A photovoltaic power station unmanned aerial vehicle inspection method and system

Abstract

The application provides a photovoltaic power station unmanned aerial vehicle inspection method and system, and relates to the technical field of photovoltaic intelligent operation and maintenance.The application collects visible light, infrared and illumination data through benchmark initialization inspection, extracts three types of features of texture, temperature field and illumination response after pretreatment, splices to generate component multi-dimensional fingerprints and establishes a benchmark library;according to the fingerprint stability, the components are divided into three levels of state, and the sparse and dense differentiated inspection routes are adaptively generated;by comparing the time sequence deviation of real-time fingerprints and benchmark fingerprints, the dominant faults such as shielding and dirt are determined according to short-time mutation, and the recessive faults such as aging, micro-hidden crack and power attenuation are determined according to continuous drift;and the benchmark data and component state are dynamically iterated and updated, the route is closed-loop optimized, and the double-layer inspection report containing the dominant faults and recessive attenuation trend is generated.The application improves the inspection efficiency and recessive defect identification rate, and realizes the predictive intelligent operation and maintenance of photovoltaic components.

Description

Technical Field

[0001] This application relates to the field of photovoltaic intelligent operation and maintenance technology, specifically a method and system for unmanned aerial vehicle (UAV) inspection of photovoltaic power plants. Background Technology

[0002] With the large-scale and intensive development of photovoltaic power plants, drone intelligent inspection has gradually replaced manual inspection and become the mainstream technical means for the operation and maintenance of photovoltaic modules. At present, most existing photovoltaic drone inspection technologies adopt fixed uniform routes and uniform sampling parameters to carry out full-field inspections. The operation mode is seriously homogenized and cannot adapt to the differentiated operating conditions of photovoltaic modules.

[0003] Existing photovoltaic drone inspection technology can only identify visible faults such as external damage and surface hot spots. It cannot capture the long-term micro-degradation trend of modules, making it difficult to provide early warning of hidden faults such as micro-cracks, encapsulation aging, and power attenuation. The inspection process is easily affected by external factors such as uneven lighting, cloud cover, and wind vibration, leading to false detections. Furthermore, the inspection involves a large number of invalid samples and repeated flights on healthy and stable modules, increasing inspection time and equipment energy consumption. Ultimately, the stability of the inspection, the accuracy of defect identification, and the overall operation and maintenance efficiency are all at a low level, which cannot meet the needs of photovoltaic power plants for efficient, accurate, and intelligent operation and maintenance. Summary of the Invention

[0004] This application aims to address the technical problems of poor stability and low accuracy of defect identification in drone inspections of photovoltaic power plants, in order to meet the needs of efficient, accurate and intelligent operation and maintenance of photovoltaic power plants.

[0005] To solve the above-mentioned technical problems, the embodiments of this application are implemented through the following aspects.

[0006] Firstly, this application provides a method for unmanned aerial vehicle (UAV) inspection of photovoltaic power plants, including: A baseline initialization inspection was carried out on all photovoltaic modules of the photovoltaic power station. Visible light images, infrared thermal images and corresponding ambient light data of each photovoltaic module were collected. The images were then preprocessed with distortion correction, noise reduction and pixel-level registration. The texture fingerprint vector, temperature field fingerprint vector and light response fingerprint vector are extracted for each photovoltaic module by a multidimensional fingerprint extraction algorithm. The three types of feature vectors are concatenated to construct a multidimensional time-series fingerprint vector exclusive to a single module, and a module-level fingerprint benchmark library for all photovoltaic modules in the station is established. Based on the stability of the benchmark fingerprint characteristics of each photovoltaic module, all photovoltaic modules are dynamically divided into three state levels: benchmark stable modules, slightly fluctuating modules, and modules with abnormalities to be monitored. Differentiated non-uniform inspection routes are generated based on the component status level. Low-density inspection paths and low sampling frame rates are used for stable baseline components, standard inspection parameters are used for components with slight fluctuations, and encrypted inspection paths and high sampling frame rates are used for components with abnormalities to be monitored.

[0007] The drone is controlled to complete automatic inspections along differentiated inspection routes, and the visible light images, infrared images and ambient light data of each photovoltaic module are collected in real time and then preprocessed in a standardized manner. A multidimensional fingerprint extraction algorithm is used to extract the real-time multidimensional fingerprint vector of this inspection. The feature deviation and time-series drift trend between the real-time fingerprint vector and the historical benchmark fingerprint vector are calculated by the time-series fingerprint deviation quantization algorithm. The fault type is distinguished based on the deviation pattern: if the fingerprint deviation suddenly increases beyond the preset short-term change threshold in a single inspection, it is judged as an obvious fault such as occlusion or dirt; if the fingerprint deviation continues to drift after multiple inspections and exceeds the long-term drift threshold, it is judged as a hidden fault such as component aging, microcracks, or power attenuation.

[0008] Based on the latest inspection data, the fingerprint baseline data and component status level of each component are iteratively updated, the next differentiated inspection route is updated in a closed loop, and a two-layer inspection report containing explicit faults and implicit degradation trends is generated to complete intelligent time-series inspection.

[0009] Furthermore, the multidimensional fingerprint extraction algorithm includes texture fingerprint extraction, which uses Local Binary Pattern (LBP) to encode the texture features of the component's visible light image. The encoding formula is as follows: Histogram normalization is performed on the encoding results to obtain a fixed-dimensional component texture fingerprint vector: .

[0010] In a further embodiment, the multidimensional fingerprint extraction algorithm includes temperature field fingerprint extraction, which uniformly divides the infrared image of a single photovoltaic module into M×N sub-regions, and calculates the average temperature, temperature variance, and regional temperature gradient of each sub-region: Integrate and normalize the temperature statistical features of all sub-regions to form a temperature field fingerprint vector: .

[0011] In a further embodiment, the multidimensional fingerprint extraction algorithm includes illumination response fingerprint extraction, which constructs environmental compensation features by using the component grayscale mean and ambient light intensity. Constructing an illumination response fingerprint vector based on multi-temporal illumination samples: .

[0012] In a further embodiment, the three types of feature vectors are concatenated to form a component-unique comprehensive multidimensional fingerprint vector: In the formula, ⊕ represents a vector concatenation operation, forming a component-specific identity fingerprint with fixed dimensions and time-series comparability.

[0013] Furthermore, the temporal fingerprint deviation quantization algorithm combines Euclidean distance and cosine similarity to jointly calculate feature deviation, obtain texture deviation, temperature deviation, and illumination deviation respectively, and fuse them to obtain the overall temporal deviation value of the component, distinguishing between instantaneous environmental fluctuations and component degradation drift.

[0014] Furthermore, the differentiated inspection route generation method includes: increasing the grid spacing and reducing the sampling frequency of the stable component; reducing the inspection grid, increasing the shooting frame rate and adding local supplementary scanning paths of the abnormal component to achieve adaptive inspection of density across the entire domain.

[0015] Furthermore, a sliding time window is used to dynamically iterate and update the fingerprint benchmark library, stabilizing components with low-frequency updates and fluctuating and abnormal components with high-frequency updates, so as to achieve adaptive updates of the fingerprint benchmark according to seasonal environment and component aging status.

[0016] Secondly, this application provides a photovoltaic power plant unmanned aerial vehicle (UAV) inspection system, comprising: The module includes a component fingerprint baseline modeling module, an image preprocessing module, a time-series fingerprint extraction module, a component status classification module, a differentiated route planning module, a time-series deviation comparison and analysis module, a two-layer fault determination module, a fingerprint iterative update module, and an inspection report generation module. The component fingerprint baseline modeling module is used to complete the initial fingerprint archiving of all components at the station. The temporal fingerprint extraction module embeds a multi-dimensional fingerprint extraction algorithm and outputs multi-dimensional feature fingerprint vectors of components. The component state classification module completes three-level state classification based on fingerprint stability. The differentiated route planning module dynamically generates sparse and dense adaptive inspection routes. The temporal deviation comparison and analysis module completes the temporal deviation calculation between real-time fingerprints and baseline fingerprints. The dual-layer fault determination module distinguishes between instantaneous explicit faults and long-term latent degradation faults. The fingerprint iterative update module dynamically updates the component baseline fingerprint and state level.

[0017] The technical advantages of this application are as follows: By constructing a three-dimensional fingerprint system for component-level texture, temperature, and illumination, false detection problems caused by uneven illumination, cloud cover, and wind vibration during single imaging are eliminated, improving the stability and accuracy of inspections. Furthermore, through a time-series drift comparison mechanism, it overcomes the limitations of traditional visual inspections, which can only identify external damage and surface hot spots, and can accurately capture long-term minor degradation trends of components, enabling early warning of hidden faults. Finally, through component-level and density-adaptive flight path strategies, invalid sampling and repeated flights in healthy component areas are reduced, lowering the inspection time and equipment energy consumption of large photovoltaic power plants. Through a dynamic fingerprint iteration update mechanism, the fingerprint benchmark adapts to seasonal changes, illumination variations, and component aging over the years. The system has self-learning and self-optimization capabilities, improving long-term operation and maintenance accuracy and enhancing the overall power generation efficiency and safe operation and maintenance level of photovoltaic power plants. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the system flow according to an embodiment of the present invention; Figure 2 This is an internal flowchart of the fingerprint iterative update module in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The multi-energy storage converter coordinated control method and system provided in this application will be described in detail below with reference to the accompanying drawings, through specific embodiments and application scenarios.

[0021] This embodiment discloses a method for unmanned aerial vehicle (UAV) inspection of photovoltaic power plants, including the following steps: S1. Route Planning and Equipment Preparation: The system retrieves topographic data, component layout coordinates, and obstacle location information for the entire photovoltaic power station area to generate an initial inspection route that covers the entire station without blind spots. The inspection drone is equipped with a visible light imaging unit, an infrared thermal imaging unit, an ambient light acquisition unit, and a high-precision RTK positioning module. It completes pre-flight equipment self-checks, camera parameter calibration, and sensor zero-point calibration to ensure that imaging and sensing parameters are consistent throughout the process.

[0022] The drone performs full-area, component-by-component imaging and data acquisition. It cruises at a constant speed along a preset route, maintaining a fixed flight altitude, speed, and pitch angle throughout the flight to ensure that the imaging field of view, shooting distance, and shooting angle are consistent for each photovoltaic module.

[0023] Using a single photovoltaic module as the smallest data acquisition unit, all photovoltaic modules within the power station are photographed one by one at fixed points. Each set of collected data contains three types of information: Visible light images: used to extract visual features such as surface texture, appearance, dirt, and damage of components; Infrared thermal imaging images: used to collect temperature features such as surface temperature distribution, local temperature changes, and hot zones of components; Ambient light data: synchronously records the ambient light intensity value of the entire field at the moment of shooting, in lux, as the original parameter for subsequent illumination compensation.

[0024] Real-time data transmission and categorized storage: Image files, illumination values, component numbers, and positioning coordinates collected by the drone are wirelessly transmitted back to the backend server in real time. The backend classifies and archives the data according to the "unique component number," binding and storing the visible light images, infrared images, and illumination data corresponding to the same component to form a one-to-one original dataset, thus completing the baseline data collection work.

[0025] Image preprocessing workflow: The original acquired images are affected by drone flight jitter, lens optical characteristics, outdoor environmental interference, and dual-camera field of view deviation, resulting in problems such as image distortion, noise interference, and misalignment of two images. They cannot be directly used for feature extraction. Therefore, three standardized preprocessing steps, namely distortion correction, image denoising, and pixel-level registration, need to be performed in sequence to unify the image benchmark and ensure the consistency of feature extraction for different inspection batches and different types of images.

[0026] Image distortion correction is crucial because both visible light cameras and infrared thermal imaging cameras on drones inherently exhibit optical distortion, primarily radial and tangential. This distortion manifests as image edge stretching, component contour distortion, and pixel position shift. Furthermore, even slight tilting during flight can cause perspective distortion. Distortion leads to inconsistent pixel coordinates for the same component at different shooting times, severely impacting the accuracy of subsequent texture and temperature feature extraction. The solution involves pre-obtaining the intrinsic parameter matrices and distortion coefficients of both cameras through camera calibration experiments, embedding these parameters into the image preprocessing module, and performing correction calculations for each original image. Based on the camera distortion coefficient, radial and tangential distortion caused by the lens is corrected to restore the true geometric shape of the image; combined with UAV flight attitude data and perspective transformation algorithm, perspective distortion caused by flight tilt is corrected; invalid edge areas of the corrected image are cropped, and only the effective imaging area of ​​the photovoltaic module is retained. After correction, the outline of the photovoltaic module is regular, the geometric position of the image pixels is restored to the true state, the pixel offset problem caused by the lens and flight attitude is eliminated, and the imaging geometric benchmark is guaranteed to be consistent.

[0027] In outdoor inspection scenarios, images are prone to two types of noise: motion noise caused by high-altitude airflow and aircraft vibration; and salt-and-pepper noise and Gaussian noise caused by cloud scattering, light changes, and inherent sensor errors.

[0028] Noise can destroy image texture details and temperature pixel values, causing distortion in fingerprint feature extraction. Therefore, noise reduction is necessary. The specific implementation method is to use a Gaussian filtering algorithm to perform global noise reduction, select a fixed-size filter convolution kernel, and perform neighborhood weighted averaging on each pixel of the distortion-corrected image. For visible light images: retain effective texture details such as component cracks, borders, and dust accumulation, and filter out minor noise and stripe interference caused by flight jitter; for infrared thermal imaging images: smooth the random jumps of temperature pixels, retain effective temperature features such as local high temperature and temperature gradient of components, and avoid noise being misjudged as hot spot faults; without losing effective features, remove various environmental and equipment noises, improve the image signal-to-noise ratio, and ensure the authenticity of texture and temperature data.

[0029] Pixel-level registration is required because the visible light camera and the infrared thermal imaging camera are two independent imaging devices. The physical deviations in their installation positions and field of view cause the visible light image and infrared image of the same component to be misaligned and the target area to not overlap.

[0030] Direct feature extraction can lead to mismatches between texture and temperature regions. Pixel-level registration maps pixel coordinates one-to-one between the two images to achieve complete alignment of regions containing the same component in both images. Specifically, a feature-point matching-based registration scheme is used, with the following steps: Stable feature points (such as strong features like the frame of the photovoltaic module, the intersection of the support, and the grid lines of the module) are extracted from the denoised visible light image and infrared image respectively. Matching operations are performed on the two sets of feature points to calculate the coordinate transformation matrix between the two images. Based on the transformation matrix, coordinate transformation and interpolation operations are performed on one of the images to ensure that every pixel position in the visible light image and the infrared image corresponds precisely, achieving pixel-level alignment. After alignment, the overlapping effective area of ​​the two images is cropped to ensure that the areas for subsequent texture feature and temperature feature extraction are completely consistent. The target areas of the visible light image and the infrared image completely overlap, and the appearance texture and temperature data of the same position of the module are accurately bound, laying the data foundation for subsequent multi-dimensional fingerprint extraction and time-series deviation comparison.

[0031] The process is as follows: Upon completion of the baseline initialization inspection, the UAV performs point-to-point imaging of all photovoltaic modules within the station along a preset full-coverage route, simultaneously acquiring visible light images, infrared thermal images, and ambient light intensity data at the corresponding time, and storing the data in groups according to the module number. Subsequently, standardized preprocessing is performed on all raw images in sequence: First, the camera calibration parameters and perspective transformation algorithm are used to complete image distortion correction, correcting the image distortion caused by lens optical distortion and flight attitude; then, a Gaussian filtering algorithm is used to remove various noise points caused by flight vibration and environmental interference, optimizing image quality; finally, pixel-level registration of the visible light image and infrared image is achieved based on feature point matching, so that the target areas of the two images accurately overlap.

[0032] After the entire preprocessing process, the image geometry, pixel position, and imaging benchmark are completely unified, and can be used for subsequent multidimensional fingerprint feature extraction and benchmark library construction.

[0033] S2. After completing the image standardization preprocessing, the system performs multi-dimensional fingerprint extraction calculations for each individual photovoltaic module, constructing three types of differentiated feature representations: texture fingerprint vector, temperature field fingerprint vector, and light response fingerprint vector. Through feature vector fusion and splicing, a unique multi-dimensional time-series fingerprint vector corresponding to each photovoltaic module is generated, ultimately establishing a refined module-level fingerprint benchmark library for the entire site. The specific functional implementation is as follows: Single-component multi-dimensional independent feature extraction function; The texture fingerprint vector extraction function uses a standardized LBP texture feature extraction algorithm to encode the micro-texture, edge structure, and gray-scale distribution of the module surface in a high-density manner for the pre-processed visible light image of the photovoltaic module.

[0034] By comparing neighboring pixel differences, performing binary encoding statistics, and processing histogram normalization, the complex visual information of two-dimensional images is compressed into a fixed-dimensional, quantifiable, and comparable texture fingerprint vector.

[0035] This vector is used to uniquely characterize the inherent structural features of the photovoltaic module's appearance. It can record subtle differences in appearance such as surface cleanliness, texture uniformity, surface cracks, aging color difference, and surface wear, transforming visual image information into digital feature data that can be recognized by machines and compared in time.

[0036] The temperature field fingerprint vector extraction function, for the pre-processed infrared thermal imaging image of the photovoltaic module, the system subdivides the effective temperature measurement area of ​​the module into a grid, and independently calculates the average temperature, temperature variance, and temperature gradient distribution characteristics of adjacent areas for each sub-region.

[0037] By performing structured statistics and normalization on the global temperature data, a fixed-dimensional temperature field fingerprint vector is formed.

[0038] This vector fully depicts the global temperature distribution and thermal field uniformity of the photovoltaic module. It does not only record the temperature at a single point, but records the temperature structure characteristics of the entire module. It can accurately characterize the microscopic temperature state of the module, such as internal heat loss, local overheating, thermal field distortion, and hidden heating anomalies, and realize digital modeling of the module's electrical aging and thermal decay state.

[0039] The illumination response fingerprint vector extraction function combines the real ambient light intensity at the imaging moment with the overall grayscale mean of the component image to construct illumination normalized response features and form illumination response fingerprint vectors.

[0040] This feature is specifically designed to counteract external environmental interference, quantify the imaging response patterns of photovoltaic modules under different light intensities, time periods, and weather conditions, and transform the "image brightness differences caused by changes in ambient light" into compensable and peelable environmental feature components, ensuring that subsequent time-series comparisons only identify changes in the module itself and not environmental changes.

[0041] The three-feature vector fusion and splicing function, after obtaining the texture fingerprint vector, temperature field fingerprint vector, and illumination response fingerprint vector corresponding to a single component, adopts a vector splicing and fusion mechanism to orderly merge the three types of feature vectors with different physical meanings and different dimensions, generating a single, fixed-dimensional, multi-physical field coupled multi-dimensional temporal fingerprint vector.

[0042] The stitched integrated fingerprint vector includes: visual structure dimension (component appearance, texture, surface condition); thermal field physical dimension (component temperature distribution, thermal uniformity, thermal defects); and environmental response dimension (light adaptation characteristics, imaging stability). This enables full-state digital identity modeling of a single photovoltaic module, giving each photovoltaic module a unique, long-term time-trackable, and quantifiable digital identity feature.

[0043] The system has a function to establish a full-site component-level fingerprint benchmark library. It performs the above-mentioned multi-dimensional fingerprint extraction and fusion calculation on all photovoltaic components in the power station one by one. Using the unique physical number of the component, RTK positioning coordinates and multi-dimensional fingerprint vector as indexes, it establishes a structured, indexable, iterable and comparable component-level time-series fingerprint benchmark library.

[0044] The benchmark library stores the original health characteristics of each component in its standard initial state, serving as a benchmark for comparison in subsequent inspections. Compared to traditional unified features for the entire station and average features at the string level, this patent adopts a single-component independent filing mode, ensuring that even minor state changes of each component can be accurately captured, providing core benchmark data support for subsequent time-series deviation calculation, hidden fault identification, and differentiated route planning.

[0045] This step achieves the first integrated digital identity construction of photovoltaic module appearance, thermal field and environmental response status by fusing texture, temperature and light modal features into a model, thus overcoming the limitations of traditional single image feature recognition. Through module-level fine-grained documentation, it upgrades from "single image recognition fault" to time-series state evolution recognition fault, providing a core technical foundation for the advanced prediction of subsequent latent aging, micro-attenuation and micro-cracks.

[0046] S3. After completing the construction of the multi-dimensional time-series fingerprint benchmark library for all photovoltaic modules in the station, each photovoltaic module possesses a unique, quantified, and multi-dimensional fused benchmark fingerprint vector. To distinguish the initial health status, characteristic stability, and potential degradation risk of different photovoltaic modules, this system performs refined dynamic level classification of all photovoltaic modules in the station based on the discrete stability, characteristic uniformity, and initial deviation of each module's own benchmark fingerprint characteristics. All photovoltaic modules are uniformly classified into three status levels: benchmark stable modules, slightly fluctuating modules, and modules with abnormalities to be monitored. This realizes the transformation of photovoltaic modules from "uniform and undifferentiated management" to "level-based, risk-based, and differentiated management." The specific functional implementation and technical principles are as follows: The core principle of the grading process is that this step does not grade based on the result of a single image failure, but rather on the stability characteristics of the component's multidimensional fingerprint itself. The system comprehensively and quantitatively evaluates the stability of the texture feature distribution, the uniformity of the temperature field features, and the consistency of the illumination response features within the fingerprint of a single component. By calculating the feature dispersion, feature offset, and feature variance level of multiple micro-samples of the same component, the system determines the current inherent stability and potential anomaly risks of the component, thereby achieving fully automatic intelligent grading without human intervention.

[0047] The three-level accurate judgment mechanism and function definition, system judgment criteria: the feature dispersion of each dimension of the component's multidimensional fingerprint vector is extremely low, the texture feature distribution is regular, the temperature field feature is uniform without local offset, the illumination response feature is linear and stable, and the overall feature deviation is within the minimum safe threshold range.

[0048] These components have clean surfaces, intact structures, and stable electrical and thermal properties. They exhibit no early aging, no latent defects, and no environmentally sensitive fluctuations, making them long-term healthy and stable components.

[0049] Subsequent inspections can significantly reduce sampling density and inspection frequency, reduce invalid duplicate sampling, and reduce the overall inspection load of the station.

[0050] The criteria for classifying components with slight fluctuations (i.e., medium-risk monitoring level components) are as follows: The component's multidimensional fingerprint vector exhibits small, discrete fluctuations, with characteristic deviations between the normal and warning thresholds. While not showing clear fault characteristics, it exhibits slight texture disturbances, minor local temperature unevenness, or slight shifts in light response compared to standard stable components. These components have no overt faults but exhibit slight state instability, potentially affected by minor dust accumulation, slight shading, installation stress, minor environmental disturbances, or initial performance fluctuations. They are considered potential risk monitoring targets. Components at this level are not considered faulty, but standard inspection and monitoring efforts must be maintained to prevent further deterioration into abnormal components.

[0051] The components to be monitored, namely high-risk key monitoring components, have high dispersion in multidimensional fingerprint vectors, distortion of temperature field features, obvious shift in texture features, and deterioration in light response stability. The overall feature deviation is close to or exceeds the warning threshold, and there is obvious feature instability. These components have shown early abnormality trends and are likely to have problems such as local shading, dirt accumulation, slight thermal unevenness, early material aging, microcrack precursors, or unstable operation. They are high-risk objects for latent faults and early defects and require subsequent intensive inspection, high-precision sampling, and key time-series tracking.

[0052] The dynamic grading technology automatically categorizes tens of thousands of photovoltaic modules according to their health, stability, and risk levels through a three-level grading system. This solves the problem of the traditional technology's extensive inspection approach, which uses the same standards, frequencies, and densities for all modules. It provides a precise basis for subsequent differentiated route planning. The three-level labels output in this step are the sole state input for generating adaptive differentiated routes, ensuring that the inspection strategy no longer relies on manual experience or fixed paths, but is entirely driven by the actual health status of the modules.

[0053] Early detection of hidden degradation trends, and identification of unstable components based on fingerprint stability differences in the absence of obvious visible faults or clear hot spot alarms, is a key preliminary step in predictive maintenance. This enables pre-classification and pre-monitoring before faults occur. It achieves dynamic and iterative intelligent classification. This classification is not a fixed static label, but a dynamic status level generated based on the stability of fingerprint features. It can be continuously updated and iterated and corrected with subsequent rounds of time-series inspections to ensure that the component level always matches the real-time health status.

[0054] The system comprehensively evaluates the initial operating status of photovoltaic modules by quantitatively analyzing the discrete stability of the multidimensional fingerprint characteristics of each module, considering multiple dimensions such as texture consistency, temperature field uniformity, and light response stability. It automatically classifies all photovoltaic modules in the station into three risk levels: benchmark stable modules, slightly fluctuating modules, and abnormal modules to be monitored.

[0055] By grading the status of components in a refined manner, the digital hierarchical identification of the health status of photovoltaic modules is realized. This provides a core basis for subsequent dynamic generation of differentiated non-uniform inspection routes, intelligent weight-reduction inspection, and intensive and precise inspection of key areas. It solves the technical drawbacks of traditional photovoltaic drone inspections, such as full coverage, low efficiency of density inspection, and indiscriminate sampling.

[0056] S4. Based on the three-level state classification results of the components, this step performs intelligent flight path planning that is state-adaptive, non-uniform, and variable in density. This differs from the homogeneous flight path generation logic of traditional photovoltaic inspections, which uses fixed grids, fixed spacing, and fixed frame rates. The system configures the inspection flight path, grid coverage density, image sampling step size, and camera shooting frame rate in a refined, zoned, and differentiated manner according to the health level of each photovoltaic module, achieving intelligent dynamic allocation of inspection resources. Specific functions implemented are as follows: For benchmark stable components with long-term stable fingerprint features, no fluctuations, and no abnormal drift trends, the system determines that such components are in good operating condition, have uniform surface conditions, and regular thermal field distribution, and do not have the risk of latent degradation or obvious failure. Therefore, the system automatically implements a lightweight sparse inspection strategy for this area. Specifically, it increases the spacing between inspection grid cells, increases the sampling step distance of the UAV flight, reduces the continuous shooting frame rate of the camera, and reduces the number of image samples and flight path coverage times per unit area. This function significantly reduces repetitive, redundant, and meaningless high-density sampling operations while ensuring continuous tracking of the basic fingerprint features of components and monitoring of sudden failures. It effectively shortens the flight mileage and inspection time of UAVs and reduces the energy consumption and equipment wear of routine inspections.

[0057] For components exhibiting slight fluctuations in fingerprint features that do not reach an abnormal threshold, the system determines that these components are generally operating normally and do not have clear fault characteristics. However, they may exhibit slight environmental sensitivity or state instability, making it unsuitable to lower the inspection standards. The system maintains the factory standard inspection grid spacing, standard flight speed, and standard sampling frame rate for these component areas, without encryption or reduction, maintaining the intensity of routine inspection and monitoring. This function ensures continuous, stable, and balanced data collection for fluctuating components, avoiding resource waste caused by over-inspection and stably capturing minute feature changes, preventing further deterioration of slight fluctuations, and achieving stable monitoring and control of components in intermediate states.

[0058] For abnormal components to be monitored that exhibit large fingerprint dispersion, significant feature offset, and early degradation trends, the system identifies these components as high-risk monitoring targets. These components are highly susceptible to hidden defects such as microcracks, thermal distortion, localized dust accumulation, and early power attenuation, which are invisible to the naked eye. The system automatically initiates a comprehensive and encrypted inspection mechanism for these areas: reducing the local inspection grid spacing, compressing the flight sampling step size, and increasing the camera's frame rate. At the same time, it generates local reciprocating scan routes to increase the number of samples and coverage angles in the same area. Through high-density, high-overlap, and multi-frame continuous sampling, the system can acquire richer and more refined texture details and temperature field distribution data, fully amplifying the component's minute abnormal features. This provides sufficient and high-precision data support for subsequent timing deviation calculations and hidden fault determination, ensuring that weak, early, and difficult-to-detect component defects can be effectively identified.

[0059] The system's global non-uniform route fusion generation function smoothly splices, merges, and optimizes the local routes corresponding to three types of differentiated inspection parameters, ultimately generating an integrated inspection route that is non-uniform across the entire station, adaptive in density, and driven by health status. The entire route does not have the uniform grid characteristics of traditional inspections, but presents an intelligent distribution characteristic of "sparse in stable areas, regular in fluctuating areas, and dense in abnormal areas", realizing the dynamic matching of inspection resources according to the health risk of components.

[0060] This step fundamentally breaks away from the inefficient operation mode of traditional drone photovoltaic inspection, which is "one-size-fits-all and equal-density coverage across the entire site." By accurately mapping the component status level with the inspection parameters, it achieves an intelligent inspection strategy that reduces the burden on normal areas and increases pressure on high-risk areas.

[0061] It not only ensures the ability to perform detailed inspections of high-risk abnormal components and significantly improves the detection rate of hidden defects, but also greatly reduces the redundancy of the overall inspection of the entire station and improves the overall operational efficiency of routine inspections of large photovoltaic power plants.

[0062] S5. After generating the differentiated non-uniform inspection route for the entire station, the system outputs an intelligent inspection path that includes variable density and parameter partitioning. The UAV, relying on its onboard RTK high-precision positioning system, attitude stabilization system, and dual-light imaging acquisition unit, performs fully automatic, non-interventional, and hierarchically adapted intelligent inspection operations along the planned differentiated inspection route. It dynamically matches flight speed, shooting distance, and sampling frame rate according to the grade parameters of different component areas, simultaneously completing high-precision real-time acquisition of visible light images, infrared thermal images, and ambient light data. All acquired raw data undergoes unified standardized preprocessing to achieve standardized, high-precision, and consistent acquisition and output of inspection data. Specific functions are implemented as follows: The UAV dynamically adapts its operational parameters based on the regional level parameters embedded in the flight path: In sparse flight path areas corresponding to the baseline stable components, the system automatically matches a faster flight speed, a larger shooting interval, and a lower sampling frame rate to quickly complete healthy area coverage in a lightweight inspection manner; In standard flight path areas corresponding to slightly fluctuating components, the UAV maintains the rated standard flight speed and standard sampling parameters to maintain the regular inspection and collection density; In dense flight path areas corresponding to abnormal components to be monitored, the UAV automatically reduces its flight speed, shortens its shooting step distance, increases its imaging frame rate, and performs local supplementary scanning and reciprocating flight to achieve high-density, multi-angle, and fine collection in high-risk areas.

[0063] The drone can take off autonomously, cruise autonomously, switch parameters autonomously, and return autonomously without human intervention, achieving fully automated hierarchical inspection operations.

[0064] During differentiated flight, the drone simultaneously collects three types of core data from each photovoltaic module through its onboard integrated sensing unit, ensuring time synchronization, spatial matching, and one-to-one binding of multi-source data to each module. Visible light image acquisition: Airborne visible light cameras capture real-time details of photovoltaic module surface texture, appearance structure, surface cleanliness, and appearance defects, obtaining high-resolution two-dimensional visual image data for subsequent accurate extraction of texture fingerprint features; Infrared thermal imaging image acquisition: Airborne infrared cameras simultaneously acquire data on the module's global temperature distribution, local temperature differences, and thermal field uniformity, accurately recording subtle temperature changes and local thermal anomalies, providing raw thermal imaging data for temperature field fingerprint calculation; Ambient light data acquisition: Airborne light sensors synchronously record ambient light intensity parameters at each imaging moment, accurately capturing dynamic changes in light during the inspection process, used for subsequent light response feature modeling and environmental interference compensation. All acquired data is wirelessly transmitted back to the backend system in real time and automatically bound and archived with the current module number, GPS / RTK positioning coordinates, and acquisition timestamp, forming a precisely corresponding module-level raw inspection dataset.

[0065] To address issues such as flight attitude disturbances, lens optical deviations, environmental noise interference, and dual-camera field-of-view misalignment in real-time data collected during UAV dynamic flight, the system applies a standardized preprocessing procedure to all inspection images from this round, identical to that used in the baseline database construction phase. This ensures that the processing rules, scales, and pixel benchmarks for this round of inspection data are consistent with those of the baseline fingerprint data, eliminating system errors caused by differences in acquisition conditions. The preprocessing specifically includes: Image distortion correction: Unified use of in-camera distortion coefficients to correct optical distortion and perspective shift during dynamic flight, restoring the true geometric shape of photovoltaic modules; Intelligent noise reduction: Removing random noise caused by flight jitter, airflow disturbance, and light scattering through filtering algorithms, improving the image signal-to-noise ratio while preserving effective defect features and temperature details; Dual image pixel-level registration: Performing precise feature matching and pixel alignment on visible light and infrared images of the same module to ensure complete overlap between the appearance texture area and the temperature field area, achieving precise spatial correspondence between visual features and thermal field features.

[0066] This step realizes an integrated intelligent inspection process that integrates differentiated flight routes, hierarchical data acquisition, multi-source synchronous sensing, and standardized data correction. By dynamically adapting the acquisition parameters of components with different health levels, it ensures that the data of abnormal components is sufficiently granular to support the identification of latent defects, while significantly reducing data redundancy in healthy areas. Through a unified and standardized preprocessing mechanism, it ensures that routine inspection data and initial baseline fingerprint data have a completely consistent calculation benchmark, providing high-quality and highly consistent raw data support for subsequent real-time fingerprint extraction, accurate time-series comparison, and quantitative analysis of minor deviations. This ensures the accuracy and reliability of time-series fingerprint comparison and fault determination from the data source.

[0067] S6. After completing the data collection and standardized preprocessing for this round of inspections, the system reuses the multi-dimensional fingerprint extraction algorithm from the benchmark database construction stage to perform feature calculation on the inspection images and environmental data, generating real-time multi-dimensional fingerprint vectors for the corresponding photovoltaic modules. Then, through a dedicated time-series fingerprint deviation quantification algorithm, the real-time fingerprint vectors are compared with the historical benchmark fingerprint vectors pre-stored for the module across the entire domain to accurately calculate the feature deviation values. At the same time, the system tracks the feature change patterns under multiple rounds of inspections, identifies time-series drift trends, and provides a quantitative basis for subsequent fault type determination. The specific functions are implemented as follows.

[0068] The system uses a single photovoltaic module as the processing unit, strictly follows the algorithm rules, parameter standards and feature dimensions used when the benchmark library was built, completes the extraction of three types of feature vectors step by step, and merges them to generate a real-time comprehensive fingerprint vector, ensuring that the calculation methods of real-time features and benchmark features are completely consistent.

[0069] Real-time texture fingerprint extraction is performed based on the visible light image after this round of preprocessing. The local binary mode algorithm with the same parameters is used to complete pixel encoding, histogram statistics and normalization processing, and output a real-time texture fingerprint vector with the same dimension as the reference vector. This vector represents the current visual state of the component, such as surface texture, cleanliness and appearance.

[0070] Real-time temperature field fingerprint extraction involves dividing the infrared image into regions according to a unified grid division rule, calculating the average temperature, temperature variance, and temperature gradient of adjacent regions for each region, integrating statistical data from all regions and normalizing them, and generating a dimension-matched real-time temperature field fingerprint vector that fully reflects the current overall temperature distribution and thermal field uniformity of the component.

[0071] Real-time illumination response fingerprint extraction combines the ambient light intensity and component image grayscale mean value collected in this round to calculate illumination response feature value. Combined with time-series sampling samples, a real-time illumination response fingerprint vector is generated to quantify the imaging response characteristics of the component under the current environmental conditions, which is used to offset feature interference caused by illumination changes.

[0072] Real-time integrated fingerprint fusion, following predetermined vector concatenation rules, sequentially merges real-time texture fingerprints, real-time temperature field fingerprints, and real-time illumination response fingerprints to generate a fixed-dimensional real-time multi-dimensional fingerprint vector. This vector fully reproduces the comprehensive operational status of the component during this round of inspections and forms a standardized data pair that can be directly compared with historical benchmark fingerprint vectors.

[0073] The system calls the time-series fingerprint deviation quantization algorithm, using the component's unique ID as an index, to retrieve the historical baseline fingerprint vector corresponding to the component in the database, and performs calculation and analysis from two dimensions: static feature deviation and dynamic time-series change.

[0074] The algorithm calculates the overall feature deviation by combining Euclidean distance and cosine similarity. It performs difference calculations on the three types of sub-feature vectors—texture, temperature field, and illumination response—to obtain the deviation value for each feature. Then, according to the weighted fusion rule, it calculates the overall feature deviation of the component. This value directly reflects the degree of deviation between the component's current state and the initial baseline state, quantifying the magnitude of state change in a single inspection.

[0075] For temporal drift trend identification, the system retrieves real-time fingerprint data and corresponding deviation records from multiple rounds of historical inspections of the component to construct a long-term series dataset. By analyzing the increase and decrease patterns, fluctuation amplitude, and continuous change direction of the deviation value, it distinguishes two change patterns: one is irregular instantaneous deviation fluctuation, which is mostly caused by external environmental factors such as clouds, gusts, and short-term changes in sunlight; the other is a temporal drift phenomenon in which the deviation value continuously increases in one direction, representing that the component's own state is gradually deteriorating. The algorithm quantifies and statistically analyzes the drift rate and drift amplitude, and marks the drift start node and development trend.

[0076] This step enables a digital and quantifiable comparison between the real-time status of components and the baseline status. A unified fingerprint extraction rule ensures that the comparison results are unaffected by differences in algorithms and parameters, resulting in high data reliability. Deviation quantification accurately pinpoints the degree to which the current state of a component deviates from the baseline. Combined with time-series trend analysis, it effectively distinguishes between environmental disturbances and component performance degradation, avoiding the vulnerability of single-image recognition to external interference. The calculated deviation amount, drift trend, and other quantitative results serve as the core basis for distinguishing between explicit and implicit faults, and also provide data support for updating component status levels and optimizing the next round of inspection routes.

[0077] S7. The system combines the fingerprint feature deviation and temporal drift trend calculated above with the pre-calibrated two-level judgment thresholds to complete fault classification and identification according to the different forms of deviation, accurately distinguishing between explicit faults caused by external factors and implicit faults caused by component degradation, and realizing automatic fault type determination. The specific functions are as follows: The system pre-determines two types of dedicated judgment thresholds by combining the power plant's on-site operating conditions, equipment parameters, and historical operating data, and then embeds them into the algorithm logic. The short-term mutation threshold is used to identify instantaneous abnormal conditions, while the long-term drift threshold is used to identify long-term gradual degradation. The two thresholds are independent of each other and each performs its own function, forming a two-layer fault judgment standard.

[0078] Based on the determination of obvious faults by a single deviation surge, when the overall fingerprint deviation of a component obtained from a single round of inspection suddenly increases and the value exceeds the preset short-term mutation threshold, the system determines that the component has an obvious fault.

[0079] Such deviations are instantaneous and sudden, with characteristic changes occurring abruptly without a prior gradual process. They are mostly caused by external factors such as foreign objects attached to the surface or environmental shading. In the context of actual photovoltaic module operation scenarios, this type of fault specifically corresponds to dust accumulation, fallen leaves, birds, or debris attached to the module surface, or temporary shading caused by clouds, buildings, or vegetation. This type of fault is a visually obvious and sudden surface problem. The identification results can be directly used as the basis for on-site maintenance and troubleshooting, guiding maintenance personnel to clean up and remove shading at specific points and quickly restore the module to normal operating status.

[0080] Based on continuous time-series drift to determine latent faults, when the deviation in a single round of inspection does not trigger the short-term sudden change threshold, but the fingerprint deviation value of the component shows a continuous, unidirectional, and slowly increasing drift trend during multiple rounds of inspection, and the cumulative drift amplitude eventually exceeds the preset long-term drift threshold, the system determines that the component has a latent fault.

[0081] This type of deviation exhibits a long-term, gradual change pattern. The degradation process is slow and the changes are subtle, failing to reveal obvious anomalies in a single image or single inspection, making it difficult to detect using traditional visual recognition methods. Considering the operating characteristics of photovoltaic modules, this type of fault corresponds to issues such as natural aging of module encapsulation materials, microcracks invisible to the naked eye in the cells, deterioration of circuit contact performance, and gradual decline in photoelectric conversion efficiency.

[0082] This type of failure stems from internal performance degradation of the component itself, characterized by continuous development and gradual worsening. The system can identify early signs of degradation through time-series drift detection, providing proactive fault warnings. This allows maintenance personnel to conduct preventative repairs, preventing minor defects from escalating into serious failures and reducing power generation losses and the risk of equipment failure.

[0083] This step relies on a combination of deviation morphology and grading thresholds to distinguish between transient faults caused by external environmental interference and long-term faults caused by component performance degradation. It breaks through the limitation of traditional single image recognition, which can only detect surface problems. On the one hand, it accurately locates obvious faults and ensures rapid handling of common defects; on the other hand, it identifies various hidden degradation problems in advance to achieve predictive maintenance. The judgment results are synchronously associated with component number, location, fault type and risk level, providing complete and reliable fault data support for inspection report preparation and maintenance task assignment.

[0084] S8. After completing the component fingerprint deviation calculation and fault type determination for this round, the system dynamically updates the benchmark fingerprint database and component health status level of all photovoltaic modules in the entire station based on the real-time multi-dimensional fingerprint data, deviation quantification results, and time-series drift trend of the latest inspection. It then automatically optimizes the differentiated inspection route parameters for the next round based on the updated component status, and finally integrates explicit fault data and implicit degradation trend data to generate a two-layer intelligent inspection report, completing a single, complete time-series closed-loop intelligent inspection process. Specific functionalities are as follows: This invention abandons the traditional static benchmark mode of "one-time database creation and permanent unchanged", and adopts a time-adaptive sliding update mechanism to update the benchmark fingerprint differently according to the actual health status of the components.

[0085] For the benchmark stabilization component, since its fingerprint features are stable over a long period of time and have no obvious drift, the system adopts a low-frequency update strategy to retain the long-term steady-state benchmark features and avoid frequent interference with the benchmark data by minor environmental fluctuations. For components with slight fluctuations and components to be monitored for abnormalities, since their states are in a dynamic process, the system initiates a high-frequency iteration mechanism. Based on the latest high-precision real-time fingerprint, combined with historical time-series fingerprint sequences, the system smoothly iterates and updates the multi-dimensional baseline fingerprint vector of the component.

[0086] By using a differentiated iterative approach, the component reference fingerprint can adapt to seasonal light changes, natural equipment aging, and long-term environmental evolution, always maintaining the latest and most accurate reference scale that best reflects the actual health status of the component. This solves the problems of comparison failure and increased false detection that occur after long-term use of traditional fixed references.

[0087] Based on the latest deviation, sudden change status, and timing drift trend of this inspection, the system re-determines and dynamically corrects the original three-level status of the components.

[0088] If a previously stable component experiences a single sudden deviation, it is temporarily marked as being in a risky state. If a stable component experiences continuous and slow drift, it is progressively upgraded to a component with slight fluctuations or an abnormal component to be monitored. If a previously fluctuating or abnormal component recovers its stable characteristics and its deviation decreases after maintenance and cleanup, the system automatically downgrades it to a stable component.

[0089] The component status level can be dynamically adjusted and synchronized in real time according to the actual status of the device, ensuring that the risk level of each component always accurately corresponds to the current operational health.

[0090] The system takes the latest component status level after iterative updates as input, automatically recalculates the whole-site inspection density strategy, and generates a brand-new differentiated inspection route for the next time in a closed loop.

[0091] For newly upgraded abnormal areas, the flight path is automatically encrypted and the sampling frame rate is increased; for areas that have recovered and stabilized, the inspection density is automatically relaxed and the sampling frequency is reduced.

[0092] This ensures that each inspection route is unique, non-repetitive, and adapts to the latest equipment status, forming a self-learning closed-loop intelligent mechanism to achieve dynamic and optimal allocation of inspection resources.

[0093] The system aggregates all inspection data from this round, constructs a two-dimensional inspection result system, and automatically generates standardized, traceable, and maintainable intelligent inspection reports. The first layer is the explicit fault result layer: it records visible faults such as shading, dust accumulation, shadows, and debris caused by a single sudden change, including fault location, fault type, deviation change amplitude, and location coordinates, to guide on-site immediate cleaning and defect elimination; the second layer is the implicit degradation trend layer: it records progressive defects such as microcracks, encapsulation aging, photoelectric conversion degradation, and thermal field uniformity deterioration caused by long-term temporal drift of the components, and outputs component health score, aging trend curve, and degradation warning level, to realize predictive operation and maintenance, batch repair, and component life assessment.

[0094] The report fully retains fingerprint deviation data, time-series change trajectories, status level records, and route iteration records, enabling the photovoltaic module's operating status to be quantified, traceable, and predictable.

[0095] This step completes the closed-loop optimization of the entire process, from data updates and status refreshes to strategy iterations and result outputs, enabling the entire inspection system to possess continuous self-learning, self-adaptive, and self-optimizing capabilities. It not only enables precise single-time fault detection but also allows for long-term tracking of the photovoltaic module's entire lifecycle status evolution, achieving an upgrade from traditional "passive fault diagnosis" to proactive predictive intelligent operation and maintenance. Ultimately, it enables high-precision, high-efficiency, low-redundancy, and predictable intelligent time-series inspections of all photovoltaic modules in the entire site.

[0096] Detailed description of the texture fingerprint extraction function: This section describes the texture feature calculation step in the multidimensional fingerprint extraction algorithm. It utilizes Local Binary Pattern (LBP) to encode and quantize the texture features of the visible light image of the photovoltaic module, transforming the image's visual texture information into a standardized, fixed-dimensional digital feature vector. This vector characterizes the surface appearance of the module, providing quantifiable texture data support for subsequent time-series comparisons and state analysis. The specific functional implementation is as follows: The algorithm calculates neighborhood features centered on any pixel within a visible light image, and sets... The coordinates of the center pixel within the image. The grayscale value of the center pixel; the number of neighboring sampling points selected. A circular neighborhood is constructed with a neighborhood radius R. Represents the number within the neighborhood The grayscale value of each sampled pixel.

[0097] Through formula The single-pixel LBP encoding is completed, where the sign function... Used to determine the grayscale relationship between neighboring pixels and the center pixel: output value 1 when the grayscale of a neighboring pixel is greater than that of the center pixel, and output value 0 otherwise.

[0098] This operation traverses the entire visible light image of the photovoltaic module pixel by pixel, converting the grayscale differences and texture brightness variations between pixels into corresponding binary code values. It fully captures the surface texture details of the module, including differences in surface cleanliness, fine surface cracks, paint wear, color difference distribution, borders and grid lines, and other visual features, thus realizing the digital encoding conversion of the texture morphology of the two-dimensional image.

[0099] After completing the LBP encoding of the entire image pixels, histogram statistics are performed on all encoded values ​​to count the frequency of occurrence of different encoded values, forming a statistical sequence that represents the global texture distribution pattern. Then, normalization operation is performed on the histogram data to eliminate numerical deviations caused by overall image brightness and local small brightness fluctuations, and to unify the data units and numerical range.

[0100] After the above processing, the final output is a fixed-dimensional texture fingerprint vector: In the formula, K is a preset fixed feature dimension, and each element in the vector corresponds to the quantization result of a type of texture feature.

[0101] Detailed description of the temperature field fingerprint extraction function: This step is an important component of the multidimensional fingerprint extraction algorithm. It relies on the infrared thermal imaging image of the photovoltaic module to quantify the temperature features, converting the temperature distribution information in the image into a standard numerical vector to characterize the thermal field state of the module. The specific implementation process and functions are as follows: First, the infrared image of a single photovoltaic module is divided into regions. According to preset rules, the complete imaging area is uniformly divided into M×N non-overlapping sub-regions. Temperature data is statistically analyzed using the sub-regions as the smallest calculation unit. Let Q be the number of temperature-measuring pixels in a single sub-region. This represents the temperature value corresponding to the q-th pixel within the sub-region.

[0102] Through formula Calculate the average temperature of each sub-region to visually reflect the overall temperature level of that local area; then use the formula... The temperature variance of a sub-region is directly calculated to assess the uniformity of local temperature distribution; a larger variance indicates more significant temperature fluctuations in that region. The formula is then used... Calculate the temperature gradient between adjacent sub-regions to accurately capture temperature abrupt changes between regions, corresponding to abnormal features such as local overheating and hot spots in the components.

[0103] After calculating the three temperature indicators for all sub-regions, all statistical results are integrated and summarized in a fixed order, and normalization is performed to unify the data dimensions and numerical range, eliminate interference caused by imaging environment and temperature measurement deviation, and finally generate a temperature field fingerprint vector with fixed dimensions.

[0104] , where L is the preset feature dimension.

[0105] The light response fingerprint extraction function is described in detail. This part belongs to the environment compensation module of the multi-dimensional fingerprint extraction algorithm. It combines the grayscale data of visible light image with the real-time ambient light intensity to construct features. It is specifically used to offset the interference of external light changes on image features and ensure that the fingerprint comparison results only reflect the state of the photovoltaic module itself.

[0106] First, iterate through the complete imaging area I of the photovoltaic module and count the grayscale values ​​of all pixels within the area. Let S be the total number of pixels in the image. For coordinates, The grayscale value of a pixel is obtained through the formula. Calculate the overall grayscale mean of the entire component image to objectively reflect the overall brightness and darkness of the component image under the current lighting conditions; Combined with synchronously collected ambient light intensity Using formula The light response characteristic value is calculated. This ratio can quantify the response law of the component's imaging effect with changes in external light intensity, transform the difference in brightness of the image caused by light intensity into a standardized compensation parameter, and remove the characteristic fluctuations caused by environmental factors.

[0107] Multiple sets of illumination response samples under different time periods and operating conditions were continuously collected. The data were integrated according to fixed rules and normalized to generate a fixed-dimensional illumination response fingerprint vector. In the formula, P is the preset feature dimension.

[0108] The feature vector splicing and comprehensive multidimensional fingerprint construction function is described in detail. This step integrates three independent feature vectors: texture, temperature field, and light response. Through vector splicing operations, a comprehensive multidimensional fingerprint vector specific to a single photovoltaic module is generated.

[0109] After obtaining the texture fingerprint vectors respectively Temperature field fingerprint vector Illumination response fingerprint vector Vector concatenation operations are performed in a predetermined order. This represents an ordered concatenation operation, and the calculation formula is: .

[0110] The calculation process maintains the original data order and values ​​of various features, only sequentially concatenating and merging three sets of feature vectors with different dimensions and physical meanings to finally obtain a comprehensive multidimensional fingerprint vector with fixed dimensions. .

[0111] The present invention also provides a photovoltaic power plant unmanned aerial vehicle (UAV) inspection system, including a component fingerprint benchmark modeling module, an image preprocessing module, a time-series fingerprint extraction module, a component status classification module, a differentiated flight path planning module, a time-series deviation comparison and analysis module, a two-layer fault judgment module, a fingerprint iterative update module, and an inspection report generation module. The component fingerprint baseline modeling module is used to complete the initial fingerprint archiving of all components at the station. The temporal fingerprint extraction module embeds a multi-dimensional fingerprint extraction algorithm and outputs multi-dimensional feature fingerprint vectors of the components. The component state classification module completes three-level state classification based on fingerprint stability. The differentiated route planning module dynamically generates sparse-adaptive inspection routes. The temporal deviation comparison and analysis module completes the temporal deviation calculation between the real-time fingerprint and the baseline fingerprint. The dual-layer fault determination module distinguishes between instantaneous explicit faults and long-term latent degradation faults. The fingerprint iterative update module dynamically updates the component baseline fingerprint and state level, such as... Figure 1 and Figure 2 As shown, the system consists of nine functional modules that work together, with each module having a clear division of labor and a close connection between processes.

[0112] The component fingerprint baseline modeling module completes the initial fingerprint profile construction; the image preprocessing module standardizes image data; the time-series fingerprint extraction module calls the algorithm to generate multi-dimensional fingerprint vectors for components; the component state classification module classifies component states into three categories based on fingerprint stability; the differentiated route planning module generates inspection routes with variable density as needed; the time-series deviation comparison and analysis module calculates the deviation and trend of real-time fingerprints and baseline fingerprints; the dual-layer fault judgment module identifies explicit faults and implicit degradation problems respectively; the fingerprint iteration update module dynamically refreshes baseline data and component states; and the inspection report generation module summarizes the results and outputs an inspection report. The entire set of modules works together to realize fully automated time-series intelligent inspection of photovoltaic power plants.

[0113] Finally, it should be noted that: The above 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 or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A photovoltaic power station unmanned aerial vehicle inspection method, characterized in that, include: A baseline initialization inspection is carried out on each photovoltaic module of the photovoltaic power station. Visible light images, infrared thermal images and ambient light data at the corresponding time are collected for each photovoltaic module. The images are then preprocessed with distortion correction, noise reduction and pixel-level registration. The texture fingerprint vector, temperature field fingerprint vector and light response fingerprint vector are extracted for each photovoltaic module by a multidimensional fingerprint extraction algorithm. The three types of feature vectors are concatenated to construct a multidimensional time-series fingerprint vector exclusive to a single module, and a module-level fingerprint benchmark library for all photovoltaic modules in the station is established. Based on the stability of the benchmark fingerprint characteristics of each photovoltaic module, all photovoltaic modules are dynamically divided into three state levels: benchmark stable modules, slightly fluctuating modules, and modules with abnormalities to be monitored. Differentiated non-uniform inspection routes are generated based on the component status level. Low-density inspection paths and low sampling frame rates are used for stable baseline components, standard inspection parameters are used for components with slight fluctuations, and encrypted inspection paths and high sampling frame rates are used for components with abnormalities to be monitored. The drone is controlled to complete automatic inspections along differentiated inspection routes, and the visible light images, infrared images and ambient light data of each photovoltaic module are collected in real time and then preprocessed in a standardized manner. A multidimensional fingerprint extraction algorithm is used to extract the real-time multidimensional fingerprint vector of this inspection. The feature deviation and time-series drift trend between the real-time fingerprint vector and the historical benchmark fingerprint vector are calculated by the time-series fingerprint deviation quantization algorithm. The fault type is distinguished according to the deviation pattern: if the fingerprint deviation of a single inspection suddenly increases and exceeds the preset short-term change threshold, it is judged as an obvious fault such as occlusion or dirt; if the fingerprint deviation of multiple consecutive inspections continues to drift and exceeds the long-term drift threshold, it is judged as a hidden fault such as component aging, microcracks, or power attenuation. Based on the latest inspection data, the fingerprint baseline data and component status level of each component are iteratively updated, the next differentiated inspection route is updated in a closed loop, and a two-layer inspection report containing explicit faults and implicit degradation trends is generated to complete intelligent time-series inspection. 2.The unmanned aerial vehicle inspection method for a photovoltaic power station according to claim 1, characterized in that: The multidimensional fingerprint extraction algorithm includes texture fingerprint extraction, which uses Local Binary Pattern (LBP) to encode the texture features of the component's visible light image. The encoding formula is as follows: Histogram normalization is performed on the encoding results to obtain a fixed-dimensional component texture fingerprint vector: 。 3.The unmanned aerial vehicle inspection method for a photovoltaic power station according to claim 1, characterized in that: The multidimensional fingerprint extraction algorithm includes temperature field fingerprint extraction, which uniformly divides the infrared image of a single photovoltaic module into M×N sub-regions and calculates the average temperature, temperature variance, and regional temperature gradient of each sub-region. Integrate and normalize the temperature statistical features of all sub-regions to form a temperature field fingerprint vector: 。 4. The unmanned aerial vehicle inspection method for a photovoltaic power station according to claim 1, characterized in that: The multidimensional fingerprint extraction algorithm includes illumination response fingerprint extraction, which constructs environmental compensation features by using the component grayscale mean and ambient light intensity; Constructing an illumination response fingerprint vector based on multi-temporal illumination samples: 。 5. The method for unmanned aerial vehicle (UAV) inspection of a photovoltaic power station according to claim 1, characterized in that: The three types of feature vectors are concatenated to form a unique, comprehensive, multi-dimensional fingerprint vector for the component. In the formula, ⊕ represents a vector concatenation operation, forming a component-specific identity fingerprint with fixed dimensions and time-series comparability.

6. The method for unmanned aerial vehicle (UAV) inspection of a photovoltaic power station according to claim 1, characterized in that: The temporal fingerprint deviation quantization algorithm combines Euclidean distance and cosine similarity to jointly calculate feature deviation, obtain texture deviation, temperature deviation, and illumination deviation respectively, and fuse them to obtain the overall temporal deviation value of the component, distinguishing between instantaneous environmental fluctuations and component degradation drift.

7. The method for unmanned aerial vehicle (UAV) inspection of a photovoltaic power station according to claim 1, characterized in that: The differentiated inspection route generation method includes: increasing the grid spacing and reducing the sampling frequency of the stable component; reducing the inspection grid, increasing the shooting frame rate and adding local supplementary scanning paths for the abnormal component, so as to realize the adaptive inspection of density across the entire domain. 8.The unmanned aerial vehicle inspection method for a photovoltaic power station of claim 1, wherein: A sliding time window is used to dynamically iterate and update the fingerprint benchmark library, stabilizing components with low-frequency updates and fluctuating and abnormal components with high-frequency updates, so as to achieve adaptive updates of the fingerprint benchmark according to seasonal environment and component aging status.

9. A photovoltaic power station unmanned aerial vehicle inspection system, characterized in that, To implement the inspection method according to any one of claims 1–8, comprising: The module includes a component fingerprint baseline modeling module, an image preprocessing module, a time-series fingerprint extraction module, a component status classification module, a differentiated route planning module, a time-series deviation comparison and analysis module, a two-layer fault determination module, a fingerprint iterative update module, and an inspection report generation module. The component fingerprint baseline modeling module is used to complete the initial fingerprint archiving of all components at the station. The temporal fingerprint extraction module embeds a multi-dimensional fingerprint extraction algorithm and outputs multi-dimensional feature fingerprint vectors of components. The component state classification module completes three-level state classification based on fingerprint stability. The differentiated route planning module dynamically generates sparse and dense adaptive inspection routes. The temporal deviation comparison and analysis module completes the temporal deviation calculation between real-time fingerprints and baseline fingerprints. The dual-layer fault determination module distinguishes between instantaneous explicit faults and long-term latent degradation faults. The fingerprint iterative update module dynamically updates the component baseline fingerprint and state level.

Images