Photovoltaic fault diagnosis positioning and maintenance suggestion pushing system

Abstract

The invention provides a photovoltaic fault diagnosis positioning and maintenance suggestion pushing system, and relates to the technical field of fault diagnosis and positioning of photovoltaic power stations. Comprises: a task information acquisition module for acquiring image data of a photovoltaic module through an unmanned aerial vehicle carrying a multi-mode sensor, performing spatial registration in combination with a GIS, and generating a fault task information set; the model matching module collects field images and user pose information through the environment modeling module, constructs a three-dimensional environment model and matches the three-dimensional environment model with a power station model to realize AR positioning; the fault analysis module analyzes the fault information through the intelligent expert database module, provides a diagnosis result and a maintenance suggestion, and pushes the diagnosis result and the maintenance suggestion to the user AR equipment; the feedback closed-loop module is used for collecting maintenance data and evaluation, and returning the maintenance data and evaluation to the expert database to optimize the knowledge graph and the diagnosis model to form a closed-loop self-learning system; the operation difficulty of operation and maintenance personnel is reduced, the standardization and safety of maintenance are improved, and the efficiency and accuracy of on-site positioning are improved.

Description

Technical Field

[0001] This invention relates to the technical field of fault diagnosis and location in photovoltaic power plants, and in particular to a photovoltaic fault diagnosis, location and maintenance suggestion push system. Background Technology

[0002] With the rapid popularization of photovoltaic power generation technology and the continuous expansion of power plant scale, efficient and precise operation and maintenance of photovoltaic modules has become a key link in ensuring the power generation efficiency and economic benefits of power plants. Photovoltaic modules are exposed to the outdoor environment for a long time and are susceptible to various faults such as hot spots, cracks, aging, and shading, posing a serious challenge to traditional inspection and diagnosis methods.

[0003] Existing technologies suffer from the following shortcomings: 1. While current mainstream methods utilize drone inspections, they typically only provide single-modal image data such as visible light and infrared, lacking effective multi-source information fusion methods. The collected data is independent of the power plant's Geographic Information System (GIS) and expert knowledge system, resulting in fragmented fault information that hinders comprehensive analysis and in-depth diagnosis. 2. Traditional methods rely on two-dimensional drawings or reports to present results, making it difficult for maintenance personnel to quickly and accurately locate faulty components in complex field environments. Even with approximate coordinates, the lack of intuitive on-site guidance leads to significant search time, and detailed fault information and maintenance guidance tied to that spatial location cannot be obtained in real time. 3. Fault diagnosis and maintenance plan development heavily rely on the personal experience of maintenance personnel, lacking standardized and intelligent decision support. Expert knowledge has not been effectively accumulated and shared, resulting in inconsistent maintenance quality, and the experience of senior engineers is difficult to quickly replicate and pass on. 4. Although augmented reality (AR) technology offers new visualization possibilities for industrial maintenance, its application in the photovoltaic field is still in its early stages. Existing solutions fail to achieve deep integration of UAV multimodal data, GIS spatial information, AR visualization, and expert knowledge base, thus failing to form a complete closed loop in the diagnostic process. Summary of the Invention

[0004] The main objective of this invention is to provide a photovoltaic fault diagnosis, location, and maintenance suggestion system, which solves the problems of low diagnostic efficiency, isolated information, insufficient location accuracy, poor guidance, excessive reliance on human experience, and insufficient applicability of the technology in the prior art.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a photovoltaic fault diagnosis, location, and maintenance suggestion push system, comprising: The task information acquisition module is used to collect image data of photovoltaic modules by using a drone equipped with a multimodal sensor, and to perform spatial registration in combination with a geographic information system to generate a fault task information set that includes the fault location, type and severity. The model matching module is used to collect on-site images and user pose information through the environment modeling module, construct a 3D environment model and match it with the power plant model, and overlay fault task information onto real components in the form of virtual markers to achieve augmented reality positioning. The fault analysis module is used to analyze fault information through the intelligent expert database module, provide diagnostic results and maintenance suggestions based on historical cases and knowledge graphs, and push them to the user's augmented reality device in a visual form. The feedback loop module is used to collect maintenance data and evaluations through the user feedback module, and then feed them back to the expert database to optimize the knowledge graph and diagnostic model.

[0006] In the preferred embodiment, the task information collection module includes: The data acquisition unit is used to acquire multimodal image data of photovoltaic modules by using a drone equipped with a visible light sensor, an infrared thermal imaging sensor, and a hyperspectral sensor, and to obtain the geographic coordinate information corresponding to the multimodal image data; The spatial registration unit is used to perform spatial registration of the multimodal image data based on the geographic coordinate information and in conjunction with the geographic information system to obtain the registered image data. The feature extraction and fusion unit is used to extract infrared thermal distribution features and visible light crack features from the registered image data using a deep learning network, and to perform weighted fusion of the infrared thermal distribution features and visible light crack features through an attention mechanism to obtain fused features; The fault identification and information generation unit is used to identify the fault type, fault location and severity of the photovoltaic module based on the fusion features, and generate a fault task information set containing the fault location, the fault type, the severity and the recommended priority. The data upload unit is used to upload the fault task information set to the cloud server for storage and management.

[0007] In the preferred embodiment, the feature extraction and fusion unit includes: The infrared portion of the registered image data is processed by a first convolutional neural network to extract infrared thermal distribution features representing areas of temperature anomalies. The visible light portion of the registered image data is processed by a second convolutional neural network to extract visible crack features representing the morphology of surface defects. For the infrared thermal distribution features and the visible light crack features, attention weights are calculated, and the attention weights are learnable parameters optimized through end-to-end training. The fused feature is obtained by weighting and summing the infrared thermal distribution feature and the visible light crack feature according to the attention weight; If the recognition accuracy of the fused features is lower than a preset threshold, the attention weights are adjusted and the features are refused.

[0008] In the preferred embodiment, the model matching module includes: The on-site data acquisition unit is used to acquire on-site environmental images and user pose information through augmented reality devices, wherein the user pose information includes position coordinates and posture angles. The environment modeling unit is used to construct a dense point cloud map by using vision-inertial odometry-based simultaneous localization and mapping technology, combining the on-site environment image and the user pose information, through feature point detection and triangulation calculation, to obtain a three-dimensional environment model; The model registration unit is used to perform iterative nearest-point registration between the three-dimensional environment model and the pre-established three-dimensional power station model to obtain the registration result; The coordinate transformation and pose solving unit is used to convert the fault location in the fault task information set into augmented reality display coordinates according to the registration result, and to solve the camera pose by minimizing the reprojection error between image feature points and 3D model points through a pose optimization algorithm. The virtual overlay unit is used to overlay virtual markers onto the real photovoltaic module according to the camera pose, so as to realize the visual location of the fault point.

[0009] In the preferred embodiment, the fault analysis module includes: An information input unit is used to obtain fault information from the fault task information set and input it into the intelligent expert database module built based on a large language model, wherein the intelligent expert database module integrates historical fault cases and photovoltaic operation and maintenance knowledge graph; The analysis and reasoning unit is used to perform semantic understanding and causal reasoning on the fault information through the intelligent expert database module, and to calculate the association weights between fault entities based on the knowledge graph, wherein the association weights are obtained by weighted summation of feature similarity function, time co-occurrence frequency function and maintenance result correlation function. The result generation unit is used to retrieve matching cases from the historical fault cases according to the association weight, and generate fault cause determination, maintenance steps and operation parameter suggestions to obtain diagnostic results and maintenance suggestions. A visualization push unit is used to convert the diagnostic results and repair suggestions into text, 3D models or animations, and push them to the user's augmented reality device for display; The confidence assessment unit is used to calculate the comprehensive confidence level by multi-model fusion if the confidence level of the diagnostic result is lower than a preset threshold, wherein the comprehensive confidence level is a weighted average of the predicted probabilities of multiple diagnostic models.

[0010] In the preferred embodiment, the feedback closed-loop module includes: The feedback data acquisition unit is used to collect maintenance logs, tool usage information, power recovery data and satisfaction evaluations through the user feedback module after fault repair to obtain user feedback data; The knowledge update unit is used to feed the user feedback data back to the intelligent expert database module and update the association weights in the knowledge graph according to the user feedback data, wherein the update is calculated by the learning rate and the weight adjustment amount. The model optimization unit is used to fine-tune the parameters of the model based on the user feedback data using an incremental learning approach. The lifespan prediction unit is used to predict the remaining lifespan of the photovoltaic modules based on the user feedback data and historical operating data using a regression model. The expression is as follows: ; in, It is a regression function. It is an error term; The closed-loop application unit is used to apply the optimized knowledge graph and the diagnostic model to subsequent fault analysis, forming a self-learning closed loop.

[0011] In the preferred embodiment, the coordinate transformation and pose solving unit includes: The three-dimensional coordinates of the fault location are obtained from the registration results, and the augmented reality display coordinates are calculated by combining the camera intrinsic parameter matrix and rotation and translation parameters, wherein the augmented reality display coordinates are the projection of the three-dimensional coordinates in the camera coordinate system; For image feature points and corresponding 3D model points, an objective function is constructed, wherein the objective function is the sum of squares of the reprojection errors under the rotation matrix plus a regularization parameter; The optimal camera pose is solved by minimizing the objective function, wherein the camera pose includes a rotation matrix and a translation vector; The position of the virtual marker is adjusted according to the optimal camera pose to ensure that the superposition accuracy reaches the centimeter level; if the deviation between the obtained camera pose and the actual pose exceeds a preset threshold, the objective function is iteratively optimized.

[0012] In the preferred embodiment, the objective function is constructed as follows: ; in, For rotation matrix, For the first Image feature points, For the corresponding 3D model points, For camera projection function, This is the regularization parameter.

[0013] In a preferred embodiment, the analysis and reasoning unit includes: The fault information is mapped to fault entities in a knowledge graph, and the relationships between entities are identified, wherein the knowledge graph is constructed using a graph neural network; Calculate the feature similarity function between the faulty entities, wherein the feature similarity function is obtained based on the cosine similarity of entity attributes; Calculate the time co-occurrence frequency function among the faulty entities, wherein the time co-occurrence frequency function is a normalized value of the number of times they co-occur in historical data; Calculate the correlation function of repair results among the faulty entities, wherein the correlation function of repair results is a correlation coefficient based on the success rate; The association weight is obtained by multiplying the feature similarity function, the time co-occurrence frequency function, and the maintenance result correlation function by their corresponding weight coefficients and then summing them. The formula is as follows: ; in, Represents the faulty entity and The correlation weight between them For feature similarity function, Let co-occurrence frequency function be used. For the correlation function of maintenance results, , , These are the corresponding weighting coefficients; Based on this association weight, causal reasoning is performed to generate diagnostic results.

[0014] In the preferred solution, the fault analysis module, specifically for fire or thermal runaway emergency response scenarios, also includes an emergency response scenario module, which includes: The multi-source heterogeneous data preprocessing unit is used to perform smoke scattering compensation and adaptive filtering on infrared thermal image sequences, and to perform robust estimation and outlier removal on electrical data; it assigns a confidence coefficient based on data missing rate, historical consistency and sensor health; it performs dynamic time alignment with a unified timestamp standard, and maps the positioning coordinates of multiple devices to the same three-dimensional coordinate system, and forms fire point feature vector and device status vector after correlation filtering and dimensionality reduction; The fire prediction unit, based on a spatiotemporal graph convolutional network and incorporating a physical constraint loss function, couples terrain-thermal-electrical multiphysics data for modeling, outputting a fire spread probability map and an electrical short-circuit risk probability map. The expression is: ; in, This provides real-time input data for multiphysics; its expression is: ; in, For terrain elevation and slope information, For thermal radiation distribution information and This provides electrical parameter information for components / branch circuits. The collaborative scheduling unit defines drones and firefighting robots as the game's main players, and fire points as the game's objects. It integrates fire point temperature / heat flux, threat distance, combustible material sensitivity, and predicted risk to construct a comprehensive fire point weight. An improved Nash equilibrium solution is then obtained through the player payoff function (fire point risk payoff minus equipment resource costs), which serves as the optimal collaborative scheduling scheme. Specifically: The set of game participants, including n drones and m firefighting robots, is expressed as: ; The fire point set expression is: ; Each subject The strategy is to select the fire points or fire point sequences to be dealt with in the current cycle. ; The fire point weight expression is: ; in For heat flux, The threat distance or cost of reaching critical equipment, Sensitivity to flammable materials ; The payoff function expression is: ; in, The resource cost for the parties involved in the game to deal with the target fire point; The cost function expression is: ; in, Remaining battery power This refers to the remaining extinguishing agent. , For reachability / processing time, parameters , , These are the time cost weight, the power consumption weight, and the extinguishing agent consumption weight, respectively.

[0015] This invention provides a photovoltaic fault diagnosis, location, and maintenance suggestion push system, comprising: a task information acquisition module, used to acquire image data of photovoltaic modules through a drone equipped with a multimodal sensor, and perform spatial registration in conjunction with a Geographic Information System (GIS) to generate a fault task information set including fault location, type, and severity; a model matching module, used to acquire on-site images and user pose information through an environment modeling module, construct a three-dimensional environment model and match it with the power station model, and overlay the fault task information onto the real components in a virtual marker manner to achieve augmented reality (AR) positioning; a fault analysis module, used to analyze fault information through an intelligent expert database module, provide diagnostic results and maintenance suggestions based on historical cases and knowledge graphs, and push them to the user's AR device in a visual form; and a feedback closed-loop module, used to collect maintenance data and evaluations through a user feedback module, and feed them back to the expert database to optimize the knowledge graph and diagnostic model, forming a closed-loop self-learning system; thus forming an efficient, self-learning intelligent operation and maintenance system, reducing the operational difficulty for operation and maintenance personnel, improving the standardization and safety of maintenance, and enhancing the efficiency and accuracy of on-site positioning. Attached Figure Description

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a system structure diagram of the present invention; Figure 2 This is a schematic diagram of the overall design of the system modules of the present invention; Figure 3 This is a closed-loop flowchart for emergency response and maintenance of photovoltaic power plants according to the present invention. Detailed Implementation

[0017] Example 1 like Figure 1-3 As shown, a photovoltaic fault diagnosis, location, and maintenance suggestion push system includes: The task information acquisition module is used to collect image data of photovoltaic modules by using a drone equipped with multimodal sensors, and to perform spatial registration in conjunction with a geographic information system to generate a fault task information set that includes the fault location, type and severity.

[0018] The model matching module is used to collect on-site images and user pose information through the environment modeling module, construct a 3D environment model and match it with the power plant model, and overlay fault task information onto real components in the form of virtual markers to achieve augmented reality positioning.

[0019] The fault analysis module analyzes fault information through the intelligent expert database module, provides diagnostic results and maintenance suggestions based on historical cases and knowledge graphs, and pushes them to the user's augmented reality device in a visual format.

[0020] The feedback loop module is used to collect maintenance data and evaluations through the user feedback module, and then feed them back to the expert database to optimize the knowledge graph and diagnostic model, forming a closed-loop self-learning system.

[0021] In this embodiment, image data is collected by drones and spatially registered using GIS to obtain a fault task information set. A three-dimensional environment model is then constructed and matched with the power plant model to achieve AR positioning. The fault information is analyzed by the intelligent expert database module, and diagnostic results and maintenance suggestions are provided based on historical cases and knowledge graphs. Finally, maintenance data and evaluations are collected by the user feedback module and fed back to the expert database to optimize the knowledge graph and diagnostic model, forming a closed-loop self-learning system. This results in an efficient and self-learning intelligent operation and maintenance system, reducing the operational difficulty for operation and maintenance personnel, improving the standardization and safety of maintenance, and enhancing the efficiency and accuracy of on-site positioning.

[0022] In the preferred embodiment, the task information collection module includes: The data acquisition unit is used to acquire multimodal image data of photovoltaic modules by using a drone equipped with a visible light sensor, an infrared thermal imaging sensor, and a hyperspectral sensor, and to obtain the geographic coordinate information corresponding to the multimodal image data; The spatial registration unit is used to perform spatial registration of multimodal image data based on geographic coordinate information and geographic information system to obtain registered image data.

[0023] The feature extraction and fusion unit is used to extract infrared thermal distribution features and visible crack features from the registered image data using a deep learning network, and to perform weighted fusion of the infrared thermal distribution features and visible crack features through an attention mechanism to obtain fused features.

[0024] The fault identification and information generation unit is used to identify the fault type, fault location and severity of photovoltaic modules based on the fusion characteristics, and generate a fault task information set containing fault location, fault type, severity and recommended priority.

[0025] The data upload unit is used to upload the fault task information set to the cloud server for storage and management.

[0026] In the preferred embodiment, the feature extraction and fusion unit includes: The infrared portion of the registered image data is processed by a first convolutional neural network to extract infrared thermal distribution features representing areas of temperature anomalies. ; in The input is the infrared thermal distribution features; another convolutional network is used to extract visual features from the visible light image.

[0027] The visible light portion of the registered image data is processed by a second convolutional neural network to extract visible crack features representing the morphology of surface defects.

[0028] Attention weights are calculated for infrared thermal distribution characteristics and visible crack characteristics. These attention weights are learnable parameters optimized through end-to-end training.

[0029] The fused features are obtained by weighting and summing the infrared thermal distribution features and visible crack features according to the attention weights: ; in, The crack is characterized by visible light. , These are the corresponding weighting coefficients.

[0030] If the recognition accuracy of the fused features is lower than the preset threshold, the attention weights are adjusted and the features are refused.

[0031] Based on the fused features, the system identifies the specific location and extent of the fault area, and the proportion of the abnormal area to the total component area. For example, if it is initially judged to be a typical hot spot effect, the system generates structured fault information that includes fault GPS coordinates, fault type, severity rating and recommended processing priority, and pushes this information to the AR device of the on-site maintenance personnel.

[0032] Once maintenance personnel arrive on-site with the AR equipment, the system's environmental modeling module begins operation. This module employs visual-inertial odometry (VIO) technology, combined with the AR device's built-in IMU sensor and RGB-D camera, to construct a real-time 3D point cloud map of the site. The system uses the ORB (Oriented FAST and Rotated BRIEF) algorithm to extract environmental feature points, which leverages the speed of the FAST corner detector. Compactness of the BRIEF descriptor The feature extraction formula is: By matching feature points and performing triangulation calculations between consecutive frames, the system constructs a dense 3D map of the site environment with centimeter-level accuracy.

[0033] In this embodiment, the task information acquisition module uses a drone equipped with a multispectral sensor to conduct efficient inspections of the photovoltaic power station, acquire multimodal image data, and perform intelligent analysis through image processing and deep learning algorithms to automatically identify and accurately determine the fault type, location, and severity of the photovoltaic modules. Combined with the high-precision GIS geographic information of the power station, the identified faulty components are accurately marked in the three-dimensional space of the power station, and a fault task information set containing fault ID, type, coordinates, severity, and recommended processing priority is generated and uploaded to the cloud server in real time.

[0034] This embodiment integrates multimodal features such as visible light and infrared by using deep learning and attention mechanisms to achieve comprehensive perception and high-precision identification of faults such as hot spots and cracks, effectively reducing the rate of missed detections and false detections, and improving the comprehensiveness and accuracy of fault identification.

[0035] The fault task information set in the cloud server is pushed to the AR device of the on-site operation and maintenance personnel in real time.

[0036] When maintenance personnel arrive at the suspected fault area, the AR device uses built-in high-precision sensors (such as cameras and IMUs) to perform real-time environmental perception and pose tracking, and build a three-dimensional environmental model of the site.

[0037] In the preferred scheme, the model matching module includes: The on-site data acquisition unit is used to acquire on-site environmental images and user pose information through AR devices, wherein the user pose information includes position coordinates and posture angles. The environment modeling unit is used to construct a dense point cloud map by using vision-inertial odometry-based simultaneous localization and mapping (V-SLAM) technology, combined with on-site environmental images and user pose information, through feature point detection and triangulation calculation, to obtain a three-dimensional environment model; The model registration unit is used to perform iterative closest point (ICP) registration between the 3D environment model and the pre-established 3D power plant model to obtain the registration result; The coordinate transformation and pose solving unit is used to convert the fault location in the fault task information set into AR display coordinates based on the registration results, and to solve the camera pose by minimizing the reprojection error between image feature points and 3D model points through a pose optimization algorithm. The virtual overlay unit is used to overlay virtual markers onto real photovoltaic modules based on the camera pose, thereby enabling the visual location of fault points.

[0038] In this embodiment, after receiving the 3D environment map, the AR scene generation module registers it with a pre-established digital twin model of the photovoltaic power station. The registration process employs the Iterative Closest Point (ICP) algorithm, which solves for the optimal spatial transformation relationship by minimizing the distance between corresponding points. After registration, the system converts the 3D coordinates of the faulty photovoltaic module into AR display coordinates. The conversion formula is:

[0039] in For the camera intrinsic parameter matrix, These are the camera's rotation and translation parameters. This represents the position of the faulty component in the world coordinate system.

[0040] On the AR device display screen of the maintenance personnel, the faulty photovoltaic module is highlighted with a red virtual border, while a detailed fault information panel is displayed. This panel includes the fault type, temperature data, historical maintenance records, and recommended maintenance steps. Through spatial registration using AR technology, the virtual marker precisely overlaps with the actual faulty module, with a positioning error controlled within 20 centimeters.

[0041] In the preferred scheme, the coordinate transformation and pose solving unit includes: The three-dimensional coordinates of the fault location are obtained from the registration results, and the AR display coordinates are calculated by combining the camera intrinsic parameter matrix and rotation and translation parameters. The AR display coordinates are the projection of the three-dimensional coordinates onto the camera coordinate system. For image feature points and corresponding 3D model points, an objective function is constructed, where the objective function is the sum of squares of the reprojection errors under the rotation matrix plus a regularization parameter; The optimal camera pose is solved by minimizing the objective function, where the camera pose includes a rotation matrix and a translation vector. The position of the virtual marker is adjusted according to the optimal camera pose to ensure that the superposition accuracy reaches the centimeter level; if the deviation between the obtained camera pose and the actual pose exceeds a preset threshold, the objective function is iteratively optimized.

[0042] In the preferred scheme, the objective function is constructed as follows: ; in For rotation matrix, For the first Image feature points, For the corresponding 3D model points, For camera projection function, This is the regularization parameter.

[0043] This embodiment uses V-SLAM, ICP registration, and pose optimization algorithms to overlay virtual fault information onto the real scene with high precision, enabling maintenance personnel to quickly locate faulty components, greatly shortening on-site search time and significantly reducing positioning errors.

[0044] This embodiment spatially matches the coordinates of faulty components contained in the fault task information set with the on-site 3D environment model, and uses AR technology to accurately overlay virtual fault information, such as the boundary box of the faulty component, component number, real-time temperature data, fault type identifier, historical fault trend chart, etc., onto the real faulty photovoltaic panel. The intuitive visualization method enables maintenance personnel to quickly and accurately locate the target fault point in the complex power plant environment, improving the efficiency and accuracy of on-site positioning.

[0045] Based on this, this embodiment introduces an intelligent expert database built on a large language model as the core diagnostic and decision support engine. The intelligent expert database deeply integrates massive historical fault case data, professional photovoltaic operation and maintenance knowledge graphs, and the latest maintenance operation specifications, and has powerful fault analysis and diagnostic reasoning capabilities.

[0046] In the preferred solution, the fault analysis module includes: The information input unit is used to obtain fault information from the fault task information set and input it into the intelligent expert database module built based on a large language model. The intelligent expert database module integrates historical fault cases and photovoltaic operation and maintenance knowledge graph. The analysis and reasoning unit is used to perform semantic understanding and causal reasoning on fault information through the intelligent expert database module, and to calculate the association weights between fault entities based on the knowledge graph. The association weights are obtained by weighted summation of feature similarity function, time co-occurrence frequency function and maintenance result correlation function. The result generation unit is used to retrieve matching cases from historical fault cases based on the association weight, and generate fault cause determination, maintenance steps and operating parameter suggestions to obtain diagnostic results and maintenance suggestions. The visualization push unit is used to push diagnostic results and maintenance suggestions not only in text form, but also in AR form through an AR interface in the form of 3D models, animated steps, virtual arrows and other visual forms to the AR devices of maintenance personnel. This greatly reduces the operation difficulty for maintenance personnel and improves the standardization and safety of maintenance.

[0047] The confidence assessment unit is used to calculate the comprehensive confidence score by fusing multiple models when the confidence score of the diagnostic result is lower than a preset threshold. The comprehensive confidence score is a weighted average of the predicted probabilities of multiple diagnostic models.

[0048] In this embodiment, based on the scenario described above, maintenance personnel accurately locate module number 15 in the 3rd row of the photovoltaic power station using AR equipment. The system has identified two fault types in this module: "hot spot effect" and "micro-cracks," and includes them as part of the fault task information set. The S72 intelligent expert database module and the S74 cloud processing and diagnosis module work together at this stage to provide maintenance personnel with in-depth diagnosis and actionable maintenance suggestions.

[0049] In the preferred embodiment, the analysis and reasoning unit includes: Fault information is mapped to fault entities in a knowledge graph, and the relationships between entities are identified. The knowledge graph is constructed using a graph neural network. Calculate the feature similarity function between faulty entities, where the feature similarity function is obtained based on the cosine similarity of entity attributes; Calculate the time co-occurrence frequency function among faulty entities, where the time co-occurrence frequency function is the normalized value of the number of times they co-occur in historical data; Calculate the correlation function of repair results between faulty entities, where the correlation function of repair results is a correlation coefficient based on the success rate; The association weight is obtained by multiplying the feature similarity function, the time co-occurrence frequency function, and the maintenance result correlation function by their respective weight coefficients and then summing them. The formula is as follows: ; in Represents the faulty entity and The correlation weight between them For feature similarity function, Let co-occurrence frequency function be used. For the correlation function of maintenance results, , , These are the corresponding weighting coefficients; Based on this, causal reasoning is used to generate diagnostic results.

[0050] In this embodiment, the intelligent expert database receives a set of fault task information from the cloud-based processing and diagnosis module. This information includes fault type (hot spot, microcrack), location (GIS coordinates and component ID), severity (e.g., temperature exceeding limit by 40°C, crack length 5cm), component model, production batch, etc. The system first uses semantic similarity matching to retrieve the most relevant historical cases and knowledge points from the vast photovoltaic fault knowledge graph built within the intelligent expert database. This knowledge graph is organized in the form of a graph neural network (GNN), where nodes represent entities (e.g., "hot spot effect," "microcrack," "welding defect," "partial shading," "component aging," "environmental factors," "production batch," "manufacturer A") and concepts (e.g., "fault type," "cause," "solution"), and edges represent relationships between entities (e.g., "cause," "is the cause," "solution includes"). Then, using the reasoning capabilities of the graph neural network, multi-hop reasoning is performed to trace the root cause of the fault.

[0051] For example, for the "hot spot effect," the system might infer common causes including "welding defects," "local shading," and "cell mismatch"; for "microcracks," it might be associated with "mechanical stress," "thermal cycling fatigue," or "manufacturing process defects." The system assesses the likelihood of different causes by calculating the path probability from "symptom" nodes to "cause" nodes in the knowledge graph, for example, using probabilistic graphical models for causal reasoning, where potential causes... and symptoms , The relationship between them can be expressed by the following formula: .

[0052] In this way, the system can quantify the contribution of different causes to the current failure and ultimately determine that "welding defects" are the main cause of hot spots, while "manufacturing process defects" combined with long-term thermal stress lead to the propagation of microcracks.

[0053] In this embodiment, the comprehensive confidence score is calculated by combining the prediction results of multiple diagnostic models. A high confidence score indicates that the diagnostic result is highly reliable. For example, a weighted average method is used, where... It is the number of diagnostic models involved in the fusion. It is the first The weights of each model, This is the model's probability of the current diagnosis: .

[0054] Based on this in-depth diagnostic result, the intelligent expert database generates a personalized and detailed maintenance operation guide for maintenance personnel, and pushes it to the maintenance personnel's AR device in a highly visual way through AR display and S71 interaction module. When maintenance personnel approach the faulty component, the AR interface will highlight virtual safety prompts, such as "Power failure warning: Please disconnect the DC circuit breaker of this component string first and use a multimeter to confirm that there is no voltage" and "Wear protective gear: Please wear insulated gloves, safety helmet and cut-resistant protective gear." The AR interface will dynamically generate a virtual list of tools required for maintenance (such as "insulated wrench", "component disassembly tool", "IV curve tester"), and display a virtual tool outline at a specific location next to the component to indicate the best placement point for easy access. For complex component replacement or wiring operations, the AR system will overlay a transparent 3D animation model above or next to the faulty component. This model will step by step demonstrate the entire process of removing screws, disconnecting wiring, removing old components, installing new components, and rewiring. Each step is accompanied by text descriptions and key parameter prompts (such as "tightening torque: 20 N·m"), and maintenance personnel can pause, fast forward, and rewind the animation. AR devices achieve seamless integration by precisely calculating the relative poses of virtual objects and the real world. For example, the display transformation matrix of a key tool or component in a virtual repair procedure animation in the AR view. It can be calculated using the following formula: ; in, It is the current camera pose matrix. It is the transformation matrix that converts the world coordinate system to the photovoltaic panel coordinate system. It is a transformation matrix that transforms the coordinate system of a photovoltaic panel to the coordinate system of a specific component. It is a transformation matrix that converts the component coordinate system to the virtual operation object coordinate system.

[0055] Furthermore, during the repair process, if testing equipment is connected, its readings can be displayed in real-time via an AR interface, such as "Voltage: 45.2V (normal range 45V-48V)" and "Current: 0.5A (normal standby state)," providing immediate feedback to repair personnel. Through this deep integration of AR and an intelligent expert database, repair personnel can not only accurately locate faults but also obtain professional-level diagnostic analysis and hands-on, visual repair guidance, improving repair efficiency and the first-time repair success rate.

[0056] This embodiment employs a multi-model fusion confidence assessment method, comprehensively considering the prediction results of different models to make the final diagnostic conclusion more scientific and reliable, assisting maintenance personnel in making more accurate judgments. The intelligent expert database used has continuous learning capabilities, continuously optimizing the diagnostic algorithm and knowledge graph by analyzing new fault data, maintenance records, and user feedback, thereby improving the accuracy and timeliness of diagnostic suggestions and realizing a closed-loop link from fault discovery, precise location, intelligent diagnosis to maintenance guidance and knowledge updates.

[0057] In the preferred embodiment, the feedback closed-loop module includes: After the fault is repaired, maintenance logs, tool usage information, power recovery data and satisfaction evaluations are collected through the user feedback module to obtain user feedback data. User feedback data is fed back to the intelligent expert database module, and the association weights in the knowledge graph are updated based on the user feedback data. The update is calculated using the learning rate and weight adjustment amount. Incremental learning is used to optimize the diagnostic model, and the model parameters are fine-tuned based on user feedback data to improve the ability to identify new fault modes. The remaining service life of photovoltaic modules is predicted by regression model based on user feedback data and historical operating data, where the remaining service life is the regression function value minus the error term. The expression is: ; in, It is a regression function. It is an error term; The optimized knowledge graph and diagnostic model are then applied to subsequent fault analysis, forming a self-learning closed loop.

[0058] In this embodiment, a regression model is used to predict the remaining service life of components, realizing a shift from post-maintenance to pre-warning. This helps to plan maintenance schedules in advance, avoid sudden downtime, and improve the reliability and economic benefits of power plant operation.

[0059] like Figure 3The diagram shown is a closed-loop flowchart for emergency response and maintenance of photovoltaic power plant faults (focusing on fire and thermal runaway). It clearly presents the entire process logic from multi-source data acquisition to post-disaster maintenance and optimization. The core descriptions are as follows: The process begins with multi-dimensional data collection, integrating data from drones (infrared, visible light, LiDAR, wind farm, electrical parameters), GIS systems (power station coordinates, topology, historical cases), and on-site robots (temperature rise detection). After data preprocessing, a multi-physics fire prediction model is used to predict fault spread and short-circuit risks. Based on the prediction results, an improved Nash equilibrium collaborative scheduling mechanism is adopted to coordinate equipment and prioritize the handling of high-risk fire points. Subsequently, a closed loop is formed through command issuance and equipment collaborative handling. Finally, fault repair and reignition prevention are completed through post-disaster risk classification and maintenance closed loop, while improving prediction accuracy, handling efficiency, and other multi-dimensional capabilities.

[0060] Furthermore, after data preprocessing, a fire prediction model based on multiphysics coupling is introduced to overcome the insufficient accuracy of predictions relying solely on wind direction, wind speed, and terrain in photovoltaic fields. A reconnaissance drone simultaneously collects terrain elevation and slope information. Information on thermal radiation distribution and component / branch electrical parameter information This generates a real-time input of topographic, thermal, and electrical multiphysics fields; the expression is: .

[0061] The operation and maintenance platform, based on a spatiotemporal graph convolutional network and incorporating a physical constraint loss function, couples and models the relationships between heat conduction feedback, terrain barrier effects, and electrical chain fault propagation, outputting a fire spread probability map and an electrical short-circuit risk probability map, expressed as follows: .

[0062] The prediction results are not only used to label regions with high propagation probability and high short-circuit probability, but also directly used as a dynamic gain term for fire priority in scheduling: when At that time, the corresponding fire point was forcibly upgraded to an emergency priority target; when At that time, the corresponding component branch is given priority for power outage isolation and its handling weight is increased. The platform updates real-time data on a rolling basis every minute and makes repeated predictions, so that the fire path and risk area are corrected in real time according to the fire situation, thereby providing continuous and reliable prior information for dispatch.

[0063] In this embodiment, during emergency response to fire or thermal runaway, the operation and maintenance platform needs to integrate heterogeneous data from multiple sources, including UAV infrared thermal imaging, visible light, LiDAR, component electrical parameters, and gas / insulation detection data collected by firefighting robots. However, such data generally suffers from high noise, difficulty in spatiotemporal alignment, feature redundancy, and high dynamism. For example, infrared thermal imaging is susceptible to interference from smoke scattering, leading to misjudgments of hot spots; electrical parameters are prone to short-term abnormal fluctuations due to electromagnetic waves; and the transmission frequencies of different devices are inconsistent with the positioning coordinate system. To ensure the reliability of subsequent fire prediction and dispatch decisions, this invention preferably performs multi-source data preprocessing before dispatch and prediction: on the one hand, smoke scattering compensation and adaptive filtering are performed on the infrared thermal image sequence, and robust estimation and outlier removal are performed on electrical data such as current, voltage, and insulation resistance to suppress environmental and equipment noise; on the other hand, based on the data missing rate, historical consistency, and sensor health of each modality, a credibility coefficient is assigned to each type of data and used for subsequent fusion. Meanwhile, the platform dynamically aligns data from different sources using a unified timestamp standard, and maps UAV RTK positioning, robot V-SLAM positioning, and GIS power plant coordinates to the same three-dimensional coordinate system, ensuring computable spatial consistency between fire points, components, and cable topology. The compact fire point feature vectors, refined through correlation filtering and dimensionality reduction, along with the standardized equipment state vectors, will serve as the input basis for fire prediction and dynamic dispatching.

[0064] Based on the aforementioned predictions and perceptions, a dynamic game-theoretic scheduling mechanism involving multi-device collaboration is preferred to overcome the limitations of fixed-path scheduling, which easily leads to resource conflicts and allocation imbalances when multiple fire points occur concurrently. Drones and firefighting robots are collectively considered as a set of game subjects, and the set of fire points is considered as the game objects. Each subject's strategy is to select the fire points or fire point sequences that need to be dealt with in the current cycle.

[0065] Let n drones and m firefighting robots be collectively defined as the set of game players, expressed as: ; The expression is: ; Each subject The strategy is to select the fire points or fire point sequences to be dealt with in the current cycle. .

[0066] The platform integrates fire point temperature / heat flux, threat distance or cost to reach critical arrays and cables, combustible material sensitivity, and the spread probability and short-circuit risk output from multiphysics predictions to construct a comprehensive fire point weight, reflecting the fire risk and urgency of response. The fire point weight expression is as follows: ; in For heat flux, The threat distance or cost of reaching critical equipment, Sensitivity to flammable materials .

[0067] Meanwhile, the main revenue function is defined as the risk return of fire point minus the cost of equipment resources. The cost term comprehensively considers the expected reach / disposal time, remaining power and equipment status constraints such as the remaining fire extinguishing agent, to ensure that dispatching not only pursues short-term fire extinguishing benefits, but also takes into account equipment endurance and mission continuity.

[0068] The profit expression is: ; The cost function expression is: ; parameter , , These are time cost weights, electricity consumption weights, and extinguishing agent consumption weights, used to adjust the influence of different cost items in the overall cost; preferably, , , All are real numbers greater than zero, and satisfy the following conditions: .

[0069] In this embodiment, maintenance personnel successfully replaced the faulty photovoltaic module following the repair suggestions provided by the AR device. After the repair, the S75 user feedback optimization module immediately activated, feeding all relevant data back to the cloud server and intelligent expert database for continuous system learning and optimization. Maintenance personnel confirmed the successful repair using the AR device or a companion smartphone app, inputting the actual repair time, the serial number of the replaced module, the tools used, any unexpected situations encountered during the repair, or any new problems discovered. Simultaneously, the maintenance personnel provided satisfaction ratings and written feedback on the diagnostic results, cause analysis, and repair suggestions provided by the intelligent expert database. For a period after the repair, the system automatically monitored the module's power output and temperature data to verify whether it had returned to normal operation. All this structured and unstructured data, such as repair logs, satisfaction ratings, and performance recovery data, was uploaded to the cloud in real time and fed back to the intelligent expert database.

[0070] The intelligent expert database utilizes the returned data to iterate and optimize itself through multiple mechanisms. Firstly, regarding the dynamic updating of the knowledge graph, if multiple maintenance feedback indicates that the correlation between a specific fault type and a certain cause is far greater than expected, the weight of the corresponding edge in the knowledge graph will be enhanced. The update formula can be expressed as: ; in, It is a physical entity To the entity edge weights, It's the learning rate. $ is a weight adjustment amount calculated based on feedback, and its value is determined comprehensively based on diagnostic accuracy, repair success rate, and user satisfaction. Simultaneously, if maintenance personnel describe novel fault phenomena or causes not recorded in the expert database in their feedback, the system's large language model will analyze this unstructured text, automatically extract new entities and relationships, and integrate them into the knowledge graph. Secondly, regarding the retraining and calibration of the diagnostic model, the deep learning model used for fault identification will periodically use these manually verified new fault image data, including original misjudgments and newly discovered faults, for incremental training or fine-tuning to correct biases in practical applications, improving the ability to identify rare or new fault patterns, and introducing a new loss function. It can be represented as: ; in, It is the loss of the original model. The loss is calculated based on feedback data. This refers to the weight of the feedback data. Furthermore, the large language model itself is optimized through a reinforcement learning mechanism. The satisfaction ratings of maintenance personnel for repair suggestions and specific textual feedback are used as reward signals for the LLM. The LLM is fine-tuned through the RLHF mechanism to make its generated diagnostic explanations, repair steps, and natural language interactive responses more accurate, clearer, and better suited to the needs of on-site maintenance personnel. Its reward function... It can be designed as follows: ; in, Verification based on repair results and Ratings from operations and maintenance personnel The punishment will be based on whether the suggestion may lead to a safety hazard. These are the corresponding weighting coefficients. Finally, with the accumulation of a large amount of fault and maintenance data, the system can build a more accurate predictive maintenance model. By analyzing the historical operating data of components, environmental factors, and maintenance records, the system can predict the remaining useful life (RUL) of components or the probability of future failures.

[0071] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A photovoltaic fault diagnosis, location, and maintenance suggestion push system, characterized in that, include: The task information acquisition module is used to collect image data of photovoltaic modules by using a drone equipped with a multimodal sensor, and to perform spatial registration in combination with a geographic information system to generate a fault task information set that includes the fault location, type and severity. The model matching module is used to collect on-site images and user pose information through the environment modeling module, construct a 3D environment model and match it with the power plant model, and overlay fault task information onto real components in the form of virtual markers to achieve augmented reality positioning. The fault analysis module is used to analyze fault information through the intelligent expert database module, provide diagnostic results and maintenance suggestions based on historical cases and knowledge graphs, and push them to the user's augmented reality device in a visual form. The feedback loop module is used to collect maintenance data and evaluations through the user feedback module, and then feed them back to the expert database to optimize the knowledge graph and diagnostic model.

2. The photovoltaic fault diagnosis, location, and maintenance suggestion push system according to claim 1, characterized in that, The task information collection module includes: The data acquisition unit is used to acquire multimodal image data of photovoltaic modules by using a drone equipped with a visible light sensor, an infrared thermal imaging sensor, and a hyperspectral sensor, and to obtain the geographic coordinate information corresponding to the multimodal image data; The spatial registration unit is used to perform spatial registration of the multimodal image data based on the geographic coordinate information and in conjunction with the geographic information system to obtain the registered image data. The feature extraction and fusion unit is used to extract infrared thermal distribution features and visible light crack features from the registered image data using a deep learning network, and to perform weighted fusion of the infrared thermal distribution features and visible light crack features through an attention mechanism to obtain fused features; The fault identification and information generation unit is used to identify the fault type, fault location and severity of the photovoltaic module based on the fusion features, and generate a fault task information set containing the fault location, the fault type, the severity and the recommended priority. The data upload unit is used to upload the fault task information set to the cloud server for storage and management.

3. The photovoltaic fault diagnosis, location, and maintenance suggestion push system according to claim 2, characterized in that, The feature extraction and fusion unit includes: The infrared portion of the registered image data is processed by a first convolutional neural network to extract infrared thermal distribution features representing areas of temperature anomalies. The visible light portion of the registered image data is processed by a second convolutional neural network to extract visible crack features representing the morphology of surface defects. For the infrared thermal distribution features and the visible light crack features, attention weights are calculated, and the attention weights are learnable parameters optimized through end-to-end training. The fused feature is obtained by weighting and summing the infrared thermal distribution feature and the visible light crack feature according to the attention weight; If the recognition accuracy of the fused features is lower than a preset threshold, the attention weights are adjusted and the features are refused.

4. The photovoltaic fault diagnosis, location, and maintenance suggestion push system according to claim 1, characterized in that, The model matching module includes: The on-site data acquisition unit is used to acquire on-site environmental images and user pose information through augmented reality devices, wherein the user pose information includes position coordinates and posture angles. The environment modeling unit is used to construct a dense point cloud map by using vision-inertial odometry-based simultaneous localization and mapping technology, combining the on-site environment image and the user pose information, through feature point detection and triangulation calculation, to obtain a three-dimensional environment model; The model registration unit is used to perform iterative nearest-point registration between the three-dimensional environment model and the pre-established three-dimensional power station model to obtain the registration result; The coordinate transformation and pose solving unit is used to convert the fault location in the fault task information set into augmented reality display coordinates according to the registration result, and to solve the camera pose by minimizing the reprojection error between image feature points and 3D model points through a pose optimization algorithm. The virtual overlay unit is used to overlay virtual markers onto the real photovoltaic module according to the camera pose, so as to realize the visual location of the fault point.

5. The photovoltaic fault diagnosis, location, and maintenance suggestion push system according to claim 1, characterized in that, The fault analysis module includes: An information input unit is used to obtain fault information from the fault task information set and input it into the intelligent expert database module built based on a large language model, wherein the intelligent expert database module integrates historical fault cases and photovoltaic operation and maintenance knowledge graph; The analysis and reasoning unit is used to perform semantic understanding and causal reasoning on the fault information through the intelligent expert database module, and to calculate the association weights between fault entities based on the knowledge graph, wherein the association weights are obtained by weighted summation of feature similarity function, time co-occurrence frequency function and maintenance result correlation function. The result generation unit is used to retrieve matching cases from the historical fault cases according to the association weight, and generate fault cause determination, maintenance steps and operation parameter suggestions to obtain diagnostic results and maintenance suggestions. A visualization push unit is used to convert the diagnostic results and repair suggestions into text, 3D models or animations, and push them to the user's augmented reality device for display; The confidence assessment unit is used to calculate the comprehensive confidence level by multi-model fusion if the confidence level of the diagnostic result is lower than a preset threshold, wherein the comprehensive confidence level is a weighted average of the predicted probabilities of multiple diagnostic models.

6. The photovoltaic fault diagnosis, location, and maintenance suggestion push system according to claim 1, characterized in that, The feedback closed-loop module includes: The feedback data acquisition unit is used to collect maintenance logs, tool usage information, power recovery data and satisfaction evaluations through the user feedback module after fault repair to obtain user feedback data; The knowledge update unit is used to feed the user feedback data back to the intelligent expert database module and update the association weights in the knowledge graph according to the user feedback data, wherein the update is calculated by the learning rate and the weight adjustment amount. The model optimization unit is used to fine-tune the parameters of the model based on the user feedback data using an incremental learning approach. The lifespan prediction unit is used to predict the remaining lifespan of the photovoltaic modules based on the user feedback data and historical operating data using a regression model. The expression is as follows: ； in, It is a regression function. It is an error term; The closed-loop application unit is used to apply the optimized knowledge graph and the diagnostic model to subsequent fault analysis, forming a self-learning closed loop.

7. The photovoltaic fault diagnosis, location, and maintenance suggestion push system according to claim 4, characterized in that, The coordinate transformation and pose solving unit includes: The three-dimensional coordinates of the fault location are obtained from the registration results, and the augmented reality display coordinates are calculated by combining the camera intrinsic parameter matrix and rotation and translation parameters, wherein the augmented reality display coordinates are the projection of the three-dimensional coordinates in the camera coordinate system; For image feature points and corresponding 3D model points, an objective function is constructed, wherein the objective function is the sum of squares of the reprojection errors under the rotation matrix plus a regularization parameter; The optimal camera pose is solved by minimizing the objective function, wherein the camera pose includes a rotation matrix and a translation vector; The position of the virtual marker is adjusted according to the optimal camera pose to ensure that the superposition accuracy reaches the centimeter level; if the deviation between the obtained camera pose and the actual pose exceeds a preset threshold, the objective function is iteratively optimized.

8. The photovoltaic fault diagnosis, location, and maintenance suggestion push system according to claim 7, characterized in that, The objective function is constructed as follows: ； in, For rotation matrix, For the first Image feature points, For the corresponding 3D model points, For camera projection function, This is the regularization parameter.

9. The photovoltaic fault diagnosis, location, and maintenance suggestion push system according to claim 5, characterized in that, The analysis and reasoning unit includes: The fault information is mapped to fault entities in a knowledge graph, and the relationships between entities are identified, wherein the knowledge graph is constructed using a graph neural network; Calculate the feature similarity function between the faulty entities, wherein the feature similarity function is obtained based on the cosine similarity of entity attributes; Calculate the time co-occurrence frequency function among the faulty entities, wherein the time co-occurrence frequency function is a normalized value of the number of times they co-occur in historical data; Calculate the correlation function of repair results among the faulty entities, wherein the correlation function of repair results is a correlation coefficient based on the success rate; The association weight is obtained by multiplying the feature similarity function, the time co-occurrence frequency function, and the maintenance result correlation function by their corresponding weight coefficients and then summing them. The formula is as follows: ； in, Represents the faulty entity and The correlation weight between them For feature similarity function, Let co-occurrence frequency function be used. For the correlation function of maintenance results, , , These are the corresponding weighting coefficients; Based on this association weight, causal reasoning is performed to generate diagnostic results.

10. The photovoltaic fault diagnosis, location, and maintenance suggestion push system according to claim 1, characterized in that, The fault analysis module, specifically for fire or thermal runaway emergency response scenarios, also includes an emergency response scenario module, which includes: The multi-source heterogeneous data preprocessing unit is used to perform smoke scattering compensation and adaptive filtering on infrared thermal image sequences, and to perform robust estimation and outlier removal on electrical data; it assigns a confidence coefficient based on data missing rate, historical consistency and sensor health; it performs dynamic time alignment with a unified timestamp standard, and maps the positioning coordinates of multiple devices to the same three-dimensional coordinate system, and forms fire point feature vector and device status vector after correlation filtering and dimensionality reduction; The fire prediction unit, based on a spatiotemporal graph convolutional network and incorporating a physical constraint loss function, couples terrain-thermal-electrical multiphysics data for modeling, outputting a fire spread probability map and an electrical short-circuit risk probability map. The expression is: ； in, This provides real-time input data for multiphysics; its expression is: ； in, For terrain elevation and slope information, For thermal radiation distribution information and This provides electrical parameter information for components / branch circuits. The collaborative scheduling unit defines drones and firefighting robots as the game's main players, and fire points as the game's objects. It integrates fire point temperature / heat flux, threat distance, combustible material sensitivity, and predicted risk to construct a comprehensive fire point weight. An improved Nash equilibrium solution is then obtained through the player payoff function (fire point risk payoff minus equipment resource costs), which serves as the optimal collaborative scheduling scheme. Specifically: The set of game participants, including n drones and m firefighting robots, is expressed as: ； The fire point set expression is: ； Each subject The strategy is to select the fire points or fire point sequences to be dealt with in the current cycle. ; The fire point weight expression is: ； in For heat flux, The threat distance or cost of reaching critical equipment, Sensitivity to flammable materials ; The profit function expression is: ； in, The resource cost for the parties involved in the game to deal with the target fire point; The cost function expression is: ； in, Remaining battery power This refers to the remaining extinguishing agent. , For reachability / processing time, parameters , , These are the time cost weight, the power consumption weight, and the extinguishing agent consumption weight, respectively.

Images