An air-ground mobile edge networking and resource coordination scheduling method for photovoltaic operation and maintenance

Abstract

The application discloses a kind of aerial mobile edge networking and resource coordination scheduling methods for photovoltaic operation and maintenance, comprising the following steps: S1, collect multi-source state data, construct dynamic inspection task graph;S2, construct resource state vector and aerial-ground platform collaborative state vector, and calculate the predicted value of pre-execution business utility;S3, construct teacher set, and generate teacher availability mask;S4, knowledge aggregation is carried out, and two-stage knowledge distillation link is constructed based on the aggregated teacher knowledge;S5, construct agent, and train;S6, cloud-side review and on-site reinspection are carried out;S7, update teacher reliability.The application can significantly improve the differentiated processing capability of inspection task, resource utilization efficiency and aerial-ground collaborative response speed in complex dynamic photovoltaic station, reduce high-risk task delay and edge queue congestion risk.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology for new energy power stations, specifically involving a method for air-ground mobile edge networking and resource collaborative scheduling for photovoltaic operation and maintenance. Background Technology

[0002] Large-scale photovoltaic (PV) power plants typically consist of a large number of widely distributed PV modules, strings, combiner boxes, inverters, box-type transformers, and power plant roads. These various devices have clearly defined spatial and electrical topological relationships, and their inspection tasks differ significantly in risk level, data scale, time constraints, and processing methods. For example, hot spots, shading, dust accumulation, and surface damage on modules can be screened extensively by drones equipped with visible light or infrared equipment; abnormal string outputs, abnormal combiner box temperature rises, and abnormal inverter status require further analysis based on power plant monitoring data; and tasks involving suspected loose wiring, localized temperature rises, cable damage, and road obstructions require close-range re-inspection by unmanned vehicles or inspection robots.

[0003] Existing photovoltaic inspection methods typically focus on UAV flight path planning and image recognition, making it difficult to differentiate between different tasks based on mission risk, data volume, link quality, and edge computing load. Uploading all inspection images and videos uniformly to the cloud can easily lead to communication link congestion, edge queue backlog, and delays in high-risk task response. Existing knowledge distillation methods mainly transfer classification results or intermediate features from high-capacity cloud models to lightweight edge models, while existing multi-agent collaborative methods primarily rely on environmental interaction for task allocation and resource control. The former struggles to directly support data uploading, communication relay, and on-site re-inspection decisions, while the latter suffers from slow training convergence and insufficient early decision stability in complex dynamic scenarios.

[0004] Therefore, a collaborative scheduling method for photovoltaic power plant inspection task chain is needed, which can transfer global knowledge from the cloud, edge area knowledge, communication resource knowledge and on-site re-inspection knowledge to the edge intelligent agent in a hierarchical manner, and use the transferred executable knowledge to guide drones, unmanned vehicles and inspection robots to complete task allocation, data uploading, computation unloading, communication relay and on-site re-inspection. Summary of the Invention

[0005] To address the challenges of wide distribution of maintenance objects, significant differences in business types, limited communication and computing resources, and insufficient collaborative efficiency between air and ground platforms in large-scale photovoltaic power plants, this invention proposes an air-ground mobile edge networking and resource collaborative scheduling method for photovoltaic operation and maintenance.

[0006] The technical solution of this invention is: a method for air-to-ground mobile edge networking and resource collaborative scheduling for photovoltaic operation and maintenance, comprising the following steps:

[0007] S1. Collect multi-source status data and construct a dynamic inspection task map;

[0008] S2. Based on the dynamic inspection task graph, construct the resource status vector and the air-ground platform collaborative status vector, and calculate the predicted business utility value before execution.

[0009] S3. Construct a set of teachers and generate a teacher availability mask;

[0010] S4. Based on the teacher availability mask, perform knowledge aggregation, and construct a two-level knowledge distillation link based on the aggregated teacher knowledge;

[0011] S5. Based on the two-level knowledge distillation link, construct an intelligent agent and train it;

[0012] S6. Based on the predicted business utility value before execution, the distillation results of the two-level knowledge distillation chain, and the training results, conduct cloud-based review and on-site re-inspection.

[0013] S7. Update teacher reliability based on cloud-based review results and on-site inspection results.

[0014] Furthermore, S1 includes the following sub-steps:

[0015] S11. Construct a photovoltaic equipment topology map based on the photovoltaic power station electronic map, equipment ledger, and power station monitoring data;

[0016] S12. Collect multi-source status data;

[0017] S13. Based on the photovoltaic equipment topology diagram and multi-source status data, construct a dynamic inspection task;

[0018] S14. Generate a photovoltaic inspection task chain and construct a dynamic inspection task diagram based on the dynamic inspection tasks.

[0019] Furthermore, in S11, the set of nodes in the photovoltaic equipment topology diagram includes photovoltaic modules, photovoltaic strings, combiner boxes, inverters, box-type transformers, road nodes, edge nodes, and communication gateways;

[0020] In S13, the task context vector of the dynamic inspection task for:

[0021]

[0022] in: Task type; For the task location; Task risk level; Set the initial priority for the task; The deadline for the task; This represents the original data volume. To compress the amount of feature data; This refers to the amount of keyframe data. Calculate the requirements for the task; For mission reliability requirements; For task dependencies or on-site re-inspection requirements;

[0023] In S14, the photovoltaic inspection task chain for:

[0024] ;

[0025] in: For wide-area screening tasks using drones; For edge node or cloud analytics tasks; Review for high-risk tasks; For close-range re-inspection tasks of unmanned vehicles or inspection robots.

[0026] Furthermore, S2 includes the following sub-steps:

[0027] S21. Based on the dynamic inspection task map, construct the resource status vector and the air-ground platform collaborative status vector.

[0028] S22. Calculate the estimated task completion delay;

[0029] S23. Calculate the predicted value of timeliness utility before execution based on the expected task completion delay and task deadline;

[0030] S24. Calculate the predicted value of data quality utility before execution;

[0031] S25. Based on the resource state vector and the air-ground platform collaborative state vector, calculate the predicted values ​​of reliability utility and success utility before execution.

[0032] S26. Calculate the business utility prediction value before execution based on the predicted values ​​of timeliness utility, data quality utility, reliability utility, and success utility before execution.

[0033] Furthermore, in S21, the resource state vector at time t... for:

[0034] ;

[0035] in: Remaining battery level on each platform; The edge node queue status; This refers to link quality or channel status. Available bandwidth; Available computing resources; The wireless resource block status; Transmit power status; This serves as a mask for the availability of relay nodes.

[0036] In S21, the air-ground platform cooperative state vector at time t for:

[0037] ;

[0038] in: For neighboring intelligent agent relationships; This indicates the current activation status of abilities and the division of roles; The task is in an overridden state; This indicates a situation of task duplication and resource contention. This refers to the current path and the adjustable range.

[0039] In S22, the estimated task completion delay is... for:

[0040] ;

[0041] in: For estimated movement delay; To sense latency; For transmission delay; To delay queuing; To calculate the delay; The on-site re-inspection was delayed;

[0042] In S23, the predicted value of the time-sensitive utility before execution. for:

[0043] ;

[0044] in: This is a timeliness utility function used to map the expected completion delay and deadline to a normalized timeliness score; For the task The deadline;

[0045] In S24, the predicted data quality utility value before execution for:

[0046] ;

[0047] in: To predict the quality of perceived data (such as resolution, sharpness, signal-to-noise ratio, etc.). Analyze the quality prediction values ​​for the model (such as fault identification confidence and classification probability); For data integrity or availability predictions (such as whether there are missing frames, compression loss, etc.);

[0048] In S25, reliability utility before execution for:

[0049] ;

[0050] in: The reliability score is calculated by combining link, edge queue, and relay status as a reliability utility function. Link quality or channel state (same as in the resource state vector) ); The edge node queue state (in the same resource state vector) ); For relay node availability mask (same as in resource state vector) );

[0051] In S25, the predicted success utility value before execution for:

[0052] ;

[0053] in: The success utility function estimates the probability of task success based on task context, resource contention, and cooperative state. For the task context vector (same as in S13) (Including task type, location, risk level, etc.)

[0054] In S25, the predicted business utility value before the execution of the task q to be executed. for:

[0055] ;

[0056] in: As the basic weight of business; For the value of task information; This is the information value adjustment coefficient; These are the predicted values ​​for each component of utility, where ; The weighting coefficients for each component of utility satisfy the following conditions: and timeliness; qual; rel; succ; success rate. It is a set of sub-utilities, namely {timeliness, data quality, reliability, and success rate}.

[0057] Furthermore, in S3, the teacher group for:

[0058] ;

[0059] in: For all teachers in the cloud; For teachers in marginalized areas; For communication resource teachers; For on-site re-inspection and correction of teachers;

[0060] In S3, the teacher availability mask for task q, teacher m, and knowledge type h. for:

[0061] .

[0062] Furthermore, S4 includes the following sub-steps:

[0063] S41. Calculate the teacher routing score based on the teacher availability mask;

[0064] S42. Calculate teacher weights based on teacher routing scores;

[0065] S43. Based on the teacher weight, perform multi-type executable knowledge aggregation to obtain aggregated teacher knowledge;

[0066] S44. Based on the aggregated teacher knowledge, construct a two-level knowledge distillation link.

[0067] Furthermore, in S41, the teacher's routing score It is a quantitative indicator used to dynamically select the most suitable teacher model, expressed as:

[0068] ;

[0069] in: For weight generation function; For task context vectors; Let be the resource state vector at time t; Let be the coordinated state vector of the air-ground platform at time t; The predicted business utility value before the execution of task q; For teacher confidence; For the uncertainty of teachers; For teachers in task types Historical reliability; The teacher availability mask for task q, teacher m, and knowledge type h;

[0070] In S42, teacher weight for:

[0071] ;

[0072] in: Let be the set of candidate teachers who can output knowledge of the h-th class; exp(·) is the exponential function;

[0073] In S43, the aggregated teacher knowledge for:

[0074] ;

[0075] in: For teacher m, the output of knowledge type h;

[0076] In S44, the loss function of the two-stage knowledge distillation link for:

[0077] ;

[0078] in: For hard label supervision loss; The distillation loss between teacher soft labels and student output; The loss is the alignment loss between teacher and student features; The strategy distillation loss is determined by task priority, processing level, and the preferences of the executing entity. Distillation loss for communication strategies that include upload granularity, relay requests, and resource reservation recommendations; This refers to the connection between the cloud and the edge. Indicates the connection from the edge to the device. The weights for the hard-label supervision loss; Weights for the distillation loss between teacher soft labels and student output; Weights for the alignment loss between teacher and student features; Weights for strategy distillation loss based on task priority, processing level, and executor preference; Weights for distillation loss in communication strategies that include upload granularity, relay requests, and resource reservation recommendations.

[0079] Furthermore, S5 includes the following sub-steps:

[0080] S51. Based on a two-level knowledge distillation link, construct an intelligent agent and build enhanced local observation for the intelligent agent;

[0081] S52. Determine the agent's output action based on enhanced local observation of the agent;

[0082] S53. Training is performed based on the output actions of the intelligent agent;

[0083] In S51, enhanced local observation of agent i for:

[0084] ;

[0085] in: This includes local location, power level, neighbors, candidate tasks, link status, and edge load; Representation of task knowledge after distillation; This is a prerequisite for task scheduling; Prior to communication resources; Use it as a mask for actionable actions; This is a vector consisting of the predicted business utility values ​​of the candidate tasks before execution.

[0086] In S52, the output action of agent i for:

[0087] ;

[0088] in: Select an action for the task; Choose the implementing entity or the collaborating entity; For data upload mode; To handle hierarchical levels; For communication resource requests or communication control actions; For path adjustment actions; This action was triggered by on-site re-inspection.

[0089] Furthermore, S7 includes the following sub-steps:

[0090] S71. Based on the cloud review results and on-site inspection results, calculate the realized business utility value after execution, and calculate the business utility deviation based on the realized business utility value after execution;

[0091] S72. Generate experience records based on business utility deviation, pre-execution business utility forecast, cloud review results, and on-site inspection results;

[0092] S73. Calculate teacher reliability based on experience records;

[0093] S74. Update teacher reliability according to task type and knowledge type;

[0094] In S71, the realized value of business utility after execution for:

[0095] ;

[0096] in: As the basic weight of business; This is the information value adjustment coefficient; For the value of task information; It is a set of individual utilities; The actual utility values ​​of each component after implementation (timeliness, data quality, reliability, success rate); These are the weighting coefficients for the utility of each component;

[0097] In S71, business utility deviation for:

[0098] ;

[0099] in: The predicted business utility value before the execution of task q;

[0100] In S72, experience records for:

[0101] ;

[0102] in, For task context vectors; Let be the resource state vector at time t; Let be the coordinated state vector of the air-ground platform at time t; The action vector in S52 is the output action of the agent at time t for task q (including task selection, execution subject, upload mode, etc.). For all intelligent agents at time Joint actions; This is the actual execution result vector;

[0103] In S73, teacher reliability for:

[0104] ;

[0105] in: This is a truncation function that restricts the input values ​​to a valid range. The indicator variable is used to indicate whether the teacher m's suggestion on the knowledge type h of task q is adopted (1 indicates adoption, 0 indicates non-adoption). To ensure consistency between the diagnostic results and the on-site re-examination results; For task completion rate or service quality satisfaction; To improve business efficiency; For the cost of time delay; For the cost of energy consumption; This comes at the cost of increased upload volume; The cost of task conflicts, duplicate executions, or incorrect coordination; The weighting of the consistency between the diagnostic results and the on-site re-examination results; Weights for task completion rate or service quality satisfaction; Weighting for the degree of improvement in business effectiveness; The weight for latency cost; The weight of energy consumption cost; Weighting for the cost of upload volume; Weights for the costs of task conflicts, duplicate executions, or incorrect coordination;

[0106] In S74, the teacher reliability is updated as follows:

[0107] ;

[0108] in: Update the step size for reliability; For the updated teacher reliability; The teacher reliability before the update.

[0109] The beneficial effects of this invention are:

[0110] (1) This invention constructs a dynamic task graph and business utility prediction mechanism. Based on the photovoltaic equipment topology, operating status, sensing results and environmental changes, an inspection task chain is dynamically generated, and the business utility is predicted before the task is executed, providing a comprehensive scoring basis for task sorting and resource scheduling.

[0111] (2) This invention introduces multi-teacher hierarchical knowledge distillation to realize cross-level transfer of executable knowledge. A heterogeneous teacher pool is constructed, consisting of global teachers in the cloud, edge area teachers, communication resource teachers, and on-site review and correction teachers. Action-level knowledge directly related to the execution of inspection tasks (task priority, data upload granularity, processing level, link selection, relay request, resource reservation, execution subject preference, on-site review trigger probability, etc.) is distilled and transferred step by step along "cloud → edge → device", enhancing the collaborative decision-making ability of the edge agent.

[0112] (3) This invention realizes the collaborative scheduling of air and ground multi-agent systems and the joint optimization of communication and computing resources. Based on local observations, business utility predictions and action priors obtained by distillation, UAVs, unmanned vehicles and inspection robots complete task selection, path adjustment, hierarchical data uploading, computation unloading, communication relay and on-site re-inspection in a distributed manner, realizing differentiated task processing and dynamic collaborative allocation of resources.

[0113] (4) This invention forms a closed-loop feedback and adaptive update of teacher reliability. The actual business utility value is calculated based on the actual delay, data quality, link reliability, energy consumption and re-inspection results after task execution, and compared with the predicted value. The reliability of each teacher under different task types is dynamically updated, and the subsequent knowledge transfer weight and scheduling strategy are adjusted to form a closed-loop optimization process.

[0114] (5) The present invention can significantly improve the differentiated processing capability, resource utilization efficiency and air-ground collaborative response speed of inspection tasks in complex and dynamic photovoltaic power stations, and reduce the risk of delay of high-risk tasks and edge queue congestion. Attached Figure Description

[0115] Figure 1 A flowchart of an air-to-ground mobile edge networking and resource collaborative scheduling method for photovoltaic operation and maintenance;

[0116] Figure 2 This is a schematic diagram of the air-ground collaborative scheduling system for photovoltaic operation and maintenance in this invention;

[0117] Figure 3 This is a graph showing the online update results of multi-teacher reliability in an embodiment of the present invention. Detailed Implementation

[0118] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0119] like Figure 1 As shown, this invention provides a method for air-to-ground mobile edge networking and resource collaborative scheduling for photovoltaic operation and maintenance, including the following steps:

[0120] S1. Collect multi-source status data and construct a dynamic inspection task map;

[0121] S2. Based on the dynamic inspection task graph, construct the resource status vector and the air-ground platform collaborative status vector, and calculate the predicted business utility value before execution.

[0122] S3. Construct a set of teachers and generate a teacher availability mask;

[0123] S4. Based on the teacher availability mask, perform knowledge aggregation, and construct a two-level knowledge distillation link based on the aggregated teacher knowledge;

[0124] S5. Based on the two-level knowledge distillation link, construct an intelligent agent and train it;

[0125] S6. Based on the predicted business utility value before execution, the distillation results of the two-level knowledge distillation chain, and the training results, conduct cloud-based review and on-site re-inspection.

[0126] S7. Update teacher reliability based on cloud-based review results and on-site inspection results.

[0127] This invention specifically relates to an air-ground collaborative inspection method for large-scale photovoltaic power plants, and more particularly to a method, system, device, and storage medium for performing vertical multi-teacher knowledge transfer between the cloud, edge, and device ends, and using the transferred task-level executable knowledge to guide drones, unmanned vehicles, inspection robots, fixed edge nodes, and mobile edge nodes in performing horizontal collaborative task allocation, hierarchical data uploading, computation offloading, communication relay, and close-range re-inspection.

[0128] This invention is applicable to centralized photovoltaic power stations, mountain photovoltaic power stations, desert photovoltaic power stations, agricultural-photovoltaic complementary power stations, and new energy power stations with large-scale array-type equipment distribution characteristics.

[0129] This invention is based on an air-to-ground mobile edge network consisting of drones, unmanned vehicles, inspection robots, fixed edge nodes, mobile edge nodes, communication gateways, and a cloud coordinator. It introduces a multi-agent collaborative scheduling mechanism enhanced by multi-teacher hierarchical knowledge distillation. By progressively migrating cloud-based global knowledge, edge region knowledge, communication resource optimization knowledge, and on-site review and correction knowledge to the edge and device ends, edge agents can combine local observations, predicted service utility values, and action priors obtained through distillation to collaboratively complete task selection, path adjustment, hierarchical data uploading, computational offloading, communication relay, resource reservation, and on-site review.

[0130] This method first constructs an equipment topology and inspection task map based on photovoltaic modules, strings, combiner boxes, inverters, box-type transformers, roads, and auxiliary facilities. It then generates dynamic inspection tasks by combining equipment operating status, sensing results, and environmental changes. Subsequently, before task execution, the system calculates predicted business utility values ​​and constructs cloud-based global teachers, edge area teachers, communication resource teachers, and on-site review and correction teachers. For different inspection tasks, the system dynamically determines the knowledge transfer weights of different teachers based on task risk, data scale, time constraints, link status, edge load, platform power consumption, relay availability, and the coordination status of air-to-ground platforms.

[0131] Unlike knowledge distillation methods that only transfer classification results, this invention transfers executable knowledge, such as task priority, data upload granularity, processing level, communication relay, resource reservation, execution subject preferences, and on-site re-inspection trigger probability, to the end-side student strategy. Drones, unmanned vehicles, and inspection robots collaboratively complete task selection, path adjustment, hierarchical data upload, computational unloading, communication resource requests, and on-site re-inspection based on local observations, predicted business utility values, and action priors obtained through distillation. After task execution, the system calculates the realized business utility value based on actual latency, data quality, link reliability, platform energy consumption, uploaded data volume, and on-site re-inspection results. It also updates the teacher reliability based on the deviation between the predicted and realized values, thereby adjusting the knowledge transfer weights and collaborative scheduling strategies in subsequent tasks.

[0132] In this embodiment of the invention, S1 includes the following sub-steps:

[0133] S11. Construct a photovoltaic equipment topology map based on the photovoltaic power station electronic map, equipment ledger, and power station monitoring data;

[0134] S12. Collect multi-source status data;

[0135] S13. Based on the photovoltaic equipment topology diagram and multi-source status data, construct a dynamic inspection task;

[0136] S14. Generate a photovoltaic inspection task chain and construct a dynamic inspection task diagram based on the dynamic inspection tasks.

[0137] This step is used to convert the equipment hierarchy, spatial array relationship, inspection path relationship, and dynamic inspection requirements in the photovoltaic power station into a processable task graph.

[0138] In this embodiment of the invention, in S11, the node set of the photovoltaic equipment topology diagram includes photovoltaic modules, photovoltaic strings, combiner boxes, inverters, box-type transformers, road nodes, edge nodes, and communication gateways.

[0139] The system constructs a photovoltaic equipment topology map based on the electronic map of the photovoltaic power station, equipment ledger, and station monitoring data.

[0140] The node set includes at least: {photovoltaic modules, photovoltaic strings, combiner boxes, inverters, box-type transformers, road nodes, edge nodes, and communication gateways}.

[0141] The relationship types include at least: the relationship between the component and the string; the relationship between the string and the combiner box; the relationship between the combiner box and the inverter; the spatial adjacency relationship of the equipment; the reachability relationship of the air inspection path; the reachability relationship of the ground re-inspection path; the wireless communication coverage relationship; and the on-site re-inspection dependency relationship.

[0142] The system synchronously collects one or more of the following data through the site monitoring system, UAV payload, unmanned vehicle sensors, inspection robot sensors, edge nodes, communication gateways, and meteorological monitoring equipment:

[0143] Photovoltaic equipment operation data: including module output power, string current, voltage, combiner box status, inverter status, box-type transformer status, historical fault records, and equipment importance;

[0144] Inspection and sensing data includes: visible light images, infrared thermal imaging images, videos, close-range high-resolution images, road obstacle images, character recognition images, and structured sensor data;

[0145] Platform operation data includes the location, speed, remaining battery power, payload capacity, computing power, storage capacity, and current task status of drones, unmanned vehicles, and inspection robots.

[0146] Communication status data includes link quality, channel gain, available bandwidth, number of radio resource blocks, transmit power, transmission rate, reliability, and relay availability.

[0147] Edge computing status data: including available computing power, queue length, estimated queuing latency, and model deployment status for fixed and mobile edge nodes;

[0148] Environmental data includes light intensity, temperature, wind speed, visibility, road conditions, and weather warnings.

[0149] The system will generate a new inspection task or adjust the priority of an existing task when any of the following conditions are met:

[0150] The operating status of photovoltaic modules, strings, combiner boxes, inverters, or box-type transformers deviates from the dynamic baseline;

[0151] The drone's wide-area screening identified hot spots, obstructions, dust accumulation, surface damage, or other suspected anomalies;

[0152] The confidence level of the analysis results from edge nodes or the cloud is insufficient, and supplementary sensing data is needed;

[0153] Suspected malfunctions require close-range re-inspection by unmanned vehicles or inspection robots;

[0154] Temporary tasks are triggered by road obstacles, extreme weather, equipment maintenance, or manual instructions.

[0155] Degraded wireless link quality, persistent edge queue congestion, or insufficient platform power necessitate a readjustment of task scheduling.

[0156] In S13, the task context vector of the dynamic inspection task for:

[0157]

[0158] in: Task type; For the task location; Task risk level; Set the initial priority for the task; The deadline for the task; This represents the original data volume. To compress the amount of feature data; This refers to the amount of keyframe data. Calculate the requirements for the task; For mission reliability requirements; For task dependencies or on-site re-inspection requirements;

[0159] In S14, the photovoltaic inspection task chain for:

[0160] ;

[0161] in: For wide-area screening tasks using drones; For edge node or cloud analytics tasks; Review for high-risk tasks; For close-range re-inspection tasks of unmanned vehicles or inspection robots.

[0162] The system organizes relevant tasks into photovoltaic inspection task chains based on equipment topology, task risks, data requirements, and analysis results. Not all four stages in the task chain are mandatory; instead, different stages are dynamically triggered, skipped, or rearranged based on task risks, model confidence levels, data quality, wireless link status, and edge computing resources.

[0163] In this embodiment of the invention, S2 includes the following sub-steps:

[0164] S21. Based on the dynamic inspection task map, construct the resource status vector and the air-ground platform collaborative status vector.

[0165] S22. Calculate the estimated task completion delay;

[0166] S23. Calculate the predicted value of timeliness utility before execution based on the expected task completion delay and task deadline;

[0167] S24. Calculate the predicted value of data quality utility before execution;

[0168] S25. Based on the resource state vector and the air-ground platform collaborative state vector, calculate the predicted values ​​of reliability utility and success utility before execution.

[0169] S26. Calculate the business utility prediction value before execution based on the predicted values ​​of timeliness utility, data quality utility, reliability utility, and success utility before execution.

[0170] This step is used to identify the differences between different inspection tasks before the task is actually executed, and to provide input for task sequencing, multi-teacher routing, and air-ground multi-agent collaborative scheduling.

[0171] In this embodiment of the invention, in S21, the resource state vector at time t... for:

[0172] ;

[0173] in: Remaining battery level on each platform; The edge node queue status; This refers to link quality or channel status. Available bandwidth; Available computing resources; The wireless resource block status; Transmit power status; This serves as a mask for the availability of relay nodes.

[0174] In S21, the air-ground platform cooperative state vector at time t for:

[0175] ;

[0176] in: For neighboring intelligent agent relationships; This indicates the current activation status of abilities and the division of roles; The task is in an overridden state; This indicates a situation of task duplication and resource contention. This refers to the current path and the adjustable range.

[0177] In S22, the expected task completion delay is calculated based on the expected movement delay, sensing delay, transmission delay, queuing delay, calculation delay, and on-site re-inspection delay.

[0178] Expected task completion delay for:

[0179] ;

[0180] in: For estimated movement delay; To sense latency; For transmission delay; To delay queuing; To calculate the delay; The on-site re-inspection was delayed;

[0181] In S23, the predicted value of timeliness utility is calculated based on the expected task completion delay and task deadline. Predicted value of timeliness utility before execution. for:

[0182] ;

[0183] in: This is a timeliness utility function used to map the expected completion delay and deadline to a normalized timeliness score; For the task The deadline;

[0184] In S24, based on the wireless link, edge node, execution platform, and on-site re-inspection conditions, the predicted values ​​of reliability utility and success utility are calculated. Predicted value of data quality utility before execution. for:

[0185] ;

[0186] in: To predict the quality of perceived data (such as resolution, sharpness, signal-to-noise ratio, etc.). Analyze the quality prediction values ​​for the model (such as fault identification confidence and classification probability); For data integrity or availability predictions (such as whether there are missing frames, compression loss, etc.);

[0187] In S25, reliability utility before execution for:

[0188] ;

[0189] in: The reliability score is calculated by combining link, edge queue, and relay status as a reliability utility function. Link quality or channel state (same as in the resource state vector) ); The edge node queue state (in the same resource state vector) ); For relay node availability mask (same as in resource state vector) );

[0190] In S25, the predicted success utility value before execution for:

[0191] ;

[0192] in: The success utility function estimates the probability of task success based on task context, resource contention, and cooperative state. For the task context vector (same as in S13) (Including task type, location, risk level, etc.)

[0193] In S25, the predicted business utility value before the execution of the task q to be executed. for:

[0194] ;

[0195] in: As the basic weight of business; For the value of task information; This is the information value adjustment coefficient; These are the predicted values ​​for each component of utility, where ; The weighting coefficients for each component of utility satisfy the following conditions: and timeliness; qual; rel; succ; success rate. It is a set of sub-utilities, namely {timeliness, data quality, reliability, and success rate}.

[0196] Business utility is a comprehensive score used for task sorting and resource scheduling, not a probability value; each sub-item of utility is normalized to [0,1] before calculation.

[0197] The basic weight of a task is determined based on the task type, failure risk, equipment importance, and scope of impact. For example: routine component image inspection tasks have low or medium basic weight; tasks involving suspected hot spots, abnormal string outputs, abnormal combiner box temperature rises, and abnormal inverter status have high basic weight; emergency investigation and high-risk on-site re-inspection tasks have high basic weight and are given priority in accessing path, communication, and computing resources; video evidence collection tasks involve a large amount of data, and their basic weight is dynamically adjusted based on the severity of the event and the requirements for evidence integrity.

[0198] In this embodiment of the invention, in S3, the teacher set for:

[0199] ;

[0200] in: For all teachers in the cloud; For teachers in marginalized areas; For communication resource teachers; For on-site re-inspection and correction of teachers;

[0201] In S3, the teacher availability mask for task q, teacher m, and knowledge type h. for:

[0202] .

[0203] This step is used to build teacher models from different sources, at different levels, and with different functions, so that end-student strategies can acquire diagnostic, computational, communication, and on-site review knowledge.

[0204] Cloud-based global teacher Deployed on a cloud server. The cloud-based global teacher outputs the following based on the photovoltaic equipment topology, historical inspection data, global task status, high-resolution images, infrared images, videos, and platform operating status: fault category probability; task risk score; task priority; recommended execution entity; recommended processing level; and probability of triggering on-site re-inspection.

[0205] Teachers in marginal areas Deployed on fixed edge nodes, site gateways, or mobile edge nodes.

[0206] Based on the device topology, regional task set, edge queue, and nearby platform status, the edge region teacher outputs: regional task priority; estimated queuing delay; edge node selection suggestions; local computation offloading suggestions; and processing level suggestions.

[0207] Communication Resources Teacher Deployed on cloud servers, edge nodes, or communication gateways.

[0208] Based on the task data volume, task time limit, reliability requirements, wireless link, available bandwidth, wireless resource blocks, transmit power, edge load, and relay status, the communication resource teacher outputs: data upload granularity suggestions; link selection suggestions; relay request suggestions; candidate relay nodes; wireless resource reservation level; estimated transmission latency; and estimated link reliability.

[0209] The communication resource teacher is used to transfer communication resource optimization experience to the end-side student strategy, enabling the end-side agent to make differentiated communication decisions based on the latency, data volume and reliability requirements of different photovoltaic inspection services.

[0210] On-site re-inspection and correction of teachers It is built based on near-field reliable results obtained from unmanned vehicles, inspection robots, or manual verification.

[0211] On-site re-inspection and correction by teachers is used for: correcting suspected fault labels; marking tasks with misjudgments and omissions at the end-point; identifying task types that need to be given higher re-inspection priority; and updating the historical reliability of different teachers under different task types.

[0212] Define the teacher set and the knowledge type set. For task q, teacher m, and knowledge type h, the system generates a teacher availability mask. When a teacher lacks the corresponding data, the model is not deployed, the link is unavailable, edge resources are insufficient, or on-site review conditions are not met, the corresponding availability mask is set to 0.

[0213] In this embodiment of the invention, S4 includes the following sub-steps:

[0214] S41. Calculate the teacher routing score based on the teacher availability mask;

[0215] S42. Calculate teacher weights based on teacher routing scores;

[0216] S43. Based on the teacher weight, perform multi-type executable knowledge aggregation to obtain aggregated teacher knowledge;

[0217] S44. Based on the aggregated teacher knowledge, construct a two-level knowledge distillation link.

[0218] This step is used to dynamically select applicable teachers and to migrate knowledge directly related to the execution of inspection tasks to the edge and device ends step by step.

[0219] In this embodiment of the invention, in S41, the teacher's routing score is... It is a quantitative indicator used to dynamically select the most suitable teacher model, expressed as:

[0220] ;

[0221] in: For weight generation function; For task context vectors; Let be the resource state vector at time t; Let be the coordinated state vector of the air-ground platform at time t; The predicted business utility value before the execution of task q; For teacher confidence; For the uncertainty of teachers; For teachers in task types Historical reliability; The teacher availability mask for task q, teacher m, and knowledge type h;

[0222] In S42, teacher weight for:

[0223] ;

[0224] in: Let be the set of candidate teachers who can output knowledge of the h-th class; exp(·) is the exponential function;

[0225] When the denominator is 0, a fallback to the local end-side policy or a preset security policy should be implemented. Teacher weights are simultaneously affected by task risk, task type, expected business utility, data scale, link quality, edge load, battery status, relay conditions, and neighboring platform status.

[0226] In S43, the aggregated teacher knowledge for:

[0227] ;

[0228] in: For teacher m, the output of knowledge type h;

[0229] The system aggregates the following knowledge based on teacher weights: fault category and risk knowledge; task priority knowledge; processing level knowledge; data upload granularity knowledge; link selection and relay knowledge; communication resource reservation knowledge; execution subject preference; and on-site re-inspection trigger knowledge.

[0230] Different knowledge types allow for different teacher weights. For discrete knowledge such as fault category, upload mode, processing level, and execution entity, the teacher output is a probability distribution or scoring vector for candidate categories; for continuous knowledge such as bandwidth, power, and resource reservation, the teacher output is a parameter suggestion vector. The aggregated results are verified against action constraints to generate prior knowledge of edge actions.

[0231] For example: for high-risk tasks involving infrared hot spots, increase the weight of cloud-based global teachers and on-site review and correction teachers; for routine area inspection tasks, increase the weight of edge area teachers; for video evidence, high reliability, or link fluctuation tasks, increase the weight of communication resource teachers; for tasks with continuous congestion at fixed edge nodes, increase the weight of end-side processing, mobile edge processing, or cloud processing recommendations.

[0232] In S44, the loss function of the two-stage knowledge distillation link for:

[0233] ;

[0234] in: For hard label supervision loss; The distillation loss between teacher soft labels and student output; The loss is the alignment loss between teacher and student features; The strategy distillation loss is determined by task priority, processing level, and the preferences of the executing entity. Distillation loss for communication strategies that include upload granularity, relay requests, and resource reservation recommendations; This refers to the connection between the cloud and the edge. Indicates the connection from the edge to the device. The weights for the hard-label supervision loss; Weights for the distillation loss between teacher soft labels and student output; Weights for the alignment loss between teacher and student features; Weights for strategy distillation loss based on task priority, processing level, and executor preference; Weights for distillation loss in communication strategies that include upload granularity, relay requests, and resource reservation recommendations.

[0235] Deploy edge student models on fixed edge nodes, site gateways, or mobile edge nodes, and deploy device-side student models on drones, unmanned vehicles, and inspection robots.

[0236] The edge student model is used for regional task screening, edge inference, and offloading suggestion generation.

[0237] The device-side student model is used for local risk identification, data upload granularity selection, task action prior generation, and communication collaboration suggestion generation.

[0238] In this embodiment of the invention, S5 includes the following sub-steps:

[0239] S51. Based on a two-level knowledge distillation link, construct an intelligent agent and build enhanced local observation for the intelligent agent;

[0240] S52. Determine the agent's output action based on enhanced local observation of the agent;

[0241] S53. Training is performed based on the output actions of the intelligent agent;

[0242] This step enables drones, unmanned vehicles, and inspection robots to achieve distributed collaboration based on local observations, operational utility predictions, and distillation results.

[0243] Configure one or more of the following agents:

[0244] Unmanned aerial vehicle (UAV) intelligent agents: used for large-scale visible light inspection, infrared thermal imaging, video acquisition, path flight, aerial relay, and preliminary end-side screening;

[0245] Intelligent agents for unmanned vehicles: used for ground path movement, close-range inspection of combiner boxes and inverters, cable and road identification, on-site re-inspection, mobile edge computing, and mobile relay;

[0246] Inspection robot intelligent agent: used for close-range perception of local equipment and execution of on-site actions;

[0247] Fixed edge nodes: used for regional task analysis, computation offloading, and queue scheduling;

[0248] Communication gateways or relay nodes: used for link selection, relay forwarding, and radio resource allocation;

[0249] Cloud coordinator: Used for teacher pool management, global task planning, and strategy training.

[0250] In S51, enhanced local observation of agent i for:

[0251] ;

[0252] in: This includes local location, power level, neighbors, candidate tasks, link status, and edge load; Representation of task knowledge after distillation; This is a prerequisite for task scheduling; Prior to communication resources; Use it as a mask for actionable actions; This is a vector consisting of the predicted business utility values ​​of the candidate tasks before execution.

[0253] In S52, the output action of agent i for:

[0254] ;

[0255] in: Select an action for the task; Choose the implementing entity or the collaborating entity; For data upload mode; To handle hierarchical levels; For communication resource requests or communication control actions; For path adjustment actions; This action was triggered by on-site re-inspection.

[0256] The communication sub-action is as follows:

[0257] in: For intelligent agent i, a communication resource request or communication control action generated at time t; Select an action for the transmission link; This is a relay request action; The number of radio resource blocks requested for allocation; For bandwidth requirements; Transmission power level; Reserve a level for communication resources.

[0258] Data upload and computation unloading modes include: no unloading mode, feature-only mode, keyframe mode, full image mode, full video mode, and redundancy mode.

[0259] The processing layers include: device, edge node, cloud, and mobile edge node.

[0260] The system adopts a centralized training and distributed execution architecture.

[0261] During the training phase, the central coordinator uses the global task graph, platform status, link status, edge queues, business utility, and task execution results to train the multi-agent collaborative strategy.

[0262] During the execution phase, drones, unmanned vehicles, and inspection robots independently output actions based on their enhanced local observations.

[0263] Multi-agent proximal policy optimization methods, value decomposition methods, heuristic methods, or hybrid optimization methods can all be used to achieve air-ground multi-agent cooperative scheduling.

[0264] In this embodiment of the invention, the photovoltaic inspection task is differentiated based on task risk, business utility prediction, distillation results, and joint actions of air and ground multi-agents.

[0265] For routine inspection tasks with low risk, low uncertainty, and low or medium business utility, the initial screening is completed by the student model on the equipment side, and only structured results or compressed features are uploaded.

[0266] For tasks with medium risk, certain time requirements, or regional correlation analysis, upload compressed features, keyframes, or complete images, and have fixed or mobile edge nodes perform the analysis.

[0267] For tasks that are high-risk, highly uncertain, involve multiple adjacent components, or affect the operating status of strings and inverters, upload complete images, infrared images, videos, or redundant data for review by a global teacher in the cloud.

[0268] For tasks that require confirming cable status, localized temperature rise in equipment, abnormalities in combiner boxes or inverters, road obstacles, and issues that are difficult to accurately identify from the perspective of a drone, unmanned vehicles or inspection robots are dispatched to perform close-range re-inspections.

[0269] Based on task risk, upload mode, task time limit, link quality, and relay status, the system determines one or more of the following communication actions: number of radio resource blocks; bandwidth requirements; transmit power level; transmission link; candidate relay nodes; resource reservation level; and transmission priority.

[0270] When high-risk tasks occur, the platform's power is insufficient, roads are blocked, edge nodes are congested, link quality deteriorates, or relays fail, the system adjusts the task execution subject, inspection path, upload granularity, processing level, or relay node.

[0271] In this embodiment of the invention, S7 includes the following sub-steps:

[0272] S71. Based on the cloud review results and on-site inspection results, calculate the realized business utility value after execution, and calculate the business utility deviation based on the realized business utility value after execution;

[0273] S72. Generate experience records based on business utility deviation, pre-execution business utility forecast, cloud review results, and on-site inspection results;

[0274] S73. Calculate teacher reliability based on experience records;

[0275] S74. Update teacher reliability according to task type and knowledge type;

[0276] This step is used to update the teacher reliability and the knowledge transfer weight in subsequent tasks based on the actual performance of the task, the results of cloud review, and the results of on-site inspection.

[0277] In S71, the realized value of business utility after execution for:

[0278] ;

[0279] in: As the basic weight of business; This is the information value adjustment coefficient; For the value of task information; It is a set of individual utilities; The actual utility values ​​of each component after implementation (timeliness, data quality, reliability, success rate); These are the weighting coefficients for the utility of each component;

[0280] In S71, business utility deviation for:

[0281] ;

[0282] in: The predicted business utility value before the execution of task q;

[0283] In S72, experience records for:

[0284] ;

[0285] in, For task context vectors; Let be the resource state vector at time t; Let be the coordinated state vector of the air-ground platform at time t; The action vector in S52 is the output action of the agent at time t for task q (including task selection, execution subject, upload mode, etc.). For all intelligent agents at time Joint actions; This is the actual execution result vector;

[0286] The actual execution result vector includes completion latency, perceived quality, upload volume, queuing latency, link reliability, platform energy consumption, task completion status, cloud verification results, and on-site verification results.

[0287] The actual execution results include at least: task completion latency; perceived data quality; uploaded data volume; edge queuing latency; link reliability; platform energy consumption; task completion status; cloud verification results; and on-site verification results of unmanned vehicles or inspection robots.

[0288] In S73, teacher reliability for:

[0289] ;

[0290] in: This is a truncation function that restricts the input values ​​to a valid range. The indicator variable is used to indicate whether the teacher m's suggestion on the knowledge type h of task q is adopted (1 indicates adoption, 0 indicates non-adoption). To ensure consistency between the diagnostic results and the on-site re-examination results; For task completion rate or service quality satisfaction; To improve business efficiency; For the cost of time delay; For the cost of energy consumption; This comes at the cost of increased upload volume; The cost of task conflicts, duplicate executions, or incorrect coordination; The weighting of the consistency between the diagnostic results and the on-site re-examination results; Weights for task completion rate or service quality satisfaction; Weighting for the degree of improvement in business effectiveness; The weight for latency cost; The weight of energy consumption cost; Weighting for the cost of upload volume; Weights for the costs of task conflicts, duplicate executions, or incorrect coordination;

[0291] In S74, the teacher reliability is updated as follows:

[0292] ;

[0293] in: Update the step size for reliability; For the updated teacher reliability; The teacher reliability before the update.

[0294] When a teacher's suggestion is not adopted The reliability remains unchanged; when the suggestion is adopted, the reliability is updated according to the actual implementation effect, and the system adjusts the teacher routing weight in subsequent tasks according to the updated teacher reliability.

[0295] When on-site re-examination results indicate a decrease in the accuracy of end-side screening, the weight of cloud-based global teachers or on-site re-examination and correction teachers should be increased.

[0296] When a certain edge node is continuously congested, the scheduling suggestion weight of the corresponding edge region teacher is reduced.

[0297] When the wireless link continues to fluctuate or the failure rate of high-risk task transmission increases, the weight of communication resource teachers should be increased, and the priority of relay requests, resource reservations, or redundant uploads should be increased.

[0298] When routine inspection tasks maintain high accuracy at the edge for an extended period, increase the proportion of edge screening and reduce the amount of raw images and videos uploaded.

[0299] In embodiments of the present invention, such as Figure 2 As shown, the system employs multiple heterogeneous teacher models, including a global planning teacher based on Q-ALNS, a global search teacher based on CQGA, a heuristic teacher based on priority and distance rules, and a utility-greedy teacher based on pre-execution utility prediction. Each teacher model provides candidate suggestions for task allocation, path order, and resource scheduling actions. The system updates the reliability of each teacher model online based on the service quality, latency, energy consumption, reliability, and actual utility after task execution.

[0300] like Figure 3 As shown, different teacher models develop different reliability weights under the influence of execution feedback. In subsequent scheduling, the system weights and fuses teacher suggestions based on this reliability, generating a multi-teacher reliability-weighted consensus. This consensus is then used as a soft label for hierarchical knowledge distillation and input into the MAPPO policy networks corresponding to UAVs, autonomous vehicles, and robots. Through this method, teacher models with better historical performance receive higher weights in subsequent task scheduling, while the influence of teacher models with poor historical performance is reduced, thereby improving the scheduling adaptability in complex photovoltaic power plant operation and maintenance tasks.

[0301] By introducing an online update mechanism for teacher reliability, this invention can dynamically adjust the influence weight of different teacher models in scheduling decisions based on actual execution feedback, avoiding the decline in scheduling performance caused by the failure of a single teacher model in a specific scenario, and improving the robustness of task allocation, path planning and resource collaborative scheduling.

[0302] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for air-to-ground mobile edge networking and resource collaborative scheduling for photovoltaic operation and maintenance, characterized in that, Includes the following steps: S1. Collect multi-source status data and construct a dynamic inspection task map; S2. Based on the dynamic inspection task graph, construct the resource status vector and the air-ground platform collaborative status vector, and calculate the predicted business utility value before execution. S3. Construct a set of teachers and generate a teacher availability mask; S4. Based on the teacher availability mask, perform knowledge aggregation, and construct a two-level knowledge distillation link based on the aggregated teacher knowledge; S5. Based on the two-level knowledge distillation link, construct an intelligent agent and train it; S6. Based on the predicted business utility value before execution, the distillation results of the two-level knowledge distillation chain, and the training results, conduct cloud-based review and on-site re-inspection. S7. Update teacher reliability based on cloud-based review results and on-site inspection results.

2. The air-to-ground mobile edge networking and resource collaborative scheduling method for photovoltaic operation and maintenance according to claim 1, characterized in that, S1 includes the following sub-steps: S11. Construct a photovoltaic equipment topology map based on the photovoltaic power station electronic map, equipment ledger, and power station monitoring data; S12. Collect multi-source status data; S13. Based on the photovoltaic equipment topology diagram and multi-source status data, construct a dynamic inspection task; S14. Generate a photovoltaic inspection task chain and construct a dynamic inspection task diagram based on the dynamic inspection tasks.

3. The air-to-ground mobile edge networking and resource collaborative scheduling method for photovoltaic operation and maintenance according to claim 1, characterized in that, In S11, the node set of the photovoltaic equipment topology diagram includes photovoltaic modules, photovoltaic strings, combiner boxes, inverters, box-type transformers, road nodes, edge nodes, and communication gateways; In S13, the task context vector of the dynamic inspection task for: ， in: Task type; For the task location; Task risk level; Set the initial priority for the task; The deadline for the task; This represents the original data volume. To compress the amount of feature data; This refers to the amount of keyframe data. Calculate the requirements for the task; For mission reliability requirements; For task dependencies or on-site re-inspection requirements; In S14, the photovoltaic inspection task chain for: ； in: For wide-area screening tasks using drones; For edge node or cloud analytics tasks; Review for high-risk tasks; For close-range re-inspection tasks of unmanned vehicles or inspection robots.

4. The air-to-ground mobile edge networking and resource collaborative scheduling method for photovoltaic operation and maintenance according to claim 1, characterized in that, S2 includes the following sub-steps: S21. Based on the dynamic inspection task map, construct resource status vector and air-ground platform collaborative status vector; S22. Calculate the estimated task completion delay; S23. Calculate the predicted value of timeliness utility before execution based on the expected task completion delay and task deadline; S24. Calculate the predicted value of data quality utility before execution; S25. Based on the resource state vector and the air-ground platform collaborative state vector, calculate the predicted values ​​of reliability utility and success utility before execution. S26. Calculate the business utility prediction value before execution based on the predicted values ​​of timeliness utility, data quality utility, reliability utility, and success utility before execution.

5. The air-to-ground mobile edge networking and resource collaborative scheduling method for photovoltaic operation and maintenance according to claim 1, characterized in that, In S21, the resource state vector at time t for: ； in: Remaining battery level on each platform; The edge node queue status; This refers to link quality or channel status. Available bandwidth; Available computing resources; The wireless resource block status; Transmit power status; This serves as a mask for the availability of relay nodes. In S21, the air-ground platform cooperative state vector at time t for: ； in: For neighboring intelligent agent relationships; This indicates the current activation status of abilities and the division of roles; The task is in an overridden state; This indicates a situation of task duplication and resource contention. This refers to the current path and the adjustable range. In step S22, the estimated task completion delay is... for: ； in: For estimated movement delay; To sense latency; For transmission delay; To delay queuing; To calculate the delay; The on-site re-inspection was delayed; In step S23, the predicted value of the timeliness of the utility before execution is determined. for: ； in: It is a time-sensitive utility function; For the task The deadline; In step S24, the predicted data quality utility value before execution is... for: ； in: To predict values ​​for perceived data quality; To analyze the predicted quality values ​​for the model; For data integrity or availability predictions; In S25, the reliability utility before execution for: ； in: For reliability utility function; This refers to link quality or channel status. The edge node queue status; This serves as a mask for the availability of relay nodes. In S25, the predicted success utility value before execution for: ； in: For success utility function; For task context vectors; In step S25, the predicted business utility value of the task q to be executed before execution is obtained. for: ； in: As the basic weight of business; For the value of task information; This is the information value adjustment coefficient; These are the predicted values ​​for the utility of each component; For each sub-item's utility, 'time' represents timeliness; 'qual' represents data quality; 'rel' represents reliability; and 'succ' represents success rate. It is a set of individual utilities.

6. The air-to-ground mobile edge networking and resource collaborative scheduling method for photovoltaic operation and maintenance according to claim 1, characterized in that, In S3, the teacher set for: ； in: For all teachers in the cloud; For teachers in marginalized areas; For communication resource teachers; For on-site re-inspection and correction of teachers; In S3, the teacher availability mask for task q, teacher m, and knowledge type h. for: 。 7. The air-to-ground mobile edge networking and resource collaborative scheduling method for photovoltaic operation and maintenance according to claim 1, characterized in that, S4 includes the following sub-steps: S41. Calculate the teacher routing score based on the teacher availability mask; S42. Calculate teacher weights based on teacher routing scores; S43. Based on the teacher weight, perform multi-type executable knowledge aggregation to obtain aggregated teacher knowledge; S44. Based on the aggregated teacher knowledge, construct a two-level knowledge distillation link.

8. The air-to-ground mobile edge networking and resource collaborative scheduling method for photovoltaic operation and maintenance according to claim 1, characterized in that, In S41, the teacher routing score for: ； in: For weight generation function; For task context vectors; Let be the resource state vector at time t; Let be the coordinated state vector of the air-ground platform at time t; This represents the predicted business utility value of the task q to be executed before execution. For teacher confidence; For the uncertainty of teachers; For teachers in task types Historical reliability; The teacher availability mask for task q, teacher m, and knowledge type h; In S42, the teacher weight for: ； in: Let be the set of candidate teachers who can output knowledge of the h-th class; exp(·) is the exponential function; In S43, the aggregated teacher knowledge for: ； in: For teacher m, output for knowledge type h; In S44, the loss function of the two-level knowledge distillation link for: ； in: For hard label supervision loss; The distillation loss between teacher soft labels and student output; The loss is the alignment loss between teacher and student features; The strategy distillation loss is determined by task priority, processing level, and the preferences of the executing entity. Distillation loss for communication strategies that include upload granularity, relay requests, and resource reservation suggestions; This refers to the connection between the cloud and the edge. Indicates the connection from the edge to the device. The weights for the hard-labeled supervision loss; Weights for the distillation loss between teacher soft labels and student output; Weights for the alignment loss between teacher and student features; Weights for strategy distillation loss based on task priority, processing level, and executor preference; Weights for distillation loss in communication strategies that include upload granularity, relay requests, and resource reservation recommendations.

9. The air-to-ground mobile edge networking and resource collaborative scheduling method for photovoltaic operation and maintenance according to claim 1, characterized in that, S5 includes the following sub-steps: S51. Based on a two-level knowledge distillation link, construct an intelligent agent and build enhanced local observation for the intelligent agent; S52. Determine the agent's output action based on enhanced local observation of the agent; S53. Training is performed based on the output actions of the intelligent agent; In S51, the enhanced local observation of agent i for: ； in: This includes local location, power level, neighbors, candidate tasks, link status, and edge load; Representation of task knowledge after distillation; This is a prerequisite for task scheduling; Prior to communication resources; Use it as a mask for actionable actions; This is a vector consisting of the predicted business utility values ​​of the candidate tasks before execution. In S52, the output action of agent i for: ； in: Select an action for the task; Choose the implementing entity or the collaborating entity; For data upload mode; To handle hierarchical levels; For communication resource requests or communication control actions; For path adjustment actions; This action was triggered by on-site re-inspection.

10. The air-to-ground mobile edge networking and resource collaborative scheduling method for photovoltaic operation and maintenance according to claim 1, characterized in that, S7 includes the following sub-steps: S71. Based on the cloud review results and on-site inspection results, calculate the realized business utility value after execution, and calculate the business utility deviation based on the realized business utility value after execution; S72. Generate experience records based on business utility deviation, pre-execution business utility forecast, cloud review results, and on-site inspection results; S73. Calculate teacher reliability based on experience records; S74. Update teacher reliability according to task type and knowledge type; In S71, the realized value of business utility after execution for: ； in: As the basic weight of business; This is the information value adjustment coefficient; For the value of task information; It is a set of individual utilities; This represents the actual utility value of each component after implementation. These are the weighting coefficients for the utility of each component; In S71, the business utility deviation for: ； in: This represents the predicted business utility value of the task q to be executed before execution. In S72, experience records for: ； in, For task context vectors; Let be the resource state vector at time t; Let be the coordinated state vector of the air-ground platform at time t; The action output by the agent for task q at time t; For all intelligent agents at time Joint actions; This is the actual execution result vector; In S73, teacher reliability for: ； in: This is a truncation function; As an indicator variable; To ensure consistency between the diagnostic results and the on-site re-examination results; For task completion rate or service quality satisfaction; To improve business efficiency; For the cost of time delay; For the cost of energy consumption; The cost of upload volume; The cost of task conflicts, duplicate executions, or incorrect coordination; The weighting of the consistency between the diagnostic results and the on-site re-examination results; Weights for task completion rate or service quality satisfaction; Weighting for the degree of improvement in business effectiveness; The weight for latency cost; The weight of energy consumption cost; Weighting for the cost of upload volume; Weights for the costs of task conflicts, duplicate executions, or incorrect coordination; In S74, the teacher reliability is updated as follows: ； in: Update the step size for reliability; For the updated teacher reliability; The teacher reliability before the update.

Images