A sensor network-based photovoltaic module intelligent operation and maintenance method and system

Abstract

The application discloses a kind of photovoltaic module intelligent operation and maintenance method and system based on sensor network, comprising the following steps: obtaining the component state, sensing state, equipment state and operation and maintenance record in photovoltaic module operation and maintenance period, establish the initial dependence edge of component, sensor and equipment node and generate trusted gate observation sequence;The trusted gate observation sequence is input into the improved graph deviation network model, and the node embedding feature and dynamic graph deviation dependence graph are obtained;State prediction and dependence reconstruction are carried out based on the dynamic graph deviation dependence graph, and an abnormal score matrix is formed;Decouple fault source according to the abnormal score matrix, generate abnormal path, fault source category and operation and maintenance action, and update credibility, dependence edge weight and model parameter in combination with execution feedback, to realize operation and maintenance result continuous correction.The application relates to the technical field of photovoltaic operation and maintenance, and realizes fault identification, source tracing and closed-loop operation and maintenance based on the improved graph deviation network.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic operation and maintenance and sensor network data analysis technology, specifically to a method and system for intelligent operation and maintenance of photovoltaic modules based on sensor networks. Background Technology

[0002] During long-term operation, photovoltaic (PV) power plants are susceptible to factors such as dust accumulation and shading, hot spots, string mismatch, loose wiring, bypass diode malfunctions, sensor drift, and inverter-side status fluctuations, leading to decreased power generation efficiency and increased operation and maintenance costs. Current PV module operation and maintenance methods typically rely on manual inspection, infrared inspection, fixed threshold alarms, or single power deviation analysis, which can detect some obvious faults. However, in scenarios involving multiple modules, multiple sensors, and multiple environmental factors, problems can easily arise, such as inaccurate identification of the source of anomalies, misinterpreting sensor drift as module failure, and misinterpreting environmental disturbances as electrical anomalies.

[0003] With the application of sensor networks in photovoltaic power plants, data such as module current, voltage, power, backsheet temperature, irradiance, wind speed, string status, and equipment status can be continuously collected, providing a data foundation for intelligent operation and maintenance of photovoltaic modules. However, traditional methods often treat sensor data as independent time series data, lacking the ability to learn the dynamic dependencies between modules, sensors, strings, and equipment. While some deep learning methods can perform anomaly detection, they typically rely on fixed topologies or fixed feature inputs, making it difficult to adaptively adjust the dependencies between modules according to changes in the power plant's operating status, and also making it difficult to simultaneously output abnormal nodes, abnormal edges, abnormal paths, and fault source categories.

[0004] Therefore, how to provide a method and system for intelligent operation and maintenance of photovoltaic modules based on sensor networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent operation and maintenance method and system for photovoltaic modules based on sensor networks. This invention fully utilizes sensor network data acquisition, node credibility calibration, an improved graph bias network model, and closed-loop feedback update technology. It details the complete process from constructing a reliable gated observation sequence from intelligent operation and maintenance data of photovoltaic modules, to generating node embedding features, dynamic graph bias dependency graphs, predicted states, dependency reconstruction results, and anomaly scoring matrices through the improved graph bias network model, and then to decoupling fault sources, generating abnormal paths and operation and maintenance actions, and updating the model based on execution feedback. It has the advantages of adaptive learning of the dynamic dependencies between photovoltaic modules, sensors, strings, and equipment, distinguishing between module-specific anomalies, connection anomalies, environmental disturbances, and sensor drift, and forming an interpretable closed-loop operation and maintenance system.

[0006] A method and system for intelligent operation and maintenance of photovoltaic modules based on sensor networks according to an embodiment of the present invention includes the following steps: S1. Obtain the module status, sensor status, equipment status and operation and maintenance records within the intelligent operation and maintenance cycle of photovoltaic modules, establish module, sensor, equipment nodes and initial dependency edges, and generate a trusted gated observation sequence based on node trustworthiness. S2. Input the intelligent operation and maintenance trusted gating observation sequence of photovoltaic modules into the improved graph bias network model, and obtain the node embedding features through node encoding and trusted gating mapping; S3. Based on the embedded feature learning of the photovoltaic module intelligent operation and maintenance node, the scenario-based dependency edges between the module, sensor and equipment are learned, and a dynamic graph deviation dependency graph is generated. S4. Based on the deviation dependency graph of the intelligent operation and maintenance dynamic diagram of photovoltaic modules, perform state prediction and dependency reconstruction to obtain the predicted state and dependency reconstruction results. S5. Calculate node deviation and edge deviation based on the predicted state, actual state and dependency reconstruction results of photovoltaic module intelligent operation and maintenance, and generate an anomaly scoring matrix. S6. Decouple the fault source based on the photovoltaic module intelligent operation and maintenance anomaly scoring matrix, and generate the anomaly path, fault source category and operation and maintenance action; S7. Based on the feedback from the intelligent operation and maintenance of photovoltaic modules, update the credibility, dependency edge weights, and dual branches to obtain the closed-loop operation and maintenance model.

[0007] Optionally, step S1 includes the following steps: S11. Read the current, voltage, power, backsheet temperature, irradiance, wind speed, string status, equipment status and inspection records during the intelligent operation and maintenance cycle of photovoltaic modules to obtain the original operation and maintenance data. S12. Organize the original operation and maintenance data according to component number, sensor number, device number and sampling time to obtain the node attribute sequence; S13. Establish initial dependency edges based on electrical connections, spatial adjacency, temperature conduction, and power correlation to obtain the initial dependency edge set; S14. Read the sampling missing rate, mutation frequency, calibration record and differences between adjacent nodes, and calculate the node reliability; S15. Correct the electrical state, temperature state, and environmental state in the node attribute sequence according to the node credibility to obtain the credible gated observation sequence.

[0008] Optionally, step S2 includes the following steps: S21. Input the components, sensors and device nodes of the trusted gated observation sequence into the node type coding layer to obtain the type coding features; S22. Input the electrical state, temperature state, environmental state, and operation and maintenance records of the trusted gated observation sequence into the state coding layer to obtain the state coding features; S23. Input the sampling time, sunshine interval and operation and maintenance cycle into the time-series coding layer to obtain the time-series coding features; S24. Input the node credibility into the credibility gating mapping layer to obtain the credibility embedding feature; S25. Input the type encoding feature, state encoding feature, time sequence encoding feature and credibility embedding feature into the model embedding end to obtain the node embedding feature.

[0009] 4. Optionally, step S3 includes the following steps: S31. Calculate the similarity of associations between component, sensor and device nodes based on node embedding features to obtain the candidate association matrix; S32. Based on the candidate association matrix, filter electrical neighbors, temperature neighbors, environmental neighbors, and operation and maintenance neighbors to obtain a set of candidate dependency edges; S33. Input the candidate dependency edge set into the improved graph bias network dependency learning layer, and combine the sensor network topology to calculate the scenario-based dependency edge weights. S34. Perform sparse selection and edge type labeling according to the scenario-based dependency edge weights to obtain the set of retained dependency edges; S35. Combine the retained dependency edge set with the node embedding features to obtain a dynamic graph deviation dependency graph that changes with the operation and maintenance status.

[0010] Optionally, step S33 includes the following steps: S331. Read the electrical coding, temperature coding, environmental coding, and confidence embedding of the nodes at both ends of the candidate dependent edge to obtain the edge weight input features; S332. Calculate the attention value of the candidate dependent edge based on the edge weight input features to obtain the first edge weight; S333. Correct the weight of the first edge based on the electrical connections, spatial adjacency, and temperature conduction relationships in the initial dependent edges to obtain the weight of the second edge; S334. Based on the mismatch records, occlusion records, hot spot records and sensor drift records of the same string, electrical mismatch weights, occlusion propagation weights, thermal anomaly diffusion weights and drift suppression weights are formed; S335. Input the second edge weight, electrical mismatch weight, occlusion propagation weight, thermal anomaly diffusion weight, and drift suppression weight into the dependency learning layer to obtain the contextualized dependency edge weight.

[0011] Optionally, step S4 includes the following steps: S41. Input the dynamic graph bias dependency graph into the graph attention aggregation layer to obtain the component node neighborhood aggregation features; S42. Input the neighborhood aggregation features into the state prediction branch, calculate the current, voltage, power and backplane temperature in the next sampling period, and obtain the predicted state; S43. Input the dynamic graph deviation dependency graph into the dependency reconstruction branch, reconstruct the edge weights and edge types of the retained dependency edges, and obtain the dependency reconstruction result. S44. Match the predicted state with the dependency reconstruction results according to the component number, string number and sampling time to obtain the dual-branch output result; S45. Write the dual-branch output results into the deviation calculation terminal of the improved graph deviation network to obtain the deviation calculation input.

[0012] Optionally, step S5 includes the following steps: S51. Align the predicted state with the actual state by nodes and by time to obtain state-aligned data. S52. Calculate the power deviation, current deviation, temperature deviation and irradiation response deviation based on the state alignment data to obtain the state deviation set; S53. Calculate edge weights, edge types, and neighborhood changes based on the dependency reconstruction results and the dynamic graph deviation dependency graph to obtain the dependency deviation set; S54. Based on the node credibility, the state deviation set and the dependency deviation set are weighted and fused to obtain the node anomaly score and the edge anomaly score. S55. Combine node anomaly scores and edge anomaly scores to generate an improved graph bias network anomaly score matrix.

[0013] Optionally, step S6 includes the following steps: S61. Compare the anomaly scoring matrix with the dynamic threshold to obtain the anomaly nodes, anomaly edges, and anomaly time periods; S62. Based on the component location, string location, and sensor location of the abnormal node, obtain the set of abnormal objects; S63. Based on the edge type, propagation direction, and edge anomaly score of the abnormal edge, trace the anomaly propagation chain to obtain the anomaly path; S64. Input the set of abnormal objects and abnormal paths into the fault source decoupling layer to obtain the categories of component body abnormality, connection abnormality, environmental disturbance or sensor drift. S65. Match cleaning, infrared inspection, wiring check, component replacement or sensor calibration actions according to the fault source category to obtain maintenance actions.

[0014] Optionally, step S7 includes the following steps: S71. Obtain execution records after cleaning, inspection, maintenance, replacement and calibration, and obtain operation and maintenance feedback data; S72. Match the operation and maintenance feedback data with abnormal nodes, abnormal edges, abnormal paths and fault source categories to obtain feedback tag data; S73. Update the node credibility, scenario-based dependency edge weight, and dynamic threshold based on the feedback tag data to obtain the first update parameter; S74. Update the state prediction branch and dependency refactoring branch based on the feedback marker data to obtain the second update parameter; S75. Write the first update parameter and the second update parameter into the improved graph deviation network model to obtain the closed-loop operation and maintenance model.

[0015] According to an embodiment of the present invention, a photovoltaic module intelligent operation and maintenance system based on a sensor network includes the following modules: The trusted gating module acquires intelligent operation and maintenance data of photovoltaic modules, initial dependency edges and node trustworthiness, and generates a trusted gating observation sequence. The node embedding module takes the trusted gated observation sequence as input to the improved graph bias network model and generates node embedding features. The dependency learning module learns the contextual dependency edges of components, sensors, and devices based on the embedded features of nodes, and generates a dynamic graph bias dependency graph. The prediction and reconstruction module performs state prediction and dependency reconstruction based on the dynamic graph deviation dependency graph, generating predicted state and dependency reconstruction results. The deviation scoring module calculates the deviation based on the predicted state, the actual state, and the dependency reconstruction results, and generates an anomaly scoring matrix. The fault decoupling module locates abnormal nodes and abnormal edges based on the abnormal scoring matrix, and generates fault source categories, abnormal paths and maintenance actions. The closed-loop update module updates the credibility, dependency edge weights, and dual branches based on execution feedback to generate a closed-loop operation and maintenance model.

[0016] The beneficial effects of this invention are as follows: By introducing an improved graph bias network model into the intelligent operation and maintenance process of photovoltaic modules, this invention integrates module nodes, sensor nodes, and equipment nodes into a learnable dependency relationship. This allows operation and maintenance analysis to no longer rely solely on a single sensor threshold or a single power change, but rather to jointly determine the source of anomalies based on node status, node reliability, and changes in edge relationships. The reliable gated observation sequence can correct observation data based on sampling gaps, anomalous jumps, calibration records, and differences between adjacent nodes, reducing the interference of sensor drift, local sampling anomalies, and environmental disturbances on fault diagnosis.

[0017] This invention employs a scenario-based dependency edge weight learning mechanism to incorporate factors such as electrical connectivity, spatial adjacency, temperature conduction, string mismatch, shading propagation, thermal anomaly diffusion, and sensor drift suppression into an improved graph deviation network model. This enables the model to dynamically generate a dynamic graph deviation dependency graph based on the intelligent operation and maintenance scenario of photovoltaic modules, enhancing its ability to express anomaly propagation relationships between modules. The state prediction branch and dependency reconstruction branch output predicted states and dependency reconstruction results, respectively, and generate node anomaly scores and edge anomaly scores at the deviation calculation end. This allows anomaly detection to not only identify "which module is abnormal" but also analyze "which dependency relationship is abnormal."

[0018] This invention further utilizes a fault source decoupling layer to jointly analyze abnormal nodes, abnormal edges, and abnormal paths. This enables the differentiation between component-level anomalies, connection anomalies, environmental disturbances, and sensor drift. It also matches these anomalies with maintenance actions such as cleaning, infrared inspection, wiring checks, component replacement, or sensor calibration, improving the interpretability and targeted execution of intelligent maintenance results for photovoltaic modules. By updating node credibility through execution feedback, scenario-based dependency edge weights, dynamic thresholds, state prediction branches, and dependency reconstruction branches, the model is continuously revised based on maintenance results, enhancing its adaptability to different sites, seasons, and operating conditions. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 A flowchart illustrating the overall process of an intelligent operation and maintenance method for photovoltaic modules based on sensor networks; Figure 2 : An improved graph bias network model and a schematic diagram of the dynamic graph bias dependency graph generation process; Figure 3 : Schematic diagram of anomaly scoring matrix, fault source decoupling and closed-loop update process. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figures 1-3 A method and system for intelligent operation and maintenance of photovoltaic modules based on sensor networks includes the following steps: S1. Obtain the module status, sensor status, equipment status and operation and maintenance records within the intelligent operation and maintenance cycle of photovoltaic modules, establish module, sensor, equipment nodes and initial dependency edges, and generate a trusted gated observation sequence based on node trustworthiness. S2. Input the intelligent operation and maintenance trusted gating observation sequence of photovoltaic modules into the improved graph bias network model, and obtain the node embedding features through node encoding and trusted gating mapping; S3. Based on the embedded feature learning of the photovoltaic module intelligent operation and maintenance node, the scenario-based dependency edges between the module, sensor and equipment are learned, and a dynamic graph deviation dependency graph is generated. S4. Based on the deviation dependency graph of the intelligent operation and maintenance dynamic diagram of photovoltaic modules, perform state prediction and dependency reconstruction to obtain the predicted state and dependency reconstruction results. S5. Calculate node deviation and edge deviation based on the predicted state, actual state and dependency reconstruction results of photovoltaic module intelligent operation and maintenance, and generate an anomaly scoring matrix. S6. Decouple the fault source based on the photovoltaic module intelligent operation and maintenance anomaly scoring matrix, and generate the anomaly path, fault source category and operation and maintenance action; S7. Based on the feedback from the intelligent operation and maintenance of photovoltaic modules, update the credibility, dependency edge weights, and dual branches to obtain the closed-loop operation and maintenance model.

[0022] In this embodiment, S1 specifically includes: Read the current, voltage, power, backsheet temperature, irradiance, wind speed, string status, equipment status, and inspection records during the intelligent operation and maintenance cycle of photovoltaic modules to obtain the raw operation and maintenance data; The original operation and maintenance data is organized according to component number, sensor number, device number and sampling time to obtain the node attribute sequence; Initial dependency edges are established based on electrical connections, spatial adjacency, temperature conduction, and power correlation to obtain the initial dependency edge set; Read the sampling missing rate, mutation frequency, calibration records, and differences between adjacent nodes, and calculate the node reliability; The electrical state, temperature state, and environmental state in the node attribute sequence are corrected according to the node credibility to obtain the credible gated observation sequence.

[0023] Data reading, node organization, and initial edge establishment in S1 are the basic input construction processes for model operation. The core is to form a reliable gated observation sequence through node reliability and use the reliable gated observation sequence as the unified input of the improved graph bias network model.

[0024] In this embodiment, the original operation and maintenance data is jointly provided by the component-level current acquisition device, voltage acquisition device, backplane temperature sensor, environmental monitoring instrument, string monitoring device, combiner box, inverter and operation and maintenance management platform.

[0025] Current data records the output current of the photovoltaic module or string at the sampling time, voltage data records the output voltage of the photovoltaic module or string at the sampling time, and power data is provided directly by the acquisition equipment or calculated from the current and voltage. Backplane temperature data records the temperature on the back of the module; irradiance data records the solar irradiance level in the area where the module is located; wind speed data records the heat dissipation environment in the area where the module is located; string status includes string open circuit, short circuit, mismatch, overcurrent, undervoltage, and communication status. Equipment status includes combiner box circuit breaker status, inverter operating status, inverter alarm status, and communication status. Inspection records include records of manual inspections, drone inspections, infrared verification, cleaning operations, wiring checks, component replacements, and sensor calibrations.

[0026] The above data are arranged continuously according to the sampling time. Each sampling point corresponds to several component nodes, sensor nodes and device nodes, forming a node attribute sequence.

[0027] In this embodiment, the component node corresponds to the photovoltaic module body, the sensing node corresponds to the current, voltage, backsheet temperature, irradiance, wind speed and environmental status acquisition unit, and the equipment node corresponds to the string, combiner box and inverter.

[0028] Initial dependency edges include electrical connection edges, spatial adjacency edges, temperature conduction edges, and power-dependent edges. Electrical connection edges record the electrical connection sequence between components, strings, combiner boxes, and inverters; Spatial adjacency edges record the component installation area, arrangement rows and columns, and relationships between adjacent components; The temperature conduction edge records the conduction relationship of backplane temperature changes between adjacent components; The power-related edge records the correlation between component power, current, and temperature changes within the same irradiation area.

[0029] The initial dependency edges are not directly used as the final anomaly judgment criteria, but instead participate in the learning of scenario-based dependency edge weights after entering the improved graph bias network model.

[0030] Through this design, the initial dependency edge provides the inherent structure of the photovoltaic power station, and the scenario-based dependency edge weights provide the dynamic relationships during operation. Together, they support the generation of the dynamic graph deviation dependency graph.

[0031] In this embodiment, the node credibility value ranges from 0 to 1.

[0032] For nodes with a sampling missing rate of no more than 1% and a mutation frequency of no more than once per 100 sampling points, the sampling confidence parameter is set between 0.90 and 1.00. For nodes with a sampling missing rate greater than 1% but not exceeding 5%, or a mutation frequency greater than once per 100 sampling points but not exceeding 3 times, the sampling confidence parameter should be set between 0.70 and 0.89. For nodes with a sampling missing rate greater than 5% or a mutation frequency greater than 3 times per 100 sampling points, the sampling confidence parameter is set between 0.30 and 0.69.

[0033] The differences between adjacent nodes are determined by the differences in power, current and backplane temperature of components in the same string and adjacent components within the same sampling window. For nodes whose differences are less than the upper limit of historical normal fluctuations, the neighborhood confidence parameter is set between 0.85 and 1.00. For nodes where the difference reaches 1.5 times the upper limit of historical normal fluctuations, the neighborhood confidence parameter is set between 0.60 and 0.84. For nodes whose deviations exceed 1.5 times the historical normal fluctuation limit, the neighborhood confidence parameter is set between 0.20 and 0.59.

[0034] For sensor nodes where calibration records were completed within 30 days and no abnormal drift was recorded, the historical reliability parameter was set between 0.90 and 1.00. For sensor nodes with calibration records exceeding 30 days but not exceeding 90 days, the historical reliability parameter is set between 0.70 and 0.89. For sensor nodes with calibration records exceeding 90 days or with drift records, the historical reliability parameter is set between 0.30 and 0.69.

[0035] The node credibility is obtained by fusing the sampled credibility parameters, the neighborhood credibility parameters, and the historical credibility parameters. The fused node credibility is directly entered into the credibility gating mapping layer.

[0036] The aforementioned node reliability range is a specific configuration of this implementation method. It can be adjusted during deployment according to the site operation and maintenance cycle and sensor calibration system, and the protection range is not limited.

[0037] In this embodiment, the trusted gated observation sequence is formed by correcting the node attribute sequence based on the node's trustworthiness.

[0038] For nodes with a confidence level higher than 0.85, the trusted gated observation sequence retains the original sampled values ​​and participates in subsequent dependency learning; For nodes with a confidence level greater than or equal to 0.60 and not higher than 0.85, the confidence-gated observation sequence is weighted and corrected using the original sampled values ​​and the estimated values ​​of adjacent nodes in the same string. For nodes with a confidence level below 0.60, the confidence-gated observation sequence reduces the node's influence on node embedding and bias scoring, and marks the node as a low-confidence node.

[0039] Low-reliability nodes are still retained in the sensor network, but drift suppression is introduced when generating scene-dependent edge weights and constructing anomaly scoring matrices to avoid misjudging component failures due to anomalies of a single sensor.

[0040] The trusted gated observation sequence includes node number, node type, sampling time, node attributes, node trustworthiness, initial dependency edge label, and low trustworthiness label. The subsequent improved graph bias network model directly reads the above fields.

[0041] refer to Figure 2 In this embodiment, S2 specifically includes: The components, sensors, and device nodes of the trusted gated observation sequence are input into the node type coding layer to obtain type coding features; The electrical state, temperature state, environmental state, and operation and maintenance records of the trusted gated observation sequence are input into the state coding layer to obtain state coding features; The sampling time, sunshine interval, and operation and maintenance cycle are input into the time-series coding layer to obtain time-series coding features; Input the node credibility into the credibility-gated mapping layer to obtain the credibility embedding feature; Type encoding features, state encoding features, time-series encoding features, and credibility embedding features are input into the model embedding end to obtain node embedding features.

[0042] The node embedding features simultaneously include node type, current state, time location, and trusted gating information, enabling the model to distinguish the different roles of component nodes, sensor nodes, and device nodes in the intelligent operation and maintenance of photovoltaic modules.

[0043] In this embodiment, the node type coding layer assigns different type codes to component nodes, sensor nodes, and device nodes respectively. The component node code reflects the operating status of the component itself, the sensor node code reflects the sampling channel and data source, and the device node code reflects the operating status of the string, combiner box, and inverter.

[0044] The state coding layer converts current, voltage, power, backplane temperature, irradiance, wind speed, string status, and device status into state coding features.

[0045] The temporal coding layer converts sampling time, sunshine interval, and operation and maintenance cycle into temporal coding features.

[0046] The Trusted Gated Mapping layer converts node trustworthiness into trustworthiness embedded features, enabling node trustworthiness to participate in dependency learning, state prediction, dependency reconstruction, and anomaly scoring within the model.

[0047] The model embedding end fuses type encoding features, state encoding features, time-series encoding features and credibility embedding features into node embedding features. The node embedding features are arranged in the order of nodes, and each node corresponds to a set of fused feature terms.

[0048] In this embodiment, the sunshine interval is specifically divided into a low-irradiance interval, a medium-irradiance interval, and a high-irradiance interval.

[0049] Irradiance less than 300W / m² is classified as low irradiance range, irradiance greater than or equal to 300W / m² and less than 700W / m² is classified as medium irradiance range, and irradiance greater than or equal to 700W / m² is classified as high irradiance range.

[0050] The operation and maintenance cycle is specifically divided into the normal operation cycle, the inspection and confirmation cycle, the maintenance execution cycle, and the retest and confirmation cycle.

[0051] The normal operating cycle corresponds to a sampling window with no maintenance work orders and no abnormal feedback; The inspection confirmation cycle corresponds to the sampling window where an inspection task has been generated but the review has not been completed; The maintenance execution cycle corresponds to the sampling window during which cleaning, wiring inspection, component replacement, or sensor calibration is being performed; The retest confirmation cycle corresponds to the sampling window for status retesting after the completion of maintenance actions.

[0052] The aforementioned interval and period encodings are fed into the temporal encoding layer, enabling the model to use different node state expressions under different sunlight conditions and operation and maintenance stages.

[0053] In the low-irradiance range, the model reduces the proportion of power bias in the outlier score; Under high irradiation range, the model increases the weights of backplane temperature deviation, power deviation and thermal anomaly diffusion in fault source decoupling; During the maintenance execution cycle and the retest confirmation cycle, the model assigns a higher closed-loop update weight to the execution feedback data.

[0054] In this embodiment, the improved graph deviation network model includes a node type encoding layer, a state encoding layer, a temporal encoding layer, a trusted gating mapping layer, a dependency learning layer, a graph attention aggregation layer, a state prediction branch, a dependency reconstruction branch, a deviation calculation end, a fault source decoupling layer, and a closed-loop update interface.

[0055] The node type coding layer receives the category tags of component nodes, sensor nodes, and device nodes, and outputs type coding features; The status coding layer receives current, voltage, power, backplane temperature, irradiance, wind speed, string status, and device status, and outputs status coding features. The temporal coding layer receives the sampling time, sunshine interval, and operation and maintenance cycle, and outputs temporal coding features; the trusted gating mapping layer receives the node trustworthiness and outputs trustworthiness embedding features. The dependency learning layer receives node embedded features and candidate dependency edges, and outputs contextualized dependency edge weights. The graph attention aggregation layer receives a dynamic graph bias dependency graph and outputs neighborhood aggregation features. The state prediction branch outputs predicted values ​​for current, voltage, power, and backplane temperature for the next sampling period; The reconstruction branch outputs the edge weight and edge type reconstruction results for the next sampling period; The deviation calculation terminal outputs an anomaly scoring matrix; the fault source decoupling layer outputs anomaly paths, fault source categories, and maintenance actions. The closed-loop update interface receives feedback marker data and writes it back to the model parameters.

[0056] In this embodiment, S3 specifically includes: The candidate association matrix is ​​obtained by calculating the association similarity between component, sensor and device nodes based on the node embedding features; Based on the candidate association matrix, electrical neighbors, temperature neighbors, environmental neighbors, and operation and maintenance neighbors are selected to obtain a set of candidate dependency edges; The candidate dependency edge set is input into the improved graph bias network dependency learning layer, and the scene-specific dependency edge weights are calculated by combining the sensor network topology. Sparse selection and edge type labeling are performed according to the scenario-based dependency edge weights to obtain the set of retained dependency edges; By combining the retained dependency edge set with the node embedding features, a dynamic graph deviation dependency graph that changes with the operation and maintenance status is obtained.

[0057] The candidate association matrix is ​​not used directly as the final dependency relationship, but rather as an intermediate input to the dependency learning layer.

[0058] The dependency learning layer recalculates the dependency edge weights based on the component's operating status, sensor reliability, initial topology, and historical operation and maintenance records, enabling the final generated dynamic graph deviation dependency graph to adjust to changes in occlusion, hot spots, mismatch, temperature diffusion, and sensor drift.

[0059] In this embodiment, the candidate association matrix is ​​obtained by calculating the similarity between the embedded features of nodes. The similarity between component nodes is used to form candidate relationships between components, the similarity between component nodes and sensor nodes is used to form candidate relationships between sampling channels, and the similarity between component nodes and device nodes is used to form candidate relationships between strings, combiner boxes, and inverters.

[0060] The candidate dependency edge set includes electrical neighbor edges, temperature neighbor edges, environmental neighbor edges, and operation and maintenance neighbor edges. Electrical neighbor edges are derived from electrical connection relationships and same-string relationships; temperature neighbor edges are derived from spatial adjacency and backplane temperature change relationships; environmental neighbor edges are derived from areas with the same irradiance and areas with similar wind speeds; and operation and maintenance neighbor edges are derived from the associations between historical inspection, cleaning, wiring inspection, and replacement records.

[0061] By using a set of candidate dependency edges, the dependency learning layer can establish a learnable bridge between a fixed topology and running data.

[0062] This embodiment further includes: Read the electrical codes, temperature codes, environmental codes, and confidence embeddings of the nodes at both ends of the candidate dependent edge to obtain the edge weight input features; The attention value of the candidate dependent edge is calculated based on the edge weight input features to obtain the first edge weight; The weights of the first edge are adjusted based on the electrical connections, spatial adjacencies, and temperature conduction relationships in the initial dependent edges to obtain the weights of the second edge. Electrical mismatch weights, occlusion propagation weights, thermal anomaly propagation weights, and drift suppression weights are formed based on the same string mismatch records, occlusion records, hot spot records, and sensor drift records. By inputting the second edge weight, electrical mismatch weight, occlusion propagation weight, thermal anomaly diffusion weight, and drift suppression weight into the dependency learning layer, the contextualized dependency edge weights are obtained.

[0063] In this embodiment, the electrical mismatch weight is determined based on the current deviation and voltage deviation of the components in the same string. When the current deviation of the components in the same string exceeds 1.3 times the historical normal current fluctuation limit and the voltage deviation exceeds 1.2 times the historical normal voltage fluctuation limit, the electrical mismatch weight is set between 0.75 and 1.00. When only one deviation exceeds the corresponding upper limit, the electrical mismatch weight is set between 0.40 and 0.74; When neither deviation exceeds the corresponding upper limit, the electrical mismatch weight is set between 0.00 and 0.39.

[0064] The shading propagation weight is determined based on the synchronous decrease in power of adjacent components and the change in irradiance. When the synchronous decrease in power of multiple adjacent components exceeds 8% and the decrease in irradiance does not exceed 3%, the shading propagation weight is set between 0.70 and 1.00. When the power synchronization decreases between 3% and 8%, the occlusion propagation weight is set between 0.35 and 0.69; When the power synchronous decrease is less than 3%, the occlusion propagation weight is set between 0.00 and 0.34.

[0065] The thermal anomaly diffusion weight is determined based on the backsheet temperature rise and the temperature difference between adjacent modules. When the backsheet temperature of a single module is more than 8°C higher than that of adjacent modules, the thermal anomaly diffusion weight is set between 0.80 and 1.00. When the temperature is 4°C to 8°C higher, the thermal anomaly diffusion weight is set between 0.40 and 0.79; when the temperature is less than 4°C higher, the thermal anomaly diffusion weight is set between 0.00 and 0.39.

[0066] The drift suppression weight is determined based on the sensor node offset and neighborhood dependency changes. When a single sensor node continues to offset and the dependent edges of adjacent components are not synchronized abnormally, the drift suppression weight is set between 0.70 and 1.00. When sensor node offset is accompanied by changes in part of the neighborhood, the drift suppression weight is set between 0.30 and 0.69; When sensor node offset is synchronized with abnormal state of multiple components, the drift suppression weight is set between 0.00 and 0.29. The above weight range is a specific configuration of this embodiment and does not limit the scope of protection.

[0067] In this embodiment, the dynamic graph deviation dependency graph is composed of node embedding features, a set of retained dependency edges, contextualized dependency edge weights, and edge type labels.

[0068] Edge type markings include electrical connection edges, temperature conduction edges, environmental synchronization edges, maintenance event edges, and sensor drift suppression edges.

[0069] The dynamic graph deviation dependency graph differs from the fixed topology graph. The fixed topology graph only records the installation relationship and electrical connection relationship of photovoltaic modules, while the dynamic graph deviation dependency graph also records the implicit dependencies learned by the model from the trusted gated observation sequence.

[0070] When photovoltaic modules are shaded, covered in dust, or experience abnormal heat diffusion, the relevant edge weights will change accordingly. When sensor drift occurs, drift suppression edge weights will reduce the impact of abnormal readings on component fault diagnosis; When a string experiences a wiring abnormality, the electrical connection edge and the mismatch weight of the same string will change.

[0071] This dynamic graph of deviation and dependency serves as the core data structure for subsequent state prediction, dependency reconstruction, and deviation scoring.

[0072] refer to Figure 3 In this embodiment, S4 specifically includes: The dynamic graph bias dependency graph is input into the graph attention aggregation layer to obtain the component node neighborhood aggregation features; The neighborhood aggregation features are input into the state prediction branch to calculate the current, voltage, power and backplane temperature in the next sampling period to obtain the predicted state. Input the dynamic graph deviation dependency graph into the dependency reconstruction branch, reconstruct the edge weights and edge types of the retained dependency edges, and obtain the dependency reconstruction result; The predicted state and the dependency reconstruction results are mapped according to component number, group number and sampling time to obtain a dual-branch output result; The output of the two branches is written into the deviation calculation terminal of the improved graph deviation network to obtain the deviation calculation input.

[0073] The state prediction branch focuses on whether the state of component nodes deviates from the normal evolution pattern, while the dependency refactoring branch focuses on whether the relationship between components, sensors, and devices deviates from the normal dependency structure.

[0074] The two branches work together, enabling the model to detect both state anomalies and relationship anomalies simultaneously.

[0075] In this embodiment, the graph attention aggregation layer aggregates the features of neighboring nodes according to the edge weights in the dynamic graph deviation dependency graph.

[0076] The state prediction branch includes a neighborhood feature fusion unit, a state mapping unit, and a multi-state output unit connected in sequence. The neighborhood feature fusion unit receives the embedded features of the component node itself and the embedded features of neighboring nodes. The state mapping unit converts the fused features into state features for the next sampling period. The multi-state output unit outputs the predicted values ​​of current, voltage, power, and backplane temperature, respectively.

[0077] The dependency reconstruction branch includes an edge feature extraction unit, an edge weight reconstruction unit, and an edge type reconstruction unit. The edge feature extraction unit receives the embedded features of the two endpoints and the contextualized dependency edge weights. The edge weight reconstruction unit outputs the edge weight reconstruction value for the next sampling period, and the edge type reconstruction unit outputs the edge type reconstruction result for the next sampling period. The state prediction branch and the dependency reconstruction branch output independently, but are merged at the deviation calculation end, so that anomaly detection is not limited to a single power deviation.

[0078] In this embodiment, the predicted states output by the state prediction branch include the predicted values ​​of the module current, module voltage, module power, and module backsheet temperature.

[0079] The dependency refactoring results output by the dependency refactoring branch include edge weight refactoring values ​​and edge type refactoring results.

[0080] The edge weight reconstruction value is used to determine whether the current dependent edge strength deviates from the normal relationship learned by the model, and the edge type reconstruction result is used to determine whether the current dependent edge category has undergone abnormal transformation.

[0081] If the power prediction deviation of a component node is large, and the corresponding edge weight reconstruction deviation is also large, then the component node is more likely to have an operational anomaly. If the reading deviation of a sensor node is large, but the corresponding dependent edge weight reconstruction deviation is not obvious, and the drift suppression weight is high, then the anomaly is more likely to be sensor drift.

[0082] Through a bi-branch design, the improved graph deviation network model can analyze anomalies from both the node state and edge relationships perspectives.

[0083] In this embodiment, S5 specifically includes: The predicted state is aligned with the actual state by nodes and by time to obtain state-aligned data. Based on the state alignment data, the power deviation, current deviation, temperature deviation, and irradiation response deviation are calculated to obtain the state deviation set; Based on the dependency reconstruction results and the dynamic graph deviation dependency graph, the edge weights, edge types, and neighborhood changes are calculated to obtain the dependency deviation set; Based on node credibility, the state deviation set and the dependency deviation set are weighted and fused to obtain node anomaly score and edge anomaly score; By combining node anomaly scores and edge anomaly scores, an improved graph bias network anomaly score matrix is ​​generated.

[0084] The state deviation set reflects the difference between the actual state of a component and the predicted state of the model, while the dependency deviation set reflects the difference between the current dependency and the refactored dependency.

[0085] Node credibility is incorporated into anomaly score fusion, ensuring that sensor nodes with low credibility do not directly increase component failure scores, while nodes with high credibility that also exhibit state deviation and dependency deviation will receive higher anomaly scores.

[0086] In this embodiment, the anomaly scoring matrix is ​​organized according to the photovoltaic module nodes and the sampling time, and simultaneously records the node anomaly score, edge anomaly score, edge type change and anomaly duration.

[0087] In the anomaly scoring matrix, node anomaly scores mainly come from deviations in power, current, temperature, and irradiation response, while edge anomaly scores mainly come from scene-dependent edge weight changes, edge type reconstruction differences, and neighborhood changes.

[0088] The dynamic threshold is determined by the abnormal score distribution of historical normal operation samples. The initial node dynamic threshold is taken as the 95th percentile of the abnormal score of historical normal nodes, and the initial edge dynamic threshold is taken as the 95th percentile of the abnormal score of historical normal edges. When samples from low-irradiance zones account for more than 60% of the current sampling window, the dynamic threshold is increased by 5% to 10%. When samples from high-irradiation zones account for more than 60% of the current sampling window and the average node confidence level is higher than 0.85, the dynamic threshold remains at its initial value. When the average node credibility is below 0.60, the dynamic threshold is increased by 10% to 20%, and the drift suppression weight is increased.

[0089] The aforementioned dynamic threshold settings can reduce false alarms caused by low irradiance fluctuations and low-reliability sensor nodes.

[0090] In this embodiment, the node anomaly scores in the anomaly scoring matrix are divided into normal intervals, attention intervals, abnormal intervals, and severe anomaly intervals.

[0091] When a node's anomaly score is below 80% of the node's dynamic threshold, it is classified as being in the normal range. When a node's anomaly score is greater than or equal to 80% of the node's dynamic threshold but lower than the node's dynamic threshold, it is classified as a monitoring interval. A node is classified as an abnormal interval when its anomaly score is greater than or equal to the node's dynamic threshold but less than 1.5 times the node's dynamic threshold. When a node's anomaly score is greater than or equal to 1.5 times the node's dynamic threshold, it is classified as a severely abnormal interval.

[0092] The edge anomaly score is also divided according to the edge dynamic threshold. The range below 80% of the edge dynamic threshold is the normal range, the range above or equal to 80% of the edge dynamic threshold but below the edge dynamic threshold is the attention range, the range above or equal to the edge dynamic threshold but below 1.5 times the edge dynamic threshold is the abnormal range, and the range above or equal to 1.5 times the edge dynamic threshold is the severe abnormal range.

[0093] The above intervals are only used for ranking in operation and maintenance decisions and do not change the continuous numerical output of the anomaly scoring matrix.

[0094] In this embodiment, S6 specifically includes: The abnormal scoring matrix is ​​compared with the dynamic threshold to obtain abnormal nodes, abnormal edges and abnormal time periods; Based on the component location, string location, and sensor location of the abnormal node, a set of abnormal objects is obtained; The abnormal propagation chain is traced based on the edge type, propagation direction, and edge abnormality score of the abnormal edge to obtain the abnormal path; Input the set of abnormal objects and abnormal paths into the fault source decoupling layer to obtain the categories of component body abnormality, connection abnormality, environmental disturbance or sensor drift. Based on the fault source category, matching actions such as cleaning, infrared inspection, wiring check, component replacement, or sensor calibration are performed to obtain maintenance actions. The fault source decoupling layer receives abnormal nodes, abnormal edges, abnormal paths, and abnormal time periods, and makes a comprehensive judgment based on component location, edge type, propagation direction, and fault records to distinguish the source of the anomaly.

[0095] In this embodiment, the fault source decoupling layer includes an abnormal object location unit, an abnormal path tracing unit, a fault source classification unit, and an operation and maintenance action matching unit.

[0096] The abnormal object location unit maps abnormal nodes to photovoltaic modules, sensors, strings, combiner boxes, or inverters; The abnormal path tracing unit determines the abnormal starting point and affected nodes based on the abnormal edge direction and edge type; The fault source classification unit outputs the component body anomaly, connection anomaly, environmental disturbance or sensor drift category based on the abnormal path, edge anomaly score, state deviation set and dependency deviation set. The operation and maintenance action matching unit outputs cleaning, infrared inspection, wiring check, component replacement or sensor calibration actions based on the fault source category.

[0097] The abnormal path includes the fault origin node, affected nodes, associated sensor nodes, edge type, and edge anomaly score. Operation and maintenance personnel can locate the component, sensor, and device locations that need to be reviewed based on the abnormal path.

[0098] In this embodiment, the criteria for classifying fault source categories are as follows: The component itself is abnormal. The abnormal score of a single component node exceeds the node dynamic threshold, the temperature deviation exceeds 4℃, the power deviation exceeds 8%, and the thermal anomaly diffusion weight reaches 0.40 or higher. Connection anomalies correspond to current or voltage deviations in multiple components within the same string group, with an electrical mismatch weight exceeding 0.75, and the anomalies are concentrated on the electrical connection side. Environmental disturbances correspond to synchronous changes in multiple adjacent components within the same sampling window, the occlusion propagation weight reaches 0.70 or higher, and the dependency reconstruction deviation does not exceed 1.2 times the edge dynamic threshold; Sensor drift corresponds to a continuous offset of a single sensing node, with a drift suppression weight of 0.70 or higher, and the state deviation and dependency deviation of adjacent components do not increase synchronously.

[0099] The above classification criteria are the specific judgment rules for the fault source decoupling layer in this embodiment, and the parameters can be corrected based on the photovoltaic power station operation and maintenance records.

[0100] In this implementation, different fault source categories correspond to different maintenance actions. Component body abnormalities correspond to infrared inspection, component appearance verification, bypass diode detection, and component replacement. For connection abnormalities, check the corresponding wiring terminals, retest the string electrical circuit, verify the combiner box terminals, and recheck the status of the inverter side. Environmental disturbances and corresponding obstruction investigations, component surface cleaning, removal of surrounding obstructions, and retesting for confirmation; Sensor drift corresponds to sensor calibration, communication link inspection, sampling channel verification, and sensor replacement.

[0101] If the abnormal score of a node in the abnormal score matrix is ​​within the attention range and the abnormal score of an edge does not exceed the dynamic threshold, the system outputs a retest action. If both the node anomaly score and the edge anomaly score are in the anomaly range, the system outputs an inspection or maintenance action. If the node anomaly score or edge anomaly score is in the severe anomaly range, the system outputs a higher priority maintenance action.

[0102] Priority is divided into three levels: Level 1 corresponds to retesting, Level 2 corresponds to inspection or cleaning, and Level 3 corresponds to maintenance, replacement, or calibration.

[0103] Level 1 priority corresponds to an anomaly score that does not exceed 1.1 times the dynamic threshold; Level 2 priority corresponds to an anomaly score that reaches 1.1 to 1.5 times the dynamic threshold; and Level 3 priority corresponds to an anomaly score that exceeds 1.5 times the dynamic threshold.

[0104] In this embodiment, S7 specifically includes: Obtain execution records after cleaning, inspection, maintenance, replacement and calibration to obtain operation and maintenance feedback data; The operation and maintenance feedback data is matched with abnormal nodes, abnormal edges, abnormal paths and fault source categories to obtain feedback tag data; The node credibility, scenario-based dependency edge weights, and dynamic thresholds are updated based on the feedback tag data to obtain the first update parameters; The second update parameter is obtained by updating the state prediction branch and dependency refactoring branch based on the feedback tag data. The first and second update parameters are written into the improved graph deviation network model to obtain the closed-loop operation and maintenance model.

[0105] The execution feedback data includes maintenance confirmation results, power changes before and after cleaning, infrared image verification results, wiring inspection results, component replacement records, sensor calibration records, and false alarm records.

[0106] After the feedback-labeled data enters the model, the node confidence, dependency edge weights, and dynamic thresholds are corrected first, and then the state prediction branch and dependency reconstruction branch are incrementally updated to avoid excessive model deviation caused by a single feedback.

[0107] In this embodiment, the feedback flag data is divided into confirmation anomaly flags, false alarm flags, repaired flags, and sensor calibration flags.

[0108] Confirm the anomaly flag to the maintenance personnel to determine if the component itself is faulty, the connection is faulty, or there is environmental disturbance. The false alarm flag corresponds to an on-site verification that no abnormalities were found in the components or equipment. After the repaired markers have been cleaned, inspected, replaced, or calibrated, the abnormal score has dropped to below 80% of the node's dynamic threshold. After calibration, the readings of the sensor nodes corresponding to the sensor calibration marks return to the normal range in the neighborhood.

[0109] The confirmed anomaly marker is used to enhance the dependency edge weights and threshold calibration of the corresponding fault category; the false alarm marker is used to increase the dynamic threshold or enhance the drift suppression weight; the repaired marker is used to reduce the persistent anomaly weight of the corresponding abnormal path; and the sensor calibration marker is used to restore the credibility of the sensor node and reduce the drift suppression weight.

[0110] During the feedback update process, the single adjustment range of node credibility does not exceed 0.15, the single adjustment range of scenario-dependent edge weight does not exceed 0.20, and the single adjustment range of dynamic threshold does not exceed 15%, in order to ensure the stability of the closed-loop update process.

[0111] In this embodiment, the training data comes from the historical operation database and operation and maintenance management platform of photovoltaic power plants. The data range includes continuously collected current, voltage, power, backsheet temperature, irradiance, wind speed, string status, combiner box status, inverter status, inspection records, cleaning records, maintenance records, component replacement records, and sensor calibration records.

[0112] The training data is divided into time windows, each containing continuous sampling points. Each sampling point records the status of component nodes, sensor nodes, and device nodes.

[0113] Tag data comes from manual inspection confirmation, infrared verification, comparison before and after cleaning, wiring inspection results, and component replacement records. Tag types include normal, component abnormal, connection abnormal, environmental disturbance, and sensor drift.

[0114] For historical normal operation data that lacks clear labels, the model uses the deviation between the predicted state and the actual state, as well as the dependency reconstruction deviation, for self-supervised training. For data with confirmed operation and maintenance results, the model adds fault source category learning and dynamic threshold calibration.

[0115] In this embodiment, the training objectives of the improved graph bias network model include state prediction error, dependency reconstruction error, edge sparsity constraint, trust gating constraint, and fault source classification error.

[0116] The state prediction error comes from the difference between the predicted state output by the state prediction branch and the actual state. The dependency refactoring error comes from the difference between the dependency refactoring result output by the dependency refactoring branch and the retained dependency edges in the dynamic graph deviation dependency graph; The edge sparsity constraint comes from the ratio limit between the number of retained dependency edges and the number of candidate dependency edges; Trusted gating constraints are derived from the matching relationship between node trustworthiness and sampling missing rate, mutation frequency, calibration records, and differences between neighboring nodes; The fault source classification error arises from the difference between the fault source category output by the fault source decoupling layer and the feedback label data.

[0117] State prediction error and dependency reconstruction error have high weights in the initial training, fault source classification error participates in the update of labeled samples, and trusted gating constraints participate in the update of all samples.

[0118] The training objectives are formed by weighted combination. The weight of state prediction error can be set to 0.35, the weight of dependency reconstruction error can be set to 0.25, the weight of edge sparsity constraint can be set to 0.10, the weight of trusted gating constraint can be set to 0.10, and the weight of fault source classification error can be set to 0.20. The above weights can be adjusted according to the data labeling ratio.

[0119] In this embodiment, the model training parameters are configured as follows: The sampling window length is set to 24 consecutive sampling points. When the sampling period is 5 minutes, one window covers 2 hours of operation. The number of candidate neighbors is set to 8 for each node; The node embedding feature is set to 64 feature channels; The dependency learning layer is set to a 2-layer graph attention structure; the state prediction branch is set to a 2-layer fully connected mapping structure. The dependency refactoring branch includes one edge weight refactoring unit and one edge type refactoring unit; Each batch of training data contains 64 time windows; The initial learning rate was set to 0.001; the number of training epochs was set to 100. Training is stopped when the aggregate loss fails to decrease for 10 consecutive rounds.

[0120] The above parameters are the training configuration in this embodiment and do not limit the scope of protection; In different sites, the sampling window length can be adjusted to 12 to 48 sampling points, the number of candidate neighbors can be adjusted to 4 to 12, the number of node embedding feature channels can be adjusted to 32, 64 or 128, and the number of training rounds can be adjusted to 50 to 200 rounds.

[0121] The aforementioned adjustable range is only used to illustrate the flexibility of model deployment and does not change the technical solution of the joint operation of trusted gating, scenario-based dependency edge weights, state prediction branches, dependency reconstruction branches, anomaly scoring matrix and closed-loop update mechanism.

[0122] In this embodiment, after the model is deployed, intelligent operation and maintenance analysis of photovoltaic modules is performed online.

[0123] Each time the system receives a new sampling window, it first generates a reliable gated observation sequence, then inputs the improved graph bias network model to generate a dynamic graph bias dependency graph, predicted state, dependency reconstruction results, and anomaly scoring matrix.

[0124] When a node or side of the anomaly scoring matrix exceeds the dynamic threshold, the fault source decoupling layer outputs the anomaly path, fault source category, and operation and maintenance action.

[0125] After maintenance personnel complete cleaning, inspection, repair, replacement or calibration, the execution record is written to the closed-loop update module. The closed-loop update module updates the node credibility, scenario-based dependency edge weights and dynamic thresholds, and performs incremental updates to the state prediction branch and dependency reconstruction branch when the sample quantity requirement is met.

[0126] The sample quantity requirement is that there must be no less than 20 feedback tag data for the same fault source category, or no less than 10 valid feedback records generated in the same component area for 7 consecutive days. Incremental update will be started after either of the above conditions is met. If the above conditions are not met, only the node credibility, dynamic threshold, and edge weight local parameters are updated, and the main parameters of the state prediction branch and the dependency reconstruction branch are not updated.

[0127] refer to Figures 1 to 3 This embodiment also provides a photovoltaic module intelligent operation and maintenance system based on sensor networks, which performs the above method.

[0128] The system includes a trusted gating module, a node embedding module, a dependency learning module, a prediction and reconstruction module, a bias scoring module, a fault decoupling module, and a closed-loop update module.

[0129] The trusted gating module acquires intelligent operation and maintenance data of photovoltaic modules, initial dependency edges and node trustworthiness, and generates a trusted gating observation sequence. The node embedding module inputs the trusted gated observation sequence into the improved graph bias network model to generate node embedding features; The dependency learning module learns the contextual dependency edges of components, sensors, and devices based on the node embedding features, and generates a dynamic graph bias dependency graph. The prediction and reconstruction module performs state prediction and dependency reconstruction based on the dynamic graph deviation dependency graph, generating predicted state and dependency reconstruction results; The deviation scoring module calculates the deviation based on the predicted state, the actual state, and the dependency reconstruction results, and generates an anomaly scoring matrix. The fault decoupling module locates abnormal nodes and abnormal edges based on the abnormal scoring matrix, and generates fault source categories, abnormal paths and maintenance actions. The closed-loop update module updates the credibility, dependency edge weights, and dual branches based on execution feedback to generate a closed-loop operation and maintenance model.

[0130] In this embodiment, the trusted gating module includes a data access unit, a node organization unit, an initial dependency edge generation unit, a credibility calculation unit, and an observation sequence generation unit.

[0131] The data access unit connects to the component-level acquisition equipment, environmental monitoring equipment, string monitoring equipment, combiner box, inverter and operation and maintenance management platform to read the intelligent operation and maintenance data of photovoltaic modules; The node sorting unit sorts the data by component number, sensor number, device number, and sampling time; The initial dependency edge generation unit establishes initial dependency edges based on electrical connections, spatial adjacency, temperature conduction, and power correlation. The credibility calculation unit calculates the node credibility based on the sampling missing rate, mutation frequency, calibration records, and differences between adjacent nodes; The observation sequence generation unit corrects the node attribute sequence according to the node credibility and outputs a credible gated observation sequence.

[0132] The data format output by the trusted gating module is directly read by the node embedding module. The data fields include node number, node type, sampling time, node attributes, node trustworthiness, and initial dependency edge label.

[0133] In this embodiment, the node embedding module includes a node type encoding unit, a state encoding unit, a timing encoding unit, a trusted gating mapping unit, and a model embedding end.

[0134] The node type encoding unit outputs type encoding features, the state encoding unit outputs state encoding features, the temporal encoding unit outputs temporal encoding features, the trustworthy gating mapping unit outputs trustworthiness embedding features, and the model embedding end fuses the above features and outputs node embedding features.

[0135] The dependency learning module includes a candidate association matrix generation unit, a candidate dependency edge filtering unit, a scenario-based dependency edge weight generation unit, a sparse selection unit, and a dynamic graph bias dependency graph generation unit.

[0136] The contextualized dependency edge weight generation unit reads the electrical mismatch weight, occlusion propagation weight, thermal anomaly diffusion weight, and drift suppression weight, and inputs these weights together with the second edge weight into the dependency learning layer to output the contextualized dependency edge weight.

[0137] The sparse selection unit retains the dependent edges related to the operation and maintenance status according to the edge weight sorting. The dynamic graph deviation dependency graph generation unit combines the retained dependent edges, edge weights and node embedding features into a dynamic graph deviation dependency graph.

[0138] In this embodiment, the prediction and reconstruction module includes a graph attention aggregation layer, a state prediction branch, and a dependency reconstruction branch.

[0139] The graph attention aggregation layer receives the dynamic graph bias dependency graph and outputs neighborhood aggregation features; The state prediction branch outputs the predicted state based on the neighborhood aggregation features; The dependency refactoring branch outputs the dependency refactoring results based on the dynamic graph deviation dependency graph. The deviation scoring module includes a state deviation calculation unit, a dependency deviation calculation unit, a trusted weighted fusion unit, and an anomaly scoring matrix generation unit.

[0140] The state deviation calculation unit calculates power deviation, current deviation, temperature deviation, and irradiation response deviation; The deviation calculation unit calculates the changes in edge weight, edge type, and neighborhood; The trusted weighted fusion unit fuses state deviations and dependency deviations according to node trustworthiness. The anomaly scoring matrix generation unit outputs node anomaly scores, edge anomaly scores, anomaly time periods, and anomaly durations.

[0141] In this embodiment, the fault decoupling module includes an abnormal object location unit, an abnormal path tracing unit, a fault source classification unit, and an operation and maintenance action matching unit.

[0142] The abnormal object location unit locates components, sensors, strings, combiner boxes, or inverters based on the abnormal node location; The abnormal path tracing unit forms abnormal paths based on the direction and type of abnormal edges; The fault source classification unit outputs component-level anomalies, connection anomalies, environmental disturbances, or sensor drift categories based on abnormal paths, state deviation sets, dependency deviation sets, and scenario-based dependency edge weights. The operation and maintenance action matching unit outputs cleaning, infrared inspection, wiring check, component replacement or sensor calibration actions based on the fault source category.

[0143] The closed-loop update module includes a feedback data access unit, a feedback tag generation unit, a credibility update unit, a boundary weight update unit, a threshold update unit, and a dual-branch update unit.

[0144] The feedback data access unit reads the operation and maintenance execution records, the feedback tag generation unit matches the execution records with abnormal nodes, abnormal edges, abnormal paths and fault source categories, the credibility update unit updates the node credibility, the edge weight update unit updates the scenario-based dependency edge weights, the threshold update unit updates the dynamic threshold, and the dual-branch update unit updates the state prediction branch and dependency reconstruction branch when the sample quantity requirement is met.

[0145] In this embodiment, the output interface of the trusted gating module transmits the trusted gating observation sequence to the node embedding module. The trusted gating observation sequence includes node number, node type, sampling time, node attributes, node trustworthiness, and initial dependency edge label. The output interface of the node embedding module passes node embedding features to the dependency learning module. The node embedding features include type encoding features, state encoding features, temporal encoding features, and credibility embedding features. The output interface of the dependency learning module passes a dynamic graph bias dependency graph to the prediction and reconstruction module. The dynamic graph bias dependency graph contains a set of retained dependency edges, contextualized dependency edge weights, and edge type labels. The output interface of the prediction and reconstruction module transmits the prediction status and dependent reconstruction results to the deviation scoring module; the output interface of the deviation scoring module transmits the anomaly scoring matrix to the fault decoupling module. The output interface of the fault decoupling module transmits abnormal nodes, abnormal edges, abnormal paths, fault source categories, and maintenance actions to the closed-loop update module. The closed-loop update module writes back node credibility, scenario-based dependency edge weights, dynamic thresholds, state prediction branch parameters, and dependency reconstruction branch parameters to the trusted gating module, dependency learning module, and prediction reconstruction module, respectively.

[0146] Through the above module connections, the system can complete reliable observation, dynamic dependency learning, prediction reconstruction, anomaly scoring, fault decoupling, and model updating around the improved graph bias network model.

[0147] In this embodiment, the system can be deployed on the edge server of the photovoltaic power station, the operation and maintenance management server, or the cloud platform.

[0148] The edge server is responsible for receiving real-time data from components and sensors, and completing the generation of trusted gated observation sequences, node embedding, dynamic graph bias dependency graph generation, and preliminary anomaly scoring. The operation and maintenance management server is responsible for decoupling fault sources, generating operation and maintenance actions, and recording execution feedback. The cloud platform is responsible for model training, parameter backup, and cross-site parameter migration.

[0149] The three types of deployment nodes exchange trusted gated observation sequences, anomaly scoring matrices, anomaly paths, fault source categories, and feedback marker data through the operation and maintenance data interface.

[0150] When the network is interrupted, the edge server can continue to output the anomaly scoring matrix and temporary operation and maintenance actions based on the local model. After the network is restored, it will synchronize and feed back the marked data with the operation and maintenance management server and the cloud platform.

[0151] This deployment method enables the system to adapt to centralized photovoltaic power plants, distributed photovoltaic arrays, and multi-regional photovoltaic module operation and maintenance scenarios.

[0152] In this embodiment, the intelligent operation and maintenance method and system for photovoltaic modules based on sensor networks solves the problem of inconsistent reliability of sensor data by using a trusted gating observation sequence; The problem of changing operational relationships between photovoltaic modules due to environmental and fault conditions can be addressed by using scenario-based dependency edge weights. The problem of single-state prediction failing to express dependency relationships is solved by using a two-branch approach of state prediction and dependency refactoring. The problem of jointly representing node anomalies and edge anomalies is solved by using an anomaly scoring matrix; The fault source decoupling layer solves the problem of difficulty in distinguishing between component abnormalities, connection abnormalities, environmental disturbances, and sensor drift. The closed-loop update module addresses the issue of operational feedback failing to inform the model.

[0153] Thus, the entire method and system operate around the improved graph bias network model, forming a complete intelligent operation and maintenance technology chain for photovoltaic modules, from reliable observation, dynamic graph bias dependency graph, prediction reconstruction, bias scoring, fault source decoupling to closed-loop update.

[0154] Example 1: To verify the feasibility of this invention in practice, it was applied to a smart operation and maintenance scenario for photovoltaic modules in a photovoltaic power station. This scenario involved 12 strings, each connected to 30 photovoltaic modules. Current, voltage, power, and backsheet temperature data acquisition nodes were deployed on the module side, while irradiance, wind speed, and ambient temperature data acquisition nodes were deployed on the environmental side. Data on the combiner box, inverter, and string switch status were integrated into the equipment side. The sampling interval was set to 5 minutes. Within one continuous operating cycle, the operation and maintenance system obtained 4320 sets of module status records, 4320 sets of environmental status records, and 126 inspection and verification records. The module number, sensor number, equipment number, and sampling time were correlated to form smart operation and maintenance data for the photovoltaic modules.

[0155] In this embodiment, the system establishes initial dependency edges based on electrical connection relationships, component installation locations, backplane temperature conduction relationships, and power synchronization change relationships, and writes each component, each sensor node, and each string device node into the initial operation and maintenance dependency graph. For nodes with a sampling missing rate higher than 3%, more than 6 single-point jumps, or calibration offset records, the system reduces the node's credibility; for nodes where the state difference between adjacent components in the same string is consistently less than a set range, the system increases the neighborhood credibility. After correcting the current, voltage, power, backplane temperature, and irradiance sequences through a trusted gating layer, a trusted gating observation sequence is formed, reducing the impact of abnormal sampling and sensor drift on subsequent model judgments.

[0156] After inputting the trusted gated observation sequence into the improved graph bias network model, the node type encoding layer encodes component nodes, sensor nodes, and device nodes respectively; the state encoding layer encodes electrical state, temperature state, environmental state, and operation and maintenance records; the time sequence encoding layer encodes sampling time, sunshine interval, and operation and maintenance cycle; and the trusted gated mapping layer converts node trustworthiness into trustworthiness embedding features. The fusion of these features yields the node embedding features, which serve as input for dynamic graph bias dependency graph learning.

[0157] During dependency learning, the model calculates the similarity of associations between components, sensors, and devices based on node embedding features, filtering electrical neighbors, temperature neighbors, environmental neighbors, and maintenance neighbors, and generating candidate dependency edges by combining the sensor network topology. The system further adjusts the weights of the candidate dependency edges based on electrical connections, spatial adjacency, temperature conduction, in-band mismatch records, occlusion records, hot spot records, and sensor drift records, forming electrical mismatch weights, occlusion propagation weights, thermal anomaly diffusion weights, and drift suppression weights, and generating contextualized dependency edge weights accordingly. After retaining the dependency edge set and combining it with node embedding features, a dynamic graph deviation dependency graph that changes with the operating state is formed.

[0158] The dynamic graph deviation dependency graph enters the state prediction branch and the dependency reconstruction branch. The state prediction branch outputs the predicted values ​​of current, voltage, power, and backplane temperature for the next sampling period; the dependency reconstruction branch outputs the dependency edge weight reconstruction value and edge type reconstruction value for the next sampling period. The system performs node alignment and time alignment between the predicted state and the actual state, calculates the power deviation, current deviation, temperature deviation, and irradiance response deviation, and calculates the edge weight change, edge type change, and neighborhood change based on the dependency reconstruction results to form node anomaly scores and edge anomaly scores, which are then combined to generate an anomaly score matrix.

[0159] During an operation and maintenance verification, the model identified that the power deviation of three modules in string 4 was continuously higher than the normal neighborhood average by 18.6%, with the backplane temperature deviations of two modules reaching 7.4℃ and 8.1℃ respectively, corresponding to an increased weight for thermal anomaly diffusion. Simultaneously, one current sensor node in string 7 exhibited continuous offset, but the power and temperature of adjacent modules did not show synchronous anomalies, resulting in an increased weight for drift suppression. The fault source decoupling layer classified the former as a module-level anomaly and generated infrared inspection and wiring retest actions, while classifying the latter as sensor drift and generating sensor calibration actions.

[0160] After completing cleaning, infrared inspection, wiring checks, component replacement, and sensor calibration, maintenance personnel write the execution records back to the system. The system matches the feedback results with abnormal nodes, abnormal edges, abnormal paths, and fault source categories, updating node credibility, scenario-based dependency edge weights, dynamic thresholds, state prediction branches, and dependency reconstruction branches. After the feedback update, the model reduces the number of false alarms for the same operating scenario, and improves the stability in distinguishing between hot spot anomalies, string mismatch anomalies, and sensor drift anomalies, forming a closed-loop maintenance model adapted to the sensor network structure of this site.

[0161] Table 1 Comparison of Smart Operation and Maintenance Data for Photovoltaic Modules

[0162] As shown in Table 1, traditional threshold-based maintenance mainly relies on single-point thresholds for power, current, or temperature. The accuracy rate for locating abnormal components is 72.4%, and the accuracy rate for classifying fault sources is 68.1%. It is prone to false alarms when sudden changes in irradiance, local occlusion, and sensor drift occur simultaneously. Ordinary time-series models can utilize historical trends, improving the location accuracy to 80.8%. However, this model does not explicitly learn the dependencies between components, sensors, and devices, resulting in an anomaly path integrity rate of only 70.6%, and insufficient explanatory power for anomaly propagation relationships.

[0163] After introducing a trusted gated observation sequence, a dynamic graph bias dependency graph, a state prediction branch, and a dependency reconstruction branch into the initial model of this invention, the accuracy of anomaly component localization reaches 89.6%, the accuracy of fault source classification reaches 87.9%, and the number of sensor drift misjudgments is reduced from 13 times in the traditional scheme to 4 times. This change shows that node trustworthiness can weaken the impact of low-trust sampling on model judgment, and scenario-based dependency edge weights can separate electrical mismatch, occlusion propagation, thermal anomaly diffusion, and sensor drift from the same batch of observation biases.

[0164] After feedback updates, the model further improved on most metrics, with anomaly component location accuracy reaching 92.8%, fault source classification accuracy reaching 91.3%, hot spot anomaly early identification time reaching 40 minutes, and maintenance action matching rate reaching 91.7%. This is because the execution feedback is incorporated into node credibility, scenario-based dependency edge weights, dynamic thresholds, and a predictive reconstruction dual-branch, enabling the model to correct subsequent judgments based on confirmed cleaning, retesting, replacement, and calibration results. The average number of reviewed work orders decreased from 28 in the traditional solution to 13, indicating that this closed-loop mechanism can reduce invalid work orders and improve the targeting of intelligent maintenance of photovoltaic modules.

[0165] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A sensor network-based intelligent operation and maintenance method for a photovoltaic module, characterized in that, Includes the following steps: S1. Obtain the module status, sensor status, equipment status and operation and maintenance records within the intelligent operation and maintenance cycle of photovoltaic modules, establish module, sensor, equipment nodes and initial dependency edges, and generate a trusted gated observation sequence based on node trustworthiness. S2. Input the intelligent operation and maintenance trusted gating observation sequence of photovoltaic modules into the improved graph bias network model, and obtain the node embedding features through node encoding and trusted gating mapping; S3. Based on the embedded feature learning of the photovoltaic module intelligent operation and maintenance node, the scenario-based dependency edges between the module, sensor and equipment are learned, and a dynamic graph deviation dependency graph is generated. S4. Based on the deviation dependency graph of the intelligent operation and maintenance dynamic diagram of photovoltaic modules, perform state prediction and dependency reconstruction to obtain the predicted state and dependency reconstruction results. S5. Calculate node deviation and edge deviation based on the predicted state, actual state and dependency reconstruction results of photovoltaic module intelligent operation and maintenance, and generate an anomaly scoring matrix. S6. Decouple the fault source based on the photovoltaic module intelligent operation and maintenance anomaly scoring matrix, and generate the anomaly path, fault source category and operation and maintenance action; S7. Based on the feedback from the intelligent operation and maintenance of photovoltaic modules, update the credibility, dependency edge weights, and dual branches to obtain the closed-loop operation and maintenance model.

2. The intelligent operation and maintenance method for photovoltaic modules based on sensor networks according to claim 1, characterized in that, S1 includes the following steps: S11. Read the current, voltage, power, backsheet temperature, irradiance, wind speed, string status, equipment status and inspection records during the intelligent operation and maintenance cycle of photovoltaic modules to obtain the original operation and maintenance data. S12. Organize the original operation and maintenance data according to component number, sensor number, device number and sampling time to obtain the node attribute sequence; S13. Establish initial dependency edges based on electrical connections, spatial adjacency, temperature conduction, and power correlation to obtain the initial dependency edge set; S14. Read the sampling missing rate, mutation frequency, calibration record and differences between adjacent nodes, and calculate the node reliability; S15. Correct the electrical state, temperature state, and environmental state in the node attribute sequence according to the node credibility to obtain the credible gated observation sequence.

3. The intelligent operation and maintenance method for photovoltaic modules based on sensor networks according to claim 1, characterized in that, S2 includes the following steps: S21. Input the components, sensors and device nodes of the trusted gated observation sequence into the node type coding layer to obtain the type coding features; S22. Input the electrical state, temperature state, environmental state, and operation and maintenance records of the trusted gated observation sequence into the state coding layer to obtain the state coding features; S23. Input the sampling time, sunshine interval and operation and maintenance cycle into the time-series coding layer to obtain the time-series coding features; S24. Input the node credibility into the credibility gating mapping layer to obtain the credibility embedding feature; S25. Input the type encoding feature, state encoding feature, time sequence encoding feature and credibility embedding feature into the model embedding end to obtain the node embedding feature. 4.The intelligent operation and maintenance method of a photovoltaic module based on a sensor network according to claim 1, characterized in that, S3 includes the following steps: S31. Calculate the similarity of associations between component, sensor and device nodes based on node embedding features to obtain the candidate association matrix; S32. Based on the candidate association matrix, filter electrical neighbors, temperature neighbors, environmental neighbors, and operation and maintenance neighbors to obtain a set of candidate dependency edges; S33. Input the candidate dependency edge set into the improved graph bias network dependency learning layer, and combine the sensor network topology to calculate the scenario-based dependency edge weights. S34. Perform sparse selection and edge type labeling according to the scenario-based dependency edge weights to obtain the set of retained dependency edges; S35. Combine the retained dependency edge set with the node embedding features to obtain a dynamic graph deviation dependency graph that changes with the operation and maintenance status.

5. The intelligent operation and maintenance method for a photovoltaic module based on a sensor network according to claim 4, characterized in that, S33 includes the following steps: S331. Read the electrical coding, temperature coding, environmental coding, and confidence embedding of the nodes at both ends of the candidate dependent edge to obtain the edge weight input features; S332. Calculate the attention value of the candidate dependent edge based on the edge weight input features to obtain the first edge weight; S333. Correct the weight of the first edge based on the electrical connections, spatial adjacency, and temperature conduction relationships in the initial dependent edges to obtain the weight of the second edge; S334. Based on the mismatch records, occlusion records, hot spot records and sensor drift records of the same string, electrical mismatch weights, occlusion propagation weights, thermal anomaly diffusion weights and drift suppression weights are formed; S335. Input the second edge weight, electrical mismatch weight, occlusion propagation weight, thermal anomaly diffusion weight, and drift suppression weight into the dependency learning layer to obtain the contextualized dependency edge weight. 6.The intelligent operation and maintenance method of a photovoltaic module based on a sensor network according to claim 1, characterized in that, S4 includes the following steps: S41. Input the dynamic graph bias dependency graph into the graph attention aggregation layer to obtain the component node neighborhood aggregation features; S42. Input the neighborhood aggregation features into the state prediction branch, calculate the current, voltage, power and backplane temperature in the next sampling period, and obtain the predicted state; S43. Input the dynamic graph deviation dependency graph into the dependency reconstruction branch, reconstruct the edge weights and edge types of the retained dependency edges, and obtain the dependency reconstruction result. S44. Match the predicted state with the dependency reconstruction results according to the component number, string number and sampling time to obtain the dual-branch output result; S45. Write the dual-branch output results into the deviation calculation terminal of the improved graph deviation network to obtain the deviation calculation input. 7.The intelligent operation and maintenance method of a photovoltaic module based on a sensor network according to claim 1, characterized in that, S5 includes the following steps: S51. Align the predicted state with the actual state by nodes and by time to obtain state-aligned data. S52. Calculate the power deviation, current deviation, temperature deviation and irradiation response deviation based on the state alignment data to obtain the state deviation set; S53. Calculate edge weights, edge types, and neighborhood changes based on the dependency reconstruction results and the dynamic graph deviation dependency graph to obtain the dependency deviation set; S54. Based on the node credibility, the state deviation set and the dependency deviation set are weighted and fused to obtain the node anomaly score and the edge anomaly score. S55. Combine node anomaly scores and edge anomaly scores to generate an improved graph bias network anomaly score matrix. 8.The intelligent operation and maintenance method of a photovoltaic module based on a sensor network according to claim 1, characterized in that, S6 includes the following steps: S61. Compare the anomaly scoring matrix with the dynamic threshold to obtain the anomaly nodes, anomaly edges, and anomaly time periods; S62. Based on the component location, string location, and sensor location of the abnormal node, obtain the set of abnormal objects; S63. Based on the edge type, propagation direction, and edge anomaly score of the abnormal edge, trace the anomaly propagation chain to obtain the anomaly path; S64. Input the set of abnormal objects and abnormal paths into the fault source decoupling layer to obtain the categories of component body abnormality, connection abnormality, environmental disturbance or sensor drift. S65. Match cleaning, infrared inspection, wiring check, component replacement or sensor calibration actions according to the fault source category to obtain maintenance actions. 9.The intelligent operation and maintenance method of a photovoltaic module based on a sensor network according to claim 1, characterized in that, S7 includes the following steps: S71. Obtain execution records after cleaning, inspection, maintenance, replacement and calibration, and obtain operation and maintenance feedback data; S72. Match the operation and maintenance feedback data with abnormal nodes, abnormal edges, abnormal paths and fault source categories to obtain feedback tag data; S73. Update the node credibility, scenario-based dependency edge weight, and dynamic threshold based on the feedback tag data to obtain the first update parameter; S74. Update the state prediction branch and dependency refactoring branch based on the feedback marker data to obtain the second update parameter; S75. Write the first update parameter and the second update parameter into the improved graph deviation network model to obtain the closed-loop operation and maintenance model. 10.The sensor network based photovoltaic component intelligent operation and maintenance method and system of claim 1, wherein, The aforementioned intelligent operation and maintenance method for photovoltaic modules based on sensor networks is characterized by comprising the following modules: The trusted gating module acquires intelligent operation and maintenance data of photovoltaic modules, initial dependency edges and node trustworthiness, and generates a trusted gating observation sequence. The node embedding module takes the trusted gated observation sequence as input to the improved graph bias network model and generates node embedding features. The dependency learning module learns the contextual dependency edges of components, sensors, and devices based on the embedded features of nodes, and generates a dynamic graph bias dependency graph. The prediction and reconstruction module performs state prediction and dependency reconstruction based on the dynamic graph deviation dependency graph, generating predicted state and dependency reconstruction results. The deviation scoring module calculates the deviation based on the predicted state, the actual state, and the dependency reconstruction results, and generates an anomaly scoring matrix. The fault decoupling module locates abnormal nodes and abnormal edges based on the abnormal scoring matrix, and generates fault source categories, abnormal paths and maintenance actions. The closed-loop update module updates the credibility, dependency edge weights, and dual branches based on execution feedback to generate a closed-loop operation and maintenance model.

Images