coupling connection edge is added between the two spatial network nodes whose product value exceeds a preset extension threshold, and the product value is assigned as the indirect edge weight. The initial spatial undirected coupled network is merged with the indirect coupled connection edges added through the coupling relationship expansion operation to generate a spatial association knowledge graph that characterizes the mutual influence relationships between target power equipment components. The actual operating environment deployment information of the target power equipment is obtained. The actual operating environment deployment information includes the geographical orientation description of the target power equipment in the deployment site and the distribution description of surrounding interference sources. Based on the actual operating environment deployment information, the edge weights of the undirected connection edges of the corresponding edge components in the spatial association knowledge graph are corrected by environmental attenuation factors.

4. The fault diagnosis and location method for power equipment operation and maintenance according to claim 1, characterized in that, The step involves inputting the temporal state transmission chain network and the spatial association knowledge graph into a preset spatiotemporal fault feature tracing model for joint inference and analysis, generating a fault source location indicator for locating the origin of the fault within the target power equipment, including: Obtain a time-series state transfer chain network that reflects the internal state transition path of the target power equipment. The time-series state transfer chain network includes multiple time-series network nodes and directed edges connecting the time-series network nodes. Each time-series network node corresponds to a sensing unit in the target power equipment, and each directed edge is accompanied by a time-series dependency transfer weight. Obtain a spatial association knowledge graph that characterizes the mutual influence relationships between components of a target power equipment. The spatial association knowledge graph contains multiple spatial network nodes and undirected connecting edges that connect the spatial network nodes. Each spatial network node corresponds to a functional component within the target power equipment, and each undirected connecting edge is accompanied by a spatial coupling strength weight. The temporal network nodes in the temporal state transfer chain network and the spatial network nodes in the spatial association knowledge graph are aligned and fused according to the one-to-one mapping relationship table between the sensor unit identifier and the target functional component identifier to generate a spatiotemporal fusion network. Each fusion node in the spatiotemporal fusion network inherits the corresponding temporal network node attributes and spatial network node attributes. For each fusion node in the spatiotemporal fusion network, a fault symptom correlation initialization process is performed. The latest signal segment of the real-time sensing response signal segment sequence of the corresponding sensing unit of the fusion node is obtained. The abnormal amplitude offset that exceeds the preset normal operating range in the latest signal segment is extracted. According to the preset abnormal level mapping table of each sensing unit, the abnormal amplitude offset is mapped to a normalized fault symptom level value, which is used as the initial fault symptom level value of the fusion node. On the spatiotemporal fusion network, a fault source probability inference operation based on a random walk mechanism is performed. Each fusion node is taken as a possible fault starting candidate point, and multiple rounds of random walk propagation process are performed. In each round of random walk propagation process, the next fusion node to walk is selected based on the joint weighted probability distribution of the temporal dependency propagation weight indicated by the directed edge of the current fusion node and the spatial coupling strength weight indicated by the undirected connection edge at each step. Record the sequence of all fusion nodes traversed from the starting candidate point to the end of the walk in each round of random walk propagation. Calculate the cumulative frequency of each fusion node being traversed in all rounds of random walk propagation. Divide the cumulative frequency of each fusion node being traversed by the product of the total number of walk rounds and the number of network summary points to obtain the fault impact range score of the fusion node. For each fusion node that serves as a starting candidate point, the initial fault symptom level values ​​of the fusion nodes that pass through during all rounds of random walk propagation from that starting candidate point are summarized. The sum of the initial fault symptom level values ​​of the fusion nodes that pass through is calculated as the initial source explanation capability score of that starting candidate point. This initial source explanation capability score is then divided by the largest initial source explanation capability score among all starting candidate points to obtain the normalized source explanation capability score of that starting candidate point. The fault impact range score value corresponding to each starting candidate point is weighted and fused with the normalized source explanation capability score value to generate a comprehensive confidence score value for the starting candidate point as the fault source location. The comprehensive confidence scores corresponding to all starting candidate points are sorted in descending order. The target functional component identifier corresponding to the starting candidate point with the highest comprehensive confidence score value is determined as the fault source location indication identifier for locating the source location of the fault inside the target power equipment. A dynamic weight decay operation is performed on the directed edges in the spatiotemporal fusion network. A time decay factor is calculated based on the signal acquisition time interval of the corresponding sensing units of the fusion nodes at both ends of the directed edge. The time decay factor is used to correct the time-dependent transit weight of the directed edge to obtain the time-decayed time-dependent transit weight.

5. The fault diagnosis and location method for power equipment operation and maintenance according to claim 1, characterized in that, The fault source location indicator is the target functional component identifier corresponding to the fault source component. The spatial association knowledge graph contains multiple spatial network nodes and undirected edges connecting the spatial network nodes, as well as the spatial coupling strength weight attached to each undirected edge. The step of generating a fault diagnosis and location report containing a unique identifier code of the fault component and a list of affected components based on the fault source location indicator and the spatial association knowledge graph, and pushing the fault diagnosis and location report to a preset operation and maintenance management terminal, includes: In the spatial association knowledge graph, locate the target spatial network node that matches the target functional component identifier corresponding to the fault source component, and use the target spatial network node as the search center point. Starting from the search center point, expand the search layer by layer outward along the undirected connection edge in the spatial association knowledge graph, record all spatial network nodes passed during the search process, and record the cumulative spatial coupling strength weight value of the undirected connection edge passed from the search center point to each searched spatial network node. The cumulative spatial coupling strength weight value is compared with the preset influence range determination threshold. The searched spatial network nodes whose cumulative spatial coupling strength weight value exceeds the preset influence range determination threshold are marked as affected component nodes. The searched spatial network nodes whose cumulative spatial coupling strength weight value does not exceed the preset influence range determination threshold are stopped from continuing to spread outward. Collect the target functional component identifiers corresponding to all spatial network nodes marked as affected component nodes, and construct a list of affected component ranges that are spatially coupled with the fault source component. Each item in the list of affected component ranges contains the affected target functional component identifier and the corresponding cumulative spatial coupling strength weight value. Obtain the mapping relationship of the target functional component identifier corresponding to the fault source component in the device coding database, query the unique identifier code of the fault component that matches the target functional component identifier corresponding to the fault source component, and construct a fault diagnosis and location report containing the unique identifier code of the fault component and the list of affected components. The fault diagnosis and location report also includes a timestamp of the report generation time. The fault diagnosis and location report is pushed to a preset operation and maintenance management terminal through a preset communication protocol interface, triggering the operation and maintenance management terminal to generate an audible and visual alarm signal containing the unique identifier code of the faulty component; The fault diagnosis and location report is encrypted during transmission. The fault diagnosis and location report is asymmetrically encrypted using the public key information preset in the operation and maintenance management terminal to generate encrypted fault diagnosis and location report ciphertext.

6. The fault diagnosis and location method for power equipment operation and maintenance according to claim 1, characterized in that, The method further includes: Obtain the set of physical failure model parameters for key components inside the target power equipment. The set of physical failure model parameters includes the aging rate constant of insulating materials, the fatigue crack propagation index of metallic conductors, the wear coefficient of mechanical parts, and the demagnetization rate parameters of magnetic materials. After generating the spatial association knowledge graph, the spatial coupling strength weights of the spatial network nodes of the corresponding components in the spatial association knowledge graph are subjected to physical degradation trend weighting correction processing based on the physical failure model parameter set, thereby generating a physically degraded spatial association knowledge graph. The real-time operation status monitoring data stream is processed for abrupt change event detection. By comparing the slope of the signal amplitude change within a sliding window with a preset abrupt change slope threshold, the abrupt change moment when the signal amplitude changes abruptly within a preset short time is identified. The time window corresponding to the abrupt change moment is marked as the abrupt change event window. The sensor response signal segment sequence corresponding to all sensor units within the abrupt change event window is extracted. The similarity or difference of the amplitude change direction of each sensor unit signal segment sequence within the abrupt change event window is analyzed. Sensor units with the same amplitude change direction are grouped into the same source response group, and sensor units with opposite amplitude change directions are grouped into the opposite source response group. Based on the division results of the same-source response group and the heterogeneous response group, the directed edges between the corresponding time-series network nodes of the sensing units in the same-source response group and between the time-series network nodes of the heterogeneous response group are adjusted in the time-series state transmission chain network. The time-series state transmission chain network after the directed edge direction adjustment and the spatial association knowledge graph after physical degradation correction are input into the spatiotemporal fault feature tracing model for joint inference and analysis to generate the fault source location indication identifier. The partial discharge spectrum data of internal components of the target power equipment were collected by a portable testing instrument during each planned shutdown and maintenance period. The partial discharge spectrum data of internal components of the equipment includes the phase distribution characteristics and discharge amplitude distribution characteristics of the discharge pulse. The phase distribution features of discharge pulses in the partial discharge spectrum data of the internal components of the device are extracted, and the phase distribution features of the discharge pulses are converted into phase-resolved partial discharge feature vectors. The phase-resolved partial discharge feature vectors are then matched with reference discharge feature vectors in a preset typical insulation defect discharge fingerprint database to determine a list of components suspected of having insulation defects. The component identifier in the list of suspected insulation defect components is compared with the fault source location indicator. If the fault source location indicator exists in the list of suspected insulation defect components, the diagnostic confidence level of the fault component in the fault diagnosis and location report is increased. The fault diagnosis and location report is synchronously pushed to the preset spare parts inventory management system, which triggers the spare parts inventory management system to retrieve the spare parts inventory quantity and storage location information that match the unique identifier code of the faulty component.

7. The fault diagnosis and location method for power equipment operation and maintenance according to claim 1, characterized in that, The method further includes: Obtain the design structure blueprint file of the target power equipment. The design structure blueprint file includes a description of the component assembly hierarchy and a description of the electrical connection lines between components. Determine the parent-child relationship of each component inside the target power equipment based on the description of the component assembly hierarchy and determine the energy flow path between each component inside the target power equipment based on the description of the electrical connection lines between components. Based on the description of the component assembly hierarchy and the description of the electrical connection lines between the components, a physical mechanism causal graph model of the target power equipment is constructed. The physical mechanism causal graph model includes component hierarchy nodes and directed edges for energy flow. The direction of the directed edges for energy flow is consistent with the direction of energy flow from the upstream component to the downstream component. After generating the temporal state transit chain network, the temporal network nodes in the temporal state transit chain network are aligned with the component-level nodes in the physical mechanism causal graph model to establish a mapping association table between data-driven temporal correlation and physical mechanism causal relationship. Extract the temporal network node identifiers at both ends of each directed edge in the temporal state transfer chain network, find the corresponding component-level node identifiers of the temporal network node identifiers at both ends according to the mapping association table, and determine whether the direction of the directed edge is consistent with the direction of the energy flow between the corresponding component-level nodes in the physical mechanism causal graph model. Directed edges with inconsistent pointing directions are marked as physically contradictory abnormal associated edges, and directed edges with consistent pointing directions are marked as physically consistent associated edges. The physically consistent associated edges are retained, and the physically contradictory abnormal associated edges are removed from the temporal state transmission chain network to obtain the target temporal state transmission chain network after physical mechanism consistency pruning. The target temporal state transmission chain network replaces the original temporal state transmission chain network and is input into the spatiotemporal fault feature tracing model for joint inference analysis and processing to generate the fault source location indication identifier. Acquire the sequence record of switch state change events inside the target power equipment. The sequence record of switch state change events includes circuit breaker opening and closing state change events, protection device action signal trigger events, and corresponding millisecond-level precise event occurrence time tags. Extract the protection device action signal triggering events and their millisecond-level precise event occurrence time tags within the preset backtracking time period in the sequence record of the switch state change events. Using the millisecond-level precise event occurrence time tag as the time reference benchmark, trace back to the abnormal fluctuation start time of the corresponding sensing response signal segment sequence of each sensing unit in the time sequence state transmission chain network. Calculate the absolute value of the time difference between the start time of the abnormal fluctuation of the sensing response signal segment sequence corresponding to each sensing unit and the millisecond-level precise event occurrence time tag of the protection device action signal trigger event. Mark the timing network node corresponding to the sensing unit with the smallest absolute value of the time difference as the monitoring node most relevant to the protection action event. Superimpose the information of the monitoring node most relevant to the protection action event into the fault diagnosis and location report to form an enhanced fault diagnosis and location report that supplements the protection action-related monitoring point information.

8. The fault diagnosis and location method for power equipment operation and maintenance according to claim 1, characterized in that, The method further includes: The target power equipment is obtained from the hourly environmental meteorological element sequence recorded by multiple external environmental meteorological monitoring stations within a continuous preset time period. The hourly environmental meteorological element sequence includes the environmental temperature value sequence, the environmental relative humidity value sequence, and the atmospheric pressure value sequence. The hourly environmental meteorological element sequence and the sensor response signal segment sequence corresponding to each sensor unit in the real-time operation status monitoring data stream are time-aligned and spliced ​​together to form an extended sensor response signal segment sequence with additional meteorological background information. The data item at each time point in the extended sensor response signal segment sequence with additional meteorological background information includes the sensor response signal value and the corresponding environmental temperature value and environmental relative humidity value. Meteorological factors and signal drift correlation analysis are performed on the extended sensor response signal segment sequence with additional meteorological background information. The temperature drift correlation coefficient between the sensor response signal value and the ambient temperature value, and the humidity drift correlation coefficient between the sensor response signal value and the ambient relative humidity value are calculated to obtain the temperature drift correlation coefficient and humidity drift correlation coefficient corresponding to each sensor unit. Based on the temperature drift correlation coefficient and humidity drift correlation coefficient corresponding to each sensor unit, when generating the time-series state transmission chain network, a meteorological drift component stripping operation is performed on the signal amplitude fluctuation pattern in the real-time operation status monitoring data stream. The temperature drift component, which is formed by multiplying the temperature drift correlation coefficient and the ambient temperature change, and the humidity drift component, which is formed by multiplying the humidity drift correlation coefficient and the ambient relative humidity change, are subtracted from the signal amplitude fluctuation pattern. The original signal amplitude fluctuation pattern is replaced by the signal amplitude fluctuation pattern after removing the meteorological drift component. The step of analyzing the transmission and evolution law of the signal amplitude fluctuation pattern along the time axis in the sequence of sensor response signal segments is re-executed to generate a meteorological stripping time-series state transmission chain network that reflects the internal state migration path of the target power equipment. The meteorological stripping time-series state transmission chain network and the spatial association knowledge graph are input into the spatiotemporal fault feature tracing model for joint inference and analysis to generate the fault source location indication. Obtain system frequency disturbance event records published by the power grid dispatch automation system in the area where the target power equipment is located. The system frequency disturbance event records include the start time of the frequency deviation from the rated value and the frequency deviation magnitude value. Based on the start time of the frequency deviation from the rated value in the system frequency disturbance event record, when generating the real-time operation status monitoring data stream, frequency disturbance events are marked on the sensor response signal segment sequence corresponding to all sensor units within the start time of the frequency deviation from the rated value and the subsequent preset time period. Obtain the vibration acceleration sensing response signal segment sequence of the rotating mechanical components in the target power equipment, perform envelope spectrum analysis on the vibration acceleration sensing response signal segment sequence, extract the amplitude sequence of the fault characteristic frequency harmonic component in the envelope spectrum, and use the amplitude sequence of the fault characteristic frequency harmonic component as an additional vibration fault feature vector. When the temporal state transfer chain network and the spatial association knowledge graph are input into the spatiotemporal fault feature tracing model, the vibration fault feature vector is also input as an auxiliary feature into the spatiotemporal fault feature tracing model for joint inference and analysis.

9. A fault diagnosis and location system for power equipment operation and maintenance, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the fault diagnosis and location method for power equipment operation and maintenance as described in any one of claims 1 to 8 by executing the machine-executable instructions.

Images