A multi-source heterogeneous data fusion high-voltage equipment fault identification method and system

Abstract

This invention discloses a method and system for high-voltage equipment fault identification based on multi-source heterogeneous data fusion, primarily relating to the field of high-voltage equipment fault identification technology. It includes: constructing a sensing and acquisition network covering high-voltage equipment and measuring points within a photovoltaic power station; acquiring multi-source heterogeneous data in real time through an edge computing architecture; extracting features from the acquired multi-source heterogeneous data to generate a time-series organized equipment state feature set; constructing a candidate association graph; constructing a local propagation model and a non-local propagation model sharing the same candidate association graph; performing joint training using historical operating data; and inputting the multi-source heterogeneous data into the trained local propagation model and non-local propagation model to obtain a set of high-voltage equipment fault identification results. The beneficial effects of this invention are: solving the technical problems of existing technologies lacking dynamic identification of different data from high-voltage equipment, resulting in lagging fault identification and low accuracy, thus achieving the technical effect of improving the reliability of fault identification.

Description

Technical Field

[0001] This invention relates to the field of high-voltage equipment fault identification technology, specifically to a high-voltage equipment fault identification method and system based on multi-source heterogeneous data fusion. Background Technology

[0002] With the continuous expansion of photovoltaic power plants, the operational reliability of high-voltage equipment such as transformers, switchgear, gas-insulated switchgear, and cable terminals directly affects the power generation efficiency and grid security of the entire power plant. Currently, fault monitoring of high-voltage equipment mainly employs partial discharge detection technologies such as ultrasonic waves, transient ground voltage, ultra-high frequency (UHF), high-frequency pulse current, and infrared thermography. Various sensors are deployed to collect status parameters such as partial discharge amplitude, frequency, total energy, and abnormal heating location, and the data is displayed and analyzed on a smart operation and maintenance management platform. Existing methods typically combine big data analytics and machine learning algorithms to predict equipment operating status and identify faults, attempting to achieve precise perception of operating conditions.

[0003] However, existing technical solutions have the following technical shortcomings in practical applications: First, traditional methods mostly adopt an isolated analysis strategy, that is, processing data from individual devices or single types of sensors separately, failing to treat the entire site as a dynamically coupled system for unified modeling, making it difficult to discover potential correlations between devices. Second, existing methods only focus on the state changes of the devices themselves, failing to capture the propagation path and coupling effect of faults between devices, making it difficult to accurately locate the source of the fault when anomalies occur across devices. Third, existing technologies lack the ability to actively learn the real health correlation structure between devices, often relying on manually preset correlation rules or static topological relationships, which cannot adapt to the dynamic changes in device coupling relationships during operation, thus limiting the accuracy and timeliness of fault identification.

[0004] Existing technologies suffer from a lack of dynamic identification of different data from high-voltage equipment, resulting in delayed and inaccurate fault identification. Summary of the Invention

[0005] This application provides a method and system for fault identification of high-voltage equipment based on the fusion of multi-source heterogeneous data, which solves the technical problems in the prior art of lacking dynamic identification of different data of high-voltage equipment, resulting in lagging fault identification and low accuracy.

[0006] In view of the above problems, this application provides a method and system for high-voltage equipment fault identification based on multi-source heterogeneous data fusion.

[0007] The first aspect of this application provides a method for fault identification in high-voltage equipment through multi-source heterogeneous data fusion, the method comprising: A sensing and acquisition network covering high-voltage equipment and measuring points within a photovoltaic power station is constructed. Multi-source heterogeneous data is collected in real time through an edge computing architecture. The multi-source heterogeneous data includes partial discharge detection data, infrared temperature detection data, environmental parameter data, and power grid interaction data of each device and measuring point. Feature extraction is performed on the collected multi-source heterogeneous data to generate a device status feature set organized in time series. Each high-voltage device and each measuring point is treated as an entity in the association graph. Based on the electrical connection relationship and physical spatial layout, the connection edges between entities are constructed to build a candidate association graph. Construct a local propagation model and a non-local propagation model that share the same candidate association graph. The local propagation model predicts the next state of an entity based on the state of its direct neighbor entities, while the non-local propagation model predicts the next state of an entity based on the state of its distant entities that have multi-step path connections with it. By using historical operational data, the local propagation model and the non-local propagation model are jointly trained to minimize the prediction error of the two models. At the same time, sparsity constraints are imposed on the weights of the connecting edges in the candidate association graph, driving the model to automatically update the edge weights during training. When the training converges, a real-time association graph reflecting the true health association between devices is obtained. By inputting multi-source heterogeneous data into the trained local propagation model and non-local propagation model, and based on the real-time correlation graph output by the model and the prediction residuals of the two models, a set of high-voltage equipment fault identification results is obtained.

[0008] In one possible implementation, the partial discharge detection data includes the amplitude, frequency, phase, and total energy of the partial discharge acquired by an ultrasonic sensor, a transient ground voltage sensor, an ultra-high frequency sensor, and a high frequency pulse current sensor. The infrared temperature detection data includes the temperature values ​​and temperature rise rate of each device and each measuring point collected by the infrared thermal imager. The environmental parameter data includes ambient temperature and humidity; The power grid interaction data includes voltage, current, and load data.

[0009] In one possible implementation, feature extraction is performed on the collected multi-source heterogeneous data to generate a device status feature set organized in time series, including: Denoising and pulse extraction are performed on partial discharge detection data to generate partial discharge trend sequences for each device and each measurement point; The system identifies the preset interest area for each device and each measuring point based on infrared temperature detection data, extracts the highest temperature value within the interest area and calculates the temperature rise rate, and generates a hot spot temperature rise rate sequence for each device and each measuring point. The partial discharge trend sequence and hot spot temperature rise rate sequence of each device and each measuring point are aligned by time to form a device status feature set.

[0010] In one possible implementation, each high-voltage device and each measuring point is treated as an entity in the association graph. Connection edges between entities are constructed based on electrical connections and physical spatial layout to build a candidate association graph, including: Each high-voltage device or each key measuring point is treated as an entity in the correlation diagram; Add electrical connection edges between entities according to the electrical primary wiring diagram, and add spatially adjacent edges between entities according to the physical spatial layout; Each electrical connection edge and spatially adjacent edge is assigned a trainable initial weight between 0 and 1 to form a weighted candidate association graph.

[0011] In one possible implementation, the local propagation model is constructed using a single-layer graph neural network, whose input is the state features of the target entity's direct neighbor entities at the current time, and whose output is the predicted state features of the target entity at the next time. The nonlocal propagation model is constructed using a K-order polynomial graph filter. Its input is the state characteristics of a distant entity that is connected to the target entity by a path of 1 to K steps at the current time, and its output is the predicted state characteristics of the target entity at the next time step. K is an integer greater than 1. The K-order polynomial graph filter includes a transfer function.

[0012] In one possible implementation, the method includes: Obtain the adjacency matrix of the candidate association graph, calculate the first to the Kth powers of the adjacency matrix to obtain the 1-step reachability matrix to the K-step reachability matrix; Assign a trainable polynomial coefficient to each reachable matrix; The transfer function of the K-order polynomial graph filter is obtained by multiplying each reachable matrix by its corresponding polynomial coefficients and summing the results.

[0013] In one possible implementation, the local propagation model and the non-local propagation model are jointly trained using historical operational data, including: Construct a joint training objective function, which includes a first term and a second term. The first term is the sum of the squared deviations between the predicted value and the actual value of the local propagation model and the sum of the squared deviations between the predicted value and the actual value of the non-local propagation model. The second term is the product of the sum of the absolute values ​​of the weights of all connecting edges in the candidate association graph and a preset sparse coefficient. The objective function is minimized using the backpropagation algorithm, and the weights of each connection edge in the candidate association graph are automatically updated during the training process. When the objective function converges, the candidate association graph at the end of training is output as the real-time association graph.

[0014] In one possible implementation, multi-source heterogeneous data is input into a trained local propagation model and a non-local propagation model. Based on the real-time correlation graph output by the models and the prediction residuals of the two models, a set of high-voltage equipment fault identification results is obtained, including: The mean and standard deviation of the weight of each connected edge during the historical normal operation period are statistically analyzed, and an anomaly detection threshold is set. The real-time collected multi-source heterogeneous data is input into the trained local propagation model and non-local propagation model, and the real-time correlation graph is output. When the current weight of any connection edge in the real-time association graph is greater than the anomaly determination threshold, it is determined that there is abnormal coupling between the two entities connected by the edge, a first high-voltage equipment fault identification result is generated, and the first high-voltage equipment fault identification result is added to the high-voltage equipment fault identification result set.

[0015] In one possible implementation, the method further includes: The multi-source heterogeneous data collected in real time is input into the trained local propagation model and non-local propagation model, and the prediction residuals of the local propagation model and non-local propagation model for each entity are calculated. When the residual predicted by the local propagation model of any entity exceeds the first residual threshold instead of the residual predicted by the local propagation model being normal, it is determined that the entity has experienced a sudden local event, a second type of warning is generated, and the abnormal entity is located. When the residual predicted by the nonlocal propagation model of any entity exceeds the second residual threshold while the residual predicted by the local propagation model is normal, it is determined that the fault has started to propagate across regions. By tracing back the contribution values ​​of each order path in the nonlocal propagation model, the multi-step propagation path from the starting entity through intermediate entities to the entity is output.

[0016] A second aspect of this application provides a high-voltage equipment fault identification system based on multi-source heterogeneous data fusion, the system comprising: The data acquisition module is used to build a sensing and acquisition network covering high-voltage equipment and measuring points in the photovoltaic power station. It collects multi-source heterogeneous data in real time through an edge computing architecture. The multi-source heterogeneous data includes partial discharge detection data, infrared temperature detection data, environmental parameter data, and power grid interaction data of each device and measuring point. The feature extraction module is used to extract features from the collected multi-source heterogeneous data and generate a device status feature set organized in time series. The candidate association graph construction module is used to treat each high-voltage equipment and each measuring point as an entity in the association graph, construct the connection edges between entities based on electrical connection relationships and physical spatial layout, and construct the candidate association graph. The model building module is used to build local propagation models and non-local propagation models that share the same candidate association graph. The local propagation model predicts the next state of an entity based on the state of its direct neighbor entities, while the non-local propagation model predicts the next state of an entity based on the state of its distant entities that have multi-step path connections with the entity. The joint training module is used to jointly train the local propagation model and the non-local propagation model using historical running data, with the goal of minimizing the prediction error of the two models. At the same time, it applies sparsity constraints to the weights of the connecting edges in the candidate association graph, driving the model to automatically update the edge weights during training. When the training converges, a real-time association graph reflecting the true health association between devices is obtained. The fault identification result acquisition module is used to input multi-source heterogeneous data into the trained local propagation model and non-local propagation model, and obtain the set of fault identification results for high-voltage equipment based on the real-time correlation graph output by the model and the prediction residuals of the two models.

[0017] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application constructs a sensing and acquisition network covering high-voltage equipment and measuring points within a photovoltaic power station. It collects multi-source heterogeneous data in real time using an edge computing architecture. This multi-source heterogeneous data includes partial discharge detection data, infrared temperature detection data, environmental parameter data, and grid interaction data from each device and measuring point. Feature extraction is performed on the collected multi-source heterogeneous data to generate a time-series organized device state feature set. Each high-voltage device and each measuring point is treated as an entity in an association graph. Connection edges between entities are constructed based on electrical connections and physical spatial layout to build a candidate association graph. A local propagation model and a non-local propagation model sharing the same candidate association graph are constructed. The local propagation model is based on the entity's direct neighbors. The system predicts the next state of an entity based on its current state. The nonlocal propagation model predicts the next state of the entity based on the states of distant entities connected to it via multi-step paths. Historical operational data is used to jointly train the local and nonlocal propagation models, aiming to minimize the prediction error between the two models. Simultaneously, sparsity constraints are imposed on the weights of the connecting edges in the candidate association graph, driving the model to automatically update edge weights during training. When training converges, a real-time association graph reflecting the true health associations between devices is obtained. Multi-source heterogeneous data is input into the trained local and nonlocal propagation models. Based on the real-time association graph output by the models and the prediction residuals of the two models, a set of high-voltage equipment fault identification results is obtained. This achieves the technical effect of improving the accuracy of fault identification through dynamic analysis of multi-source heterogeneous data. Attached Figure Description

[0018] Appendix Figure 1This is a schematic diagram of a high-voltage equipment fault identification method based on multi-source heterogeneous data fusion provided in an embodiment of the present invention.

[0019] Appendix Figure 2 This is a schematic diagram of a high-voltage equipment fault identification system based on multi-source heterogeneous data fusion provided in an embodiment of the present invention.

[0020] The labels shown in the attached diagram: The system includes a data acquisition module 11, a feature extraction module 12, a candidate association graph construction module 13, a model construction module 14, a joint training module 15, and a fault identification result acquisition module 16. Detailed Implementation

[0021] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims. It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to these processes, methods, products, or devices.

[0022] Example 1, as shown in the appendix Figure 1 As shown, this application provides a method for fault identification of high-voltage equipment based on multi-source heterogeneous data fusion, wherein the method includes: A sensing and acquisition network covering high-voltage equipment and measuring points within a photovoltaic power station is constructed. Multi-source heterogeneous data is collected in real time through an edge computing architecture. The multi-source heterogeneous data includes partial discharge detection data, infrared temperature detection data, environmental parameter data, and power grid interaction data of each device and measuring point. It should be noted that the sensing and acquisition network refers to a sensor cluster deployed on various high-voltage equipment and key measuring points within a photovoltaic power station. This cluster is connected via industrial fieldbus or wireless communication to form a unified data acquisition system. Its coverage includes primary equipment such as transformers, switchgear, gas-insulated switchgear, and cable terminals, as well as temperature measuring points and discharge measuring points installed on each piece of equipment.

[0023] Determine the list of high-voltage equipment requiring monitoring and the locations of key monitoring points. Taking a 50MWp photovoltaic power station as an example, its key high-voltage equipment includes: two 110kV main transformers (#1 and #2), six high-voltage switchgear (#1 to #6), three 110kV GIS bays (incoming line bay, main transformer bay, and busbar equipment bay), and four sets of SVG cable terminal joints (three phases per set). For each transformer, an ultrasonic sensor is installed in the middle of its main tank to detect internal discharge, and a high-frequency pulse current sensor is installed at the end screen of the bushing to detect discharge along the bushing. For each switchgear, transient ground voltage sensors are installed on its front, middle, and rear panels to detect surface discharge. For each GIS bay, an ultra-high frequency sensor is installed at its gas chamber observation window to detect internal insulator discharge. For each SVG cable terminal joint, a high-frequency pulse current sensor is installed at its grounding wire to detect grounding discharge. Simultaneously, infrared thermal imagers are installed at key heat-generating locations of each piece of equipment, such as the base of the transformer bushing, the upper and lower contacts of the switchgear, the GIS busbar connection, and the cable terminal crimping points.

[0024] All sensors are connected to nearby edge acquisition units via shielded cables. Each acquisition unit is responsible for data acquisition and preprocessing from 10-15 sensors. The edge acquisition units use ARM-based embedded processors running a real-time operating system to continuously sample sensor signals: ultrasonic signals at 1MHz, transient ground voltage signals at 50MHz, ultra-high frequency signals at 500MHz, high-frequency pulse current signals at 100MHz, infrared thermal imagers output one frame of thermal image per second, environmental temperature and humidity sensors output data every 10 seconds, and power grid interaction data is acquired from the power plant monitoring system once per second via the Modbus protocol. The edge acquisition units perform real-time processing of the raw data, such as bandpass filtering of partial discharge signals to remove power frequency interference, temperature extraction from infrared thermal images, and packaging the processed feature data according to a unified timestamp.

[0025] Feature extraction is performed on the collected multi-source heterogeneous data to generate a device status feature set organized in time series. Furthermore, the partial discharge detection data includes the amplitude, frequency, phase, and total energy of the partial discharge collected by an ultrasonic sensor, a transient ground voltage sensor, an ultra-high frequency sensor, and a high frequency pulse current sensor. The infrared temperature detection data includes the temperature values ​​and temperature rise rate of each device and each measuring point collected by the infrared thermal imager. The environmental parameter data includes ambient temperature and humidity; The power grid interaction data includes voltage, current, and load data.

[0026] Furthermore, feature extraction is performed on the collected multi-source heterogeneous data to generate a device status feature set organized in time series, including: Denoising and pulse extraction are performed on partial discharge detection data to generate partial discharge trend sequences for each device and each measurement point; The system identifies the preset interest area for each device and each measuring point based on infrared temperature detection data, extracts the highest temperature value within the interest area and calculates the temperature rise rate, and generates a hot spot temperature rise rate sequence for each device and each measuring point. The partial discharge trend sequence and hot spot temperature rise rate sequence of each device and each measuring point are aligned by time to form a device status feature set.

[0027] This step systematically extracts features from the raw multi-source data uploaded by the edge acquisition unit. Taking a gas chamber of a 110kV GIS device as an example, the raw signal acquired by its UHF sensor first enters the feature extraction process. For partial discharge data, wavelet transform soft thresholding is used for noise reduction. Wavelet decomposition separates the real discharge pulses in the signal from the background noise. After setting the high-frequency noise component to zero, it is reconstructed, effectively removing random noise and periodic interference in the environment. The denoised signal is then used for pulse detection using a dual threshold method. An amplitude threshold and a time window threshold are set. Only when the signal amplitude exceeds the preset amplitude threshold and the pulse duration exceeds the preset time window threshold is it determined to be a valid partial discharge pulse, thus eliminating false triggering by spike noise. For each detected discharge pulse, the peak amplitude is extracted to represent the discharge intensity, the phase represents the position of the discharge relative to the power frequency voltage for discharge type identification, and the energy represents the overall impact of the discharge. The discharge frequency per unit time is statistically analyzed to represent the activity level of the discharge. After continuous recording for a period of time, a partial discharge characteristic sequence with minute intervals is obtained, including the partial discharge amplitude sequence, partial discharge frequency sequence, partial discharge phase sequence, and partial discharge total energy sequence. The values ​​in each sequence are arranged in chronological order, reflecting the evolution of discharge activity.

[0028] For infrared temperature detection data, a pre-defined area of ​​interest is first automatically located on each frame of the thermal image using an image recognition algorithm. This includes areas prone to heat generation, such as the connection between the GIS gas chamber and the bushing, the upper and lower contact positions of the switchgear, and the crimping points of cable terminals. These areas are pre-marked in the thermal image using a coordinate frame. The temperature values ​​of all pixels within this area are extracted, and the maximum value is taken as the hotspot temperature at that moment, as the highest temperature point is often the most likely location for anomalies. One frame of thermal image is continuously acquired every minute to obtain a hotspot temperature time series. A first-order difference calculation is performed on the temperature series, i.e., the temperature of the next moment is subtracted from the temperature of the previous moment, to obtain a temperature rise rate series. Positive values ​​in this series represent temperature increases, and negative values ​​represent temperature decreases. The absolute value reflects the severity of temperature changes, enabling early detection of abnormal heating trends.

[0029] For environmental parameter data and power grid interaction data, ambient temperature and humidity, as well as power grid voltage, current, and load data are directly recorded by timestamp. These data serve as background parameters to correct the judgment benchmark for equipment status. Finally, all data sources are aligned according to a unified time coordinate, for example, using one minute as a time point, and partial discharge characteristics, temperature characteristics, environmental parameters, and power grid parameters are combined into a multi-dimensional feature vector. For each high-voltage device and each measuring point, an independent set of such feature time series is generated to jointly construct the equipment status feature set. This step, through systematic feature extraction, transforms massive amounts of raw sensor data into structured feature time series, significantly reducing data dimensionality while retaining key equipment status information, providing high-quality input data for subsequent graphical model learning. The extracted partial discharge trend sequence and hotspot temperature rise rate sequence can reflect the dynamic evolution of discharge activity and heating activity, respectively, providing sensitive indicators for identifying latent faults.

[0030] Each high-voltage device and each measuring point is treated as an entity in the association graph. Based on the electrical connection relationship and physical spatial layout, the connection edges between entities are constructed to build a candidate association graph. In one embodiment, an entity refers to the basic building block of the association graph, corresponding to each specific physical device or each independent monitoring location in this technical solution. For example, the #1 main transformer is considered as an entity, and the A-phase cable connector of the #2 switchgear is considered as an independent measuring point entity. An association graph is a mathematical structure composed of entities as nodes and possible influence relationships between entities as edges, used to characterize the coupling relationship network between elements within the site. The specific construction process is described below.

[0031] Construct a local propagation model and a non-local propagation model that share the same candidate association graph. The local propagation model predicts the next state of an entity based on the state of its direct neighbor entities, while the non-local propagation model predicts the next state of an entity based on the state of its distant entities that have multi-step path connections with it. Two propagation models sharing the same candidate association graph are constructed to capture fault propagation patterns at different spatial scales. The local propagation model adopts a single-layer graph neural network architecture, whose core design principle is that the state update of each entity depends only on the current state of its direct neighbor entities. Taking a switch cabinet entity as an example, its direct neighbor entities in the candidate association graph include adjacent switch cabinets connected by electrical connection edges and neighboring devices connected by spatial adjacency edges. The local propagation model first obtains the state characteristics of the target entity at the current moment, and simultaneously obtains the state characteristics of all direct neighbor entities at the current moment. Then, it fuses this neighbor information with the target entity's own information through the aggregation function of the graph neural network. Common aggregation methods include mean aggregation, attention-weighted aggregation, or gated aggregation, where the attention mechanism can adaptively allocate weights according to the influence of different neighbors on the target entity. After aggregation, a nonlinear transformation is applied to output the predicted state characteristics of the target entity at the next moment. The single-layer design of the local propagation model determines that its receptive field is limited to a one-hop neighbor and cannot perceive the influence of farther distances. This characteristic makes it highly sensitive to local sudden events, and the local propagation model can respond quickly when a sudden increase in partial discharge occurs in the equipment.

[0032] The nonlocal propagation model employs a high-order polynomial graph filter. Its core design principle is that the state update of each entity needs to consider the influence of distant entities connected to the target entity via multi-step paths. The polynomial graph filter aggregates multi-hop neighbor information by performing a polynomial expansion on the adjacency matrix of the candidate association graph. Specifically, for a preset order K, the model calculates the first to K powers of the adjacency matrix, corresponding to entity relationships reachable in one, two, and K hops, respectively. Then, a trainable polynomial coefficient is assigned to each order, and the reachability matrices of each order are multiplied by their corresponding coefficients and summed to obtain the transfer function. During runtime, the current state features of all entities constitute the state matrix, which is transformed through the transfer function to output the predicted state features for the next time step. The value of K is typically preset to 3 to 5 based on the device type and fault propagation characteristics. For example, a higher order might be used for thermal diffusion processes to capture long-distance effects, while a lower order might be used for discharge propagation. Nonlocal propagation models can perceive the cumulative propagation effect of faults along equipment links through multi-level path aggregation. For example, if a transformer overheats due to overload, the overheating can be transmitted step by step through electrical connections and spatial adjacency, eventually leading to abnormal temperature in a remote switchgear.

[0033] By using historical operational data, the local propagation model and the non-local propagation model are jointly trained to minimize the prediction error of the two models. At the same time, sparsity constraints are imposed on the weights of the connecting edges in the candidate association graph, driving the model to automatically update the edge weights during training. When the training converges, a real-time association graph reflecting the true health association between devices is obtained. Furthermore, the local propagation model and the non-local propagation model are jointly trained using historical operational data, including: Construct a joint training objective function, which includes a first term and a second term. The first term is the sum of the squared deviations between the predicted value and the actual value of the local propagation model and the sum of the squared deviations between the predicted value and the actual value of the non-local propagation model. The second term is the product of the sum of the absolute values ​​of the weights of all connecting edges in the candidate association graph and a preset sparse coefficient. The objective function is minimized using the backpropagation algorithm, and the weights of each connection edge in the candidate association graph are automatically updated during the training process. When the objective function converges, the candidate association graph at the end of training is output as the real-time association graph.

[0034] Preferably, historical operational data is used to jointly optimize and train the two models, enabling the models to automatically learn the real health correlations between devices. First, a training dataset is prepared, selecting three consecutive months of historical operational data, containing the state feature vectors of all entities at each time point. The data is divided into training, validation, and test sets in chronological order. The data at each time point includes the partial discharge characteristics, temperature characteristics, environmental parameters, and power grid parameters of all entities at the current moment, as well as the actual state values ​​of all entities at the next moment, serving as supervision labels.

[0035] A joint training objective function is constructed, consisting of two parts. The first part is the prediction error term, which calculates and sums the squared deviations between the predicted values ​​of the local propagation model and the actual values ​​of the non-local propagation model. Taking a specific time point as an example, the local propagation model outputs one set of predicted values ​​for the next time step for all entities, while the non-local propagation model outputs another set. The mean squared error between these two sets of predicted values ​​and the actual state value at the next time step is calculated separately, and the two error values ​​are summed to obtain the prediction error at that time point. The prediction errors at all time points are summed or averaged to obtain the overall prediction error term. The second part is the sparsity constraint term, which calculates the sum of the absolute values ​​of the weights of all connected edges in the candidate association graph and multiplies it by a preset sparsity coefficient. The sparsity coefficient is a hyperparameter, tuned using a validation set, typically ranging from 0.001 to 0.1, used to balance the weights between prediction accuracy and graph sparsity. The overall objective function is the sum of the prediction error term and the sparsity constraint term.

[0036] The training process employs mini-batch gradient descent, randomly selecting consecutive time-series segments from the training set as a batch. For each batch, the following operations are performed: The current entity state features are input into the local propagation model and the non-local propagation model. Both models compute based on the currently shared candidate association graph, outputting the predicted state for the next time step. The sum of squared deviations between the predicted and actual values ​​of the two models is calculated, and a sparsity constraint term is added to obtain the total loss value for the current batch. The gradient of the loss value with respect to all trainable parameters in the model is calculated using the backpropagation algorithm, including the graph neural network parameters of the local propagation model, the polynomial coefficients of the non-local propagation model, and the weight of each connection edge in the candidate association graph. All parameters are updated according to the learning rate to reduce the loss value. This process is repeated multiple times across the entire training dataset.

[0037] During training, the sparsity constraint continues to play a role: for edges unrelated to the actual health relationships between devices, their weights gradually approach zero during gradient updates; for edges reflecting actual relationships, their weights are maintained or enhanced. For example, an edge connecting adjacent switchgear added based on spatial layout might have its weight adjusted from an initial 0.3 to 0.4 if historical data indicates a correlation between the state changes of these two switchgears; conversely, if the data shows no substantial correlation, its weight might drop below 0.01, approaching zero. After dozens of iterations of training, when the loss value on the validation set no longer decreases or the decrease becomes gradual, training is considered converged and stopped. At this point, the candidate relationship graph at the end of training is output as the optimal relationship graph, also known as the real-time relationship graph.

[0038] Through joint training, on the one hand, the two models learn collaboratively based on a shared graph structure. The local propagation model ensures sensitivity to direct correlations, while the nonlocal propagation model provides the ability to capture indirect influences, complementing each other. On the other hand, sparsity constraints automatically perform feature selection, filtering out truly important device correlations from candidate correlation graphs containing prior knowledge and eliminating redundant or spurious connections. The resulting real-time correlation graph condenses the health coupling patterns between devices contained in historical data, reflecting the true mutual influence relationships between devices within the site, and providing a reliable benchmark graph structure for subsequent online monitoring and fault identification.

[0039] By inputting multi-source heterogeneous data into the trained local propagation model and non-local propagation model, and based on the real-time correlation graph output by the model and the prediction residuals of the two models, a set of high-voltage equipment fault identification results is obtained.

[0040] Furthermore, multi-source heterogeneous data is input into the trained local propagation model and non-local propagation model. Based on the real-time correlation graph output by the models and the prediction residuals of the two models, a set of high-voltage equipment fault identification results is obtained, including: The mean and standard deviation of the weight of each connected edge during the historical normal operation period are statistically analyzed, and an anomaly detection threshold is set. The real-time collected multi-source heterogeneous data is input into the trained local propagation model and non-local propagation model, and the real-time correlation graph is output. When the current weight of any connection edge in the real-time association graph is greater than the anomaly determination threshold, it is determined that there is abnormal coupling between the two entities connected by the edge, a first high-voltage equipment fault identification result is generated, and the first high-voltage equipment fault identification result is added to the high-voltage equipment fault identification result set.

[0041] Furthermore, it also includes: The multi-source heterogeneous data collected in real time is input into the trained local propagation model and non-local propagation model, and the prediction residuals of the local propagation model and non-local propagation model for each entity are calculated. When the residual predicted by the local propagation model of any entity exceeds the first residual threshold instead of the residual predicted by the local propagation model being normal, it is determined that the entity has experienced a sudden local event, a second type of warning is generated, and the abnormal entity is located. When the residual predicted by the nonlocal propagation model of any entity exceeds the second residual threshold while the residual predicted by the local propagation model is normal, it is determined that the fault has started to propagate across regions. By tracing back the contribution values ​​of each order path in the nonlocal propagation model, the multi-step propagation path from the starting entity through intermediate entities to the entity is output.

[0042] In one embodiment, a trained dual-model approach is used for online fault identification, collaboratively detecting equipment anomalies through two mechanisms: real-time correlation graph changes and dual-model prediction residuals. First, a statistical baseline for normal operation is established. Data from the past three months of normal equipment operation is selected, and the time-series mean and standard deviation of the weights for each connection edge in the candidate correlation graph are calculated. Taking the connection edge between the transformer and the GIS incoming line bay as an example, its historical weight sequence is statistically analyzed to obtain the mean μ and standard deviation σ. An anomaly threshold is set to μ + 3σ, meaning weights exceeding three times the standard deviation are considered anomalies. Simultaneously, for each entity, the distribution of its local propagation model residuals and non-local propagation model residuals during normal operation is statistically analyzed, and a first residual threshold and a second residual threshold are set, typically the 95th or 99th quartile of the historical residual distribution.

[0043] During the online operation phase, real-time multi-source heterogeneous data is input into the trained local propagation model and non-local propagation model. Based on the entity state features of the current input and the optimal association graph structure obtained during training convergence, the model calculates the real-time association graph at the current moment, i.e., the current weight value of each connection edge. Simultaneously, both models output predicted values ​​for the state of each entity at the next moment. These predicted values ​​are compared with the actual observed values ​​collected at the next moment, and the prediction residuals of the local propagation model and non-local propagation model for each entity are calculated.

[0044] The first type of fault identification is performed based on the real-time correlation graph. Each connection edge is checked individually to see if its current weight exceeds its anomaly detection threshold. If the current weight of an edge is greater than μ+3σ, for example, the edge weight between the transformer and the GIS incoming line bay increases from the historical average of 0.78 to 0.95, exceeding the threshold of 0.93, then an abnormal coupling is determined between the two entities connected by that edge. The first high-voltage equipment fault identification result is generated as "An abnormal coupling exists between the transformer and the GIS incoming line bay, which may indicate that a fault is propagating between the two devices," and this result is added to the high-voltage equipment fault identification result set.

[0045] Type II and Type III fault identification is performed based on the prediction residuals of the dual-model approach. For each entity, the relationship between its local propagation model prediction residuals and non-local propagation model prediction residuals and their respective thresholds is compared. If the local propagation model prediction residual of an entity exceeds the first residual threshold, while the non-local propagation model prediction residual is within the normal range, it indicates that the entity has experienced a sudden change that the local model cannot explain, but the non-local model can partially explain it by considering multiple paths, which is consistent with the characteristics of a sudden local event. For example, if a switchgear experiences a sudden increase in partial discharge, the local model, which only considers direct neighbors, will make inaccurate predictions and exceed the residual limit; the non-local model may capture some indirect effects through multiple paths, resulting in a relatively smaller residual. In this case, it is determined that the entity has experienced a sudden local event, generating a Type II warning: "Sudden partial discharge pulse in switchgear, immediate inspection recommended," and locating the abnormal entity as the switchgear.

[0046] If the residual predicted by the nonlocal propagation model of an entity exceeds the second residual threshold, while the residual predicted by the local propagation model is within the normal range, it indicates that the state change of the entity cannot be explained by the far-end effects of multi-step paths. However, the local model's prediction based on direct neighbors is relatively accurate. This usually means that the anomaly is propagating across regions, and the current entity is the affected party rather than the source. For example, if the temperature of a cable terminal rises abnormally, the local model predicts accurately and has a small residual because the temperature of its direct neighbors is normal; the nonlocal model attempts to explain this through multi-step paths but fails, resulting in a large residual. At this point, it is determined that the fault has begun to propagate across regions. A reverse tracing mechanism is initiated to analyze the contribution value of each order path in the nonlocal propagation model to the entity. The coefficients of each order path in the polynomial filter and the corresponding reachability matrix are extracted, and the contribution weight of each path to the prediction result is calculated. The top few paths with the largest contributions are selected, the sequence of intermediate entities traversed by the path is analyzed, and the multi-step propagation path from the starting entity through intermediate entities to the target entity is output, such as "transformer → GIS bay → cable terminal," indicating that the fault may originate from the transformer and propagate to the cable terminal through the GIS bay.

[0047] Furthermore, each high-voltage device and each measuring point is treated as an entity in the association graph. Connection edges between entities are constructed based on electrical connections and physical spatial layout, resulting in a candidate association graph, including: Each high-voltage device or each key measuring point is treated as an entity in the correlation diagram; Add electrical connection edges between entities according to the electrical primary wiring diagram, and add spatially adjacent edges between entities according to the physical spatial layout; Each electrical connection edge and spatially adjacent edge is assigned a trainable initial weight between 0 and 1 to form a weighted candidate association graph.

[0048] Furthermore, the local propagation model is constructed using a single-layer graph neural network. Its input is the state features of the target entity's direct neighbor entities at the current time, and its output is the predicted state features of the target entity at the next time. The nonlocal propagation model is constructed using a K-order polynomial graph filter. Its input is the state characteristics of a distant entity that is connected to the target entity by a path of 1 to K steps at the current time, and its output is the predicted state characteristics of the target entity at the next time step. K is an integer greater than 1. The K-order polynomial graph filter includes a transfer function.

[0049] Furthermore, including: Obtain the adjacency matrix of the candidate association graph, calculate the first to the Kth powers of the adjacency matrix to obtain the 1-step reachability matrix to the K-step reachability matrix; Assign a trainable polynomial coefficient to each reachable matrix; The transfer function of the K-order polynomial graph filter is obtained by multiplying each reachable matrix by its corresponding polynomial coefficients and summing the results.

[0050] In one embodiment, the local propagation model employs a single-layer graph neural network architecture, whose network structure includes an input layer, an attention aggregation layer, and an output layer. Its design philosophy is that the next state of each device is primarily determined by the current state of its directly adjacent devices, conforming to the physical laws of partial discharge pulses propagating directly along electrical connections and sudden faults instantly affecting neighboring devices.

[0051] The input layer is used to obtain the state feature vector of the target entity at the current moment, with dimension d, and the set of state feature vectors of all direct neighbor entities at the current moment, with each neighbor feature vector having the same dimension d. If the target entity is a transformer, its state feature vector is formed by concatenating the partial discharge trend sequence and the hotspot temperature rise rate sequence after time alignment, and includes a total of 6 feature dimensions: partial discharge amplitude, frequency, phase, total energy, hotspot temperature, and temperature rise rate.

[0052] The attention aggregation layer is used to calculate the attention coefficient between each neighboring entity and the target entity. The specific calculation process is as follows: First, the feature vector of the target entity is concatenated with the feature vectors of its neighboring entities to obtain a 2d-dimensional concatenated vector; then, this concatenated vector is input into a single-layer feedforward neural network, which contains a trainable weight matrix W. att (Dimension 1×2d) and the LeakyReLU activation function are used to output unnormalized attention scores. Finally, the attention scores of all neighbors of the target entity are normalized using the Softmax function to obtain attention weights, which are between 0 and 1 and the sum of all neighbor weights is 1. The feature vectors of each neighbor entity are then summed according to their corresponding attention weights to obtain the neighbor aggregated feature vector.

[0053] The output layer concatenates the target entity's own feature vector with the aggregated feature vectors of its neighbors to obtain a 2d-dimensional combined feature vector. This combined feature vector is then input into a fully connected layer, which contains a trainable weight matrix W. out (Dimension d×2d) and bias vector b out The model then undergoes a nonlinear transformation using the ReLU activation function, outputting the predicted state feature vector of the target entity at the next time step, with dimension d. The design of the single-layer graph neural network ensures that the model only focuses on direct neighbors, exhibits high sensitivity to local sudden events, and has a small number of parameters and high computational efficiency, making it suitable for real-time online prediction.

[0054] The nonlocal propagation model is constructed using a high-order polynomial graph filter. Its design philosophy is that the next state of each device may be influenced by distant devices connected to the target device via multiple paths. This aligns with the physical laws that heat diffusion requires transmission through intermediate devices, mechanical stress is transmitted along the structure, and the cumulative effect of faults is amplified step by step. The order is preset based on the device type and fault propagation characteristics; for heat diffusion processes, it is typically third to fifth order to capture long-distance effects, while for discharge propagation, it may be second to third order.

[0055] The nonlocal propagation model of the polynomial graph filter is constructed using a K-order polynomial graph filter. Preferably, the value of K is preset to 3, which is determined based on the physical laws of fault propagation in high-voltage equipment of photovoltaic power plants. Since faults are usually transmitted step by step through electrical connections, the impact is significantly attenuated after three steps.

[0056] First, construct an adjacency matrix A from the candidate association graph. The matrix size is M×M, where M is the total number of entities, which is the sum of the number of high-voltage equipment and the number of key measuring points. Matrix element A ij The rule for determining the value of A is: if there is an electrical connection edge or a spatially adjacent edge between entity i and entity j, then A ij Take the current trainable weight value of the edge, ranging from 0 to 1; if there is no connecting edge, then A ij Set to 0; and the diagonal element A ij Take 0.

[0057] Then, calculate the powers of the adjacency matrix. Calculate A to the power of 1 to obtain the one-step reachability matrix, where the elements indicate whether there is a path of length 1 from entity i to entity j, and the value is the product of the weights of each edge on the path. Calculate A to the power of 2 to obtain the two-step reachability matrix, where the elements indicate whether there is a path of length 2 from entity i to entity j, and the value is the product of the weights of the two edges. Calculate A to the power of 3 to obtain the three-step reachability matrix.

[0058] Each reachable matrix is ​​assigned a trainable polynomial coefficient, denoted as θ1, θ2, and θ3, with each coefficient initially set to 1 / M. These coefficients are then optimized and updated via backpropagation during training. The transfer function of the polynomial graph filter is defined as: H = θ1·A + θ2·A² + θ3·A³.

[0059] During model execution, the state feature vectors of all entities at the current time step are arranged in entity order to form a state matrix X (with dimensions M×d). The output of the nonlocal propagation model is: Xnext = H·X, which is the transfer function multiplied by the state matrix on the left to obtain the predicted state feature matrix of all entities at the next time step, with dimensions M×d.

[0060] For any target entity, its predicted state is a weighted combination of the current states of all entities reachable by one, two, and three steps, with the weights determined by the polynomial coefficients and path weights. During model runtime, the current state features of all devices constitute a state matrix. The output of the nonlocal propagation model is the result of left-multiplying this state matrix by the transfer function. The next time step state of each device is a weighted combination of the current states of all devices reachable by one to multiple steps.

[0061] Example 2, based on the same inventive concept as the high-voltage equipment fault identification method using multi-source heterogeneous data fusion in the foregoing examples, as shown in the appendix. Figure 2 As shown, this application provides a high-voltage equipment fault identification system based on multi-source heterogeneous data fusion. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0062] Data acquisition module 11 is used to construct a sensing and acquisition network covering high-voltage equipment and measuring points in the photovoltaic power station. It collects multi-source heterogeneous data in real time through an edge computing architecture. The multi-source heterogeneous data includes partial discharge detection data, infrared temperature detection data, environmental parameter data, and power grid interaction data of each device and measuring point. Feature extraction module 12 is used to extract features from the collected multi-source heterogeneous data and generate a device status feature set organized in time series. The candidate association graph construction module 13 is used to treat each high-voltage equipment and each measuring point as entities in the association graph, construct the connection edges between entities according to the electrical connection relationship and physical spatial layout, and construct the candidate association graph. The model building module 14 is used to build a local propagation model and a non-local propagation model that share the same candidate association graph. The local propagation model predicts the next state of an entity based on the state of its direct neighbor entities, while the non-local propagation model predicts the next state of an entity based on the state of its distant entities that have multi-step path connections with the entity. The joint training module 15 is used to jointly train the local propagation model and the non-local propagation model using historical running data, with the goal of minimizing the prediction error of the two models. At the same time, it applies sparsity constraints to the weights of the connecting edges in the candidate association graph, driving the model to automatically update the edge weights during training. When the training converges, a real-time association graph reflecting the true health association between devices is obtained. The fault identification result acquisition module 16 is used to input multi-source heterogeneous data into the trained local propagation model and non-local propagation model, and obtain the set of fault identification results for high-voltage equipment based on the real-time correlation graph output by the model and the prediction residuals of the two models.

[0063] Furthermore, the partial discharge detection data includes the amplitude, frequency, phase, and total energy of the partial discharge collected by an ultrasonic sensor, a transient ground voltage sensor, an ultra-high frequency sensor, and a high frequency pulse current sensor. The infrared temperature detection data includes the temperature values ​​and temperature rise rate of each device and each measuring point collected by the infrared thermal imager. The environmental parameter data includes ambient temperature and humidity; The power grid interaction data includes voltage, current, and load data.

[0064] Furthermore, the feature extraction module 12 is used to perform the following steps: Denoising and pulse extraction are performed on partial discharge detection data to generate partial discharge trend sequences for each device and each measurement point; The system identifies the preset interest area for each device and each measuring point based on infrared temperature detection data, extracts the highest temperature value within the interest area and calculates the temperature rise rate, and generates a hot spot temperature rise rate sequence for each device and each measuring point. The partial discharge trend sequence and hot spot temperature rise rate sequence of each device and each measuring point are aligned by time to form a device status feature set.

[0065] Furthermore, the candidate association graph construction module 13 is used to perform the following steps: Each high-voltage device or each key measuring point is treated as an entity in the correlation diagram; Add electrical connection edges between entities according to the electrical primary wiring diagram, and add spatially adjacent edges between entities according to the physical spatial layout; Each electrical connection edge and spatially adjacent edge is assigned a trainable initial weight between 0 and 1 to form a weighted candidate association graph.

[0066] Furthermore, the local propagation model is constructed using a single-layer graph neural network. Its input is the state features of the target entity's direct neighbor entities at the current time, and its output is the predicted state features of the target entity at the next time. The nonlocal propagation model is constructed using a K-order polynomial graph filter. Its input is the state characteristics of a distant entity that is connected to the target entity by a path of 1 to K steps at the current time, and its output is the predicted state characteristics of the target entity at the next time step. K is an integer greater than 1. The K-order polynomial graph filter includes a transfer function.

[0067] Furthermore, the steps performed by the system also include: Obtain the adjacency matrix of the candidate association graph, calculate the first to the Kth powers of the adjacency matrix to obtain the 1-step reachability matrix to the K-step reachability matrix; Assign a trainable polynomial coefficient to each reachable matrix; The transfer function of the K-order polynomial graph filter is obtained by multiplying each reachable matrix by its corresponding polynomial coefficients and summing the results.

[0068] Furthermore, the joint training module 15 is used to perform the following steps: Construct a joint training objective function, which includes a first term and a second term. The first term is the sum of the squared deviations between the predicted value and the actual value of the local propagation model and the sum of the squared deviations between the predicted value and the actual value of the non-local propagation model. The second term is the product of the sum of the absolute values ​​of the weights of all connecting edges in the candidate association graph and a preset sparse coefficient. The objective function is minimized using the backpropagation algorithm, and the weights of each connection edge in the candidate association graph are automatically updated during the training process. When the objective function converges, the candidate association graph at the end of training is output as the real-time association graph.

[0069] Furthermore, the fault identification result acquisition module 16 is used to perform the following steps: The mean and standard deviation of the weight of each connected edge during the historical normal operation period are statistically analyzed, and an anomaly detection threshold is set. The real-time collected multi-source heterogeneous data is input into the trained local propagation model and non-local propagation model, and the real-time correlation graph is output. When the current weight of any connection edge in the real-time association graph is greater than the anomaly determination threshold, it is determined that there is abnormal coupling between the two entities connected by the edge, a first high-voltage equipment fault identification result is generated, and the first high-voltage equipment fault identification result is added to the high-voltage equipment fault identification result set.

[0070] Furthermore, the fault identification result acquisition module 16 is used to perform the following steps: The multi-source heterogeneous data collected in real time is input into the trained local propagation model and non-local propagation model, and the prediction residuals of the local propagation model and non-local propagation model for each entity are calculated. When the residual predicted by the local propagation model of any entity exceeds the first residual threshold instead of the residual predicted by the local propagation model being normal, it is determined that the entity has experienced a sudden local event, a second type of warning is generated, and the abnormal entity is located. When the residual predicted by the nonlocal propagation model of any entity exceeds the second residual threshold while the residual predicted by the local propagation model is normal, it is determined that the fault has started to propagate across regions. By tracing back the contribution values ​​of each order path in the nonlocal propagation model, the multi-step propagation path from the starting entity through intermediate entities to the entity is output.

[0071] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0072] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0073] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for fault identification of high-voltage equipment through multi-source heterogeneous data fusion, characterized in that, The method includes: A sensing and acquisition network covering high-voltage equipment and measuring points within a photovoltaic power station is constructed. Multi-source heterogeneous data is collected in real time through an edge computing architecture. The multi-source heterogeneous data includes partial discharge detection data, infrared temperature detection data, environmental parameter data, and power grid interaction data of each device and measuring point. Feature extraction is performed on the collected multi-source heterogeneous data to generate a device status feature set organized in time series. Each high-voltage device and each measuring point is treated as an entity in the association graph. Based on the electrical connection relationship and physical spatial layout, the connection edges between entities are constructed to build a candidate association graph. Construct a local propagation model and a non-local propagation model that share the same candidate association graph. The local propagation model predicts the next state of an entity based on the state of its direct neighbor entities, while the non-local propagation model predicts the next state of an entity based on the state of its distant entities that have multi-step path connections with it. By using historical operational data, the local propagation model and the non-local propagation model are jointly trained to minimize the prediction error of the two models. At the same time, sparsity constraints are imposed on the weights of the connecting edges in the candidate association graph, driving the model to automatically update the edge weights during training. When the training converges, a real-time association graph reflecting the true health association between devices is obtained. By inputting multi-source heterogeneous data into the trained local propagation model and non-local propagation model, and based on the real-time correlation graph output by the model and the prediction residuals of the two models, a set of high-voltage equipment fault identification results is obtained.

2. The high-voltage equipment fault identification method based on multi-source heterogeneous data fusion as described in claim 1, characterized in that, The partial discharge detection data includes the amplitude, frequency, phase, and total energy of the partial discharge collected by ultrasonic sensors, transient ground voltage sensors, ultra-high frequency sensors, and high frequency pulse current sensors. The infrared temperature detection data includes the temperature values ​​and temperature rise rate of each device and each measuring point collected by the infrared thermal imager. The environmental parameter data includes ambient temperature and humidity; The power grid interaction data includes voltage, current, and load data.

3. The high-voltage equipment fault identification method based on multi-source heterogeneous data fusion as described in claim 1, characterized in that, Feature extraction is performed on the collected multi-source heterogeneous data to generate a device status feature set organized in time series, including: Denoising and pulse extraction are performed on partial discharge detection data to generate partial discharge trend sequences for each device and each measurement point; The system identifies the preset interest area for each device and each measuring point based on infrared temperature detection data, extracts the highest temperature value within the interest area and calculates the temperature rise rate, and generates a hot spot temperature rise rate sequence for each device and each measuring point. The partial discharge trend sequence and hot spot temperature rise rate sequence of each device and each measuring point are aligned by time to form a device status feature set.

4. The high-voltage equipment fault identification method based on multi-source heterogeneous data fusion as described in claim 1, characterized in that, Treating each high-voltage device and each measuring point as an entity in the association graph, constructing connection edges between entities based on electrical connections and physical spatial layout, and building a candidate association graph, including: Each high-voltage device or each key measuring point is treated as an entity in the correlation diagram; Add electrical connection edges between entities according to the electrical primary wiring diagram, and add spatially adjacent edges between entities according to the physical spatial layout; Each electrical connection edge and spatially adjacent edge is assigned a trainable initial weight between 0 and 1 to form a weighted candidate association graph.

5. The high-voltage equipment fault identification method based on multi-source heterogeneous data fusion as described in claim 1, characterized in that, The local propagation model is constructed using a single-layer graph neural network. Its input is the state features of the target entity's direct neighbors at the current time, and its output is the predicted state features of the target entity at the next time. The nonlocal propagation model is constructed using a K-order polynomial graph filter. Its input is the state characteristics of a distant entity that is connected to the target entity by a path of 1 to K steps at the current time, and its output is the predicted state characteristics of the target entity at the next time step. K is an integer greater than 1. The K-order polynomial graph filter includes a transfer function.

6. The high-voltage equipment fault identification method based on multi-source heterogeneous data fusion as described in claim 5, characterized in that, include: Obtain the adjacency matrix of the candidate association graph, calculate the first to the Kth powers of the adjacency matrix to obtain the 1-step reachability matrix to the K-step reachability matrix; Assign a trainable polynomial coefficient to each reachable matrix; The transfer function of the K-order polynomial graph filter is obtained by multiplying each reachable matrix by its corresponding polynomial coefficients and summing the results.

7. The high-voltage equipment fault identification method based on multi-source heterogeneous data fusion as described in claim 1, characterized in that, Joint training of the local propagation model and the non-local propagation model using historical operational data includes: Construct a joint training objective function, which includes a first term and a second term. The first term is the sum of the squared deviations between the predicted value and the actual value of the local propagation model and the sum of the squared deviations between the predicted value and the actual value of the non-local propagation model. The second term is the product of the sum of the absolute values ​​of the weights of all connecting edges in the candidate association graph and a preset sparse coefficient. The objective function is minimized using the backpropagation algorithm, and the weights of each connection edge in the candidate association graph are automatically updated during the training process. When the objective function converges, the candidate association graph at the end of training is output as the real-time association graph.

8. The high-voltage equipment fault identification method based on multi-source heterogeneous data fusion as described in claim 1, characterized in that, Multi-source heterogeneous data is input into the trained local propagation model and non-local propagation model. Based on the real-time correlation graph output by the models and the prediction residuals of the two models, a set of high-voltage equipment fault identification results is obtained, including: The mean and standard deviation of the weight of each connected edge during the historical normal operation period are statistically analyzed, and an anomaly detection threshold is set. The real-time collected multi-source heterogeneous data is input into the trained local propagation model and non-local propagation model, and the real-time correlation graph is output. When the current weight of any connection edge in the real-time association graph is greater than the anomaly determination threshold, it is determined that there is abnormal coupling between the two entities connected by the edge, a first high-voltage equipment fault identification result is generated, and the first high-voltage equipment fault identification result is added to the high-voltage equipment fault identification result set.

9. The high-voltage equipment fault identification method based on multi-source heterogeneous data fusion as described in claim 8, characterized in that, Also includes: The multi-source heterogeneous data collected in real time is input into the trained local propagation model and non-local propagation model, and the prediction residuals of the local propagation model and non-local propagation model for each entity are calculated. When the residual predicted by the local propagation model of any entity exceeds the first residual threshold instead of the residual predicted by the local propagation model being normal, it is determined that the entity has experienced a sudden local event, a second type of warning is generated, and the abnormal entity is located. When the residual predicted by the nonlocal propagation model of any entity exceeds the second residual threshold while the residual predicted by the local propagation model is normal, it is determined that the fault has started to propagate across regions. By tracing back the contribution values ​​of each order path in the nonlocal propagation model, the multi-step propagation path from the starting entity through intermediate entities to the entity is output.

10. A high-voltage equipment fault identification system based on multi-source heterogeneous data fusion, characterized in that, The system is used to implement the high-voltage equipment fault identification method based on multi-source heterogeneous data fusion as described in any one of claims 1-9, and the system comprises: The data acquisition module is used to build a sensing and acquisition network covering high-voltage equipment and measuring points in the photovoltaic power station. It collects multi-source heterogeneous data in real time through an edge computing architecture. The multi-source heterogeneous data includes partial discharge detection data, infrared temperature detection data, environmental parameter data, and power grid interaction data of each device and measuring point. The feature extraction module is used to extract features from the collected multi-source heterogeneous data and generate a device status feature set organized in time series. The candidate association graph construction module is used to treat each high-voltage equipment and each measuring point as an entity in the association graph, construct the connection edges between entities based on electrical connection relationships and physical spatial layout, and construct the candidate association graph. The model building module is used to build local propagation models and non-local propagation models that share the same candidate association graph. The local propagation model predicts the next state of an entity based on the state of its direct neighbor entities, while the non-local propagation model predicts the next state of an entity based on the state of its distant entities that have multi-step path connections with the entity. The joint training module is used to jointly train the local propagation model and the non-local propagation model using historical running data, with the goal of minimizing the prediction error of the two models. At the same time, it applies sparsity constraints to the weights of the connecting edges in the candidate association graph, driving the model to automatically update the edge weights during training. When the training converges, a real-time association graph reflecting the true health association between devices is obtained. The fault identification result acquisition module is used to input multi-source heterogeneous data into the trained local propagation model and non-local propagation model, and obtain the set of fault identification results for high-voltage equipment based on the real-time correlation graph output by the model and the prediction residuals of the two models.

Images