Distributed photovoltaic power station multi-machine cooperative intelligent inspection system

Abstract

The embodiment of the specification provides a distributed photovoltaic power station multi-machine cooperative intelligent inspection system, through the cooperation architecture of a cloud intelligent scheduling center, a distributed fixed airfield network and a drone cluster, the restriction of single machine endurance on the inspection radius is broken, a single system can cover tens to hundreds of megawatt level scattered photovoltaic power station groups, and wide range full-automatic unattended inspection is realized. The cloud scheduling center identifies the flyable window according to multi-source meteorological sensing data, automatically generates an inspection task and performs multi-machine load balancing distribution combined with a genetic algorithm, and then dynamically plans an optimal flight route by fusing meteorological prediction and no-fly zone information, so that the task execution efficiency and flight safety are significantly improved.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of unmanned aerial vehicle (UAV) inspection technology, and in particular to a multi-UAV collaborative intelligent inspection system for distributed photovoltaic power stations. Background Technology

[0002] Distributed photovoltaic power stations are widely distributed across various scenarios such as rooftops, farmhouses, and barren mountains, characterized by small individual capacity and dispersed spatial layout. Existing inspection methods mostly rely on manual operation of single drones or single-point automated airports. Due to the limited endurance of multi-rotor drones and the narrow coverage area of ​​a single flight, it is difficult to complete continuous inspections of all stations on a large geographical scale. Task scheduling mostly adopts manual dispatching or timed triggering, which cannot dynamically allocate inspection resources according to real-time weather changes and station operating status, resulting in delayed fault response and uneven resource utilization. Inspection data from each station is stored independently with inconsistent formats, lacking the conditions for centralized aggregation and overall analysis, making it difficult to support cross-station equipment status trend analysis and operation and maintenance decisions. In addition, the onboard data processing capacity is insufficient, and a large number of raw images need to be fully transmitted back for offline analysis, which not only consumes communication bandwidth but also prolongs the defect discovery and review cycle.

[0003] Therefore, a better solution is urgently needed. Summary of the Invention

[0004] In view of this, the embodiments of this specification provide a multi-machine collaborative intelligent inspection system for distributed photovoltaic power plants to solve the technical defects existing in the prior art.

[0005] According to a first aspect of the embodiments of this specification, a multi-machine collaborative intelligent inspection system for distributed photovoltaic power plants is provided, characterized in that it includes a cloud-based intelligent dispatch center, a distributed fixed airport network, and a drone swarm; The cloud-based intelligent dispatch center includes a station information management module, a meteorological sensing and forecasting module, a task generation and sorting module, a multi-machine task allocation module, a dynamic path planning module, and a data aggregation and analysis module. The site information management module is used to establish and maintain a spatial database containing information on the geographical location, capacity, operating status, and priority of each photovoltaic power station; The meteorological sensing and forecasting module is used to acquire multi-source meteorological data and predict flight windows for future periods; The task generation and sorting module is used to generate an inspection task queue based on the periodic inspection plan and real-time fault alarms, and to sort the tasks by priority. The multi-aircraft task allocation module is used to allocate tasks in the inspection task queue to each fixed airport node using a genetic algorithm; The dynamic path planning module is used to generate the optimal inspection route for each fixed airport node by combining meteorological forecast information and no-fly zone information; The distributed fixed airport network includes multiple fixed airport nodes, each of which includes a protective cabin, communication antenna, edge computing unit, automatic charging device, apron, and meteorological monitoring sensors; The drone swarm consists of multiple multi-rotor drones and / or vertical take-off and landing fixed-wing drones. The drones are equipped with visible light cameras, infrared thermal imagers and onboard computing units. The onboard computing units run a lightweight artificial intelligence defect recognition model to perform real-time defect recognition on the images collected during flight. The edge computing unit is used to preprocess the identified abnormal images. The preprocessed abnormal images and associated metadata are then uploaded to the data aggregation and analysis module via the communication antenna. The data aggregation and analysis module runs a deep learning verification model to perform defect verification and fine classification on the received abnormal images and generate inspection reports.

[0006] In one possible implementation, when the multi-aircraft task allocation module uses a genetic algorithm to allocate tasks, the fitness function comprehensively considers the total inspection completion time, the task load balance of each fixed airport node, the total distance of UAV transfer, and the time window satisfaction rate. The constraints include that the difference in the total number of tasks allocated to each node does not exceed a preset proportion threshold, and fault-type inspection tasks are completed within a preset response time limit.

[0007] In one possible implementation, the dynamic path planning module generates inspection routes based on the Traveling Salesman Problem algorithm, using the fixed airport node's own location as the start and return points, and the assigned photovoltaic power station locations to be inspected as the necessary path points. During the path solving process, the dynamic path planning module integrates future weather forecast information provided by the weather perception and prediction module. When weather conditions exceeding the safe flight threshold of the UAV are predicted in a certain area, the module dynamically adjusts the access sequence of the power stations in that area or postpones the corresponding tasks to subsequent batches. At the same time, it acquires the boundary coordinate data of permanent and temporary no-fly zones, so that the planned route automatically avoids the no-fly zone boundaries.

[0008] In one possible implementation, the airborne computing unit runs a lightweight artificial intelligence defect recognition model to identify hot spot defects, crack defects, and shading defects of photovoltaic modules in real time, and marks the identified abnormal images with defect type, defect location coordinates, recognition confidence level, and shooting timestamp as associated metadata.

[0009] In one possible implementation, the preprocessing operations performed by the edge computing unit include image denoising, image enhancement, and secondary verification filtering of anomalous images to reduce the false alarm rate.

[0010] In one possible implementation, the protective nest of the fixed airport node is divided into multiple functional layers from top to bottom: the top layer is the communication antenna layer, which is equipped with 4G / 5G antennas, WiFi antennas and satellite navigation antennas; the second layer is the edge computing layer, which is equipped with edge computing units; the third layer is the automatic charging layer, which is equipped with battery compartments, robotic arms and charging interfaces; and the bottom layer is the apron layer, which is equipped with visual positioning markers, wireless charging contacts and a liftable platform.

[0011] In one possible implementation, the deep learning verification model running in the data aggregation and analysis module adopts a deep residual network architecture to verify and classify defects in abnormal images uploaded by the edge computing unit. The fine classification subdivides hot spot defects into local hot spots, bypass diode fault hot spots, and junction box abnormal hot spots, and subdivides crack defects into through cracks, tree cracks, and edge cracks. The data aggregation and analysis module also performs defect development trend analysis based on historical inspection data of the same site and issues early warnings for defects with accelerated deterioration trends.

[0012] In one possible implementation, the site information management module is also used to dynamically update the real-time operating data of each site. The real-time operating data includes the current power generation, inverter operating status, and string current abnormality alarms. When generating inspection tasks, the task generation and sorting module obtains real-time operating data to help determine the timing of the generation of periodic inspection tasks and the triggering conditions for fault alarms.

[0013] In one possible implementation, vertical take-off and landing fixed-wing UAVs in the UAV swarm are used to perform large-scale rapid inspection tasks in areas with sparsely distributed airfields and large coverage areas, while multi-rotor UAVs are used to perform detailed inspection tasks in areas with high airfield density.

[0014] In one possible implementation, meteorological monitoring sensors are deployed outside fixed airport nodes to continuously collect real-time ground meteorological data at their location. The collected data includes wind speed, wind direction, temperature, humidity, solar irradiance, and rainfall. The meteorological perception and prediction module integrates real-time ground meteorological data from each fixed airport node, upper-air meteorological forecast data from public meteorological service platforms, and forecast data from meteorological radar and satellites to generate meteorological forecasts for future periods and identify flight windows based on the meteorological forecasts.

[0015] This system, through a collaborative architecture of a cloud-based intelligent dispatch center, a distributed fixed airport network, and a drone swarm, breaks the limitation of single-drone endurance on inspection radius. This allows a single system to cover distributed photovoltaic power station clusters ranging from tens to hundreds of megawatts, achieving large-scale, fully automated, unattended inspections. The cloud-based dispatch center identifies flight windows based on multi-source meteorological sensing data, automatically generates inspection tasks, and uses a genetic algorithm for multi-drone load balancing. It then dynamically plans optimal routes by integrating weather forecasts and no-fly zone information, significantly improving task execution efficiency and flight safety. An airborne lightweight AI model identifies component defects in real time during flight, only transmitting abnormal images and associated metadata. Edge computing units preprocess and perform secondary verification to effectively reduce false alarms, and then a cloud-based deep learning model performs high-precision verification and fine classification. This forms a two-level data processing mechanism of real-time edge screening and cloud-based deep analysis, saving communication bandwidth while ensuring the accuracy of defect identification. Site information, inspection tasks, and defect data are uniformly aggregated in the cloud to generate inspection reports, providing integrated support for the full lifecycle management of distributed photovoltaic assets. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a multi-machine collaborative intelligent inspection system for a distributed photovoltaic power station, provided in one embodiment of this specification. Detailed Implementation

[0017] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0018] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0019] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0020] This specification provides a multi-machine collaborative intelligent inspection system for distributed photovoltaic power plants, which will be described in detail in the following embodiments.

[0021] See Figure 1 , Figure 1 This diagram illustrates a system schematic of a distributed photovoltaic power station multi-drone collaborative intelligent inspection system according to an embodiment of this specification, specifically including a cloud-based intelligent dispatch center, a distributed fixed airport network, and a drone swarm. The cloud-based intelligent dispatch center includes a site information management module, a meteorological sensing and prediction module, a task generation and sorting module, a multi-drone task allocation module, a dynamic path planning module, and a data aggregation and analysis module. The site information management module is used to establish and maintain a spatial database containing information on the geographical location, capacity, operating status, and priority of each photovoltaic power station. The meteorological sensing and prediction module is used to acquire multi-source meteorological data and predict flight windows for future periods. The task generation and sorting module is used to generate an inspection task queue based on regular inspection plans and real-time fault alarms and to prioritize them. The multi-drone task allocation module is used to allocate tasks in the inspection task queue to each fixed airport node using a genetic algorithm. The dynamic path planning module is used to generate optimal inspection routes for each fixed airport node by combining meteorological forecast information and no-fly zone information. The distributed fixed airport network includes multiple fixed airport nodes, each of which includes a protective aircraft nest, a communication antenna, an edge computing unit, an automatic charging device, a parking apron, and meteorological monitoring sensors. The drone swarm comprises multiple multi-rotor drones and / or vertical takeoff and landing fixed-wing drones. Each drone carries a visible light camera, an infrared thermal imager, and an onboard computing unit. The onboard computing unit runs a lightweight AI defect identification model for real-time defect identification of acquired images during flight. An edge computing unit preprocesses the identified anomalous images. The preprocessed anomalous images and associated metadata are then uploaded to a data aggregation and analysis module via a communication antenna. The data aggregation and analysis module runs a deep learning verification model to perform defect verification and fine-grained classification on the received anomalous images and generate an inspection report.

[0022] Among them, the cloud-based intelligent dispatch center can refer to the central control and data processing system deployed on the cloud server, which is responsible for the overall scheduling, resource management and data analysis of the entire inspection system. For example, it can establish a secure communication link with each fixed airport node through the Internet, receive the status information and inspection data of each node, and issue task instructions and route data.

[0023] The site information management module can refer to a functional unit used to centrally manage the basic information and real-time operation data of all photovoltaic power stations to be inspected. For example, this module establishes a data interface with the data acquisition device or monitoring system of each power station, automatically synchronizes the geographical coordinates, installed capacity, component type, commissioning time, historical fault records, current power generation, inverter status and string current data of the power station, and associates this information with the geographic information system to form a visualized site information model.

[0024] The meteorological perception and prediction module can refer to a functional unit used to integrate multi-source meteorological observation and forecast data and identify the flight conditions of UAVs. For example, this module obtains real-time ground meteorological data by accessing meteorological monitoring sensors deployed at various fixed airport nodes, and at the same time accesses forecast data from public meteorological service platforms. It comprehensively generates the spatiotemporal distribution prediction results of meteorological elements such as wind speed, solar irradiance, and rainfall for the next two hours, and identifies the flight windows that meet the flight safety thresholds based on this.

[0025] The task generation and sorting module refers to the functional unit responsible for automatically generating inspection tasks and sorting them according to their urgency and importance. For example, this module maintains a regular inspection plan table. When the preset inspection cycle is reached, it automatically generates regular inspection tasks. At the same time, it monitors the fault alarm information stream of each power station in real time. Once a fault alarm is received, it immediately generates a fault inspection task and prioritizes all tasks to be executed by comprehensively considering factors such as task type, fault level, and station capacity, forming an orderly inspection task queue.

[0026] The multi-aircraft task allocation module can refer to a functional unit that uses optimization algorithms to reasonably allocate inspection tasks to each fixed airport node. For example, this module obtains the number of available drones, power status and maintenance status of each fixed airport node, and uses a genetic algorithm as the core solver to search for a task allocation scheme that minimizes the total inspection completion time and the transfer distance, under the constraints of load balancing of tasks at each node and fault response time limit.

[0027] The dynamic path planning module can be defined as a functional unit that generates the optimal flight route for each UAV from takeoff to return. For example, after receiving the task allocation results, this module plans the order of visiting each power station based on the Traveling Salesman Problem algorithm for the task set undertaken by each fixed airport node. At the same time, it integrates the future weather information provided by the weather forecast module to dynamically adjust the visiting sequence to avoid areas with severe weather, and obtains airspace management information to make the route automatically bypass no-fly zones. Finally, it generates the optimal route that includes waypoint coordinates and estimated power consumption.

[0028] Fixed airport nodes can refer to unmanned drone storage and support sites deployed within the geographical area of ​​a photovoltaic power station cluster. For example, a closed hangar facility made of steel or composite materials with all-weather protection capabilities, which integrates charging equipment, computing equipment and communication equipment, can automatically complete the storage, charging, mission data transmission and reception and take-off of drones. It is a key intermediate node connecting the cloud dispatch center and the inspection execution layer.

[0029] Protective drone nests can refer to enclosed cabin structures that provide parking, charging, and safety protection for drones. For example, they can be box-type structures built with weathering steel and heat insulation materials, with automatic opening and closing top covers or side doors, internal temperature and humidity control devices to maintain a suitable operating environment for the equipment, and rainproof, dustproof, and anti-theft functions.

[0030] Edge computing units can refer to local computing devices deployed within fixed airport nodes, such as an embedded artificial intelligence computing board equipped with a graphics processor or neural network processor. It can run lightweight image processing algorithms and defect recognition models locally to quickly preprocess and initially screen image data unloaded by drones after their return, thereby reducing the amount of data transmitted to the cloud and lowering response latency.

[0031] Automatic charging devices can refer to mechatronic devices that can automatically replenish the power batteries of drones. For example, a subsystem consisting of a battery compartment array, a multi-axis robotic arm, and a charging management circuit can be used. The robotic arm can remove the depleted battery from the drone and put it into the battery compartment for charging, while simultaneously removing the fully charged battery from the battery compartment and putting it into the drone, thus realizing automatic battery replacement and charging cycle management.

[0032] The data aggregation and analysis module can refer to a functional unit deployed in the cloud that centrally stores and deeply analyzes all site inspection data. For example, this module maintains a cloud database to uniformly store abnormal images and metadata uploaded by each node, and runs a verification model based on deep neural networks to perform secondary confirmation and more granular classification of defects. Finally, it automatically generates an inspection report containing a defect list, location markings, and maintenance suggestions based on the analysis results.

[0033] The present invention will be further described below through a detailed embodiment: In a county-wide distributed photovoltaic power station cluster inspection application, routine periodic inspections are required for 120 rooftop photovoltaic power stations with a total installed capacity of 80 megawatts, distributed over an area of ​​approximately 50 square kilometers.

[0034] This distributed photovoltaic power station multi-machine collaborative intelligent inspection system has deployed a complete inspection network within the county. The cloud-based intelligent dispatch center is deployed on a virtual server in the municipal cloud computing center, communicating with each fixed airport node via dedicated fiber optic lines. The distributed fixed airport network comprises ten fixed airport nodes. The location of each node was determined after analyzing the power station distribution density and terrain conditions. Each node has an effective coverage radius of approximately seven kilometers, covering a surrounding power station cluster of five to ten megawatts. Each fixed airport node is equipped with one multi-rotor drone, and all drones together form a drone swarm.

[0035] During the system power-on initialization phase, the site information management module establishes a data interface with the existing operation monitoring systems of each power station to automatically acquire the geographical coordinates, roof orientation and tilt angle, module string layout, installed capacity, commissioning date, and historical fault and maintenance records of all 120 power stations. This information is then integrated to establish a spatial database. Simultaneously, the site information management module continuously synchronizes real-time operating data from each power station, including string current values ​​and inverter power output values, to dynamically update the current operating status of each station.

[0036] The meteorological perception and prediction module initiates predictive analysis for the daily inspection window at midnight. This module acquires real-time data on wind speed, temperature, humidity, solar irradiance, and rainfall from meteorological monitoring sensors deployed atop each fixed airport node. Simultaneously, it obtains hourly gridded weather forecast data from the local meteorological bureau via an internet interface. After integrating this multi-source meteorological data, the module generates predicted wind speed and rainfall distribution maps for each time period and sub-region of the day. By comparing the predicted values ​​with safe flight thresholds for drones, it identifies flyable periods within the coverage area of ​​each fixed airport node that meet flight conditions.

[0037] The task generation and sorting module generates tasks based on a preset inspection schedule. This schedule sets differentiated inspection cycles according to the capacity and historical fault frequency of each power station. When the time interval since the last inspection of a power station reaches its preset cycle, the module automatically generates a scheduled inspection task. Simultaneously, the task generation and sorting module continuously monitors the power station's operational data stream. When an abnormal drop in string current is detected at a power station, triggering an inverter alarm signal, the module immediately generates a fault inspection task and assigns it the highest priority. For all pending inspection tasks, the module comprehensively compares the task type, fault severity level, power station capacity, and remaining days until the next planned inspection to calculate a comprehensive priority score for each task, and arranges them in descending order of score to form the daily inspection task queue.

[0038] After receiving the inspection task queue, the multi-drone task allocation module initiates the genetic algorithm solution process. First, it acquires the current status data of each fixed airport node, confirming that each node's drones are in standby mode and their batteries are fully charged. Then, it randomly generates an initial population containing several individuals, each representing a task allocation scheme. During the fitness calculation phase, the module calculates the maximum time required for each node to complete its task under the given allocation scheme, based on the distance matrix from each node to its assigned power station and the estimated inspection time per power station. It also calculates the standard deviation of the task allocation for each node to assess load balancing and the total flight distance of all drones between different power stations. The fitness value is a weighted average of the above factors. For fault inspection tasks, the module checks whether the scheduled execution time is within the fault response time limit; if it exceeds the time limit, the fitness is penalized. After iterative evolution through selection, crossover, and mutation, the algorithm converges and outputs the final task allocation scheme.

[0039] The dynamic path planning module receives a list of assigned power stations for each fixed airport node. Using the fixed airport node as the start and end point, and the power station to be inspected as intermediate access points, it constructs an initial path using a successive insertion method. Then, it improves the path using a two-element exchange optimization strategy to find the access order that minimizes the total flight distance. When generating the flight path, the dynamic path planning module retrieves the weather forecast distribution map generated by the weather perception and prediction module. If a power station in the path is found to be in a predicted wind speed exceeding the limit during the planned access period, the access order is adjusted, postponing the access to that power station until the predicted wind speed decreases. Simultaneously, the module retrieves no-fly zone data issued by the local airspace management department. If a section of the path crosses the no-fly zone boundary of a small airport, the module inserts a transition waypoint outside the no-fly zone edge, causing the flight path to detour along the no-fly zone boundary.

[0040] After receiving the flight path data, the drone automatically launches from its protective nest at the corresponding fixed airport node. The lift platform on the helipad rises, and the drone initiates its takeoff. Following the flight path, the drone arrives at the first target power station, descends to an inspection hovering state, and its onboard visible light camera scans and photographs the rooftop photovoltaic array with a preset overlap rate, while an infrared thermal imager simultaneously acquires thermal radiation images. The onboard computing unit instantly runs a lightweight AI defect identification model on each captured image frame. When the model determines that a frame contains hotspot features and the identification confidence level exceeds a preset threshold, the image is marked as an abnormal image, and the defect type (hotspot), the pixel coordinates of the defect in the image, the identification confidence level, and the capture timestamp are recorded. This information is stored as associated metadata bound to the image. Images judged as normal by the model are temporarily stored directly in the onboard storage space.

[0041] After completing its inspection of all distribution power stations, the drone automatically returns and lands on the helipad at the designated airport node. The edge computing unit reads all abnormal images and associated metadata from the drone's storage device via a high-speed data interface. The edge computing unit performs noise reduction and contrast enhancement preprocessing on each abnormal image and runs a verification model with a slightly larger parameter scale to perform secondary filtering on the initial identification results, eliminating obviously false alarm images. The retained confirmed abnormal images, along with their metadata, are compressed and packaged. The communication antenna uploads the packaged data to the data aggregation and analysis module of the cloud-based intelligent dispatch center via a 4G / 5G network.

[0042] The data aggregation and analysis module receives data packets from all ten fixed airport nodes, and after unpacking, obtains all confirmed anomaly images for this round of inspections. These images are then input into a deep learning verification model deployed on a graphics processing unit cluster. The model first performs secondary verification of defects on each image, eliminating false alarms remaining from edge computing unit verification. For images with confirmed defects, the model further refines their classification, identifying the specific cause type of hotspots and outputting the precise location bounding box of the defect. Simultaneously, the data aggregation and analysis module retrieves past inspection records for the power station from the database, performs trend comparisons on defects recurring at the same location, and generates warning markers for defects with accelerated deterioration rates. Finally, the module synthesizes the above analysis results and automatically generates a summary report for this round of inspections. The report includes a list of power stations covered by the inspection, a detailed list of all discovered defects, a location map of each defect on the component, a defect development trend analysis curve, and maintenance priority recommendations for each defect.

[0043] The beneficial effects of this embodiment in this specification include at least the following: By constructing a three-tier architecture system consisting of a cloud-based intelligent dispatch center, a distributed fixed airport network, and a drone swarm, fully automated and comprehensive inspection of dispersed photovoltaic power station clusters is achieved. Its advantages lie in the combination of centralized cloud-based dispatch and autonomous execution by distributed nodes, overcoming the limitation of single drone endurance on inspection range and enabling collaborative inspection of multiple stations over a large area; the full-process automation from automatic task generation and genetic algorithm optimization allocation to dynamic route planning eliminates manual dispatching and control; the edge computing unit performs image pre-screening locally, only transmitting abnormal data back to the cloud, significantly saving communication bandwidth and reducing cloud processing load; the cloud-based deep learning verification model and the airborne lightweight model form a two-level defect identification system, balancing real-time performance and identification accuracy. The system as a whole achieves unattended closed-loop operation, significantly improving inspection efficiency and fault response speed, and reducing maintenance labor costs.

[0044] According to the aforementioned system, when the multi-aircraft task allocation module uses a genetic algorithm to allocate tasks, the fitness function comprehensively considers the total inspection completion time, the task load balance of each fixed airport node, the total distance of UAV transfer, and the time window satisfaction rate. The constraints include that the difference in the total number of tasks allocated to each node does not exceed a preset proportion threshold, and fault-type inspection tasks are completed within a preset response time limit.

[0045] The total inspection completion time can refer to the length of time from the first drone taking off to the last drone completing the inspection and landing. For example, when multiple drones perform inspection tasks in parallel, each fixed airport node independently executes its assigned task set. The node with the longest time determines the total completion time of the entire inspection batch. This indicator reflects the overall execution efficiency of the task allocation scheme.

[0046] Node task load balancing can be a quantitative indicator that measures the degree of difference in task allocation among fixed airport nodes. For example, it can be represented by the number of power stations to be inspected or the dispersion of the estimated total inspection time. High load balancing means that the task load of each node is similar, avoiding the problem of uneven resource utilization where some nodes are overloaded while others are idle.

[0047] The total transfer distance of drones can refer to the sum of the transfer flight distances of all drones during the inspection mission from the current power station to the next power station. For example, when a drone visits multiple power stations located in different locations in sequence, the spatial interval between the power stations constitutes the transfer distance. This indicator reflects the rationality of the spatial layout of the task allocation scheme.

[0048] The time window satisfaction rate refers to the proportion of inspection tasks scheduled to be performed within the weather-permitted flight window to the total number of tasks. For example, the inspection tasks of each power station need to be completed within the weather-permitted flight window of the corresponding area. The higher the time window satisfaction rate, the better the task allocation scheme complies with weather constraints.

[0049] The preset ratio threshold can refer to the upper limit of the difference in the amount of tasks allocated to each node. For example, it can be stipulated that the difference between the maximum and minimum values ​​of the number of power stations to be inspected or the estimated total working hours allocated to each fixed airport node must not exceed the threshold. Its function is to eliminate those allocation schemes that cause serious imbalances in the workload between nodes during the algorithm evolution process.

[0050] The preset response time limit can refer to the maximum time allowed for a fault-type inspection task from the triggering of an alarm to the arrival of the drone at the target power station to start the inspection. For example, a shorter response time limit is set for high-priority fault tasks. The algorithm uses this time limit as a hard constraint to check and ensure that the fault task is not delayed due to being assigned to a node that is too far away or the task queue is too crowded.

[0051] Continuing with the aforementioned example of inspecting county-level distributed photovoltaic power station clusters, the specific processing procedure when the genetic algorithm is activated in the multi-machine task allocation module for task allocation is as follows.

[0052] The module first generates an initial population. The population size is set to a certain number of individuals. Each individual uses an integer encoding to represent a complete task allocation scheme. The encoding length is equal to the total number of inspection tasks to be assigned. Each position in the encoding corresponds to an inspection task, and the integer value at that position represents the target fixed airport node number to which the task is assigned. For example, if there are currently twenty inspection tasks and ten fixed airport nodes, then each individual is an integer sequence of length twenty, where each element takes the value between one and ten. Individuals in the initial population are generated randomly, while some individuals generated by heuristic rules are added randomly to ensure the diversity of the initial population and to include some genes for better solutions.

[0053] During the fitness calculation phase, the module decodes the corresponding task allocation scheme for each individual in the population. The module obtains the spatial coordinates of each power station from the site information management module and the position coordinates of each fixed airport node, calculating the distance matrix. For each node, based on its assigned task list and distance matrix, the module uses the nearest neighbor heuristic to estimate the total inspection time required for the node to complete all assigned tasks, including hovering and imaging time over each power station and transfer flight time between power stations. The maximum total inspection time among all nodes is taken as the total inspection completion time indicator.

[0054] The module calculates the number of tasks assigned to each fixed airport node and uses the standard deviation of the task count as an indicator of node task load balancing. The smaller the standard deviation, the more balanced the task distribution.

[0055] The module calculates the flight distance of all transfer segments on the estimated flight path of each node and sums them to obtain the total transfer distance of the UAV.

[0056] The module retrieves flight window data for each region provided by the meteorological sensing and prediction module for the current day. It checks whether the arrival time of each inspection task at its corresponding node, following the estimated path, falls within the flight window of the target area. The proportion of tasks satisfying the window constraints out of the total number of tasks is calculated to obtain the time window satisfaction rate.

[0057] The fitness function integrates the above four indicators. The module converts the reciprocals of the three indicators—total inspection completion time, load balancing, and total relocation distance—into positive indicators, and then performs a weighted sum with the time window satisfaction rate. The weight coefficients of each indicator are set according to actual application requirements to form the final fitness value of each individual indicator.

[0058] During the constraint check phase, the module calculates the difference between the maximum and minimum number of tasks for each node and determines whether it exceeds a preset proportional threshold. If it does, a penalty value is assigned to that individual, significantly reducing its fitness value. Simultaneously, the module checks the estimated execution time of all fault inspection tasks in the allocation scheme to determine whether the time interval from the current moment to the estimated arrival at the target power station is within the preset response time limit. If any fault task times out, a fitness penalty is also applied to that individual.

[0059] The module then performs a selection operation. A tournament selection strategy is adopted, in which several individuals are randomly selected from the population to form a tournament group. The individual with the highest fitness value in the group is selected to enter the next generation of the population. This process is repeated until the size of the next generation of the population reaches the preset value.

[0060] In the crossover operation, the module pairs individuals selected for the next generation population and performs partial mapping crossover with a preset crossover probability. Two crossover points are randomly selected on the coding sequences of the two parent individuals, and the gene segments between the crossover points are exchanged. Then, the duplicate genes caused by the exchange are repaired through the mapping relationship to generate two new offspring individuals.

[0061] During the mutation operation, the module perturbs the coding sequence of each individual with a preset low probability. A position in the individual's coding sequence is randomly selected, and the node number at that position is randomly changed to another different valid node number to maintain the genetic diversity of the population and avoid getting trapped in local optima.

[0062] After the above iterative cycle of selection, crossover, and mutation, when the preset maximum number of iterations is reached or the fitness value of the best individual no longer shows significant improvement after several consecutive generations, the module terminates the evolutionary process, outputs the task allocation scheme represented by the individual with the highest fitness value in the population as the final result, and sends it to the dynamic path planning module for further processing.

[0063] The beneficial effects of this embodiment in this specification include at least the following: by employing a fitness function in the multi-machine task allocation module that comprehensively considers the total inspection completion time, load balancing, total transfer distance, and time window satisfaction rate, and by applying constraints such as node task volume differences and fault response time limits, the genetic algorithm can search for a task allocation scheme that achieves the optimal balance among multiple objectives and constraints. Its beneficial effect lies in avoiding the load imbalance phenomenon that may occur in traditional manual task dispatching, where some nodes experience task accumulation while others remain idle. Simultaneously, it ensures that emergency fault-related tasks receive priority response and are handled within the specified time limit, significantly improving the overall execution efficiency and resource utilization of multi-machine collaborative inspections.

[0064] According to the aforementioned system, the dynamic path planning module generates inspection routes based on the Traveling Salesman Problem algorithm. It uses the fixed airport node's own location as the start and return points, and the assigned photovoltaic power station locations to be inspected as mandatory path points. During the path solving process, the dynamic path planning module integrates future weather forecast information provided by the weather perception and prediction module. When weather conditions exceeding the safe flight threshold for drones are predicted in a certain area, the module dynamically adjusts the access sequence of power stations in that area or postpones the corresponding tasks to later batches. Simultaneously, it acquires the boundary coordinate data of permanent and temporary no-fly zones, enabling the planned route to automatically bypass the no-fly zone boundaries.

[0065] Among them, the Traveling Salesman Problem (TSP) solution algorithm can refer to the optimization algorithm used to solve the problem of finding the shortest loop from the starting point to the starting point after visiting all points exactly once, given a series of points and the distances between them. Examples include the nearest neighbor algorithm, the insertion algorithm, the two-element swap optimization algorithm, or the dynamic programming algorithm. In this module, it is used to determine the optimal order in which the UAV starts from the fixed airport node and visits each target power station in sequence.

[0066] The safe flight threshold for drones refers to the maximum weather conditions that a drone can withstand under safe operating conditions. This can be categorized as a flight envelope consisting of parameters such as maximum permissible wind speed, maximum permissible rainfall intensity, and minimum permissible light conditions. When predicted weather conditions exceed these thresholds, the drone risks losing control, crashing, or being unable to complete its inspection mission.

[0067] The visit sequence can refer to the time arrangement in which the drone visits each target power station in sequence. For example, in the path planning results, each power station is assigned an estimated arrival time and departure time. The visit sequence determines when the drone appears in a specific spatial location at a specific time.

[0068] Continuing with the aforementioned example of inspecting distributed photovoltaic power station clusters in a county, the dynamic path planning module, after receiving the task allocation results for each fixed airport node from the multi-aircraft task allocation module, plans the route for each node one by one. Taking one fixed airport node as an example, this node is assigned inspection tasks for eight rooftop photovoltaic power stations located in different locations.

[0069] The module first determines that the starting and ending points of the path planning are the coordinates of the fixed airport node, and the coordinates of the eight power stations to be inspected are the intermediate access points that must be passed through, forming a typical problem of finding the optimal inspection order. The module uses the insertion method to construct the initial path: starting from the starting point, the distance to each power station is calculated, and the power station with the closest distance is selected as the first access point; then, the minimum path increment after inserting each position of the current path into the remaining power stations is calculated, and all power stations are inserted into the path in turn to form the initial inspection order.

[0070] After obtaining the initial access order, the module uses a two-element swap optimization strategy to iteratively improve the order: it attempts to swap the access order of any two power stations in the path in turn, and calculates the total length of the path after the swap; if the swap shortens the total length, the swap is retained; otherwise, the original order is restored. This process is repeated until swapping any two elements no longer shortens the path, at which point the optimal access order of the static path planning is obtained.

[0071] After completing static path planning, the module introduces meteorological constraints for dynamic adjustment. The module retrieves time-segmented meteorological forecast grid data generated by the meteorological perception and prediction module for the coverage area of ​​the fixed airport node. Following the planned optimal access sequence, the module calculates the arrival and departure times of the UAV at each power station based on the estimated flight time of each flight segment and the estimated inspection time of each power station. The access time of each power station is compared with the predicted meteorological data for its corresponding spatial location. When it is found that the predicted wind speed in the area where the fourth power station is located exceeds the safe flight threshold for the UAV during its estimated access time, the module initiates an adjustment strategy. The module swaps the fourth power station with the sixth power station (where meteorological conditions are favorable in subsequent access times) in the path sequence, recalculates each waypoint and arrival time, and re-verifies the meteorological conditions. After several trial adjustments, an access sequence that satisfies all meteorological constraints is formed.

[0072] For power plant tasks that cannot avoid severe weather even after adjustments, the module removes the power plant inspection task from the current batch, marks it as a delayed task, and automatically resubmits the task to the task queue of the task generation and sorting module, so that it can be scheduled for execution in a later, more suitable weather window batch.

[0073] Simultaneously, the module acquires airspace management data. This data includes a list of permanent no-fly zone boundary coordinates covering the inspection area and the coordinates of temporary no-fly zone boundaries effective for the day. The module performs spatial verification on the planned flight routes, checking whether the connecting lines of each flight segment intersect with the no-fly zone polygon. When a flight segment is detected to have crossed the boundary of a small airport's no-fly zone, the module inserts transition waypoints at two corner points outside the no-fly zone boundary, replacing the straight-line flight segment with a path that detours along the outer edge of the no-fly zone boundary, ensuring that the entire flight route remains outside the no-fly zone.

[0074] Finally, the dynamic path planning module outputs complete inspection route data for the fixed airport node, including the three-dimensional coordinate sequence of take-off and landing waypoints, power station access waypoints, and no-fly zone detour transition waypoints, as well as the estimated flight time and estimated power consumption between adjacent waypoints. This route data is packaged and sent to the communication antenna of the corresponding fixed airport node via the cloud, and then forwarded to the flight control system of the standby UAV.

[0075] The beneficial effects of this embodiment in this specification include at least the following: By deeply integrating the Traveling Salesman Problem solution with real-time weather forecast information and no-fly zone spatial constraints in the dynamic path planning module, the generated inspection routes can proactively avoid areas with severe weather risks and no-fly zones while pursuing the shortest flight distance. Its beneficial effects are that the strategy of dynamically adjusting the access sequence enables the system to complete inspection tasks to the greatest extent possible under variable weather conditions, avoiding the cancellation of entire batches of tasks due to local weather exceeding standards; the ability of routes to automatically bypass no-fly zones ensures the legality and compliance of flight operations and airspace safety. The combination of these two factors significantly improves the success rate and operational reliability of the system in complex environments and restricted airspace conditions.

[0076] According to the aforementioned system, the lightweight artificial intelligence defect recognition model running on the airborne computing unit is used to identify hot spot defects, crack defects and shading defects of photovoltaic modules in real time, and to mark the defect type, defect location coordinates, recognition confidence level and shooting timestamp of the identified abnormal images as associated metadata.

[0077] Among them, lightweight artificial intelligence defect recognition models can refer to artificial intelligence inference models that have been compressed and optimized and can run in real time on embedded platforms with limited computing resources. For example, they can adopt convolutional neural network structures with depth separable convolution or channel pruning. Their model parameters and computational load are much smaller than those of general cloud models, and they can complete image inference at a speed of tens of frames per second on the onboard computing unit of UAVs.

[0078] Defect types can refer to classification labels for abnormal states of photovoltaic modules. For example, abnormal local temperature rise on the module surface can be classified as hot spot defects, through-hole or tree-like cracks in the module glass or cells can be classified as crack defects, and severe dust accumulation on the module surface due to foreign objects covering it can be classified as shading defects.

[0079] Defect location coordinates can refer to the spatial positioning information of the defect in the acquired original image. For example, the center point coordinates and width and height dimensions of the defect area bounding box represented by horizontal and vertical pixel coordinates with the upper left corner of the image as the origin. This is used to mark the specific location of the defect on the component in the subsequent inspection report.

[0080] Recognition confidence can refer to the quantitative value of how confident a lightweight model is in its recognition output. For example, the probability value output by the model after inferring about an image. The higher the value, the more reliable the model believes the recognition result is. It is an important basis for subsequent screening of abnormal images and setting alarm thresholds.

[0081] Associated metadata can refer to non-image-based structured descriptive information bound to each abnormal image, such as defect type string, defect location coordinates, confidence level, and shooting timestamp. This data is stored in a structured format, which facilitates subsequent database retrieval, statistical analysis, and report generation.

[0082] Continuing with the aforementioned example of inspecting county-level distributed photovoltaic power station clusters, after the drone flies to the target photovoltaic power station along the planned route and enters hovering inspection mode, its onboard visible light camera scans the photovoltaic module array line by line with preset forward overlap and lateral overlap rates, while the infrared thermal imager simultaneously acquires thermal radiation images of the corresponding area. The image data stream is continuously input into the onboard computing unit at a high frame rate.

[0083] The lightweight AI defect recognition model deployed on the airborne computing unit had completed offline training and quantization compression before system deployment. The training data came from tens of thousands of labeled photovoltaic module images accumulated from historical inspections, covering samples of normal modules and various defective modules. The model structure was built based on depthwise separable convolution, and the number of channels was pruned to maintain high recognition accuracy while compressing the number of parameters to a size that could be fully loaded into the memory of the airborne computing unit, with single-frame inference time kept in the millisecond range.

[0084] During the inspection process, the visible light images first undergo basic preprocessing steps, including scaling to the resolution required by the model input and normalizing pixel values. The preprocessed images are then fed into the model frame by frame for forward inference. The model's output layer contains multiple prediction branches, which are used to determine whether the image contains defective targets, the category of the defective targets, and the bounding box coordinates of the defective targets in the image.

[0085] When the model detects that the current frame contains a defect target and the maximum output category probability exceeds the corresponding preset threshold, the system classifies the frame as an abnormal image. The model extracts the defect type label from the output, such as the identified category being a hot spot; extracts the center pixel coordinates and width and height values ​​of the defect bounding box as the defect location coordinates; uses the category probability value output by the model as the recognition confidence; and reads the precise shooting time of the current frame from the camera's exposure timestamp register as the shooting timestamp.

[0086] The four data items—defect type, defect location coordinates, identification confidence level, and capture timestamp—are organized into a structured byte stream according to a preset key-value pair format, serving as the associated metadata for the abnormal image. A one-to-one mapping relationship is established between the associated metadata and the corresponding compressed image data through a data structure in memory, and then written to the task data directory of the onboard storage device. For normal frames where the model does not detect any defective targets, no metadata record is generated; the image data is written to another temporary storage partition, awaiting the completion of the inspection process and determining whether to retain or automatically overwrite it based on storage space occupancy.

[0087] Through this process, in a typical single inspection flight, the drone collects tens of thousands of images. The lightweight model filters out a small number of abnormal frames and generates complete associated metadata, providing accurate input data for subsequent secondary filtering by edge computing units and fine-grained verification by cloud-based deep learning models.

[0088] The beneficial effects of this embodiment in this specification include at least the following: By deploying a lightweight artificial intelligence defect recognition model on the airborne computing unit, real-time on-board defect recognition is achieved. This benefits because defect detection is moved from the traditional offline post-processing stage to the flight inspection process, enabling the UAV to complete tagging and metadata generation immediately upon acquiring abnormal images. This avoids the blind storage and backhaul of all image data, laying the foundation for subsequent bandwidth optimization strategies that only transmit abnormal data. Simultaneously, the structured associated metadata records the defect type, location, confidence level, and time information, providing accurate data support for defect tracing, location, and full lifecycle management.

[0089] According to the aforementioned system, the preprocessing operations performed by the edge computing unit include image denoising, image enhancement, and secondary verification filtering of abnormal images to reduce the false alarm rate.

[0090] Image denoising refers to the operation of removing or suppressing noise components in an image using digital image processing algorithms. For example, Gaussian filtering, median filtering, or nonlocal mean denoising algorithms can be used to eliminate random noise points introduced by sensor dark current, signal transmission interference, or low lighting conditions, thereby improving the visual quality of the image and the accuracy of subsequent analysis.

[0091] Image enhancement refers to operations that improve the visual effect of an image by adjusting its attributes such as contrast, brightness, and sharpness. For example, an adaptive histogram equalization algorithm can be used to enhance the visibility of defect details in low-contrast areas, or a Laplacian sharpening operator can be used to enhance the clarity of defect edge contours, making defect features more prominent.

[0092] Secondary verification filtering can refer to performing an independent verification and screening operation on the abnormal images initially identified by the airborne terminal at the edge side. For example, running an auxiliary classification model with higher accuracy but slightly higher computational cost can re-determine the existence of defects in the abnormal images marked by the airborne terminal, filtering out images that the auxiliary model judges as false alarms, and only uploading images that have passed the double verification to the cloud, thereby reducing the network transmission burden and cloud processing pressure.

[0093] The false alarm rate refers to the proportion of images that are actually not defective among all images judged as abnormal by the onboard model. For example, the model may generate false alarms due to factors such as changes in lighting, dirt on the surface of components, or reflections on the water surface. Reducing the false alarm rate means increasing the reliability of alarms and reducing the transmission and processing of invalid data.

[0094] Continuing with the aforementioned example of inspecting county-level distributed photovoltaic power station clusters, after the drone completes its inspection mission, automatically returns, and lands on the helipad of a fixed airport node, the abnormal images and associated metadata accumulated in the onboard storage device are read by the edge computing unit through a high-speed data interface.

[0095] The edge computing unit first establishes a processing pipeline. For each frame of an image marked as an anomalous image, the edge computing unit reads its original recognition results from the associated metadata, including the defect type and recognition confidence level, as a reference for subsequent processing.

[0096] The first step in the processing pipeline is image denoising. Since photovoltaic inspection images are typically captured outdoors under natural light conditions, they may be affected by atmospheric scattering, dust on the lens surface, and sensor thermal noise. Random noise may be scattered across some flat areas in the images, and these noise points can sometimes be misjudged as minor defects by the onboard lightweight model. The edge computing unit uses a non-local means denoising algorithm to process each frame of abnormal images. This algorithm searches for image blocks similar to the target pixel's neighborhood across the entire image range and uses a weighted average of similar blocks to estimate the true value of the target pixel. While smoothing noise, it can better preserve the structural texture and edge information of the defect area, avoiding blurring the key features of the true defect due to denoising.

[0097] The second step in the processing pipeline is image enhancement. The edge computing unit applies adaptive histogram equalization to the denoised image, dividing it into several sub-regions. Local contrast enhancement is then applied to each sub-region, significantly amplifying subtle cracks or minor hot spots that were previously low in contrast and difficult for the human eye to detect. Subsequently, an unsharpened mask sharpening algorithm is applied. This involves subtracting the original image from the Gaussian blurred image to obtain a detail layer, which is then weighted and superimposed back onto the original image to enhance the sharpness of defect edges.

[0098] The third step in the processing pipeline is secondary verification filtering. An edge computing unit hosts a verification model with a larger parameter scale than the onboard lightweight model. This verification model employs a more complex feature extraction backbone network, and its input image resolution is several times that of the onboard model, enabling it to capture finer-grained texture features. The edge computing unit inputs denoised and enhanced anomalous images into the verification model for inference. The verification model outputs a binary classification result (whether a defect exists or not) and a corresponding confidence score for each input image. The edge computing unit compares the confidence score output by the verification model with a preset secondary filtering threshold: images with a confidence score higher than the threshold are considered to have a real defect and are retained; images with a confidence score lower than the threshold are considered false positives from the onboard model and are filtered out and discarded. For example, an image that was falsely reported as an occlusion defect by the onboard model due to surface dirt on a component, after denoising and enhancement, amplifies the texture difference between the dirt area and the actual occlusion, allowing the verification model to correctly identify it as a normal area and remove it from the upload queue.

[0099] After the above three preprocessing steps, the edge computing unit re-associates and packages the retained confirmed abnormal images with their corresponding updated metadata, compresses them, and uploads them to the cloud via a communication antenna. Simultaneously, the edge computing unit records statistical data for this filtering process, including the number of original abnormal images, the number of filtered images, and the number of images retained for upload, for subsequent system performance evaluation and model iteration optimization.

[0100] The beneficial effects of this embodiment in this specification include at least the following: by introducing a preprocessing pipeline including denoising, enhancement, and secondary verification filtering into the edge computing unit, the quality of the abnormal images initially generated at the airborne end is improved and false alarms are filtered out. The beneficial effects are that the denoising and enhancement operations effectively improve the signal quality of the image, making weak defect signals more clearly identifiable; the secondary verification filtering utilizes the richer computing resources on the edge side to run a more accurate verification model, eliminating false alarms generated by the lightweight airborne model, significantly reducing the overall false alarm rate of the system, reducing the occupation of communication bandwidth by invalid data and the waste of cloud computing resources, and improving the efficiency and alarm reliability of the entire inspection data processing link.

[0101] According to the aforementioned system, the protective nest of the fixed airport node is divided into multiple functional layers from top to bottom: the top layer is the communication antenna layer, which is equipped with 4G / 5G antennas, WiFi antennas and satellite navigation antennas; the second layer is the edge computing layer, which is equipped with edge computing units; the third layer is the automatic charging layer, which is equipped with battery compartments, robotic arms and charging interfaces; the bottom layer is the apron layer, which is equipped with visual positioning markers, wireless charging contacts and a liftable platform.

[0102] Among them, satellite navigation antenna can refer to antenna devices used to receive signals from global navigation satellite systems, such as antenna modules that receive navigation and positioning signals broadcast by satellite constellations such as GPS and BeiDou, providing high-precision reference position coordinates for fixed airport nodes themselves, and can also be used as a high-precision positioning differential reference station for UAVs.

[0103] A battery compartment can refer to an array-type storage device used to store and manage the charging and discharging of multiple drone power batteries. For example, it can be a rack-type structure with multiple independent charging slots, each with an independent charging control circuit and temperature monitoring sensor, which can automatically execute the charging program for the placed batteries and automatically switch to trickle maintenance mode after being fully charged.

[0104] A robotic arm can refer to an electromechanical actuator used to automatically complete the removal and insertion of drone batteries. For example, a multi-degree-of-freedom articulated robotic arm with a special gripper at the end that adapts to the shape and snap-fit ​​structure of the drone battery can accurately locate the drone battery compartment and perform the battery insertion and removal operations under visual guidance.

[0105] Visual positioning markers refer to high-contrast marking patterns placed on the surface of the helipad to assist drones in achieving precise visual guidance for landing. For example, black and white patterns in the form of Apriltag or QR codes are used. The drone takes pictures of and identifies the markers using an onboard downward-facing camera, calculates its own position offset and altitude relative to the center point of the helipad, and achieves automatic landing positioning at the centimeter level.

[0106] Wireless charging contacts can refer to charging interfaces that automatically establish electrical connections when a drone lands. For example, a row of elastic pins or metal contact plates set on the surface of the landing pad can be tightly pressed against the charging electrodes at the corresponding positions on the bottom of the drone's landing gear under landing pressure, establishing a stable high-current charging circuit without the need for manual plugging or unplugging of any cables.

[0107] A liftable platform can refer to a carrying platform at the bottom of the helipad that is driven by an electric push rod or screw lifting mechanism and can be raised and lowered in the vertical direction. For example, before the drone takes off, the platform is raised to the top opening of the protective nest for the drone to take off, and after the drone lands, the platform is lowered into the closed cabin so that the drone can be charged and stored in a protected environment.

[0108] Continuing with the aforementioned example of inspecting a county-level distributed photovoltaic power station cluster, the fixed airport node is deployed on an open area in the center of a photovoltaic power station cluster. The protective nest adopts an outdoor standardized container design, with the outer shell composed of weather-resistant steel plates and heat-insulating sandwich panels, and an electrically operated swing-open skylight on the top.

[0109] The internal structure of the protective nest is arranged vertically in layers according to function. The top layer is the communication antenna layer. This layer has a set of omnidirectional communication antennas installed on the outer top of the protective nest: two 4G / 5G antennas establish redundant links with the mobile communication networks of different operators to ensure that the communication connection with the cloud intelligent dispatch center can still be maintained when the signal of a single operator's network is poor; a high-gain WiFi antenna is used to provide a high-speed local data transmission link between the node and the drone during the take-off and landing phase; a satellite navigation antenna, including a receiving antenna and a differential reference station antenna, provides the node with sub-meter accuracy self-positioning coordinates and broadcasts real-time differential correction data to the drone through the WiFi link to assist the drone in achieving high-precision autonomous landing.

[0110] Below the communication antenna layer is the edge computing layer. This layer is an independent electromagnetically shielded compartment housing edge computing units. Each edge computing unit is an industrial-grade embedded motherboard integrating a graphics processor and a central processing unit, passively cooled by heat dissipation fins connected to the protective housing shell via thermally conductive silicone. The edge computing units connect to the antenna communication modules of the communication antenna layer and the underlying automatic charging control board via Ethernet interfaces, running edge task management and image preprocessing software.

[0111] Below the edge computing layer lies the automatic charging layer. This layer occupies the largest volumetric space within the protected drone housing, and its core components are a four-axis robotic arm that moves along a horizontal guide rail and an eight-slot battery compartment array. Each slot in the battery compartment is an independent drawer-type module with built-in intelligent charging management circuitry, supporting high-rate fast charging. The robotic arm's end effector integrates a vision camera and a customized pneumatic gripper, capable of recognizing and securely grasping the standard snap-fit ​​structure on the drone's battery casing. When the drone lands and descends to the docking height of the charging layer, the robotic arm moves to the drone's battery compartment and performs a series of automatic battery swapping operations: removing the old battery, placing it in an empty slot for charging, and retrieving the new battery from a fully charged slot and inserting it into the drone's battery compartment. The entire battery swapping process is completed within minutes, requiring no human intervention.

[0112] The bottom layer of the protective nest is the helipad layer. The helipad is a circular aluminum alloy platform with a high-contrast visual positioning mark pattern sprayed in the center of the platform surface. This pattern encodes the unique identifier of this fixed airport node, and the UAV's onboard vision system can simultaneously achieve identity verification and spatial positioning by recognizing this pattern. A set of elastic wireless charging contacts distributed in a ring on the platform surface makes precise contact with the electrodes at the bottom of its landing gear after the UAV lands, providing continuous trickle charging during the recharging phase after battery swapping, keeping the battery fully charged and ready for use. The entire helipad is mounted on a scissor-type electric lifting mechanism, with a servo motor driving a lead screw for smooth lifting and lowering. When a patrol mission start command is received, the lifting platform rises to be level with the top skylight of the protective nest, the UAV unlocks and takes off; after the UAV completes the patrol, returns, lands, and locks, the lifting platform descends into the enclosed cabin, the skylight closes, and the UAV is in an all-weather protected environment.

[0113] The beneficial effects of this embodiment in this specification include at least the following: By vertically layering the internal structure of the protective nest of the fixed airport node according to function, communication, computing, charging and battery swapping, and take-off and landing functions are intensively integrated within a limited space. Its advantages lie in that the layered architecture minimizes electromagnetic interference, thermal interference, and physical interference between subsystems, which is beneficial for stable operation and maintenance replacement of the equipment; the automatic battery swapping robotic arm, combined with the multi-slot battery compartment design, enables the immediate swapping and charging cycle of UAV batteries, eliminating traditional charging waiting time and ensuring the high-frequency continuous sortie capability of UAVs; the combination of visual positioning markers and a liftable platform enables precise automatic landing of UAVs and all-weather enclosed protection, making the fixed airport node a truly unmanned and reliable frontline support unit.

[0114] According to the aforementioned system, the deep learning verification model running in the data aggregation and analysis module adopts a deep residual network architecture to perform defect verification and fine classification on abnormal images uploaded by the edge computing unit. Fine classification subdivides hotspot defects into local hotspots, bypass diode fault hotspots, and junction box abnormal hotspots, and subdivides crack defects into through cracks, tree-like cracks, and edge cracks. The data aggregation and analysis module also analyzes defect development trends based on historical inspection data from the same site, issuing early warnings for defects showing accelerated deterioration trends.

[0115] Among them, the deep residual network architecture can refer to a deep network structure that introduces cross-layer identity connections in a convolutional neural network. For example, by adding a skip connection path that directly connects the input to the output in each residual block composed of several convolutional layers, the network can still be effectively trained when the number of layers is greatly increased, avoiding the gradient vanishing problem of deep networks. This enables the extraction of more abstract and discriminative feature representations from high-resolution photovoltaic inspection images.

[0116] Localized hot spots can refer to small areas of heat spots on photovoltaic modules where the temperature rises relatively concentrated. For example, the localized heating caused by local impurities or slight microcracks inside a single cell can appear as small point-like high-temperature areas in infrared images.

[0117] Bypass diode failure hotspots refer to the phenomenon of abnormal heating of the corresponding battery string due to short circuit or open circuit of the bypass diode. For example, after the bypass diode of a certain string component fails, the battery string cannot be bypassed when it is partially blocked or shielded, and the entire string of batteries becomes a load and heats up. In infrared images, it appears as a strip-shaped high temperature area of ​​the entire string or most of the string.

[0118] Abnormal hot spots in the junction box can refer to hot spots inside the junction box on the back of the module that cause abnormal heating due to poor contact, poor soldering, or diode aging. For example, heat inside the junction box is conducted to the front glass, and in infrared images, it usually appears on one side edge of the module, showing a sheet-like or clump-like high-temperature feature.

[0119] A through crack can refer to a long crack that runs through the entire cell or multiple cells, such as a fracture caused by a severe impact or uneven mechanical stress on the module. In visible light images, it appears as a continuous dark line across the cell and is often accompanied by significant power degradation.

[0120] Dendritic cracks can refer to multiple irregular cracks that branch out from a single point, such as radial or dendritic crack networks triggered by tiny impact points. The lengths of the crack branches vary, and the overall structure appears as a tree-like or spiderweb-like distribution in visible light images.

[0121] Edge cracks can be short cracks that appear at the edge of the cell and typically extend parallel to or perpendicular to the edge direction. They can be caused by edge stress concentration or long-term thermal cycling during module lamination. In visible light images, they appear as short, thin line defects originating at the edge of the cell.

[0122] Defect development trend analysis can refer to the process of judging the future evolution direction of the same defect based on the sequence of changes in size, shape or temperature characteristics in multiple inspections. For example, by comparing the length change records of a crack in several consecutive months of inspections, its propagation rate can be calculated to determine whether the crack is in a stable state or is accelerating its propagation, and the time point at which it may affect the performance of the component can be predicted accordingly.

[0123] Continuing with the aforementioned example of county-level distributed photovoltaic power station cluster inspection, the data aggregation and analysis module of the cloud-based intelligent dispatch center is deployed on a high-performance computing server equipped with multiple graphics processors. The module runs a deep learning verification model pre-trained using a massive dataset of photovoltaic inspection anomaly images.

[0124] This deep learning kernel model employs a deep residual network as its backbone feature extraction network. The network's input layer receives denoised and enhanced anomalous images uploaded from edge computing units. The image first passes through an initial feature extraction layer consisting of convolutional kernels and pooling layers, progressively reducing the image resolution while increasing the number of feature channels. The data then flows into multiple cascaded residual blocks. Each residual block contains two or three convolutional layers connected by batch normalization and activation functions. The output of a residual block equals the output of its internal convolutional paths plus the identity mapping of the residual block's input. This skip connection design allows gradients to be directly propagated back to shallower layers via identity paths during backpropagation training, supporting the construction of extremely deep network structures with hundreds of layers. During the inference phase, the network extracts edge texture features, local shape features, and global semantic features layer by layer, ultimately outputting a high-dimensional feature vector at the fully connected layer.

[0125] The verification model's output includes two parallel classification task branches. The first branch is the defect verification and confirmation branch. This branch receives the feature vector output from the backbone network, passes it through a fully connected layer and an activation function, and outputs a binary classification result to determine whether the input image actually contains a defect. When the output probability exceeds the verification and confirmation threshold, the model determines that the image is a confirmed defect; otherwise, it is considered a false alarm and is excluded.

[0126] For defect images that have passed verification, their feature vectors are fed into the second branch, the fine-classification branch. The fine-classification branch is a multi-class classifier. For hotspot-type defects, the branch classifies them within the hotspot sub-class space, further subdividing the defects into local hotspots, bypass diode fault hotspots, and junction box abnormal hotspots. The model completes this fine-classification by comprehensively judging features such as the geometric shape of the high-temperature region in the infrared image, its spatial distribution on the component, and the surrounding temperature gradient pattern. For example, the high-temperature region of a bypass diode fault hotspot is usually a narrow strip along the battery string direction and consistent with the battery string layout of the component; local hotspots appear as isolated small-area circular or irregular high-temperature points; and the high-temperature region of a junction box abnormal hotspot is concentrated on one side edge of the component, with the temperature gradient decreasing from the edge towards the interior of the component.

[0127] For crack-type defects, the fine classification branch further subdivides them into through cracks, dendritic cracks, and edge cracks. The model distinguishes between these types by analyzing features such as the continuity of dark linear features in visible light images, the degree of branching, and their relative position to the edge of the solar cell. Through cracks appear as single or multiple parallel continuous linear defects spanning the main grid lines of the solar cell; dendritic cracks present as multiple irregular branches radiating outwards from the center point; edge cracks originate at the edge of the solar cell and extend inwards but typically do not cross the main grid lines.

[0128] After reviewing and finely classifying all abnormal images in the current batch, the data aggregation and analysis module initiates the defect development trend analysis process. The module retrieves historical inspection defect records from the database for the same site and component location. When a defect is found to have been recorded in multiple consecutive inspections with matching location coordinates, the module extracts the defect's size parameters or peak temperature parameters from each record, constructing a parameter sequence that changes over time. The module calculates the changes in parameters between adjacent inspection weeks and analyzes the rate of change. If the rate of change shows a progressively increasing trend—that is, the monthly increase in defect size is getting larger or the rate of temperature peak increase is accelerating—the module determines that the defect has an accelerating deterioration trend, marks it as an accelerated deterioration defect, and generates maintenance recommendations with an early warning level. Finally, the module integrates the defect review results, fine classification labels, trend analysis conclusions, and maintenance recommendations into a formatted inspection report output.

[0129] The beneficial effects of this embodiment in this specification include at least the following: By deploying a deep learning verification model based on a deep residual network in the cloud, high-precision defect confirmation and fine-grained classification of abnormal images uploaded from the edge side are achieved. Its beneficial effects lie in the fact that the large-capacity feature extraction capability of the deep residual network enables it to capture subtle defect details and category differences that are difficult for lightweight airborne models to discern, achieving accurate subclassification of hot spots and crack defects. This provides maintenance personnel with more specific information on the causes of failures, helping to formulate more targeted maintenance plans. Simultaneously, defect development trend analysis based on historical inspection data can provide early warnings of rapidly deteriorating defects, transforming passive fault repair into proactive preventative maintenance. This avoids power generation losses and safety accidents caused by sudden deterioration of defects, significantly improving the full lifecycle management level of photovoltaic power plant assets.

[0130] According to the aforementioned system, the site information management module is also used to dynamically update the real-time operating data of each site. The real-time operating data includes the current power generation, inverter operating status, and string current abnormality alarms. When generating inspection tasks, the task generation and sorting module obtains real-time operating data to help determine the timing of periodic inspection task generation and fault alarm triggering conditions.

[0131] The current power generation can refer to the actual output power value of the photovoltaic power station at the current moment. For example, the power data collected in real time by the power meter at the grid connection point of the power station or the output power register of the inverter reflects the real-time power generation capacity of the power station. Its value is affected by the combined effects of solar irradiance, module temperature, module cleanliness and potential faults.

[0132] The inverter's operating status can refer to the current working mode or operating status of the photovoltaic inverter. For example, the inverter's internal state machine may be in a specific state such as grid-connected power generation, standby, fault shutdown, or power-limited operation. This state is continuously monitored by the inverter's own diagnostic system and reported to the monitoring backend.

[0133] A string current anomaly alarm can refer to an alarm signal detected by the string-level monitoring unit that the string current deviates from the normal expected value. For example, an alarm is triggered when the current value of a certain string is significantly lower than that of other strings connected in parallel or lower than the expected current value calculated based on the current irradiance. This alarm usually indicates that the string may have problems such as shading, component failure or poor connector contact.

[0134] Continuing with the aforementioned example of county-level distributed photovoltaic power station cluster inspection, the station information management module maintains a continuous data synchronization connection with the data acquisition gateway of each photovoltaic power station during system operation, thereby acquiring the real-time operational data stream of each power station.

[0135] The station information management module subscribes to real-time data topics from the monitoring systems of each power station using a standardized telemetry data protocol. For each station, the subscribed data items include: the current power generation value at the grid-connected metering point of the power station, with a data refresh cycle of minutes; the operating status words reported by each inverter in the power station, where different codes of the status words represent whether the inverter is currently in grid-connected power generation mode, standby mode, fault protection mode, or power limiting mode; and the current measurement values ​​of each DC string circuit and the corresponding abnormal alarm flags. When the string monitoring unit detects that the string current is lower than the preset benchmark value by more than a threshold, the abnormal alarm flag is set.

[0136] The station information management module writes the received real-time operation data into the corresponding station record in the spatial database. Each station record in the database contains a time-series storage unit, which stores the recent historical curves of continuously changing parameters such as power generation in chronological order; it also contains a status cache unit, which records the latest values ​​and change timestamps of the inverter's current operating status and string abnormal alarm status.

[0137] When executing its task generation logic, the task generation and sorting module proactively queries the site information management module for real-time operating data of each site as a basis for judgment. For the generation of periodic inspection tasks, the task generation and sorting module not only checks whether the time interval since the last inspection has reached the preset inspection cycle, but also makes a comprehensive judgment based on the current power generation data of the site. For example, if a power station has reached the cycle threshold but the current power generation of the power station is close to the theoretical power value predicted based on the current irradiance, and all inverters are in normal grid-connected status, indicating that the power station is operating well, the module can appropriately delay the generation priority of the periodic inspection task, allocating limited inspection resources to sites that require more attention. Conversely, if a power station has not fully reached the inspection cycle, but the recent power generation time curve shows a continuous downward trend, and this decline cannot be fully explained by changes in irradiance, the module will generate the periodic inspection task of the power station in advance and increase its priority.

[0138] For fault inspection tasks, the task generation and sorting module continuously scans the string current anomaly alarm flag. When a string current anomaly alarm is detected in a certain site and is set, the module does not immediately generate a fault inspection task. Instead, it first verifies the authenticity of the alarm. The module queries the current operating status of the inverter where the string is located. If the inverter itself is in fault protection mode, the string anomaly is likely caused by an inverter fault rather than a component fault. The module classifies this alarm as an inverter-related alarm and handles it through the equipment maintenance process. If the inverter is in normal grid-connected status but the string current remains low, the module further compares the current historical curves of this string with those of other non-alarm strings under the same inverter. Only after confirming that the difference is a persistent difference rather than a momentary fluctuation caused by temporary cloud cover does the module trigger the generation of the corresponding fault inspection task. This verification mechanism effectively filters out false alarms caused by inverter-side problems and short-term environmental fluctuations, ensuring that UAV inspection resources are accurately deployed to sites where component-level faults actually exist.

[0139] The beneficial effects of this embodiment in this specification include at least the following: Through the dynamic updating of real-time operating data of each station by the station information management module, and the comprehensive utilization of power generation, inverter status, and string current alarms by the task generation and sorting module during task generation, a shift from passive, timed inspections to status-driven intelligent inspections is achieved. Its beneficial effect lies in the fact that task generation no longer mechanically relies entirely on a fixed cycle, but rather integrates the actual health status information of the power station, allowing inspection resources to be prioritized for power stations where operating data shows abnormal signs, improving the timeliness of potential fault detection. Simultaneously, the multi-dimensional verification mechanism for string current anomaly alarms effectively eliminates false alarms caused by inverter-side faults and transient environmental interference, reducing the number of invalid inspection deployments, saving UAV lifespan and system operating costs, and further improving overall inspection efficiency.

[0140] According to the aforementioned system, vertical take-off and landing fixed-wing UAVs in the UAV swarm are used to perform large-scale rapid inspection tasks in areas with sparsely distributed airfields and large coverage areas, while multi-rotor UAVs are used to perform refined inspection tasks in areas with high airfield density.

[0141] Vertical takeoff and landing (VTOL) fixed-wing UAVs can refer to composite configuration aircraft that combine the vertical takeoff and landing capabilities of rotorcraft with the high-speed cruise capabilities of fixed-wing aircraft. For example, UAVs with tilt-rotor or tail-seat configurations take off and land vertically in rotor mode during the takeoff and landing phase, and switch to fixed-wing mode during the cruise phase to fly long distances at a faster cruise speed and with lower energy consumption. The endurance can reach several hours, and the coverage range of a single sortie can reach tens of kilometers.

[0142] Large-scale rapid inspection missions can refer to inspection missions whose main objective is to quickly obtain an overview of the overall status of various stations within a large geographical area and to investigate key anomalies. For example, aerial surveys can be conducted on multiple sparsely distributed ground power stations or large industrial and commercial rooftop power stations that are far apart from each other. This requires the aircraft to have high cruising speed and long endurance to cover as many stations as possible in a single sortie.

[0143] Multi-rotor drones can refer to rotorcraft that achieve flight attitude control and lift provision through the speed difference of multiple rotors, such as quadcopter or hexcopter drones. They have the ability to hover precisely, maneuver at low speeds, and observe at close range, making them suitable for line-by-line scanning and detailed photography of photovoltaic modules in dense areas.

[0144] Refined inspection tasks refer to inspection tasks that require high-resolution imaging and close-range inspection of every detail of photovoltaic modules. For example, low-altitude slow-speed scanning of tightly arranged module arrays in rooftop photovoltaic or agricultural photovoltaic power stations requires the aircraft to have excellent low-speed stability and hovering accuracy to ensure safe flight in narrow spaces and to acquire high-definition images of each module.

[0145] Continuing with the aforementioned example of inspecting distributed photovoltaic power station clusters in a county, the distribution of distributed photovoltaic power stations within the county exhibits a clear spatial differentiation: The county seat and its suburbs have a large number of industrial, commercial, and residential rooftop photovoltaic power stations, which are densely packed but each station has a relatively small capacity, and the power stations are close to each other, demonstrating a high degree of intensification; while in the agricultural and hilly areas surrounding the county, there are several larger-capacity but more distant ground-mounted distributed power stations and agricultural-photovoltaic complementary power stations, which are sparsely distributed and have poor road conditions, making manual vehicle inspections time-consuming.

[0146] To address the aforementioned scenario characteristics, the system's drone swarm configuration incorporates two configurations. The swarm comprises multiple multi-rotor drones deployed at fixed airport nodes in urban and suburban areas. These multi-rotor drones possess precise hovering and low-speed maneuverability, are equipped with high-resolution visible light cameras and mid-wave infrared thermal imagers, and feature gimbals with multi-axis stabilization and precise angle control capabilities. When performing inspection tasks on densely packed rooftop photovoltaic systems in urban areas, the multi-rotor drones can quickly reach multiple nearby target power stations after taking off from fixed airport nodes. They can navigate through complex building clusters, maintaining precise hovering over pre-set shooting points above each rooftop, and using flight paths with sufficient overlap to perform line-by-line scanning of the component arrays on the sloping roof surfaces, ensuring that the details of each component are clearly discernible in the images.

[0147] Several vertical takeoff and landing (VTOL) fixed-wing UAVs are deployed simultaneously in the cluster, stationed at two fixed airport nodes covering the outer areas of the county. The VTOL fixed-wing UAVs are tilt-rotor configurations; during takeoff and landing, the four rotors tilt upwards to provide lift, enabling vertical takeoff and landing. During cruise, the rotors tilt forward to switch to fixed-wing flight mode. This type of aircraft has a significantly higher cruise speed than multi-rotor UAVs, and its single-spin-off endurance can be several times that of multi-rotor UAVs. A single mission can cover multiple airfields within a range of tens of kilometers.

[0148] Upon receiving an inspection mission from sparsely distributed ground power stations on the outskirts of a county, the VTOL fixed-wing UAV takes off vertically from a liftable platform at a fixed airport node, reaches a safe altitude, and then smoothly tilts its rotors to a horizontal position, entering a high-speed fixed-wing cruise mode. Following an optimized route generated by the dynamic path planning module, it flies in straight segments over various ground power stations and agricultural-solar hybrid power stations distributed across a vast area. Upon approaching each power station, the VTOL fixed-wing UAV reduces its cruise speed and adjusts its course, using its onboard wide-field-of-view visible light camera and infrared thermal imager to rapidly scan and photograph the entire area of ​​the power station using a strip push-broom method. Although the ground resolution of a single frame image is lower than that of a multi-rotor hovering high-resolution image, its high frame rate continuous shooting and wide coverage capability are sufficient to complete a panoramic survey of a megawatt-level ground power station within minutes, quickly identifying abnormal areas with large-area abnormal temperature rises or obvious component damage. The coordinates of these abnormal areas are marked, and targeted, detailed review tasks are automatically generated, which are then confirmed by subsequent low-altitude, high-resolution re-photographing by a multi-rotor UAV.

[0149] Through the coordinated operation of the aforementioned multi-configuration UAVs, the system has achieved a differentiated inspection strategy for power plant clusters with different spatial distribution characteristics: dense areas are directly and meticulously covered by multi-rotor UAVs, while sparse areas are quickly surveyed by vertical take-off and landing fixed-wing UAVs to locate anomalies, and then verified by multi-rotor UAVs at specific points, achieving the best balance between overall inspection efficiency and defect detection accuracy.

[0150] The beneficial effects of this embodiment in this specification include at least the following: by simultaneously configuring multi-rotor UAVs and VTOL fixed-wing UAVs in a UAV swarm, and by differentiating their roles based on the density of power station distribution and the nature of the mission, the system can flexibly adapt to diverse distributed photovoltaic layout scenarios, ranging from dense urban rooftops to sparse suburban ground-based power stations. The beneficial effects are that the VTOL fixed-wing UAVs, with their long endurance and high cruising speed, compensate for the shortcomings of multi-rotor UAVs in terms of endurance and efficiency when covering large sparse areas, eliminating the cost of manual vehicle patrols or individual deployment of multi-rotor UAVs to remote power stations; while the multi-rotor UAVs, with their hovering and low-speed maneuverability, ensure the need for detailed imaging in dense and complex environments. The division of labor and cooperation between the two types of aircraft significantly expands the power station capacity and geographical range that a single system can cover, achieving global optimization of inspection efficiency and detection accuracy.

[0151] According to the aforementioned system, meteorological monitoring sensors are deployed outside fixed airport nodes to continuously collect real-time ground meteorological data at their locations. The collected data includes wind speed, wind direction, temperature, humidity, solar irradiance, and rainfall. The meteorological perception and prediction module integrates real-time ground meteorological data from each fixed airport node, upper-air meteorological forecast data from public meteorological service platforms, and forecast data from meteorological radar and satellites to generate meteorological forecasts for future periods and identify flight windows based on these forecasts.

[0152] Real-time ground meteorological data can refer to actual observation data that reflects the physical state of the near-surface atmosphere, obtained directly by meteorological sensors at the location of fixed airport nodes. Examples include instantaneous wind speed and direction of flow measured by anemometers installed on meteorological poles at fixed airport nodes, atmospheric temperature and relative humidity measured by temperature and humidity sensors, solar irradiance measured by a sunroof meter, and minute-level rainfall intensity measured by a rain gauge. These data represent the real local meteorological conditions of each node.

[0153] Upper-air meteorological forecast data can refer to atmospheric parameter forecast products at various altitude levels calculated and released by professional meteorological agencies through numerical weather prediction models. For example, hourly gridded weather forecast data covering the target area obtained from public meteorological service platforms includes forecast values ​​of wind speed components, geopotential height, temperature, and relative humidity at different pressure levels, providing necessary meteorological background information for the cruise flight of vertical take-off and landing fixed-wing UAVs in higher airspace.

[0154] Forecast data from weather radar and satellites can refer to the spatiotemporal evolution information of convective weather systems and cloud and rain areas based on weather radar echoes and meteorological satellite cloud images. For example, weather radar mosaic reflectivity forecast products can predict the movement path and intensity changes of strong convective rainfall areas tens of minutes in advance, and satellite cloud image cloud top brightness temperature forecast products can predict the movement and evolution trends of large-scale cloud systems. These data provide support for identifying unsuitable flight areas on a wide spatial scale.

[0155] A flight window refers to a combination of time periods and spatial areas that are deemed to meet the conditions for safe flight of a drone after comprehensively considering all meteorological factors. For example, when the predicted wind speed in a certain area is lower than the drone's maximum wind resistance, the predicted rainfall is lower than the allowable value corresponding to the waterproof rating, and the lighting conditions meet the minimum requirements for the normal operation of the visual positioning and obstacle avoidance systems, the spatial area and time period are marked as a flight window, which is the basic constraint for subsequent mission scheduling and path planning.

[0156] Continuing with the aforementioned example of inspecting a county-level distributed photovoltaic power station cluster, ten fixed airport nodes are distributed across different locations within the county. Each fixed airport node has an integrated meteorological monitoring sensor assembly installed on the exterior of its protective enclosure. This assembly is mounted on a meteorological pole approximately three meters high extending from the top of the protective enclosure to avoid interference from near-ground buildings and equipment on airflow. The sensor assembly includes: an ultrasonic anemometer that measures wind speed and direction using the time difference of ultrasonic pulse propagation in the air; an integrated temperature and humidity probe placed inside a radiation-proof ventilation hood to measure air temperature and relative humidity; a total radiation meter to measure the total solar irradiance at a horizontal level; and a tipping bucket rain gauge to monitor the onset, intensity, and cumulative rainfall in real time.

[0157] Meteorological monitoring sensors at each fixed airport node operate continuously at a fixed data acquisition frequency, generating a set of real-time ground meteorological data every minute, including wind speed, wind direction, temperature, humidity, solar irradiance, and rainfall. This data is then transmitted in real-time via 4G / 5G networks through the communication antennas of the fixed airport nodes to the meteorological sensing and prediction module in the cloud-based intelligent dispatch center.

[0158] The meteorological perception and prediction module runs a meteorological forecasting and analysis program daily. The module first aggregates real-time surface meteorological data from all ten fixed airport nodes, forming a high-density, ten-minute-level surface observation network covering the entire county. This surface data reflects the true state of the local micro-meteorological environment, with an accuracy higher than that of numerical weather prediction models extrapolating near-surface data.

[0159] Meanwhile, the meteorological perception and prediction module calls the public meteorological data application programming interface of the provincial meteorological service center through the Internet interface to obtain the following data products: hourly surface and upper-air meteorological element forecast data covering the county area, forecasting wind speed components, temperature, relative humidity, precipitation and cloud cover at each grid point in the future period; weather radar network mosaic echo forecast products covering the county and surrounding areas, providing radar echo intensity movement forecasts updated every tens of minutes for the next few hours; and cloud top brightness temperature and cloud classification forecast products based on geostationary meteorological satellites, providing hourly forecasts of the movement and evolution trends of large-scale cloud systems.

[0160] The meteorological sensing and prediction module fuses the aforementioned multi-source data within a unified geographic grid and time coordinate system. For wind speed, the module uses real-time measured wind speed values ​​from each fixed airport node as ground truth to correct the real-time bias of the wind speed field predicted by the numerical model. The corrected wind speed prediction grid more accurately reflects the actual wind field affected by local topography and surface roughness. For precipitation, the module combines radar echo extrapolation forecasts and satellite cloud image forecasts to track the movement path and intensity evolution of precipitation echo cells on a minute-level time scale, generating time-segmented precipitation probabilities and predicted rainfall intensities for the coverage area of ​​each fixed airport node.

[0161] Based on the fused meteorological forecast results, the meteorological perception and forecasting module performs flight condition assessments on a grid-by-grid and time-by-time basis. The assessment criteria include: whether the predicted wind speed is below the upper limit required for the normal operation of the lightweight AI defect identification model; whether the predicted rainfall is zero or below a trace threshold; and whether the irradiance meets the minimum requirements for the normal operation of the visible light camera imaging and visual positioning system. Spatiotemporal grid cells that simultaneously meet all the above conditions are marked as flight windows, forming a regional and time-by-time flight window distribution map covering the county. This distribution map is sent in structured data form to the task generation and sorting module and the multi-aircraft task allocation module, serving as the core meteorological constraint boundary for task execution scheduling and route planning.

[0162] The beneficial effects of this embodiment in this specification include at least the following: by deploying meteorological monitoring sensors at various fixed airport nodes to form a distributed local real-time observation network, and by integrating multi-source meteorological information such as public numerical forecasts, weather radar, and satellite data in the cloud, high spatiotemporal resolution and accurate prediction of future meteorological conditions in the inspection area are achieved. Its beneficial effects lie in the fact that distributed ground-based real-time observation effectively compensates for the deficiencies of public weather forecasts in characterizing local micro-meteorology, and the introduction of radar and satellite data enhances the short-term forecasting capability for sudden convective weather. The resulting flight window provides a reliable meteorological safety boundary for the entire system's task scheduling decisions, enabling inspection tasks to be precisely scheduled for execution within safe time periods and areas. This maximizes the system's available operational time window while ensuring UAV flight safety, significantly improving the system's task execution efficiency and success rate under variable weather conditions.

[0163] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0164] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0165] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A distributed photovoltaic power station multi-machine cooperative intelligent inspection system, characterized in that, This includes cloud-based intelligent dispatch centers, distributed fixed airport networks, and drone swarms; The cloud-based intelligent dispatch center includes a station information management module, a meteorological sensing and forecasting module, a task generation and sorting module, a multi-machine task allocation module, a dynamic path planning module, and a data aggregation and analysis module. The site information management module is used to establish and maintain a spatial database containing information on the geographical location, capacity, operating status, and priority of each photovoltaic power station. The meteorological sensing and prediction module is used to acquire multi-source meteorological data and predict flight windows for future periods. The task generation and sorting module is used to generate an inspection task queue based on the periodic inspection plan and real-time fault alarms and sort them by priority. The multi-aircraft task allocation module is used to allocate tasks in the inspection task queue to each fixed airport node using a genetic algorithm. The dynamic path planning module is used to generate the optimal inspection route for each fixed airport node by combining meteorological forecast information and no-fly zone information. The distributed fixed airport network includes multiple fixed airport nodes, each of which includes a protective cabin, a communication antenna, an edge computing unit, an automatic charging device, a helipad, and a meteorological monitoring sensor. The drone swarm includes multiple multi-rotor drones and / or vertical take-off and landing fixed-wing drones. The drones are equipped with visible light cameras, infrared thermal imagers and onboard computing units. The onboard computing units run a lightweight artificial intelligence defect recognition model to perform real-time defect recognition on the images collected during flight. The edge computing unit is used to preprocess the identified abnormal images, and the preprocessed abnormal images and associated metadata are uploaded to the data aggregation and analysis module through the communication antenna; The data aggregation and analysis module runs a deep learning verification model to perform defect verification and fine classification on the received abnormal images and generate inspection reports.

2. The system of claim 1, wherein, When the multi-machine task allocation module uses a genetic algorithm to allocate tasks, the fitness function comprehensively considers the total inspection completion time, the task load balance of each fixed airport node, the total distance of UAV transfer, and the time window satisfaction rate. The constraints include that the difference in the total number of tasks allocated to each node does not exceed a preset proportion threshold, and fault-type inspection tasks are completed within a preset response time limit.

3. The system of claim 1, wherein, The dynamic path planning module generates inspection routes based on the Traveling Salesman Problem algorithm, using the fixed airport node's own location as the start and return points, and the assigned photovoltaic power station locations to be inspected as mandatory path points. During the path solving process, the dynamic path planning module integrates future weather forecast information provided by the weather perception and prediction module. When weather conditions exceeding the safe flight threshold of the UAV are predicted in a certain area, the module dynamically adjusts the access sequence of the power stations in that area or postpones the corresponding tasks to subsequent batches. At the same time, it acquires the boundary coordinate data of permanent and temporary no-fly zones, enabling the planned route to automatically bypass the no-fly zone boundaries.

4. The system of claim 1, wherein, The lightweight artificial intelligence defect recognition model running on the airborne computing unit is used to identify hot spot defects, crack defects and shading defects of photovoltaic modules in real time, and to mark the defect type, defect location coordinates, recognition confidence level and shooting timestamp of the identified abnormal images as the associated metadata.

5. The system of claim 1, wherein, The preprocessing operations performed by the edge computing unit include image denoising, image enhancement, and secondary verification filtering of the abnormal image to reduce the false alarm rate.

6. The system according to claim 1, characterized in that, The protective nest of the fixed airport node is divided into multiple functional layers from top to bottom: the top layer is the communication antenna layer, which is equipped with 4G / 5G antennas, WiFi antennas and satellite navigation antennas; the second layer is the edge computing layer, which is equipped with the edge computing unit; the third layer is the automatic charging layer, which is equipped with battery compartment, robotic arm and charging interface; the bottom layer is the apron layer, which is equipped with visual positioning markers, wireless charging contacts and lifting platform.

7. The system according to claim 1, characterized in that, The deep learning verification model running in the data aggregation and analysis module adopts a deep residual network architecture to perform defect verification and fine classification on the abnormal images uploaded by the edge computing unit. The fine classification subdivides hot spot defects into local hot spots, bypass diode fault hot spots, and junction box abnormal hot spots, and subdivides crack defects into through cracks, tree-like cracks, and edge cracks. The data aggregation and analysis module also performs defect development trend analysis based on historical inspection data of the same site and issues early warning prompts for defects with accelerated deterioration trends.

8. The system according to claim 1, characterized in that, The site information management module is also used to dynamically update the real-time operation data of each site. The real-time operation data includes the current power generation, inverter operating status, and string current abnormality alarm. When generating inspection tasks, the task generation and sorting module obtains the real-time operation data to help determine the timing of generating periodic inspection tasks and the triggering conditions for fault alarms.

9. The system according to claim 1, characterized in that, The vertical take-off and landing fixed-wing UAVs in the UAV cluster are used to perform large-scale rapid inspection tasks in areas with sparsely distributed airfields and large coverage areas, while the multi-rotor UAVs are used to perform refined inspection tasks in areas with high airfield density.

10. The system according to claim 1, characterized in that, The meteorological monitoring sensors are deployed outside the fixed airport nodes to continuously collect real-time ground meteorological data at their location. The collected data includes wind speed, wind direction, temperature, humidity, solar irradiance, and rainfall. The meteorological perception and prediction module integrates the real-time ground meteorological data from each fixed airport node, upper-air meteorological forecast data from the public meteorological service platform, and forecast data from meteorological radar and satellites to generate meteorological forecast results for future periods and identify flight windows based on the meteorological forecast results.

Images