Unmanned aerial vehicle-based photovoltaic power station inspection method and system

Abstract

The application relates to the photovoltaic operation and maintenance technical field and discloses a photovoltaic power station inspection method and system based on a UAV, which comprises the following steps: acquiring a target photovoltaic component topology graph and image acquisition time, combining a light source position to calculate surface reflection intensity distribution, generating a unit-level reflection risk field matrix, eliminating high-risk areas in the matrix, screening and determining a unit-level diagnosable visual angle window set by using pixel resolution and obstacle avoidance distance, planning a partition inspection route and a sensor pose sequence based on the topology graph and the window set to collect initial image data, calculating an imaging quality score, extracting a backup shooting task from the window set to generate backup image data when the imaging quality score is lower than a threshold, spatially registering the two groups of image data to extract a dynamic highlight area, separating reflection artifacts to extract an abnormal candidate area, mapping geographical coordinates and updating the route and the matrix. The application realizes anti-reflection adaptive planning and self-evolution dynamic closed-loop diagnosis of the UAV inspection of the photovoltaic power station.

Description

Technical Field

[0001] This application relates to the field of photovoltaic operation and maintenance technology, and in particular to a method and system for inspecting photovoltaic power plants based on unmanned aerial vehicles (UAVs). Background Technology

[0002] With the global shift towards clean energy, the photovoltaic (PV) power generation industry has experienced explosive growth, and the scale of PV power plant construction is constantly expanding, often occupying vast areas of land. During the long-term operation of these power plants, PV modules inevitably develop defects such as microcracks, hot spots, glass breakage, and junction box malfunctions due to environmental corrosion, mechanical stress, or electrical aging. Given the sheer number of modules, traditional manual inspection methods relying on handheld infrared or visible light equipment are not only time-consuming and labor-intensive but also present significant blind spots and safety hazards. Therefore, using drones equipped with high-resolution visible light and thermal imaging cameras for automated aerial inspections has become the mainstream technology for improving the efficiency of PV power plant operation and maintenance. Drones can rapidly traverse above the power plant along pre-set GPS routes, collecting massive amounts of image data from the module surfaces, which is then processed by backend computer vision algorithms for defect identification.

[0003] However, existing drone-based photovoltaic (PV) inspection technologies have encountered severe bottlenecks in practical engineering applications due to "optical blindness" and "open-loop failure." PV modules are covered with high-transmittance glass panels, which are highly susceptible to strong specular reflection under intense sunlight. Traditional inspection route planning typically employs fixed grid-like flight and vertical overhead shooting strategies, completely neglecting the dynamic changes in solar altitude and azimuth angles over time, as well as the spatial differences in module installation tilt angles. This results in the drone's camera field of view inevitably cutting into the reflective light cone at specific flight times and locations, leading to large areas of overexposed artifacts in the captured images. This intense reflection directly obscures real cracks or hot spots on the PV panel, causing numerous missed detections by the backend visual algorithms. Even some advanced systems attempting to reduce reflection through hardware means such as polarizing filters struggle to cope with complex and intersecting light field environments, ultimately resulting in the failure of entire batches of image data.

[0004] Furthermore, current photovoltaic inspection systems generally lack adaptive quality perception and system self-learning evolution capabilities during flight. In actual flight, drones are often affected by gusts, GPS drift, or transient changes in lighting caused by cloud cover, easily resulting in blurred or underexposed images. Traditional systems often only discover image quality defects after data transmission, necessitating re-flight and retesting, wasting manpower and time. Simultaneously, the backend detection system struggles to distinguish between dynamic reflective bright spots caused solely by parallax and actual physical damage, easily leading to false alarms. More critically, when the initial topology database of the power plant contains errors, such as component installation tilt angles, existing inspection systems cannot use observed deviations to correct the underlying database. This one-way execution mode, lacking closed-loop feedback, traps the inspection system in a dead end of frequent false alarms, repeated re-flights, and an inability to self-optimize, severely hindering the advancement of unmanned operation and maintenance of photovoltaic power plants towards higher levels of fully autonomous intelligence. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a method and system for inspecting photovoltaic power plants based on unmanned aerial vehicles (UAVs). This system adaptively avoids interference from sunroof reflections under complex spatiotemporal and lighting conditions, and achieves dynamic closed-loop feedback of image quality and automatic iterative updates of the data model during flight. This enables efficient and accurate acquisition of artifact-free inspection images of photovoltaic modules to support high-reliability defect diagnosis.

[0006] Firstly, this application provides a method for inspecting a photovoltaic power station based on an unmanned aerial vehicle (UAV), the method comprising: S1: Obtain the topology map and image acquisition time of the target photovoltaic module, extract the normal vector of the target photovoltaic module based on the topology map, and calculate the reflection intensity distribution on the surface of the target photovoltaic module by combining the light source position determined by the image acquisition time, and generate a unit-level reflection risk field matrix; S2: Eliminate high-risk areas in the unit-level reflection risk field matrix to obtain candidate observation spatial domains. Use preset pixel resolution and obstacle avoidance distance to filter candidate observation spatial domains and determine the unit-level diagnostic viewing window set of the target photovoltaic module. S3: Based on the topology map and the set of unit-level diagnostic view windows, plan the partition inspection route and the corresponding sensor pose sequence, and control the UAV to execute the partition inspection route to collect initial image data. S4: Calculate the imaging quality score of the initial image data. When the imaging quality score is lower than the preset threshold, extract alternative viewpoints from the set of unit-level diagnosable viewpoint windows to generate a reshooting task. Control the drone to execute the reshooting task to obtain reshooting image data. S5: Spatial registration is performed between the initial image data and the re-shot image data to extract the dynamic highlight area. The reflection artifact area is separated by calculating the correlation between the dynamic highlight area and the unit-level reflection risk field matrix. After excluding the reflection artifact area, the abnormal candidate area is extracted. The abnormal candidate area and the reflection artifact area are mapped to geographic coordinate information respectively, and the zonal inspection route and the unit-level reflection risk field matrix are updated based on the corresponding geographic coordinate information.

[0007] Secondly, this application provides a photovoltaic power station inspection system based on unmanned aerial vehicles (UAVs), the system comprising: The reflection risk module is used to obtain the topology map and image acquisition time of the target photovoltaic module, extract the normal vector of the target photovoltaic module based on the topology map, and calculate the reflection intensity distribution on the surface of the target photovoltaic module by combining the light source position determined by the image acquisition time, and generate a unit-level reflection risk field matrix. The viewing angle filtering module is used to remove high-risk areas in the unit-level reflection risk field matrix to obtain candidate observation spatial domains. It uses preset pixel resolution and obstacle avoidance distance to filter candidate observation spatial domains and determine the set of unit-level diagnosable viewing angle windows for the target photovoltaic module. The route planning module is used to plan the zone inspection route and the corresponding sensor pose sequence based on the topology map and the set of unit-level diagnostic view windows, and control the UAV to execute the zone inspection route to collect initial image data. The reshooting task module is used to calculate the imaging quality score of the initial image data. When the imaging quality score is lower than a preset threshold, it extracts alternative viewpoints from the set of unit-level diagnosable viewpoint windows to generate a reshooting task and controls the drone to execute the reshooting task to obtain reshooting image data. The anomaly update module is used to spatially register the initial image data with the re-captured image data to extract the dynamic highlight area. It separates the reflection artifact area by calculating the correlation between the dynamic highlight area and the unit-level reflection risk field matrix. After excluding the reflection artifact area, it extracts the anomaly candidate area. It maps the anomaly candidate area and the reflection artifact area to geographic coordinate information respectively, and updates the zonal inspection route and the unit-level reflection risk field matrix based on the corresponding geographic coordinate information.

[0008] Compared with the prior art, the beneficial effects of the present invention are at least as follows: 1. First, this application proposes an adaptive static anti-reflective flight path planning mechanism based on physical optics, which eliminates the damage to image quality caused by specular reflection on the surface of photovoltaic modules from the source. By integrating solar astronomical ephemeris, the three-dimensional topology of the modules and the bidirectional reflection distribution function, this application can accurately calculate and generate a unit-level reflection risk field matrix, thereby transforming the resolution requirements of optical imaging and the obstacle avoidance constraints of UAV flight into spatial screening conditions. This technical solution ensures that each generated zone inspection flight path and the synchronized sensor pose sequence not only ensure that the camera optical axis is always accurately focused on the target, but also ensures that the UAV always operates within a safe and anti-reflective "diagnostic viewing window", solving the engineering problem of large-area overexposure and the masking of small defect features caused by traditional grid-based blind flight.

[0009] 2. Secondly, this application constructs a closed-loop control system for dynamic image quality evaluation and real-time reshooting during flight, which improves the anti-interference capability and engineering operation efficiency of UAV inspection in complex natural environments. Traditional inspections often only discover image blurring or exposure abnormalities caused by wind drift or sudden changes in lighting after the mission is completed, resulting in the need to reschedule the flight for a second shot. However, this application can immediately perform multi-dimensional scoring of sharpness, exposure, and target coverage after each image is acquired. Once it falls below the threshold, the back projection technology of photogrammetry is used to accurately map the two-dimensional image defects back to the three-dimensional physical space, and the non-reflective waypoints are directly retrieved from the previously reserved alternative safety perspectives to generate a local reshooting task. This technical solution combines static prior planning with dynamic real-time correction, compressing the traditionally calculated reshoot cost per day to a few seconds of automatic adjustment during flight, ensuring that every frame of data brought back to the ground meets high-quality inspection standards.

[0010] 3. Finally, this application endows the inspection system with advanced capabilities for identifying and refining data from multiple perspectives and for self-learning and evolution, breaking the reliance of traditional detection algorithms on manual verification. By spatially registering and calculating brightness differences between the initial and re-shot images of the same component, this application utilizes the dynamic physical characteristics of specular reflection changing with viewing angle to successfully remove reflection artifacts that are easily confused with real hot spots, reducing the false alarm rate of defects. Furthermore, this application reverse-maps the finally confirmed abnormal data back to the three-dimensional geographic coordinate system, which not only automatically improves the route priority of frequently faulty areas in the next inspection, but also uses the actual reflection deviation to infer and correct the incorrectly registered component installation tilt angle in the underlying topology map. Through the digital twin feedback mechanism of continuously calibrating the initial theoretical model with on-site measured data, the entire inspection system has increasingly accurate self-optimization capabilities, realizing truly unmanned, intelligent, and closed-loop operation and maintenance of photovoltaic power plants throughout their entire life cycle. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of the photovoltaic power station inspection method based on unmanned aerial vehicles (UAVs) in the embodiments of this application; Figure 2 This is a structural diagram of the drone-based photovoltaic power station inspection system in this application embodiment. Detailed Implementation

[0013] This application provides a method and system for inspecting photovoltaic power plants based on unmanned aerial vehicles (UAVs). The terms "first," "second," "third," "fourth," etc. (if present)," in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0014] Example 1: For ease of understanding, the specific process of the embodiments of this application is described below, such as... Figure 1 The embodiment of this application shows a method for inspecting a photovoltaic power station based on a drone, which includes: S1: Obtain the topology map and image acquisition time of the target photovoltaic module, extract the normal vector of the target photovoltaic module based on the topology map, and calculate the reflection intensity distribution on the surface of the target photovoltaic module by combining the light source position determined by the image acquisition time, and generate a unit-level reflection risk field matrix.

[0015] In S1, obtaining the topology map of the target photovoltaic module and extracting the normal vector of the target photovoltaic module includes: reading the initial layout data of the photovoltaic power station, which contains the unique identifier of the target photovoltaic module, latitude and longitude coordinates, altitude information, and installation tilt angle and azimuth angle; calculating the orientation of the target photovoltaic module surface in three-dimensional space based on the installation tilt angle and azimuth angle to obtain the normal vector information; and fusing the initial layout data and the normal vector information to construct the topology map.

[0016] Specifically, S1 generates a unit-level reflection risk field matrix, including: calculating the solar altitude angle and solar azimuth angle based on image acquisition time information and the latitude and longitude coordinates of the target photovoltaic module, and using these as light source position information to construct a solar direction vector; calculating the spatial angle between the solar direction vector and the normal vector information, inputting the spatial angle into a bidirectional reflection distribution function model based on Fresnel's law of reflection, and calculating the specular reflection intensity distribution on the surface of the target photovoltaic module in different emission directions; delineating a three-dimensional spatial region around the target photovoltaic module and dividing it into multiple grid nodes, calculating the reflection risk value of each grid node based on the specular reflection intensity distribution, and generating a three-dimensional spatial matrix as a unit-level reflection risk field matrix.

[0017] Specifically, in the inspection of photovoltaic power plants based on drones, in order to solve the technical problem that the surface image of photovoltaic modules is locally overexposed and effective defect features are submerged due to the reflection of the sunroof, this embodiment provides a specific implementation method for step S1. It aims to pre-quantify the distribution of specular reflection risk in the airspace around the module by accurately calculating the spatial orientation of the photovoltaic module and the direction of the incident sunlight, so as to provide a quantitative basis for avoiding strong reflection interference and selecting safe and effective observation angles.

[0018] First, a topology map containing high-precision 3D spatial information is constructed, and the normal vector of the target photovoltaic module is extracted from it. Specifically, structured initial layout data is read from the photovoltaic power plant's management system or design documents. Each data entry contains at least the unique identifier of the target photovoltaic module, its latitude and longitude coordinates, altitude, installation tilt angle, and installation azimuth angle. The installation tilt angle is defined as the angle between the photovoltaic module's light-receiving surface and the horizontal ground, and the installation azimuth angle is defined as the angle between the projection of the module's surface normal onto the horizontal plane and the due north direction. Based on these two angle parameters, the orientation of the module's surface in 3D space is calculated using trigonometric functions to obtain the unit normal vector information. A vertical... In the angular coordinate system, the positive X-axis is defined as due east, the positive Y-axis as due north, and the positive Z-axis as the vertically upward direction. The components of the normal vector can then be directly derived from the following formulas: the X-component equals the sine of the installation tilt angle multiplied by the sine of the installation azimuth angle; the Y-component equals the sine of the installation tilt angle multiplied by the cosine of the installation azimuth angle; and the Z-component equals the cosine of the installation tilt angle. The calculated normal vector information is then fused with the spatial positioning data, such as the latitude and longitude coordinates and altitude of the components. Simultaneously, the electrical connections between components, the shading relationships between adjacent components, and the terrain elevation data obtained from the digital elevation model are supplemented to construct a unified topological map representing the spatial distribution and geometric attributes of the photovoltaic power station.

[0019] After obtaining the topology map and normal vector information, the location of the light source is further determined by combining the image acquisition time, thereby generating a unit-level reflection risk field matrix. This unit-level reflection risk field matrix is ​​a three-dimensional spatial matrix, whose rows and columns correspond to the X, Y, and Z coordinates in space, respectively. The value of each element in the matrix represents the normalized value of specular reflection intensity that will be experienced when looking at the target photovoltaic module from that spatial location. The generation process first requires calculating the solar altitude angle and solar azimuth angle using astronomical algorithms based on the geographical coordinates of the photovoltaic power station and the image acquisition time information. The solar altitude angle is the angle between the sunlight and the ground plane, and the solar azimuth angle is the angle between the sunlight and the ground plane. The angle between the projection onto the ground plane and due north is collectively referred to as the light source position information. In the calculation, the current date is first converted into Julian days, then the ecliptic longitude and latitude of the sun are calculated, and the right ascension and declination are derived. Finally, the local latitude and local sidereal time are combined to convert it into horizontal coordinates, and corrections such as the Earth's orbital ecliptic, obliquity of the ecliptic, and atmospheric refraction are introduced into the calculation. Based on the obtained altitude and azimuth angles, the direction of sunlight is constructed as a unit vector, whose X component is equal to the cosine of the solar altitude angle multiplied by the sine of the solar azimuth angle, the Y component is equal to the cosine of the solar altitude angle multiplied by the cosine of the solar azimuth angle, and the Z component is equal to the sine of the solar altitude angle.

[0020] Then, the spatial angle between the direction vector of sunlight and the aforementioned normal vector is calculated. Substituting the components of the two vectors into the vector dot product formula and obtaining the inverse cosine value gives the spatial angle. The size of the spatial angle directly determines the principal direction of the specular reflection. The smaller the angle, the closer the sunlight is to perpendicular incidence. At this time, the direction of specular reflection is close to the normal vector, and the risk area of ​​strong reflection is concentrated near the normal vector. The larger the angle, the more the reflection direction deviates significantly from the normal vector.

[0021] After obtaining the spatial angle, it is input into a bidirectional reflection distribution function model based on Fresnel's law of reflection to calculate the specular reflection intensity distribution on the component surface in different outgoing directions. The bidirectional reflection distribution function is a mathematical model describing the optical reflection characteristics of an object's surface; its inputs are the incident light direction and the observation direction, and its output is the relative intensity of the reflected light. In this embodiment, the model is constructed based on Fresnel's law of reflection and micro-surface element theory. Fresnel's law of reflection reveals the energy distribution relationship when light undergoes reflection and refraction at interfaces of media with different refractive indices. For the glass cover of the photovoltaic module surface... The reflectivity of a plate increases with the angle of incidence. The micro-element theory treats the actual rough surface as being composed of a large number of randomly oriented tiny planes. Each micro-element independently reflects light according to Fresnel's law, forming a reflection lobe with a certain degree of diffusion on a macroscopic scale. In specific calculations, the basic reflectivity at the angle of incidence is first determined by applying Fresnel's formula based on the spatial angle and the glass refractive index. Then, a Gaussian distribution function characterizing the surface roughness is introduced to describe the statistical dispersion of the micro-element normal vector. By integrating the reflection directions in the hemispherical space, a conical reflection intensity distribution map centered on the ideal specular reflection direction is output.

[0022] Finally, a three-dimensional space covering the entire potential flight area of ​​the UAV is delineated around the target photovoltaic module, and discretized into regular grid nodes with a spatial resolution of meters. For each grid node, the line-of-sight vector pointing from the node to the target photovoltaic module is calculated, and it is determined whether the line-of-sight vector falls within the specular reflection cone defined by the bidirectional reflection distribution function model. If the line-of-sight vector falls within the cone, a normalized reflection risk value is assigned based on the corresponding reflection intensity in that direction. If it does not fall within the cone, the reflection risk value of the node is zero. After assigning values ​​to all nodes, a unit-level reflection risk field matrix corresponding to the module is formed. Since the installation tilt angle, azimuth angle, and solar position of different photovoltaic modules are different, this matrix is ​​dynamically calculated and updated for each module during the inspection period.

[0023] Through the above process, this step organically integrates astronomical illumination calculation, three-dimensional spatial geometric modeling of components, and optical reflection characteristics, generating a spatial distribution map of reflection risk at the unit level. This provides prior knowledge for adaptive inspection path planning that avoids strong light interference, helping to improve the usability of imaging data from the source. In this application, the subsequent removal of high-reflection-risk areas based on this matrix can reduce the probability of abnormal features of photovoltaic modules being obscured due to specular reflection, thereby ensuring the reliable detection of minute defects such as microcracks and hot spots during the inspection process.

[0024] S2: Eliminate high-risk areas in the unit-level reflection risk field matrix to obtain candidate observation spatial domains. Use preset pixel resolution and obstacle avoidance distance to filter candidate observation spatial domains and determine the unit-level diagnostic viewing window set of the target photovoltaic module.

[0025] S2 includes: traversing the unit-level reflection risk field matrix, removing grid nodes with reflection risk values ​​greater than a preset risk threshold, and forming a candidate observation airspace from the grid nodes that were not removed; obtaining the camera intrinsic parameters of the camera carried by the UAV, performing inverse optical imaging calculations in combination with the preset pixel resolution to obtain the upper limit of the effective shooting distance that meets the pixel resolution requirements, using the UAV's safe obstacle avoidance distance as the lower limit of the effective shooting distance, and forming an effective shooting distance interval from the lower limit of the effective shooting distance to the upper limit of the effective shooting distance; calculating the spatial distance from each grid node in the candidate observation airspace to the surface of the target photovoltaic module, retaining the grid nodes whose spatial distances fall within the effective shooting distance interval to form a preliminary screening observation airspace, and using the preliminary screening observation airspace as a set of unit-level diagnostic viewpoint windows.

[0026] Specifically, after generating the unit-level reflection risk field matrix representing the distribution of specular reflection risk around the photovoltaic module in S1, if the entire spatial range of this matrix is ​​directly used as the candidate observation position of the UAV, although the risk value in some areas is low, the image details may still be insufficient due to the shooting distance being too far, or the collision risk may be caused by the shooting distance being too close. To solve this technical problem, this step jointly quantifies the reflection risk with imaging resolution and flight safety constraints, and extracts the set of unit-level diagnostic viewpoint windows that can actually be shot from the reflection risk field matrix, providing highly reliable spatial constraints for subsequent flight path planning.

[0027] First, high-risk areas are eliminated from the unit-level reflection risk field matrix to generate candidate observation spatial domains. Each grid node in the reflection risk field matrix is ​​assigned a normalized reflection risk value, which represents the specular reflection intensity that the target photovoltaic module will experience when photographed from that node location. A risk threshold is preset, which is determined based on the imaging device's tolerance to overexposure and the dynamic range of the subsequent image processing algorithm. All grid nodes in the reflection risk field matrix are traversed, and nodes with reflection risk values ​​greater than the preset risk threshold are eliminated and no longer considered as shooting locations. All the nodes that are not eliminated together constitute an irregular three-dimensional spatial region, i.e., the candidate observation spatial domain. This operation pre-excludes observation locations that will inevitably cause severe overexposure, resulting in the inability to recover the surface defects of the module such as cracks and hot spots, thus fundamentally ensuring the basic imaging quality of the candidate locations.

[0028] Subsequently, the effective shooting distance range is determined based on pixel resolution and safe obstacle avoidance distance. Pixel resolution is defined as the actual physical size corresponding to a single pixel in the image, and is a key indicator for assessing whether minute defects on the surface of photovoltaic modules can be identified. To achieve the detection of millimeter-level or sub-millimeter-level defects, the pixel resolution during actual shooting must be controlled within a preset threshold. This step obtains the camera intrinsic parameters of the camera mounted on the UAV, including at least the lens focal length, the physical size of the sensor imaging area, and the number of pixels in that direction. Using the geometric relationship of optical imaging, the shooting distance and pixel resolution can be correlated. Pixel resolution equals shooting distance multiplied by the physical size of a single pixel of the sensor, then divided by the lens focal length. The physical size of a single pixel is calculated by dividing the width of the sensor imaging area by the number of horizontal pixels. By reversing this relationship, given the upper limit of the target pixel resolution, the farthest shooting distance that meets the requirements can be calculated, which is the upper limit of the effective shooting distance. For example, when the pixel resolution is required to not exceed a certain value, the upper limit of the effective shooting distance is equal to that value multiplied by the lens focal length and divided by the physical size of a single pixel of the sensor. At the same time, in order to ensure that the drone does not collide with photovoltaic modules, brackets or terrain during flight, the drone's safe obstacle avoidance distance needs to be used as the lower limit of the effective shooting distance. The safe obstacle avoidance distance is determined comprehensively based on the drone's positioning accuracy, control response delay and safety margin, and is the minimum spatial distance allowed between the drone and the target module. The lower limit and the upper limit calculated above together constitute a continuous effective shooting distance range.

[0029] Next, the candidate observation airspace is filtered using the effective shooting distance range to form the final set of diagnostic view windows. For each grid node in the candidate observation airspace, the spatial distance from the node to the surface of the target photovoltaic module is calculated. This spatial distance can be the Euclidean distance from the node's spatial coordinates to the plane of the module surface, with the geometric center of the module or a corner point as the reference point. The calculated spatial distance is compared with the effective shooting distance range. Only when the distance is between the lower and upper limits is the grid node retained. All nodes retained after distance filtering constitute a preliminary observation airspace, which serves as the unit-level diagnostic view window set for the current target photovoltaic module. Each observation position in this set simultaneously satisfies three constraints: reflection risk is below the allowable threshold, imaging resolution is better than the preset requirements, and the flight distance exceeds the lower limit of safe obstacle avoidance. Due to the different installation orientations and surrounding environments of different photovoltaic modules, each module will generate a unique set of diagnostic view windows, thus providing precise spatial guidance for adaptive inspection of each module.

[0030] This step directly embeds optical pixel resolution constraints and flight safety constraints into the spatial screening process, making the generated set of diagnostic viewpoint windows inherently engineering-operable. It avoids the risk of collisions caused by forcibly approaching components to meet resolution requirements, or the loss of details due to excessive conservatism and excessive distance from components. It inherits the pre-calculated results of the reflection risk field and provides a clean and safe feasible domain for the conflict-free planning of subsequent multi-component inspection routes, enabling the entire inspection method to efficiently and safely acquire high-quality images that meet diagnostic accuracy in practical applications.

[0031] S3: Based on the topology map and the set of unit-level diagnostic view windows, plan the partition inspection route and the corresponding sensor pose sequence, and control the UAV to execute the partition inspection route to collect initial image data.

[0032] S3 includes: using a clustering algorithm to divide all target photovoltaic modules into multiple inspection zones; calculating the diagnostic priority value of each inspection zone based on the historical fault frequency of the target photovoltaic modules; determining the traversal order of all inspection zones according to the diagnostic priority value in descending order; extracting position nodes as initial track points from the corresponding unit-level diagnostic viewpoint window set within each inspection zone according to the traversal order; generating a local track point sequence based on a preset image overlap rate; smoothing the local track point sequence using a curve fitting algorithm to generate a zone inspection route; calculating the direction vector of each local track point in the local track point sequence pointing to the center of the corresponding target photovoltaic module; converting the direction vector into pitch and yaw angles in the camera coordinate system; and combining it with a preset roll angle to generate a sensor pose sequence synchronized with the zone inspection route.

[0033] Specifically, after obtaining the topological connection relationship of each photovoltaic module and the corresponding set of unit-level diagnostic view windows, although the problem of where to shoot a single module to balance reflection safety and imaging clarity has been solved, the photovoltaic power station has a large number of modules and a complex spatial distribution. If the shooting position is planned in isolation for each module, it is easy to cause the UAV to have an excessively long flight range and run out of power before it can complete all the inspection tasks. At the same time, it is difficult to ensure key coverage of areas with high historical fault incidence. This application divides all target photovoltaic modules into several inspection zones according to spatial clustering, sets the priority traversal order according to the fault history, and then uses view windows and image overlap constraints to generate a smooth and flyable zone inspection route and a synchronized sensor pose sequence within each zone, so that the UAV can complete the initial image data acquisition of the entire station in sequence and efficiently.

[0034] During implementation, a clustering algorithm is first used to divide all target photovoltaic modules into multiple inspection zones. The clustering algorithm referred to here is an unsupervised clustering calculation method based on the spatial coordinates of the photovoltaic modules on the topology map. Specifically, the k-means clustering algorithm can be used, with the three-dimensional spatial coordinates of the center point of each module as the input feature. An appropriate number of clusters is set so that the modules in each zone are relatively concentrated in space and the size of the zone matches the coverage capability of a single flight of the UAV. Through clustering, the numerous modules that were originally distributed throughout the entire power plant are organized into several spatial regions that can be traversed in one go, thereby decomposing the large-scale inspection task into multiple sub-tasks of controllable size.

[0035] After obtaining the inspection zones, the diagnostic priority value of each inspection zone is calculated based on the historical fault frequency of the target photovoltaic modules. The historical fault frequency refers to the ratio of the number of times each module has been detected and confirmed to have faults such as microcracks, hot spots, and junction box abnormalities in the past period to the total number of detections, reflecting the potential risk of the module continuing to fail. To avoid statistical bias caused by different numbers of modules in different zones, the average historical fault frequency of each module in each inspection zone is first taken, and then the average fault frequency of each zone is normalized so that the diagnostic priority values ​​of different zones are within the same dimension range. The normalized diagnostic priority value is the diagnostic priority value of the inspection zone, and its magnitude represents the urgency of the overall fault investigation of the zone. Then, the traversal order of all inspection zones is determined according to the diagnostic priority value from large to small, ensuring that zones with high historical fault density and high potential abnormal risk are inspected first.

[0036] Following the predetermined traversal order, route planning is performed one by one within each inspection zone. For the currently being processed inspection zone, spatial location nodes are extracted as initial track points from the set of unit-level diagnostic view windows corresponding to each component within that zone. Specifically, for each target photovoltaic module within the zone, the grid node corresponding to the geometric center of the window can be selected from its set of diagnostic view windows, or multiple nodes can be sampled according to the spatial distribution of the window at a preset step size to form the initial track point set for that zone. Each of these initial track points meets the requirements of low reflection risk, high imaging resolution, and safe obstacle avoidance distance, but they have not yet formed a continuous track.

[0037] Next, a local waypoint sequence is generated based on the preset image overlap rate. The image overlap rate is the proportion of overlap between images acquired by the UAV at adjacent waypoints along the flight direction or scanning direction. Its value setting needs to meet the feature matching density requirements of subsequent image stitching and 3D reconstruction algorithms. For the initial waypoint set, the spatial step size between adjacent waypoints is calculated using the outer rectangle size of each component, the camera field of view, and the preset expected overlap rate. Then, along the component arrangement direction or the preset scanning direction, a dense waypoint sequence that satisfies the overlap constraint is interpolated or adjusted based on the initial waypoints. Each node of this sequence simultaneously satisfies the shooting safety constraint, resolution constraint, and image continuity constraint, forming the local waypoint sequence of the inspection zone.

[0038] After obtaining the local track point sequence, a curve fitting algorithm is used to smooth it and generate a zone inspection route. The curve fitting algorithm is a method to fit a discrete spatial point sequence into a continuous and differentiable spatial curve. It can be implemented using cubic B-spline curves or Bézier curves. The algorithm uses the nodes in the local track point sequence as control points or shape points to generate a spatial curve that passes through or approximates these nodes and satisfies the condition of continuous curvature. This processing can avoid the energy waste and instability of the UAV due to sudden stops and turns when flying in a straight line between nodes, and also ensure that the route does not deviate from the allowable shooting range in space.

[0039] After the flight path is generated, the camera shooting attitude needs to be determined for each local track point on the flight path, and a sensor pose sequence synchronized with the zonal inspection flight path is generated. For each local track point in the local track point sequence, the direction vector pointing to the center of the corresponding target photovoltaic module is calculated. This vector is the target direction that the camera optical axis should be aligned with when shooting at that point. This direction vector is transformed from the world coordinate system or the local coordinate system of the power station to the camera coordinate system of the UAV, and the camera pitch angle and yaw angle are obtained. The pitch angle is the angle between the optical axis and the horizontal plane, and the yaw angle is the angle between the projection of the optical axis on the horizontal plane and the reference direction. Combined with the preset roll angle, which is usually set to zero to ensure that the image horizon is aligned with the edge of the module, the three Euler angles together constitute the complete attitude command of the sensor at the track point. The position information and attitude information of each point on the entire flight path are arranged in sequence to obtain the sensor pose sequence synchronized with the zonal inspection flight path.

[0040] Finally, the image acquisition system on the UAV flew over each inspection zone in sequence according to the planned inspection route and sensor pose sequence, triggering the camera to acquire images at each track point to complete the acquisition of initial image data for the entire station. Throughout the process, the route planning and pose generation worked closely together, ensuring that the camera was accurately aligned with the target component at each exposure moment and took pictures from a safe position inside the diagnostic view window.

[0041] By employing a strategy of first partitioning, then sorting, and finally planning, the complex photovoltaic power station inspection task is broken down into multiple sub-tasks executed sequentially according to the urgency of the fault. This reduces the time and computing power consumption of single route planning and enables drones to prioritize coverage of high-risk areas within their limited flight time. Simultaneously, the joint planning of flight path and attitude ensures that the camera always operates under safe and optimal shooting conditions during flight, providing stable, complete, and high-resolution initial image data for subsequent anomaly detection. This fundamentally improves the practical usability and defect detection efficiency of unmanned photovoltaic power station inspections.

[0042] S4: Calculate the imaging quality score of the initial image data. When the imaging quality score is lower than the preset threshold, extract alternative viewpoints from the set of unit-level diagnosable viewpoint windows to generate a reshoot task, and control the drone to execute the reshoot task to obtain reshoot image data.

[0043] S4 calculates the imaging quality score of the initial image data, including: performing convolution processing on the initial image data using the Laplacian operator and calculating the variance of the convolution response values ​​as the sharpness score; generating a brightness histogram of the initial image data and calculating the distribution ratio of pixels in the brightness histogram within a preset intermediate grayscale range as the exposure score; using an edge detection algorithm to identify the outline of the target photovoltaic module in the initial image data, calculating the number of target pixels enclosed by the outline, and using the ratio of the number of target pixels to the total number of pixels in the initial image data as the target coverage score; and using multiple preset weighting coefficients to perform a weighted average calculation of the sharpness score, exposure score, and target coverage score to obtain the imaging quality score.

[0044] S4 involves extracting alternative viewpoints to generate a reshoot task, including: when the imaging quality score is lower than a preset threshold, performing grayscale threshold segmentation on the initial image data to extract the pixel coordinate range of the local image region that does not meet the quality requirements; obtaining the camera intrinsic parameters and sensor pose corresponding to the sensor pose sequence in the initial image data, back-projecting the pixel coordinate range onto the three-dimensional coordinate system of the target photovoltaic module to obtain the local region coordinates; extracting alternative viewpoints that were not selected as initial waypoints from the set of unit-level diagnosable viewpoint windows, calculating the reflection risk value of each alternative viewpoint, and using the alternative viewpoint with the lowest reflection risk value as the reshoot waypoint coordinates; calculating the direction vector from the reshoot waypoint coordinates to the center of the local region coordinates to obtain the sensor pose required for reshoot, and encapsulating it as a local reshoot task and inserting it into the current inspection route of the UAV.

[0045] Specifically, in order to detect and correct problems such as individual image blurring, overexposure, underexposure, or component deviation from the center of the field of view caused by wind disturbance, positioning deviation, or transient changes in lighting during flight, this application initiates online image quality evaluation and closed-loop reshooting after the initial image data acquisition at each waypoint is completed.

[0046] The image quality score is obtained by combining three indicators: The first is sharpness, which uses the Laplacian operator to perform spatial filtering convolution on the initial image data and calculates the variance of the convolution response value of the entire image. The Laplacian operator, as a second-order differential operator, can enhance edge and texture details in the image; its response variance directly corresponds to the richness of high-frequency information in the image, with a larger variance indicating higher sharpness. The second is exposure, which first generates a grayscale histogram of the initial image data and then counts the proportion of pixels distributed within a preset middle grayscale range of the histogram to the total number of pixels in the image. A higher proportion indicates more normal image exposure, avoiding large-area shifts towards completely black or completely white areas. The third is target coverage, which uses gradient-based algorithms such as Canny edge detection to identify target photovoltaic modules in the image. The outer contour is obtained and its circumscribed range is acquired. Then, the total number of pixels contained within the contour is counted and the ratio is calculated with the total number of pixels in the initial image data. This ratio directly reflects whether the area occupied by the inspected component in the whole image is sufficient. The Canny operator can stably and completely extract edges by calculating image gradients, non-maximum suppression, and double threshold hysteresis connection, thus providing an accurate contour basis for area statistics. The three scores have different dimensions and need to be normalized before weighted combination, uniformly mapping them to the same scoring scale. The normalization method adopts a linear mapping between each score value and its normal reference upper limit in engineering statistics. The weight coefficients are adjusted through repeated testing. The sharpness score weight is 0.4, the exposure score and the target coverage score are each 0.3. The three are weighted and added together to obtain the image quality score.

[0047] The calculation results are compared with a preset threshold, which depends on the minimum acceptable threshold of image information fidelity for subsequent defect detection algorithms. When the score is lower than this threshold, it indicates that the current image is unusable, and reshooting is initiated. The first step of reshooting is to determine where the problem lies, that is, the coordinates of the local area. For abnormal areas found in the evaluation analysis, the corresponding problem area localization method is adopted: for overexposed or underexposed areas caused by exposure abnormalities, grayscale threshold segmentation based on brightness threshold is directly performed on the initial image, and the row and column boundaries of pixel patches with grayscale values ​​exceeding the normal range are recorded as the pixel coordinate range of the local image area; while for areas with insufficient sharpness... For blurred regions, threshold segmentation is performed on the distribution map reflecting the image gradient intensity to identify low-resolution pixel patches with gradient responses below a given threshold. The pixel coordinates within the contour range of these patches are used as the pixel coordinate range of the local image region. Next, the camera intrinsic parameters and sensor pose sequences used when the image was captured are queried to find the sensor poses that strictly correspond to the image frame. These parameters are used to construct a back-projection mapping relationship from the two-dimensional image plane to the three-dimensional component surface. Using this mapping, the pixel coordinate range of the image plane is converted into the actual spatial coordinate range in the three-dimensional coordinate system of the target photovoltaic component, thus obtaining the local region coordinates.

[0048] Having identified the location of the problem area on the component, the focus shifted to finding solutions from the set of diagnostic viewpoint windows at the unit level. This set of windows had already generated multiple safe and reachable candidate shooting views for each component during the initial flight path planning. This time, the window blocks not selected as initial waypoints were directly retrieved as candidate views. Each candidate viewpoint had its corresponding reflection risk value calibrated during the initial construction phase. The viewpoint with the lowest reflection risk value was selected, and its spatial coordinates were set as the coordinates of the reshoot waypoint. This approach ensured in one step that the reshoot waypoint was also located in the safe airspace with the lowest reflection risk. Then, the coordinates from... The vector pointing to the center of the local area coordinates for the reshoot waypoint is the direction the camera's optical axis should be aligned with during the reshoot. By converting this direction vector to the camera coordinate system, the pitch and yaw angles are decomposed and combined with the roll angle which is preset to zero to form the sensor pose required for the reshoot. This pose and the reshoot waypoint coordinates are encapsulated into a local reshoot task command. The flight control system inserts the reshoot waypoint into the currently executing inspection route in the task queue, and the UAV is scheduled to perform the reshoot immediately after the current waypoint task is completed. After the reshoot is completed, the UAV automatically returns to the original route to continue the subsequent tasks.

[0049] The above process expands the dimension of image quality control from the initial static airspace planning to dynamic real-time feedback and correction during flight. It can automatically detect and correct individual data defects caused by on-site interference that is difficult to completely avoid in actual flight. This solves the problem of having to arrange the entire flight to return and retake due to the failure of a single image in traditional inspections. Instead, it allows for autonomous reshooting within a few seconds during flight. This online adaptive capability also reuses the diagnostic viewpoint information obtained in the early stage without introducing additional computational load. As a result, it ensures from a mechanism perspective that every component image that finally enters the defect analysis stage is reliable and complete, laying a solid data foundation for accurate anomaly detection without human intervention.

[0050] S5: Spatial registration is performed between the initial image data and the re-shot image data to extract the dynamic highlight area. The reflection artifact area is separated by calculating the correlation between the dynamic highlight area and the unit-level reflection risk field matrix. After excluding the reflection artifact area, the abnormal candidate area is extracted. The abnormal candidate area and the reflection artifact area are mapped to geographic coordinate information respectively, and the zonal inspection route and the unit-level reflection risk field matrix are updated based on the corresponding geographic coordinate information.

[0051] S5 involves extracting candidate regions for anomalies, including: extracting matching feature points between the initial image data and the re-enhanced image data; solving the homography transformation model based on the matching feature points to resample the re-enhanced image data, obtaining a registered image aligned with the initial image data; calculating the brightness difference between the initial image data and the registered image pixel by pixel; extracting connected components with brightness differences greater than a preset difference threshold as dynamic highlight regions; back-projecting the dynamic highlight regions onto a three-dimensional physical coordinate system; querying the corresponding reflection risk value in the unit-level reflection risk field matrix; calculating the correlation matching degree between the reflection risk value and the brightness difference of the dynamic highlight regions; marking regions with correlation matching degrees higher than a preset matching threshold as reflection artifact regions; excluding reflection artifact regions in the registered image to obtain unlabeled connected components; calculating the texture feature value and brightness statistics of each unlabeled connected component to initially identify anomalies; and verifying the consistency of the initially identified anomalies by combining the historical inspection data of neighboring components recorded in the topology graph, determining the verified regions as candidate regions for anomalies.

[0052] In S5, mapping to geographic coordinate information and updating the zonal inspection route and unit-level reflection risk field matrix based on geographic coordinate information includes: extracting the centroid pixel coordinates of anomaly candidate regions on the two-dimensional image; using camera intrinsic parameters, sensor pose, and track point coordinates when acquiring the image to convert the centroid pixel coordinates into geographic coordinate information in absolute physical space; constructing a geometric bounding box for photovoltaic modules based on latitude and longitude coordinates; performing spatial intersection judgment between geographic coordinate information and the geometric bounding box to lock the target photovoltaic module and extract the corresponding unit identification information, and inputting it along with the geographic coordinate information into the storage module to generate anomaly records; using the same mapping method to obtain the geographic coordinate information of the separated reflection artifact regions, and inputting it along with the corresponding unit identification information into the storage module to generate reflection artifact records; calculating the spatial distribution density of anomaly records, and increasing the diagnostic priority of the zonal where the target photovoltaic module is located based on the spatial distribution density to update the configuration parameters of the zonal inspection route; comparing the actual reflection location in the reflection artifact records with the prediction results of the unit-level reflection risk field matrix to obtain the prediction deviation, correcting the installation tilt angle in the topology map based on the prediction deviation, and updating the unit-level reflection risk field matrix based on the corrected installation tilt angle.

[0053] Specifically, after obtaining the initial and retaken images of the same component through S4, S5 begins to extract information differences from these two multi-view images. The overexposed bright spots caused by specular reflection are extremely sensitive to the shooting angle and often appear intermittently in the first two shots, while physical damage such as cracks and hot spots does not change with the viewing angle. This application utilizes this phenomenon to spatially register the images before and after the retake, and remove the reflection artifacts that only appear due to the change in viewing angle. After eliminating these false bright signals, the real component defect area is accurately extracted, and finally the defect coordinates are mapped to the geographic information and power plant topology database. At the same time, the flight path priority and reflection prediction model are corrected, making future inspections more and more accurate.

[0054] In practice, the initial image data and the re-captured image data are first spatially registered to generate a registered image that is perfectly aligned with the initial image. Significant geometric features, such as corner points with high eigenvalues ​​in the gradient covariance matrix, are extracted from both images as matching feature points. Feature descriptors of the gradient and texture distribution around each feature point are calculated. Candidate matching point pairs are selected by comparing the distances between feature descriptors. Then, a random sampling consensus algorithm is used to remove mismatched points with inconsistent geometric positions. The remaining reliable matching point pairs are used to solve the homography transformation model. This homography transformation model uses a homography matrix to express the projection mapping relationship of the same physical plane in space under the poses of two cameras. The matrix parameters are solved from the coordinates of the matching point pairs using the weighted least squares method and applied to the re-captured image data. Through coordinate mapping and bilinear interpolation resampling, a registered image aligned with the initial image data at the pixel level is obtained.

[0055] Next, the brightness difference is calculated pixel by pixel on the two aligned images. The absolute value is used to represent the brightness fluctuation caused by the change of viewing angle. Pixels with brightness differences exceeding the preset difference threshold are marked, and connected component aggregation is performed on these marked pixels. Each continuous block formed is a dynamic highlight area. They are called dynamic because their highlight state changes with the shooting posture, which is exactly the typical behavior of reflection artifacts.

[0056] To confirm that these dynamically highlighted areas are genuine reflection artifacts rather than other transient changes, they need to be correlated with a pre-established unit-level reflection risk field matrix. First, the two-dimensional contour of each dynamically highlighted area, combined with the camera intrinsic parameters, sensor pose, and trackpoint coordinates at the time of initial image acquisition, is back-projected onto the three-dimensional physical coordinate system of the component surface through a rigid body transformation from the camera coordinate system to the world coordinate system, thus locating the actual position of the area. Then, the spatial grid points of the unit-level reflection risk field matrix corresponding to the initial trackpoint at the time of image capture are retrieved, and the reflection risk value covering the area is queried. Finally, multi-dimensional features are integrated to... The correlation matching degree is calculated, which involves not only judging the spatial distribution of dynamic bright areas and the geometric overlap of high reflection risk areas, but also quantitatively assessing the boundary sharpness and internal brightness saturation characteristics of bright areas. These indicators are normalized and weighted according to empirically determined weights, such as a correlation coefficient weight of 0.6, an absolute value of reflection risk weight of 0.2, and a regional feature consistency weight of 0.2. The weighted sum is then used to obtain the final correlation matching degree. When the matching degree is higher than the preset matching threshold, it means that the optical behavior of the area highly conforms to the specular reflection law, and it is marked as a reflection artifact area, recorded separately, and prepared for removal.

[0057] After excluding all labeled reflection artifact regions from the registered image, the remaining unlabeled connected components in the field of view are then analyzed using texture and brightness to identify true anomalies. The gray-level co-occurrence matrix of each unlabeled connected component is calculated, and texture features such as contrast and entropy are extracted. At the same time, the mean and standard deviation of the brightness of the pixels within each component are statistically analyzed. Normal component surfaces have uniform and stable textures. Any region that causes the above statistics to deviate significantly from the normal baseline will be initially identified as an abnormal region. For this preliminary result, the historical inspection data of neighboring components recorded in the topology map are retrieved for consistency verification. If neighboring regions, especially components in the same string, also show similar abnormal behavior in the historical records, or if the efficiency of the same string also shows efficiency degradation during the same period, then the credibility of the initially identified abnormal region is enhanced through cross-verification of spatial correlation and electrical correlation, and it is finally confirmed as an abnormal candidate region.

[0058] Once the candidate anomaly region is identified, it is immediately mapped to geographic information and used to update the background parameters. First, the average value of all pixel coordinates in each candidate anomaly region is calculated to obtain its centroid pixel coordinates on the 2D image. Based on the camera intrinsic parameters, sensor pose, and absolute coordinates of the track point when the image was captured, the centroid pixel coordinates are sequentially transformed through the image coordinate system, camera coordinate system, and UAV body coordinate system, and finally into the world geodetic coordinate system to obtain the geographic coordinate information corresponding to the anomaly point, namely longitude, latitude, and altitude.

[0059] After obtaining the geographic coordinates, spatial intersection determination is performed between them and the pre-generated geometric bounding boxes of each photovoltaic module in the topology database. The geometric bounding boxes are spatial ranges pre-constructed in the topology map based on the module's installation location and external dimensions. By determining which bounding box the anomaly point coordinates fall into, the target photovoltaic module with the anomaly is located, and its unit identification information, including string number and module serial number, is extracted. The unit identification and geographic coordinate information are combined to generate an anomaly record containing a timestamp, anomaly type, and module location, which is then written to the storage module. Similarly, for the reflection artifact areas marked and separated in the previous steps, the same coordinate transformation and bounding box intersection determination method is used to obtain their three-dimensional geographic coordinate information and corresponding unit identification information, which are then entered into the storage module to generate independent reflection artifact records.

[0060] These two types of accumulated records will be used to intelligently optimize different parameters of the inspection system. For abnormal records, by statistically analyzing the spatial distribution density of abnormal records within each zone, hotspots with frequent recent faults can be immediately identified. Based on this, the diagnostic priority of the zone containing the target photovoltaic module is automatically increased, and this priority value is updated in the configuration parameters of the zone's inspection route, ensuring that high-risk areas are prioritized and scanned more thoroughly in the next inspection. At the same time, by extracting the actual reflection location from the reflection artifact records, it is compared with the previously predicted risk value in the unit-level reflection risk field matrix to obtain the prediction deviation. For example, if a certain location repeatedly reflects at an angle predicted to be low-risk, it is very likely that the installation tilt angle recorded for that module in the topology diagram is incorrect. Based on this prediction deviation, the corresponding installation tilt angle in the topology diagram is automatically corrected, and the geometric relationship between the sun, the module surface, and the camera in the formula is recalculated based on the new tilt angle, updating the unit-level reflection risk field matrix and smoothing out the mismatch between perception and reality. In this way, the entire inspection knowledge system evolves itself with each flight, continuously improving execution efficiency and diagnostic accuracy.

[0061] This step enables the inspection system to separate artifacts from data and use these artifacts to make corrections. It transforms tedious image comparisons into an effective means of distinguishing between genuine and fake defects. By utilizing every on-site measured reflection event and anomaly, it dynamically resets flight path priorities and calibrates the 3D reflection model. This eliminates the excessive reliance on manual review in large-scale power plant inspections and elevates the repetitive inspection process into a closed-loop diagnostic system with self-learning and optimization capabilities. This is the core of ensuring efficient, low-cost, and long-term accurate operation and maintenance of photovoltaic power plants.

[0062] Through the coordination between S1 to S5, this application realizes a full-chain intelligent inspection closed loop, from static optical prediction obstacle avoidance, dynamic route joint planning, in-flight quality closed-loop control, to the final multi-view false detection and self-evolution of the underlying data model. This effectively eliminates the problems of missed detection and false detection caused by specular reflection interference and environmental changes, and greatly improves the accuracy of unmanned operation and maintenance of photovoltaic power plants and the system's self-optimization capability.

[0063] Example 2: The above describes a method for inspecting a photovoltaic power station based on a drone, as described in the embodiments of this application. The following describes a system for inspecting a photovoltaic power station based on a drone, as described in the embodiments of this application. Figure 2 As shown in the figure, a photovoltaic power station inspection system based on a drone in this application embodiment includes: The reflection risk module is used to obtain the topology map and image acquisition time of the target photovoltaic module, extract the normal vector of the target photovoltaic module based on the topology map, and calculate the reflection intensity distribution on the surface of the target photovoltaic module by combining the light source position determined by the image acquisition time, and generate a unit-level reflection risk field matrix.

[0064] The viewing angle filtering module is used to remove high-risk areas in the unit-level reflection risk field matrix to obtain candidate observation spatial domains. It uses preset pixel resolution and obstacle avoidance distance to filter candidate observation spatial domains and determine the set of unit-level diagnostic viewing angle windows for the target photovoltaic module.

[0065] The route planning module is used to plan the zone inspection routes and corresponding sensor pose sequences based on the topology map and the set of unit-level diagnostic view windows, and control the UAV to execute the zone inspection routes to collect initial image data.

[0066] The reshooting task module is used to calculate the imaging quality score of the initial image data. When the imaging quality score is lower than a preset threshold, it extracts alternative viewpoints from the set of unit-level diagnosable viewpoint windows to generate a reshooting task and controls the drone to execute the reshooting task to obtain reshooting image data.

[0067] The anomaly update module is used to spatially register the initial image data with the re-captured image data to extract the dynamic highlight area. It separates the reflection artifact area by calculating the correlation between the dynamic highlight area and the unit-level reflection risk field matrix. After excluding the reflection artifact area, it extracts the anomaly candidate area. It maps the anomaly candidate area and the reflection artifact area to geographic coordinate information respectively, and updates the zonal inspection route and the unit-level reflection risk field matrix based on the corresponding geographic coordinate information.

[0068] Through the collaborative efforts of the aforementioned components, this application further enhances the accuracy of unmanned operation and maintenance of photovoltaic power plants and the system's self-optimization capabilities.

[0069] In summary, this application provides a method and system for inspecting photovoltaic power plants based on unmanned aerial vehicles (UAVs). This method integrates astronomical illumination models, multi-view geometry, and automated control theory. First, by generating a unit-level reflection risk field matrix and combining it with spatial resolution and safety constraints, a precise, reflective, and diagnosable viewing window is created for the UAV, thereby planning the optimal zoned inspection route and camera pose. During execution, the system introduces an online quality evaluation and dynamic back-projection re-shooting mechanism to instantly repair inferior images affected by wind disturbances and light variations, avoiding high re-flight costs. Furthermore, by spatially registering and subtracting multi-view images, highly deceptive reflection artifacts are successfully removed, and confirmed anomaly data is used to feed back into the topology database, achieving automatic correction of route priority and basic tilt angle parameters. This application differs from the traditional open-loop blind flight operation mode of photovoltaic inspection, endowing the system with the ability to eliminate artifacts and self-evolve as a digital twin, providing a complete solution with significant commercial value and technological barriers for the efficient, accurate, and unmanned operation and maintenance of large-scale photovoltaic power plants.

[0070] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0071] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0072] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for inspecting photovoltaic power plants based on unmanned aerial vehicles (UAVs), characterized in that, The method includes: S1: Obtain the topology map and image acquisition time of the target photovoltaic module, extract the normal vector of the target photovoltaic module based on the topology map, and calculate the reflection intensity distribution on the surface of the target photovoltaic module by combining the light source position determined by the image acquisition time, and generate a unit-level reflection risk field matrix; S2: Eliminate high-risk areas in the unit-level reflection risk field matrix to obtain candidate observation spatial domains. Use preset pixel resolution and obstacle avoidance distance to filter candidate observation spatial domains and determine the unit-level diagnostic viewing window set of the target photovoltaic module. S3: Based on the topology map and the set of unit-level diagnostic view windows, plan the partition inspection route and the corresponding sensor pose sequence, and control the UAV to execute the partition inspection route to collect initial image data. S4: Calculate the imaging quality score of the initial image data. When the imaging quality score is lower than the preset threshold, extract alternative viewpoints from the set of unit-level diagnosable viewpoint windows to generate a reshooting task. Control the drone to execute the reshooting task to obtain reshooting image data. S5: Spatial registration is performed between the initial image data and the re-shot image data to extract the dynamic highlight area. The reflection artifact area is separated by calculating the correlation between the dynamic highlight area and the unit-level reflection risk field matrix. After excluding the reflection artifact area, the abnormal candidate area is extracted. The abnormal candidate area and the reflection artifact area are mapped to geographic coordinate information respectively, and the zonal inspection route and the unit-level reflection risk field matrix are updated based on the corresponding geographic coordinate information.

2. The photovoltaic power station inspection method based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, S1 obtains the topology of the target photovoltaic module and extracts the normal vector of the target photovoltaic module, including: The initial layout data of the photovoltaic power station is read. The initial layout data includes the unique identifier of the target photovoltaic module, latitude and longitude coordinates, altitude information, and installation tilt angle and azimuth angle. The orientation of the target photovoltaic module surface in three-dimensional space is calculated based on the installation tilt angle and azimuth angle to obtain the normal vector information. The initial layout data and normal vector information are fused to construct the topology map.

3. The photovoltaic power station inspection method based on unmanned aerial vehicles (UAVs) according to claim 2, characterized in that, S1 generates a unit-level reflection risk field matrix, including: Based on the image acquisition time information and the latitude and longitude coordinates of the target photovoltaic module, the solar altitude angle and solar azimuth angle are calculated, and used as the light source position information to construct the solar light direction vector; Calculate the spatial angle between the direction vector of sunlight and the normal vector, input the spatial angle into the bidirectional reflection distribution function model based on Fresnel's law of reflection, and calculate the specular reflection intensity distribution on the surface of the target photovoltaic module in different emission directions; A three-dimensional spatial region is delineated around the target photovoltaic module and divided into multiple grid nodes. The reflection risk value of each grid node is calculated based on the specular reflection intensity distribution, and a three-dimensional spatial matrix is ​​generated as the unit-level reflection risk field matrix.

4. The photovoltaic power station inspection method based on unmanned aerial vehicles (UAVs) according to claim 3, characterized in that, S2 includes: Traverse the cell-level reflection risk field matrix and remove grid nodes whose reflection risk value is greater than the preset risk threshold. The remaining grid nodes form the candidate observation space. The camera intrinsic parameters of the camera on the drone are obtained, and optical imaging is reverse-calculated in combination with the preset pixel resolution to obtain the upper limit of the effective shooting distance that meets the pixel resolution requirements. The safe obstacle avoidance distance of the drone is used as the lower limit of the effective shooting distance. The effective shooting distance range is composed of the lower limit of the effective shooting distance and the upper limit of the effective shooting distance. Calculate the spatial distance from each grid node in the candidate observation space to the surface of the target photovoltaic module. Grid nodes whose spatial distance falls within the effective shooting distance range are retained to form the initial screening observation space. The initial screening observation space is used as a set of unit-level diagnostic view windows.

5. The photovoltaic power station inspection method based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, S3 includes: Clustering algorithms are used to divide all target photovoltaic modules into multiple inspection zones. Based on the historical failure frequency of the target photovoltaic modules, the diagnostic priority value of each inspection zone is calculated. The traversal order of all inspection zones is determined according to the diagnostic priority value from largest to smallest. According to the traversal order, within each inspection zone, position nodes are extracted from the corresponding set of diagnosable view windows at the unit level as initial track points, and a local track point sequence is generated based on the preset image overlap rate. A curve fitting algorithm is used to smooth the local track point sequence to generate a zone inspection route. The direction vector of each local track point in the local track point sequence pointing to the center of the corresponding target photovoltaic module is calculated. The direction vector is converted into pitch angle and yaw angle in the camera coordinate system. Combined with the preset roll angle, a sensor pose sequence synchronized with the zone inspection route is generated.

6. The photovoltaic power station inspection method based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, S4 calculates the imaging quality score of the initial image data, including: The initial image data is convolved using the Laplacian operator, and the variance of the convolution response value is used as the sharpness score. A brightness histogram of the initial image data is generated, and the distribution ratio of pixels in the brightness histogram within the preset intermediate gray range is used as the exposure score. An edge detection algorithm is used to identify the outline of the target photovoltaic module in the initial image data, and the number of target pixels enclosed by the outline is calculated. The ratio of the number of target pixels to the total number of pixels in the initial image data is used as the target coverage score. The image quality score is obtained by using a weighted average of the sharpness score, exposure score, and target coverage score using multiple preset weighting coefficients.

7. The photovoltaic power station inspection method based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, S4 extracts alternative viewpoints to generate reshoot tasks, including: When the imaging quality score is lower than a preset threshold, grayscale thresholding is performed on the initial image data to extract the pixel coordinate range of the local image region that does not meet the quality requirements. Obtain the camera intrinsic parameters corresponding to the initial image data and the corresponding sensor pose in the sensor pose sequence, and back-project the pixel coordinate range to the three-dimensional coordinate system of the target photovoltaic module to obtain the local region coordinates; Extract alternative viewpoints that were not selected as initial waypoints from the set of diagnosable viewpoint windows at the unit level, calculate the reflection risk value of each alternative viewpoint, and use the alternative viewpoint with the lowest reflection risk value as the coordinates of the reshoot waypoint. Calculate the direction vector from the waypoint coordinates of the reshoot to the center of the local area coordinates to obtain the sensor pose required for the reshoot, and encapsulate it as a local reshoot task and insert it into the current inspection route of the UAV.

8. The photovoltaic power station inspection method based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, S5 extracts anomaly candidate regions, including: Extract matching feature points between the initial image data and the re-shot image data, solve the homography transformation model based on the matching feature points to resample the re-shot image data, and obtain a registered image aligned with the initial image data. Calculate the brightness difference between the initial image data and the registered image pixel by pixel, and extract the connected components with brightness differences greater than a preset difference threshold as dynamic highlight regions. The dynamic highlight area is back-projected onto the three-dimensional physical coordinate system. The corresponding reflection risk value is queried in the unit-level reflection risk field matrix. The correlation matching degree between the reflection risk value and the brightness difference of the dynamic highlight area is calculated. Areas with a correlation matching degree higher than the preset matching threshold are marked as reflection artifact areas. Unlabeled connected regions are obtained by excluding reflection artifact regions in the registered image. The texture feature value and brightness statistics of each unlabeled connected region are calculated to initially identify abnormal regions. The consistency of the initially identified abnormal regions is verified by combining the historical inspection data of the neighboring components recorded in the topology graph. Regions that pass the verification are determined as abnormal candidate regions.

9. The photovoltaic power station inspection method based on unmanned aerial vehicles (UAVs) according to claim 2, characterized in that, In S5, information is mapped to geographic coordinates, and the zonal inspection routes and unit-level reflection risk field matrices are updated based on these geographic coordinates, including: Extract the centroid pixel coordinates of the anomaly candidate region on the 2D image, and use the camera intrinsic parameters, sensor pose and track point coordinates when the image is acquired to convert the centroid pixel coordinates into geographic coordinate information in absolute physical space. The geometric bounding box of the photovoltaic module is constructed based on latitude and longitude coordinates. The geographic coordinate information is spatially intersected with the geometric bounding box to lock the target photovoltaic module and extract the corresponding unit identification information. The geographic coordinate information is then entered into the storage module to generate an anomaly record. The same mapping method is used to obtain the geographic coordinate information of the separated reflection artifact area, and the corresponding unit identification information is entered into the storage module to generate a reflection artifact record. The spatial distribution density of abnormal records is statistically analyzed, and the diagnostic priority of the zone where the target photovoltaic module is located is increased based on the spatial distribution density, so as to update the configuration parameters of the zone inspection route. The prediction deviation is obtained by comparing the actual reflection location in the reflection artifact record with the prediction result of the unit-level reflection risk field matrix. The installation tilt angle in the topology diagram is corrected based on the prediction deviation, and the unit-level reflection risk field matrix is ​​updated based on the corrected installation tilt angle.

10. A drone-based photovoltaic power plant inspection system, used to implement the drone-based photovoltaic power plant inspection method as described in any one of claims 1-9, characterized in that, The system includes: The reflection risk module is used to obtain the topology map and image acquisition time of the target photovoltaic module, extract the normal vector of the target photovoltaic module based on the topology map, and calculate the reflection intensity distribution on the surface of the target photovoltaic module by combining the light source position determined by the image acquisition time, and generate a unit-level reflection risk field matrix. The viewing angle filtering module is used to remove high-risk areas in the unit-level reflection risk field matrix to obtain candidate observation spatial domains. It uses preset pixel resolution and obstacle avoidance distance to filter candidate observation spatial domains and determine the set of unit-level diagnosable viewing angle windows for the target photovoltaic module. The route planning module is used to plan the zone inspection route and the corresponding sensor pose sequence based on the topology map and the set of unit-level diagnostic view windows, and control the UAV to execute the zone inspection route to collect initial image data. The reshooting task module is used to calculate the imaging quality score of the initial image data. When the imaging quality score is lower than a preset threshold, it extracts alternative viewpoints from the set of unit-level diagnosable viewpoint windows to generate a reshooting task and controls the drone to execute the reshooting task to obtain reshooting image data. The anomaly update module is used to spatially register the initial image data with the re-captured image data to extract the dynamic highlight area. It separates the reflection artifact area by calculating the correlation between the dynamic highlight area and the unit-level reflection risk field matrix. After excluding the reflection artifact area, it extracts the anomaly candidate area. It maps the anomaly candidate area and the reflection artifact area to geographic coordinate information respectively, and updates the zonal inspection route and the unit-level reflection risk field matrix based on the corresponding geographic coordinate information.

Images