A method for inspecting hidden cracks of a photovoltaic module by using PL infrared fusion time sequence compensation

Abstract

The present application belongs to the technical field of photovoltaic power station intelligent operation and maintenance and fault detection, and particularly relates to a PL infrared fusion time sequence compensation photovoltaic module hidden crack inspection method. The method comprises: acquiring photovoltaic power station flight route planning data, unmanned aerial vehicle platform parameters, environmental monitoring parameters and incremental learning model, performing task configuration to generate a dual-mode acquisition task configuration structure; preprocessing PL images to generate a preprocessed image data structure; aligning time stamps and detecting channel health status, extracting a temperature field stable region to construct a time sequence temperature compensation model; performing frame-by-frame temperature compensation and feature extraction to generate a pixel-level aligned dual-mode image data structure; and identifying hidden cracks through geometric index normalization and convolutional neural network inference. The present application effectively improves the accuracy and robustness of hidden crack detection, overcomes environmental temperature fluctuation interference, realizes automatic intelligent inspection, and optimizes performance through continuous learning.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance and fault detection technology of photovoltaic power plants, and particularly relates to a method for inspecting hidden cracks in photovoltaic modules using PL infrared fusion time-compensation. Background Technology

[0002] Microscopic cracks, known as "hidden cracks," are easily generated in photovoltaic (PV) modules during manufacturing, transportation, and installation. These defects are difficult to detect initially but can lead to performance degradation and even hot spot failure over time, severely impacting power generation efficiency and system safety. Currently, infrared thermal imaging is a common method for identifying PV module faults, but it is insensitive to the subtle temperature changes caused by hidden cracks and is easily affected by environmental temperature fluctuations, light variations, and weather conditions, resulting in insufficient identification accuracy. While photoluminescence detection technology is sensitive to hidden crack features, it faces challenges in practical inspections, including strong background light interference during the day, weak signals, and low detection efficiency. Existing detection methods often rely on single detection modes and lack multi-feature collaborative verification mechanisms, making it difficult to achieve accurate and reliable identification of PV module hidden cracks. Therefore, there is an urgent need for a high-precision hidden crack inspection method that can effectively integrate multi-modal information, suppress environmental interference, and possess feature cross-verification capabilities to meet the large-scale, high-efficiency operation and maintenance needs of PV power plants.

[0003] Furthermore, in the fields of intelligent operation and maintenance of photovoltaic power plants and non-destructive testing of photovoltaic modules, existing solutions for concealed inspection of photovoltaic modules based on UAV platform parameters and photovoltaic power plant flight path planning data typically employ manual equipment inspection, fixed-position infrared thermal imaging, or single-modal detection based on photoluminescence images. These methods collect image data through simple flight path planning or position configuration, and then use threshold segmentation, template matching, or empirical rules to determine and record abnormal areas on the module surface. However, these methods suffer from limitations such as a lack of unified management of image acquisition timing and environmental monitoring parameters, susceptibility of the infrared temperature field to fluctuations in the environmental temperature field, and a lack of unified module unit numbering and microcrack evolution recording mechanisms between multiple inspections. Existing methods largely rely on on-site maintenance personnel to set inspection timing and equipment parameters based on experience. After acquiring photoluminescence and infrared thermal images, they perform independent preprocessing and defect judgment. Then, they manually match suspected defect areas with photovoltaic power plant asset information. In large-scale photovoltaic power plants with significant changes in environmental monitoring parameters, frequent changes in UAV platform attitude, and complex photovoltaic power plant flight path planning data, the following problems are prone to occur: microcrack-related defects are unstable in different batches of inspection results, are easily masked by changes in background temperature field and light fluctuations, and the inspection data is difficult to form a structured record at the module unit level. The microcrack status is also difficult to track and compare over a long period of time. These methods are insufficient to meet the stable realization of a microcrack identification result structure and incremental learning update model structure that can support photovoltaic module microcrack inspection and maintenance decisions. Regarding the joint processing of building a time-series temperature compensation model based on photovoltaic power plant flight path planning data, UAV platform parameters, and environmental monitoring parameters, and generating a hidden crack identification result structure through convolutional neural network inference processing based on pixel-aligned dual-modal image data, existing technologies generally lack integrated design among the synchronous acquisition of photoluminescence images and infrared temperature fields, construction of the time-series temperature compensation model structure, generation of pixel-aligned dual-modal image data structure, and output of the hidden crack identification result structure. They often simplify the processing of local data from a single modality or a single time section, lacking the support of incremental learning and updating of the model structure. The closed-loop update mechanism makes it difficult to establish a continuous and consistent process in photovoltaic power plants for unmanned aerial vehicle (UAV) platform-based inspection and condition assessment of photovoltaic modules for hidden cracks. This process involves data acquisition driven by the dual-modal acquisition task configuration structure, generation of PL preprocessed image data structure, construction of time-series temperature compensation model structure, generation of pixel-level aligned dual-modal image data structure, output of hidden crack identification results structure, and maintenance of incremental learning and updating model structure. Consequently, the hidden inspection results of photovoltaic modules are difficult to form a unified, traceable, and easily callable structured expression in the time and space dimensions, which adversely affects the preparation of operation and maintenance plans and fault prevention in photovoltaic power plants. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for inspecting microcracks in photovoltaic modules using PL infrared fusion timing compensation, comprising: Acquire photovoltaic power station flight path planning data, UAV platform parameters, environmental monitoring parameters and incremental learning update model structure, register task numbers, construct module mapping relationships, verify basic safety boundaries and organize and bind synchronous triggering and attitude compensation control parameters to generate a dual-modal acquisition task configuration structure; Based on the dual-modal acquisition task configuration structure, index binding, image integrity verification, format standardization and flat field correction, local contrast enhancement and standard size cropping are performed to generate PL preprocessed image data structure. Based on the PL preprocessed image data structure, timestamp alignment, channel health status detection, preliminary marking of abnormal segments, extraction of stable temperature field regions, background reference point screening, and regional temperature statistical processing are performed to generate a time-series temperature compensation model structure. Based on the temporal temperature compensation model structure, frame-by-frame temperature compensation calculation, temperature field smoothing resampling, abnormal pixel marking update, hot spot region contour, temperature gradient field and component edge contour feature extraction and pixel mapping relationship calculation are performed to generate a pixel-level aligned dual-modal image data structure. Based on a pixel-level aligned dual-modal image data structure, candidate extraction of PL dark line regions, calculation of infrared local temperature gradient and binding of component unit numbers are performed. Geometric length, width, spatial overlap and temperature difference indices are extracted, normalized rule mapping and convolutional neural network inference processing are performed to generate the hidden crack identification result structure.

[0005] Furthermore, the photovoltaic power station route planning data, UAV platform parameters, environmental monitoring parameters, and incremental learning update model structure include: The photovoltaic power station route planning data includes the photovoltaic field boundary, module array layout, waypoint sequence of each inspection route, inspection altitude, heading angle, overlap rate configuration, and preset inspection time window. The parameters of the UAV platform include UAV model identification, payload capacity, remaining battery capacity estimation method, maximum speed, maximum climb and descent rate, allowable attitude angle range, gimbal control accuracy, and the installation position and field-of-view geometry of the photoluminescence imaging module and infrared thermal imaging module. The environmental monitoring parameters include the current solar irradiance of the site, ambient temperature, array surface temperature and background level, wind speed and direction, air humidity, and weather event labels. The incremental learning update model structure records the version identifier of the currently active hidden crack identification model, the distribution characteristics of the sample data used during model training, the model's requirements for PL image spatial resolution and dynamic range, the model's requirements for infrared temperature resolution and temperature change sensitivity, as well as the model's recommended requirements for acquisition coverage redundancy, flight path spacing, and viewpoint diversity.

[0006] Furthermore, the process of verifying basic security boundaries also includes: The basic safety boundary verification process includes reading flight path planning data and UAV platform parameters, comparing the planned flight altitude with the minimum recommended flight altitude in the UAV platform parameters (if it is lower than the minimum recommended flight altitude, it is considered too low); comparing the planned turning radius with the minimum allowed turning radius in the UAV platform parameters (if it is smaller than the minimum allowed turning radius, it is considered too small); comparing the planned flight speed with the maximum cruising speed in the UAV platform parameters (if it exceeds the maximum cruising speed, it is considered too high); and marking risk indicators for flight path sub-tasks with such conditions. Combined with wind speed and direction labels, wind resistance capability verification records are added to the relevant flight path sub-tasks. A set of no-fly polygon areas is generated for geometric detection, and waypoints with intrusion risk are marked as waypoints requiring adjustment.

[0007] Furthermore, the process of synchronizing and binding the attitude compensation control parameters also includes: The processing and binding of synchronous triggering and attitude compensation control parameters includes reading the time resolution requirements for synchronous acquisition of PL and IR, reading the expected speed and length segment information of each route, reading the maximum triggering frequency and load processing capacity, combining and constructing synchronous triggering parameter placeholder fields, generating attitude compensation parameter placeholder fields, and binding the synchronous triggering parameter placeholder fields and attitude compensation parameter placeholder fields to each route subtask record respectively.

[0008] Furthermore, the processes of flat field correction, local contrast enhancement, and standard size cropping also include: Flat field correction processing includes loading a flat field reference file and performing pixel-level response correction through a lookup table. Local contrast enhancement processing involves dividing the data into several local windows, grouping each window, and applying different contrast stretching strategies. The standard size cropping process involves cropping the image into several standard size sub-images and assigning a component number, row and column index, and line segment group identifier to each sub-image.

[0009] Furthermore, the process of extracting stable temperature regions, selecting background reference points, and statistically processing regional temperatures also includes: The temperature field stable region extraction process includes removing candidate segments with hidden anomalies, smoothing the remaining segments and counting the fluctuation amplitude, and registering segments with fluctuation amplitudes below a preset stability threshold as temperature field stable regions. Background reference point screening process includes reading the brightness distribution and dark line feature summary of the component-level sub-image, judging the intact component candidate areas with uniform brightness and no obvious dark lines or local abnormal shadows through simple geometric rules and brightness thresholds, and cross-matching with temperature field stable areas to select background reference point candidates. Regional temperature statistical processing includes calculating the equilibrium level of temperature readings at different time scales, their upward or downward trends, and their correlation with environmental monitoring parameters.

[0010] Furthermore, the process of extracting hotspot region contours, temperature gradient fields, and component edge contour features, and determining pixel mapping relationships also includes: Hotspot region contour extraction processing includes constructing candidate hotspot regions, determining the boundaries of high-temperature regions, and obtaining fine contour curves; The temperature gradient field construction process involves calculating the temperature gradient changes in the horizontal and vertical directions to form a temperature gradient field description. The component edge contour extraction process includes edge detection and edge tracking to obtain the component's outer contour curve and the position of the internal dividing lines. The pixel mapping relationship retrieval process involves obtaining a family of spatial mapping functions and mapping each pixel position in the PL subgraph to the corresponding position in the infrared image.

[0011] Furthermore, the process of extracting candidate dark line regions in the PL, calculating the local infrared temperature gradient, and binding the component unit number also includes: The PL dark line region candidate extraction process involves dividing the image into blocks, constructing a set of sub-graphs with a single component unit as the smallest processing unit, and detecting elongated regions whose brightness is significantly lower than the background brightness of the same component unit through brightness histogram statistics, local contrast analysis, and connectivity analysis based on neighborhood relationships. Connected regions that satisfy length, width, and elongation constraints are identified as PL dark line candidate regions. The infrared local temperature gradient calculation process includes local analysis of the temperature field within a preset neighborhood range around each candidate region of PL dark line, calculating the temperature variation trend with spatial location along the long and short sides of the component to obtain local temperature gradient information, and constructing a comparison window inside and outside the candidate region to calculate the difference between the average temperature, maximum temperature and background temperature. The component unit number binding process includes reading the configuration data about the component array topology and component unit division in the dual-modal acquisition task configuration structure, overlaying and matching the geometric outer rectangle of the PL dark line candidate area with the component unit boundary, and recording the component number and component unit number.

[0012] Furthermore, the process of extracting geometric length, width, spatial overlap, and temperature difference indices, mapping normalization rules, and processing them through convolutional neural network inference also includes: The extraction and processing of geometric length and width indicators includes calculating the extension distance of the dark line candidate region along the direction of the main structure inside the component and the coverage area in the vertical direction; The spatial overlap index extraction process includes statistical analysis of the coverage area ratio of candidate regions within their respective component units and across adjacent component units; The temperature difference index extraction process includes calculating the temperature difference between the average temperature within the same candidate region and the average temperature of the stable region of the component unit, and comparing the temperature difference between the candidate region and the background window. Furthermore, the process of normalization rule mapping and convolutional neural network inference also includes: The normalization rule mapping process involves mapping indicators with different dimensions and different value ranges to a unified feature space, and performing cropping and mapping by referring to the indicator distribution range obtained from historical inspection tasks. The convolutional neural network inference processing involves inputting a local sub-image of photoluminescence and a local sub-image of infrared temperature as the image channel input to the network front end. It automatically learns the spatial correspondence between the morphology of PL dark lines and infrared temperature anomalies. It then concatenates the normalized geometric length, width, spatial overlap, and temperature difference indices with the high-dimensional feature vector output by the convolutional layer to provide a prediction result and confidence score for whether the candidate region is a hidden crack.

[0013] The key innovations of this invention include: (1) Based on the pixel-level aligned dual-modal image data structure, combined with PL dark line region candidate extraction, infrared local temperature gradient calculation and component unit number binding, the multi-dimensional features after the extraction of geometric length, width, spatial overlap and temperature difference index are input into the convolutional neural network for inference processing through normalization rule mapping to generate the hidden crack recognition result structure, thereby constructing a multi-modal feature fusion judgment link based on pixel-level coordinates within the same component unit.

[0014] (2) Based on the PL preprocessed image data structure, through timestamp alignment, channel health status detection and abnormal segment preliminary marking, a time-series temperature compensation model structure is generated on the basis of temperature field stable area extraction, background reference point screening and regional temperature statistical processing. In frame-by-frame temperature compensation operation, temperature field smoothing resampling, abnormal pixel marking update, hot spot area contour, temperature gradient field and component edge contour feature extraction and pixel mapping relationship calculation, a pixel-level aligned dual-modal image data structure is output, forming a time-series temperature compensation and spatial registration integrated processing link for photovoltaic power station scenarios.

[0015] (3) Starting from acquiring photovoltaic power station route planning data, UAV platform parameters, environmental monitoring parameters and incremental learning update model structure, a dual-modal acquisition task configuration structure is generated through task number registration, module mapping relationship construction, basic safety boundary verification and synchronous triggering and attitude compensation control parameter sorting and binding processing. This drives the generation of PL preprocessing image data structure, time-series temperature compensation model structure, pixel-level aligned dual-modal image data structure and hidden crack identification result structure in sequence. Based on component position information restoration, linkage layer layout and annotation layer generation, hidden crack sample annotation, training data construction and model parameter update and version registration processing, an incremental learning update model structure is generated, thereby constructing a closed-loop link connecting acquisition, preprocessing, compensation, identification and model update in photovoltaic power station.

[0016] The following are its main beneficial effects: (1) By simultaneously utilizing the candidate extraction results of PL dark line regions and the infrared local temperature gradient on the pixel-level aligned dual-modal image data structure, and binding them with the component unit number to extract geometric length, width, spatial overlap and temperature difference indicators, and then sending them into the convolutional neural network for inference processing through normalization rule mapping, the present invention expresses the photoluminescence dark line features and temperature field changes in the hidden crack identification result structure with unified coordinates and structured fields. Compared with the practice of threshold judgment and manual comparison based solely on photoluminescence images or infrared thermal images in the background technology, it can realize the hidden crack judgment with the joint participation of multimodal features at the same component unit and the same pixel-level position, making the hidden crack identification result structure more stable in distinguishing false defects and interference factors under complex illumination and temperature field change conditions, which is convenient for subsequent calling and verification by component unit dimension in the photovoltaic power station operation and maintenance system.

[0017] (2) By completing timestamp alignment and channel health status detection under the drive of PL preprocessing image data structure, abnormal segments are initially marked, and regional temperature statistics are generated to form a time-series temperature compensation model structure based on temperature field stable area extraction and background reference point screening. Then, the model is used to perform frame-by-frame temperature compensation calculation and smooth resampling on the infrared temperature field. At the same time, combined with abnormal pixel marking update and hot spot area contour, temperature gradient field and component edge contour feature extraction and pixel mapping relationship to obtain the output pixel-level aligned dual-modal image data structure, this invention can separate the overall temperature field shift caused by environmental temperature field fluctuations from the local temperature features related to microcracks under the circumstances of changes in photovoltaic power station environmental monitoring parameters and UAV platform attitude changes. Compared with the simple threshold judgment method in the background technology, the temperature gradient and hot spot area features on which the subsequent microcrack identification result structure depends can better reflect the component body state and reduce the impact of environmental changes on the stability of microcrack judgment.

[0018] (3) By uniformly managing photovoltaic power station route planning data, UAV platform parameters, environmental monitoring parameters and incremental learning update model structure in the dual-modal acquisition task configuration structure, and constructing a PL preprocessing image data structure, time-series temperature compensation model structure, pixel-level aligned dual-modal image data structure and hidden crack identification result structure by registering task numbers and module mapping relationships, and then generating the incremental learning update model structure by restoring component position information, linking layer layout and generating annotation layers, as well as hidden crack sample annotation, training data construction and model parameter update and version registration processing, this invention links the hidden crack identification result structure obtained from multiple inspections in photovoltaic power stations with photovoltaic power station route planning data and the spatial position of photovoltaic components in the long term. Compared with the practice in the background technology where the inspection results are independent and lack a unified component unit number and hidden crack evolution recording mechanism, the hidden crack sample annotation and training data construction can be continuously accumulated under a unified data model, which is convenient to be applied in the subsequent dual-modal acquisition task configuration structure through model parameter update and version registration processing. Thus, in the scenario of large-scale inspection and long-term operation and maintenance of photovoltaic power stations, a traceable, iterative and automatically callable incremental learning update model structure is formed. Attached Figure Description

[0019] Figure 1 This application provides a technical architecture diagram of a PL infrared fusion timing compensation photovoltaic module microcrack inspection method. Figure 2 A flowchart illustrating the implementation of a PL infrared fusion timing compensation photovoltaic module microcrack inspection method provided in this application embodiment; Figure 3 This is a schematic diagram of a PL infrared fusion timing compensation photovoltaic module microcrack inspection method provided in an embodiment of this application. Detailed Implementation

[0020] Figure 1 This application provides a technical architecture diagram of a PL infrared fusion timing compensation photovoltaic module microcrack inspection method, as shown in the embodiments below. Figure 1 As shown, the photovoltaic module microcrack inspection method proposed in this invention comprises a system architecture consisting of a multimodal data synchronous acquisition module, a time-series temperature compensation module, a dual-feature matching and recognition model module, an intelligent interaction and visualization system module, and a continuous optimization and knowledge management module, connected in sequence. The multimodal data synchronous acquisition module achieves hardware-synchronous acquisition and pixel-level alignment of photoluminescence images and infrared temperature data through a UAV integrated platform; the time-series temperature compensation module suppresses environmental temperature field fluctuations based on dynamic background reference point technology, generating stable temperature field data; the dual-feature matching and recognition model module fuses the geometric features of PL images and the thermal features of infrared images, and combines convolutional neural networks for collaborative reasoning; the intelligent interaction and visualization system module provides multi-level visualization interfaces and parameter interaction functions, supporting the linkage analysis and report generation of recognition results; and the continuous optimization and knowledge management module relies on an incremental learning mechanism to continuously optimize the model and accumulate knowledge. These modules work together to form a closed-loop intelligent inspection system, providing support for the accurate and automated detection of microcracks in photovoltaic modules.

[0021] Figure 2 A flowchart illustrating the implementation of a PL infrared fusion timing compensation photovoltaic module microcrack inspection method provided in this application embodiment is shown below. Figure 2As shown, the main steps include: system initialization and data acquisition steps: after the inspection system starts, it performs equipment self-test and parameter initialization, loads the preset flight path and detection parameters, and the UAV flies according to the predetermined path. Through the hardware synchronous triggering mechanism, it controls the photoluminescence imaging module and infrared thermal imager to collect data and simultaneously records the position, attitude and environmental parameters to generate a complete monitoring data archive; data preprocessing and registration compensation steps: the acquired PL images are subjected to flat field correction, noise filtering and contrast enhancement processing, the infrared temperature data is subjected to outlier removal and smoothing filtering, and the environmental temperature field fluctuation interference is eliminated based on the established time-series temperature compensation model. Then, the precise pixel-level spatial correspondence between the PL image and the infrared image is established through the feature point matching algorithm; multi-feature extraction and fusion analysis steps: the geometric morphology, distribution law and intensity features of the hidden cracks are extracted from the preprocessed PL image, and the temperature gradient distribution and hot spot area range are extracted from the compensated infrared image. Based on the changing trend characteristics, a feature association analysis model is constructed. By calculating the spatial matching degree and feature correlation index, the collaborative verification and deep fusion of dual-modal features are achieved, generating an association analysis model. The intelligent identification and interactive verification steps, based on the dual-feature fusion analysis results, employ a multi-level confidence assessment mechanism to identify and classify potential hidden crack areas. An interactive analysis tool is provided through a visual interface, supporting engineers to perform multi-dimensional verification and parameter adjustment of the identification results. The knowledge update and performance optimization steps involve the standardized storage and management of manually confirmed hidden crack samples, and the initiation of an incremental learning process. Based on newly added sample data, the feature extraction algorithm and identification model parameters are optimized to continuously improve the system's identification performance. Finally, the report generation and data management steps automatically generate standardized inspection analysis reports containing hidden crack distribution, feature parameter statistics, and maintenance suggestions, and establish a historical database system to provide data support for preventative maintenance and operational optimization of photovoltaic power plants.

[0022] Example 1: Refer to Figure 1 This is a schematic diagram of a PL infrared fusion timing compensation photovoltaic module microcrack inspection method provided by an embodiment of the present invention. The process may include at least steps S100-S500: S100: Acquire photovoltaic power station route planning data, UAV platform parameters, environmental monitoring parameters and incremental learning update model structure, register task numbers, construct module mapping relationships, verify basic safety boundaries and organize and bind synchronous triggering and attitude compensation control parameters to generate a dual-modal acquisition task configuration structure. S200, based on the dual-modal acquisition task configuration structure, performs index binding, image integrity verification, format standardization and flat field correction, local contrast enhancement and standard size cropping processing to generate PL preprocessed image data structure; S300, based on the PL preprocessing image data structure, performs timestamp alignment, channel health status detection, preliminary marking of abnormal segments, extraction of stable temperature field regions, background reference point screening and regional temperature statistical processing, and generates a time-series temperature compensation model structure. S400, based on the time-series temperature compensation model structure, performs frame-by-frame temperature compensation calculation, temperature field smoothing resampling, abnormal pixel marking update, and hot spot region contour, temperature gradient field and component edge contour feature extraction and pixel mapping relationship calculation to generate pixel-level aligned dual-modal image data structure. S500, based on a pixel-level aligned dual-modal image data structure, performs candidate extraction of PL dark line regions, calculation of infrared local temperature gradient and binding of component unit numbers, as well as extraction of geometric length, width, spatial overlap and temperature difference indices, normalization rule mapping and convolutional neural network inference processing, to generate a hidden crack identification result structure.

[0023] Step S100 includes at least steps S110-S130: S110: Acquire photovoltaic power station route planning data, UAV platform parameters, environmental monitoring parameters and incremental learning update model structure, register task numbers, construct module mapping relationships, verify basic safety boundaries and organize and bind synchronous triggering and attitude compensation control parameters to generate a dual-modal acquisition task configuration basic dataset. In this embodiment, photovoltaic power plant operation and maintenance personnel import or edit photovoltaic power plant route planning data through a ground control terminal. This data includes information such as the photovoltaic field boundary, module array layout, waypoint sequence for each inspection route, inspection altitude, heading angle, overlap rate configuration, and preset inspection time windows. Unmanned Aerial Vehicle (UAV) platform parameters are provided by the UAV platform management module. These parameters include at least the UAV model identifier, payload capacity, remaining battery capacity estimation method, maximum speed, maximum climb and descent rates, permissible attitude angle range, gimbal control accuracy, and the installation location and field-of-view geometry of the photoluminescence (PL) imaging module and infrared (IR) thermal imaging module. Environmental monitoring parameters are collected in real-time or periodically by the environmental monitoring subsystem. These parameters include at least the current solar irradiance of the field area, ambient temperature, array surface temperature background level, wind speed and direction, air humidity, and weather event tags such as dust storms and precipitation that may affect imaging quality. The incremental learning and updating model structure is generated by the identification model management module in the preceding steps. This structure records the currently active hidden crack recognition model version identifier, the distribution characteristics of the sample data used during model training, the model's requirements for PL image spatial resolution and dynamic range, the model's requirements for infrared temperature resolution and temperature change sensitivity, and the model's recommended requirements for acquisition coverage redundancy, flight path spacing, and viewpoint diversity. Four types of inputs constitute the minimum parameter set for configuring the dual-modal acquisition task. Flight path planning data, UAV platform parameters, and environmental monitoring parameters are mandatory inputs. The incremental learning and updating model structure is the core input for closed-loop updates; if it is missing, the system will only execute according to the default model constraints.

[0024] Specifically, upon receiving the aforementioned input, the task configuration module automatically creates a task number for this dual-modal inspection task on the ground control terminal. This task number is generated by concatenating the power station identifier, date, batch sequence number, and task type, and serves as the primary key field for all subsequent data records and configuration items. During task number registration, the task configuration module automatically scans the route groups and sub-segments in the route planning data, establishing a relationship between each route and the task number, and generating a route sub-task identifier. This identifier records the route number, the associated component array number, and the planned execution order. After task number registration is complete, the task configuration module associates the fields related to payload capacity and flight performance in the UAV platform parameters with each route sub-task for subsequent determination of whether the range and power supply meet the requirements of the sub-task. When maintenance personnel confirm the creation or update of a task through the interface, the task configuration module triggers a complete task number registration process and records the task creation timestamp and operator identifier for later traceability.

[0025] Furthermore, the module mapping relationship is automatically constructed by the module mapping subunit in the dual-modal task configuration module. Based on the installation location, field of view, and geometric relationship with the UAV body coordinate system of the PL imaging module and IR thermal imaging module recorded in the UAV platform parameters, the module mapping subunit establishes a one-to-one mapping relationship between each route subtask in the route planning data and its corresponding imaging module instance, gimbal control channel, flight control trajectory control channel, and environmental monitoring sampling channel. For route subtasks marked as requiring high-precision hidden crack inspection, the module mapping subunit prioritizes the PL imaging mode with higher spatial resolution and adds redundant shooting labels to the mapping relationship for subsequent configuration of more intensive exposure triggering rhythms. When a change in the model version number in the incremental learning update model structure is detected, the module mapping subunit automatically reloads the model's constraints on resolution and viewpoint diversity, performs batch verification of the module mapping relationships of existing route subtasks, provides adjustment suggestions for configuration items that do not meet the current model requirements, and after confirmation by maintenance personnel in the interface, the system automatically updates the mapping records and writes them into the task configuration base dataset. After the module mapping relationship is established, the task configuration module generates a multi-dimensional mapping table between the task number, the route subtask identifier, and the module instance, providing a complete structured input for the subsequent synchronization triggering and attitude compensation algorithms.

[0026] During the basic safety boundary verification process, the basic safety boundary verification subunit reads the flight path planning data and UAV platform parameters, and comprehensively compares the planned flight altitude, maximum lateral offset and turning radius in the flight path planning with the UAV's allowed attitude angle range, minimum recommended flight altitude and maximum cruise speed. Risk markers are added to flight path sub-tasks that have excessively low altitude, small turning radius or excessively high flight speed, and the specific reasons for the risks are recorded as structured safety prompt fields.

[0027] Specifically, the basic safety boundary verification process includes reading flight path planning data and UAV platform parameters, comparing the planned flight altitude with the minimum recommended flight altitude in the UAV platform parameters, and considering the altitude as too low if it is lower than the minimum recommended flight altitude; comparing the planned turning radius with the minimum allowed turning radius in the UAV platform parameters, and considering the turning radius as too small if it is smaller than the minimum allowed turning radius; comparing the planned flight speed with the maximum cruising speed in the UAV platform parameters, and considering the flight speed as too high if it exceeds the maximum cruising speed; marking risk indicators for flight path sub-tasks with such conditions, and adding wind resistance verification records to the relevant flight path sub-tasks in conjunction with wind speed and wind direction labels, generating a set of no-fly polygon areas for geometric detection, and marking waypoints with intrusion risk as waypoints that need to be adjusted.

[0028] Simultaneously, the basic safety boundary verification subunit, combining wind speed and direction labels from environmental monitoring parameters, adds wind resistance capability verification records to relevant flight path sub-tasks in areas where wind speed exceeds a set threshold, and writes the suggested upper limit of flight speed and pitch angle limit into the task configuration base dataset. If environmental monitoring parameters indicate the presence of weather events such as precipitation, dust storms, or severe fog, the basic safety boundary verification subunit marks the task status for the corresponding time window in the task configuration base dataset as requiring manual review and generates a corresponding scheduling flag field, which is manually confirmed by the scheduling system before task issuance. Based on this, the basic safety boundary verification subunit generates a set of no-fly polygon areas for the photovoltaic module array area, performs geometric detection on the intersection of the flight path and the no-fly area, marks waypoints with intrusion risk as needing adjustment, and generates records containing waypoint indexes and adjustment suggestions.

[0029] The synchronous triggering and attitude compensation control parameter organization and binding process in this step mainly completes the establishment of the parameter framework and the binding of task granularity. The synchronous triggering parameter organization subunit reads the temporal resolution requirements for PL and IR synchronous acquisition from the incremental learning update model structure, reads the expected speed and route length segment information for each route from the route planning data, and reads the maximum triggering frequency and payload processing capacity from the UAV platform parameters. It combines the three types of information to construct synchronous triggering parameter placeholder fields and sets different trigger density levels for different route sub-tasks. The attitude compensation control parameter organization subunit generates attitude compensation parameter placeholder fields based on the gimbal control accuracy, field-of-view center offset of the PL imaging module and IR thermal imaging module in the UAV platform parameters, combined with the suggested values ​​of heading angle and pitch angle in the route planning data. These fields include pitch angle offset, roll angle limit, gimbal centering strategy, etc. During the binding process, the task configuration module binds the aforementioned synchronization trigger parameter placeholder fields and attitude compensation parameter placeholder fields to each flight path subtask record, and establishes a primary key association with the task number and module mapping table. This allows subsequent steps to directly access the entire parameter framework related to synchronization triggering and attitude compensation based on the task number when receiving the dual-modal acquisition task configuration basic dataset. After the task configuration module completes task number registration, module mapping relationship construction, basic safety boundary verification, and synchronization triggering and attitude compensation control parameter organization and binding, it generates the dual-modal acquisition task configuration basic dataset. This dataset is registered in the data storage module as a structured product, containing task number, flight path subtask identifier, multimodal module mapping relationship, safety boundary record, and synchronization triggering and attitude compensation parameter placeholder fields. Logically, it serves as the sole input source for S120 to extract flight path information, module working mode, and attitude sampling interval from the dual-modal acquisition task configuration basic dataset, and also provides task-level configuration basis for PL image preprocessing and quality control in stage S200.

[0030] S120. Extract flight segment information, module working mode and attitude sampling interval from the dual-modal acquisition task configuration basic dataset, organize synchronous trigger parameters and generate attitude compensation control parameters, and generate a set of synchronous trigger and attitude compensation control parameters. In this embodiment, the synchronization triggering and attitude compensation configuration module first reads the target task record by task number from the dual-modal acquisition task configuration basic dataset generated in step S110, and extracts the flight line segment information field. The flight line segment information field includes the waypoint coordinate sequence of each flight line sub-task, the coverage area identifier of the flight line in the photovoltaic module array, the planned flight altitude, the planned heading angle, the flight line execution priority, the expected flight time window, and the azimuth relationship with the row and column directions of the module array. The synchronization triggering and attitude compensation configuration module performs topological sorting of the flight line segment information, grouping adjacent flight line sub-tasks that cover the same module array area into flight line segment groups, and generating a unified flight line segment group identifier and execution order label for each flight line segment group. In this way, the system can uniformly plan at the flight line segment group granularity when subsequently generating triggering rhythm and attitude compensation strategies.

[0031] The module operating mode field reads the operating mode labels of the PL imaging module and the IR thermal imaging module from the dual-modal acquisition task configuration base dataset. These operating mode labels include preset level definitions for PL image exposure time, imaging frame rate, spectral range, and gain level, and preset level definitions for IR temperature acquisition frame rate, temperature measurement range, temperature precision, and automatic range control strategy. When parsing the module operating mode field, the synchronization triggering and attitude compensation configuration module maps each mode label to a specific set of hardware and software configuration parameters. For example, it maps a PL imaging mode to a combination of fixed exposure time level, fixed gain level, and fixed imaging frame rate, and maps an IR temperature measurement mode to a combination of fixed temperature measurement range and frame rate. The attitude sampling interval field is derived from the attitude sampling strategy recorded in the task configuration base dataset. This field describes the time interval or flight interval at which the flight control module reads the UAV attitude data during flight path execution, used for subsequent attitude compensation of the PL and IR imaging results. Multiple flight path sub-tasks within the same flight path segment group maintain consistent temporal resolution of attitude data through a unified attitude sampling interval.

[0032] During the synchronization trigger parameter processing, the synchronization trigger and attitude compensation configuration module combines the line segment information, module operating mode, and attitude sampling interval to calculate the trigger rhythm of PL image exposure and IR image acquisition. Specifically, based on the length of the line segment group, the planned speed, and the frame rate parameter in the module operating mode, the synchronization trigger and attitude compensation configuration module derives the available trigger slot sequence for each line segment group during execution and maps these trigger slots to the range of the line segment group. For line segment groups requiring high-density acquisition, the synchronization trigger and attitude compensation configuration module allocates more PL exposure trigger events in the trigger slot sequence according to the acquisition coverage redundancy suggestion in the incremental learning update model structure, and intersperses IR acquisition trigger events in between; for line segment groups of ordinary inspection, trigger events are allocated according to the default redundancy. The synchronization trigger and attitude compensation configuration module also refers to the attitude sampling interval field to align the time point of the trigger event with the attitude sampling time point as much as possible, so that the adjacent attitude measurement values ​​can be directly used during subsequent attitude compensation without additional interpolation. During the process of organizing trigger parameters, if it is detected that the combination of the route segment group length and the planned speed results in too few trigger slots, the synchronous trigger and attitude compensation configuration module will regenerate the trigger rhythm by reducing the frame rate or adjusting the trigger interval, and record the reason for the adjustment in the task configuration basic dataset.

[0033] Attitude compensation control parameters are generated by the attitude compensation subunit within the same configuration module. The attitude compensation subunit reads the component array row and column directions and photovoltaic module tilt angle data from the flight path information, combines them with the gimbal mounting geometry from the UAV platform parameters, and derives the ideal camera optical axis pointing direction—the target observation direction for the PL imaging module and IR thermal imaging module—based on the heading angle and pitch suggestion values ​​for each flight path group. The attitude compensation subunit calculates the gimbal pitch offset, roll angle correction, and gimbal centering strategy parameters based on the deviation between this ideal observation direction and the UAV's body attitude, and generates the attitude compensation control parameters. In some implementations, the attitude compensation subunit can also update the model structure based on incremental learning to meet the model's requirements for diverse viewing angles, configuring multi-view shooting strategies for some flight path groups. This involves setting different pitch and yaw offsets for different flight path segments in the gimbal control parameters to obtain multi-angle PL and IR images. During the attitude compensation control parameter generation process, the attitude compensation subunit smooths out angle jumps that might cause excessive gimbal control shocks, constraining the attitude change between adjacent trigger events within the allowable range of the UAV platform parameters.

[0034] After completing the compilation of synchronization trigger parameters and the generation of attitude compensation control parameters, the synchronization trigger and attitude compensation configuration module combines the two to generate a synchronization trigger and attitude compensation control parameter set. This set, indexed by the task number and flight path group identifier, records the trigger time sequence, trigger type (PL exposure or IR acquisition), corresponding attitude compensation target, gimbal control parameters, and associated attitude sampling point identifiers for each flight path group. The synchronization trigger and attitude compensation control parameter set is written into the task configuration storage module and provided to the flight control module and payload control module through the task distribution interface, serving as the sole configuration source for synchronization trigger and attitude compensation execution during the execution phase. Simultaneously, this parameter set serves as input to S130, referenced by S130 during the acquisition task binding and module working sequence generation process, and as the basis for attitude constraints during subsequent PL infrared image spatial registration and feature alignment in S400.

[0035] S130. Bind the acquisition task to the synchronous triggering and attitude compensation control parameter set, generate the module working timing and solidify the acquisition safety constraints to generate a dual-modal acquisition task configuration structure. In this embodiment, the data acquisition task scheduling and configuration solidification module reads the parameter records corresponding to the target task from the set of synchronous triggering and attitude compensation control parameters generated in S120 according to the task number, and creates a new data acquisition task configuration entry in the ground control terminal. The data acquisition task binding subunit first binds each group of synchronous triggering and attitude compensation control parameters to the corresponding line subtask record in the dual-modal data acquisition task configuration basic dataset according to the line segment group identifier, forming a hierarchical binding relationship of task level, line segment group level, and trigger event level. The data acquisition task binding subunit also writes the model version identifier in the incremental learning update model structure into the data acquisition task configuration entry, which is used as the basis for determining which version of the identification model and acquisition strategy was used when retrospectively reviewing the data acquired by the task. When the maintenance personnel perform the "task issuance" operation in the interface or the system scheduling module automatically reaches the predetermined time window according to the power station inspection plan, the data acquisition task scheduling and configuration solidification module marks the bound data acquisition task configuration entry as pending execution according to the existing task status, and triggers the module work sequence generation process.

[0036] After acquiring the task binding results, the module timing generation subunit generates a module timing table according to the task number and flight path execution order. The module timing table divides the UAV into multiple stages, including pre-flight self-check, climb phase, inspection flight path execution phase, return phase, and landing phase, further refining the inspection flight path execution phase to time slices at the flight path segment group level. For each time slice, the module timing generation subunit generates a specific module operation command sequence based on the trigger event sequence and attitude compensation target recorded in the synchronization trigger and attitude compensation control parameter set. This sequence includes the startup time, warm-up time, exposure or acquisition trigger time, attitude compensation execution command issuance time, and environmental monitoring parameter sampling time for the PL imaging module and IR thermal imaging module. In some implementations, the module timing generation subunit also inserts mid-course return assessment checkpoints for long-range missions based on the battery capacity estimation model in the UAV platform parameters, and configures endurance assessment logic for each checkpoint in the timing table. Once the module's operational timing sequence is generated, a time-ordered module operational timing structure is formed. This structure is not only distributed to the flight control module and payload control module for execution during flight, but also a copy is stored in the data processing system to provide a reference for subsequent time alignment and abnormal event marking of the PL preprocessed image data structure and infrared temperature data.

[0037] During the data acquisition and safety constraint solidification process, the data acquisition task scheduling and configuration solidification module integrates the safety boundary records generated from the basic safety boundary verification in S110 with the currently generated module working sequence structure. The safety constraint solidification subunit attaches safety constraint labels to each flight path segment group and each type of trigger event, such as minimum flight altitude limits, maximum cruise speed limits, maximum permissible roll and pitch angle deviation ranges, no-entry zone identifiers, and pause trigger strategies triggered under specific environmental monitoring parameter conditions. The safety constraint solidification subunit transforms these safety constraint labels into a set of control rules that can be directly executed by the flight control module and payload control module. The rule set specifies automatic handling strategies when the actual flight parameters of the UAV conflict with the boundary conditions in the safety constraint labels, such as pausing PL exposure trigger events, reducing airspeed, switching to a safe altitude, or interrupting the current flight path and returning to base. The safety constraint solidification subunit also records the rule set version number and generation timestamp in the data acquisition task configuration entry, forming a safety constraint version management record, which facilitates subsequent auditing and post-mortem analysis of problematic tasks. In an optional implementation, when the incremental learning update model structure indicates that the model has an increased demand for samples from certain extreme viewpoints, the safety constraint solidification subunit will adjust the attitude constraint labels of some flight line segments within the allowable range of the safety boundary, so that the system can obtain more diverse viewpoint samples while protecting flight safety.

[0038] After completing the acquisition task binding, module working sequence generation, and acquisition safety constraint solidification, the acquisition task scheduling and configuration solidification module encapsulates the task-level binding relationship, timing structure, and safety rule set into a unified dual-modal acquisition task configuration structure. This dual-modal acquisition task configuration structure is stored in the task configuration storage module using the task number as the primary key. When the task is issued, it is transmitted from the ground control terminal to the UAV flight control module and payload control module via a data link, serving as the sole configuration basis for the joint acquisition of PL imaging and IR thermal imaging during the execution phase. Simultaneously, within the data processing system, the dual-modal acquisition task configuration structure serves as input for step S210, where it is used to obtain the dual-modal acquisition task configuration structure, perform index binding, image integrity verification, and format standardization processing. Furthermore, it provides a unified task configuration reference for subsequent steps, including PL image preprocessing, infrared temperature sequence preprocessing, temporal temperature compensation modeling, and PL infrared spatial registration and feature alignment. In summary, the technical effects of this step are as follows: By completing task-level parameter binding, module working sequence generation, and security constraint solidification in a single step, the task configuration elements related to dual-modal acquisition are organized into a structured dual-modal acquisition task configuration structure. This enables subsequent acquisition and processing stages to run under the same configuration baseline, improves the consistency of multi-module collaborative acquisition and the traceability of task configuration, and establishes a stable closed-loop relationship between incremental updates of the identification model and on-site acquisition strategies.

[0039] Step S200 includes at least steps S210-S230: S210. Obtain the dual-modal acquisition task configuration structure, perform index binding, image integrity verification and format standardization processing, and obtain the PL original image preprocessing result set; In this embodiment, after the UAV completes a photovoltaic array flight path mission, the data receiving and preprocessing subsystem receives the photoluminescence image data acquired during the mission from the onboard storage unit or wireless transmission link via the data access module. The photoluminescence image is in grayscale or pseudo-color format output by the Photoluminescence (PL) imaging module. The original file also contains metadata such as flight path number, timestamp, shutter parameters, gain level, and shooting posture. After receiving the image data, the data access module first reads the corresponding acquisition task configuration entry from the dual-modal acquisition task configuration structure generated in S130 according to the task number and flight path segment group identifier. This entry records the PL imaging mode, trigger rhythm configuration, and attitude compensation strategy to be executed for each flight path segment in this mission. Based on this, the data access module constructs a task-level index table, establishing an index record for each frame of PL image according to the task number, flight path segment identifier, trigger sequence number, and timestamp, and establishing a mapping relationship with the module working timing table in the dual-modal acquisition task configuration structure, thereby forming a bidirectional index between the original PL image file and the acquisition task configuration.

[0040] Specifically, after completing the task-level index construction, the index binding subunit compares the file header information of each PL image frame with the trigger time sequence in the dual-modal acquisition task configuration structure one by one. Based on the similarity of timestamps and the matching relationship of trigger sequence numbers, it binds the image frame with the corresponding trigger event record. If a trigger event is found to have no corresponding frame in the image data, a missing frame marker field is generated in the index table, and the missing frame event is registered in the log recording module. For cases where multiple candidate images exist within the same time window, the index binding subunit compares configuration fields such as shutter parameters, gain levels, and imaging modes, selecting the frame that is completely consistent with the recorded values ​​in the dual-modal acquisition task configuration structure as the valid binding object. The remaining frames are treated as redundant frames, and redundant labels are marked in the index table for subsequent analysis. Through the above process, the index table output by the index binding subunit structurally records the correspondence between each PL image frame and the task number, flight path group identifier, trigger event, and attitude sampling point identifier. This index table serves as one of the basic data for subsequent image integrity verification and format standardization processing, and is also directly written into the PL raw image preprocessing result set.

[0041] Image integrity verification is performed by an integrity verification subunit. This subunit reads the file path and metadata of each PL image frame from the index binding results, loads the image content frame by frame, and verifies the image size, number of data channels, and file length. If the image size is found to be inconsistent with the expected resolution in the dual-modal acquisition task configuration structure, or if the number of image channels is not within the predefined set of grayscale single-channel and pseudo-color three-channel, the integrity verification subunit marks the image as having an abnormal size and records the abnormality type in the preprocessing result set. The integrity verification subunit also performs truncation detection on the end of the image file, marking images with a file length significantly less than a preset threshold as suspected interrupted files. At the image content level, the integrity verification subunit statistically analyzes the brightness histogram distribution range of the entire image. If it detects that most pixels are concentrated near the lowest or highest grayscale, it determines that the image is severely underexposed or severely overexposed, registers the underexposed or overexposed flags in the preprocessing result set, and stores a brightness statistical summary. This subunit reads environmental monitoring parameters and UAV attitude information fields simultaneously. When there is a significant attitude change or strong wind label in the corresponding time period of a certain frame image, a flight attitude risk mark will be added to the frame record to provide a reference for the subsequent quality screening stage.

[0042] Format specification processing is performed by the format specification subunit. This subunit, referencing the imaging mode definition in the dual-modal acquisition task configuration structure, uniformly converts PL images from different sources into a preset internal standard format. The specific processing includes unifying the image encoding format to lossless compression or uncompressed format, unifying the color space to a single-channel intensity space, and, in scenarios with color encoding, retrieving the original intensity values ​​through a lookup table and standardizing the original intensity range to a predetermined bit depth. Furthermore, the format specification subunit corrects boundary pixels related to the gimbal cropping area in the image based on the PL imaging field of view recorded in the task configuration, ensuring strict alignment of the image matrix with the preset sampling grid. For cases where the detected data channel order is inconsistent with the standard, the format specification subunit restores the standard channel order through channel rearrangement and records the rearrangement history in the preprocessing result set. After the format standardization process is completed, the system encapsulates the index binding result, integrity verification result, and standardized image after format conversion into a PL original image preprocessing result set. This result set is the input basis for subsequent processing in the main step S200. At the same time, in the process, it serves as the direct data source for S220 to extract image brightness distribution, noise level, and occlusion from the PL original image preprocessing result set. In the subsequent S300, it serves as an intermediate state before the generation of the PL preprocessed image data structure and participates in timestamp alignment and channel health status detection.

[0043] S220. Extract the image brightness distribution, noise level and occlusion from the PL original image preprocessing result set, update the flat field correction parameters and select the noise filtering strategy to generate the PL quality screening image set. In this embodiment, the quality analysis and screening submodule reads each frame of the format-standardized PL image from the PL raw image preprocessing result set obtained in S210, in order of task number and flight line segment group identifier, along with the corresponding integrity check label, environmental monitoring summary, and attitude risk marker. The quality analysis and screening submodule is internally configured with a brightness analysis unit, a noise assessment unit, and an occlusion detection unit. These three units jointly analyze different quality dimensions of the PL image and write the analysis results into a quality feature record table. The brightness analysis unit constructs a multi-block brightness statistical grid on the entire PL image, dividing the image into several regular sub-blocks. Each sub-block separately calculates the average gray level, extreme gray level, and gray level distribution dispersion, forming a brightness distribution feature vector. The noise assessment unit selects relatively smooth texture areas in each sub-block and calculates the variation amplitude of gray level differences between adjacent pixels, forming a noise level feature. The occlusion detection unit performs aggregate analysis on large-area low-brightness areas and high-frequency texture abrupt change areas based on the PL imaging characteristics, identifying abnormal areas caused by sensor occlusion, dirt, or strong reflection, and outputs occlusion mask estimation results. The above analysis process is automatically triggered by the quality analysis and screening submodule. After the data access module completes the writing of a batch of PL raw image preprocessing result sets, the quality analysis and screening submodule detects the new task number entering the analysis queue and begins to analyze all image frames under that task frame by frame.

[0044] Furthermore, the flat-field correction parameter update logic is based on the image brightness distribution feature vector output by the brightness analysis unit. The system pre-maintains a flat-field correction reference file in the PL imaging module, which records the response baseline corresponding to the pixel position of each sensor under standard conditions. The quality analysis and screening submodule compares the brightness distribution statistics in the current task with the flat-field correction reference file, marks scenes with systematic brightness shifts in each sub-block, and determines whether the shift trend is a global uniform shift or a local area shift. For global uniform shifts, the flat-field correction parameter update logic updates the flat-field correction parameters by adjusting the overall gain correction factor; for local area shifts, it overlays a local correction term on the flat-field reference file and records the component array position corresponding to the correction area. The updated flat-field correction parameters do not directly overwrite the original reference file, but are stored in the parameter management module as a new version, and the flat-field correction parameter version number associated with the current image frame is written into the PL quality screening image set, thereby constructing a clear parameter version management link.

[0045] The noise filtering strategy selection is driven by the noise level features output by the noise assessment unit. The noise assessment unit first classifies the noise patterns of the entire image based on the noise level distribution of each sub-block, distinguishing between different patterns such as low noise, uniform medium noise, and local strong noise. The quality analysis and screening submodule also considers environmental monitoring parameters and the imaging mode settings in the acquisition task configuration. In flight segments with large illumination fluctuations or drastic attitude changes, the noise pattern classification weights are shifted towards local strong noise. After synthesizing the above information, the noise filtering strategy selection logic assigns different denoising schemes to each noise pattern. For example, a lightweight smoothing filtering scheme is used for low-noise images, a spatial smoothing and edge protection parallel scheme is used for uniform medium noise images, and a regional strong filtering scheme guided by occlusion masks is used for locally strong noise images. The options for different schemes are recorded in the quality feature record table as part of the image quality screening results and are invoked in subsequent flat field correction and local contrast enhancement stages.

[0046] The occlusion detection results directly affect the PL quality screening image set generation logic. After obtaining the occlusion mask estimate, the occlusion detection unit compares the proportion of the occluded area in the entire image with a preset threshold. When the proportion of the occluded area exceeds the threshold, the image is marked as severely occluded and will not be included in the next step of the PL quality screening image set. When the proportion of the occluded area is in the middle range, the image is added to the selectable image set, and the quality label records whether the occluded area is located at the upper edge, lower edge, or middle region of the component array, providing location information for subsequent engineering personnel to review. When the proportion of the occluded area is low, it is considered acceptable occlusion. After integrating the brightness analysis results, noise evaluation results, and occlusion detection results, the quality analysis and screening submodule generates a quality level label and a screening decision label for each frame of image. Only images that pass the integrity check and whose quality level meets the preset standard will be fully written into the PL quality screening image set.

[0047] In one implementation example, the photovoltaic power plant operation and maintenance platform is deployed in the station control center's computer room, and the platform runs the quality analysis and screening submodule. After a drone completes a module array inspection task, it uploads PL image data in batches to the station control center's storage server via a high-speed wireless link. After receiving the completion event, the data access module generates a set of preprocessed PL raw images for the task and puts them into the screening queue. When the quality analysis and screening submodule detects a new task in the queue, it automatically starts the brightness analysis, noise assessment, and occlusion detection processes without manual intervention; after the task is completed, it automatically generates a set of PL quality screening images and sends a quality analysis completion notification to the upstream task scheduling module. Thus, in this embodiment, the PL quality screening process runs automatically entirely based on internal system trigger conditions, which are mainly driven by the data access completion event and changes in the task number. The final generated set of PL quality screening images contains the image content itself, quality labels, and parameter version information. This set serves as the sole input source for S230 flat field correction and local contrast enhancement processing, and also provides samples for subsequent optimization of flat field reference files and noise filtering strategies.

[0048] S230. Perform flat field correction, local contrast enhancement and standard size cropping on the PL quality screening image set to generate PL preprocessed image data structure; In this step, the PL image preprocessing module reads the image frames that have passed quality screening from the PL quality screening image set output by S220 in order of task number, line segment group identifier, and trigger sequence number, along with the accompanying quality labels and parameter version information. The PL image preprocessing module internally includes a flat field correction subunit, a local contrast enhancement subunit, and a standard size cropping subunit. These three subunits process each frame in the PL quality screening image set sequentially, and record the intermediate states generated during processing in a local cache for troubleshooting and parameter playback.

[0049] The flat-field correction subunit first loads the corresponding version of the flat-field reference file from the parameter management module based on the flat-field correction parameter version number recorded for each frame of the image. The flat-field reference file records the sensor response baseline and local correction terms at each location, indexed by pixel coordinates. The flat-field correction subunit associates each pixel in the image with the corresponding record in the flat-field reference file, obtains the correction weight for that pixel through a lookup table, and completes pixel-level response correction, eliminating spatial fixed-mode response differences in the sensor array. For areas marked by the quality analysis and screening submodule as having local brightness shifts, the flat-field correction subunit also adds local correction weights to bring the brightness level of the corrected area back to the baseline range defined by the reference file. The entire flat-field correction process is executed sequentially in the image memory, without changing the spatial dimensions of the image, only adjusting the pixel intensity values; simultaneously, the flat-field correction subunit writes a summary of the average brightness and contrast of the entire image before and after processing into the preprocessing record, providing a reference for subsequent local contrast enhancement.

[0050] After receiving the flat-field corrected image data, the local contrast enhancement subunit divides the entire image into several local windows, each covering several battery cell regions, based on the PL imaging characteristics and the manifestation patterns of hidden cracks. The subunit first groups the local windows based on the PL image brightness distribution. For windows with low brightness and slow brightness gradient changes, a stronger contrast stretching strategy is applied; for windows with high brightness or obvious texture structures, an edge protection strategy is used to prevent the contrast enhancement process from damaging the edge morphology of the hidden crack dark lines. Based on the noise pattern classification labels generated for each frame in stage S220, the subunit selects different contrast enhancement curves for different noise patterns. For example, a slight smoothing operation is introduced for medium-noise images to reduce noise amplification after enhancement; for locally strong noise images, contrast enhancement is only applied to the unmasked areas under the guidance of an occlusion mask. After local contrast enhancement is completed, the subunit records the enhancement parameter level and enhancement version number in the image metadata, facilitating playback of the enhancement process during subsequent model training.

[0051] After acquiring the PL image processed with flat-field correction and local contrast enhancement, the standard-size cropping subunit crops the entire image into several standard-size sub-images based on the component array layout information, image field of view, and standard cell size recorded in the dual-modal acquisition task configuration structure. Each standard-size sub-image corresponds to one or more cell units in the photovoltaic module. The standard-size cropping subunit assigns a component number, row and column index, and flight path segment group identifier to each sub-image according to the component number mapping table in the task configuration. For scenarios with slight parallax and attitude deviation, the standard-size cropping subunit refers to the attitude compensation target recorded in the module working timing table in S130 and fine-tunes the cropping window position to ensure that the sub-image covers the entire cell area as much as possible while maintaining component boundary alignment. In this embodiment, the standard-size cropping subunit also supports generating multi-view sub-images for the same component unit. When a multi-view shooting strategy exists in the task configuration, a separate sub-image is generated for each viewpoint, and viewpoint labels are recorded in the metadata. After standard-size cropping is completed, the PL image preprocessing module organizes all standard-size sub-images under the same task into a structured dataset. The dataset not only stores the sub-image content itself, but also metadata such as component number, row and column index, viewpoint label, processing parameter version number, and quality label.

[0052] Through the above processing, the PL image preprocessing module transforms the PL quality-screened image set into a PL preprocessed image data structure. This PL preprocessed image data structure serves as a key intermediate product in the method of this invention. On one hand, in terms of process structure, it is the sole input source for S300 when performing timestamp alignment, channel health status detection, and preliminary marking of abnormal segments based on the PL preprocessed image data structure. On the other hand, in cross-main-step association, it serves as the basic data for the PL side during subsequent infrared temperature field processing and dual-modal spatial registration, providing a unified, standardized, and fully annotated information foundation for subsequent hidden crack feature extraction and convolutional neural network inference. The technical effect of this step can be summarized as follows: By completing the unified processing of PL images from task configuration association, integrity verification, quality screening to flat field correction, local contrast enhancement, and standard-size cropping in consecutive steps from S210 to S230, the original PL images are organized into a PL preprocessed image data structure with a clear structure, controllable quality, and well-defined component-level positioning. This provides a stable and reliable data foundation for subsequent time-series temperature compensation modeling, dual-modal spatial registration, and hidden crack feature recognition, and significantly reduces the interference of original imaging defects on the downstream recognition process.

[0053] Step S300 includes at least steps S310-S330: S310. Obtain the PL preprocessed image data structure and the original infrared temperature sequence data, perform timestamp alignment, channel health status detection and preliminary marking of abnormal segments, and obtain the infrared raw temperature preprocessing result set. In this embodiment, the temporal alignment and channel health analysis subsystem first reads the acquisition time stamp, component array position, flight line segment group identifier, and processing parameter version information of each standard-size sub-image from the PL preprocessed image data structure output from the previous main step S200, in order of task number and component number. Simultaneously, it accesses the raw infrared temperature sequence data corresponding to the same task number from the infrared acquisition and storage module. This raw infrared temperature sequence data records the infrared imaging frames or area temperature sampling values ​​arranged in chronological order, along with metadata such as acquisition time stamp, infrared channel number, infrared temperature measurement mode level, and field of view coverage. When the temporal alignment and channel health analysis subsystem detects that both the PL preprocessed image data structure and the raw infrared temperature sequence data for the same task number have arrived, it triggers the batch processing flow of this step through the task scheduling module, without requiring manual intervention, thus forming an automated temporal processing starting point. The timestamp alignment subunit first constructs the internal timeline of the task, mapping the image time tags in the PL preprocessed image data structure to a unified timeline according to component number and flight segment group identifier. Then, it maps the acquisition time tags in the raw infrared temperature sequence data to the same timeline, forming a component-level time curve on the timeline for each combination of component number and flight segment group. Following a preset time proximity window, the timestamp alignment subunit searches for infrared samples near the acquisition time of each sub-image in the PL preprocessed image on each component-level time curve. It establishes one-to-one or one-to-many matching relationships between infrared samples with time intervals within the allowable range and their corresponding PL sub-images, generating a missing match marker in cases where no suitable infrared sample can be found. This matching relationship is written into an intermediate mapping table and recorded along with the component number and flight segment group identifier in the raw infrared temperature preprocessing result set.

[0054] Specifically, after completing timestamp alignment, the channel health status detection subunit aggregates sample sequences from the raw infrared temperature sequence data according to channel number. Each channel record corresponds to a physical temperature measurement point on the infrared imaging sensor or a spatially aggregated area temperature measurement point. Within each channel record, the subunit calculates the fluctuation range and instantaneous jump characteristics of the temperature reading over the entire mission execution time, and combines this with the infrared temperature measurement mode setting, environmental monitoring parameters, and flight attitude label at the time of acquisition to determine the channel's operating status under the current mission. When a channel's reading remains constant throughout the mission or only exhibits a very small number of non-zero values, the subunit marks that channel as a suspected frozen channel. When a channel experiences an abnormal temperature jump within a very short time while neighboring channels do not show similar changes, that channel is marked as a suspected spike noise channel. When a channel has normal readings at the beginning of the mission but has no data or abnormal data for a long period afterward, it is marked as a suspected mid-mission failure channel. All the above markers are written into the infrared raw temperature preprocessing result set through structured state fields, while retaining the original temperature sequence content, so as to make a comprehensive judgment in the subsequent extraction of stable temperature fields and background reference point screening.

[0055] Furthermore, the preliminary anomaly segment marking logic is initiated after the channel health status detection is completed. The preliminary anomaly segment marking subunit scans the temperature change trajectory from the raw infrared temperature sequence data according to component number and time sequence, performs segmentation processing on each component-level time series, and constructs a time segmentation structure composed of multiple continuous time segments. For each time segment, the preliminary anomaly segment marking subunit compares the temperature fluctuation amplitude within the segment with the temperature level difference between the adjacent segments, and refers to event labels recorded in the environmental monitoring parameters such as rapid cloud cover, strong winds, or ambient temperature jumps, to mark segments that show sudden temperature rises or falls but do not correspond to environmental event labels, marking these segments as hidden anomaly candidate segments. In this process, the preliminary anomaly segment marking subunit also considers the channel health status field to avoid misjudging the channel's own hardware failure behavior as abnormal temperature field behavior. After the initial marking of abnormal segments is completed, the infrared raw temperature preprocessing result set forms three key information layers: timestamp alignment mapping layer, channel health status recording layer, and abnormal segment initial marking layer. This result set is used by S320 to extract temperature field stable regions, noise fluctuation regions, and intact component candidate regions from the infrared raw temperature preprocessing result set. It also serves as the basic data structure for other sub-steps within S300 throughout the entire process. In addition, it serves as one of the inputs for infrared data quality judgment before S400 performs frame-by-frame temperature compensation calculations in cross-main step association.

[0056] S320. Extract stable temperature field regions, noise fluctuation segments, and candidate regions of intact components from the infrared raw temperature preprocessing results set, perform background reference point screening and regional temperature statistics, and generate a time-series temperature compensation reference point set. In this embodiment, the temperature field stable region identification and background reference point screening subsystem reads each component-level time series and its corresponding preliminary abnormal segment markers and channel health status records from the infrared raw temperature preprocessing result set output by S310. Combined with the component-level image quality labels in the PL preprocessing image data structure, it performs a comprehensive analysis of stability behavior in both spatial and temporal dimensions. The temperature field stable region identification unit first segments the component-level time series according to the segment boundaries given by the preliminary abnormal segment markers. Segments marked as potential hidden anomalies are removed from the candidate set, retaining only the remaining segments as temperature field stable candidate segments. Then, within each candidate segment, temperature readings are smoothed and fluctuation amplitudes are statistically analyzed. Segments with fluctuation amplitudes below a preset stability threshold are registered as temperature field stable regions. This process is automatically triggered by the system when it detects an update to the infrared raw temperature preprocessing result set and the channel health status detection is completed. The scheduling module initiates the temperature field stable region identification task sequentially according to the task number, achieving automatic screening of infrared data stability in batch inspection tasks.

[0057] After identifying stable temperature regions, the noise fluctuation segment analysis subunit finely divides the excluded time segments, initially classifying the areas within the coverage of abnormal segments into three categories based on temperature fluctuation patterns: short-term high-frequency noise segments, slow drift segments, and structural change segments. Short-term high-frequency noise segments typically exhibit consistent spikes or fluctuations in temperature readings over a very short period; slow drift segments show slow, unidirectional temperature changes over a longer period; and structural change segments exhibit persistent temperature differences corresponding to the dark line areas in the component data structure of the PL preprocessing image. The noise fluctuation segment analysis subunit records the segment boundary, fluctuation feature summary, and associated component number for each category, which are used as exclusion criteria or separate statistical objects in subsequent regional temperature statistical processing. For segments marked as spike noise segments, the system completely excludes them when constructing the time-series temperature compensation reference point set; for slow drift segments and structural change segments, they are used as reference inputs for trend analysis in the regional temperature statistics module.

[0058] The identification of intact component candidate regions is performed by the intact region selection subunit. This subunit reads the brightness distribution and dark line feature summaries of component-level sub-images from the PL preprocessed image data structure. Before inference by the hidden crack identification model, it uses simple geometric rules and brightness thresholds to determine component regions exhibiting uniform brightness and without obvious dark lines or local abnormal shadows in the current task, which are tentatively designated as intact component candidate regions. The intact region selection subunit cross-matches the component numbers and sub-image position indices of these regions with the temperature field stable regions in the infrared raw temperature preprocessing result set, selecting regions that have stable regions in the time dimension and intact component characteristics in the spatial dimension as background reference point candidates. For each background reference point candidate, the background reference point selection subunit records its pixel coordinates or region index on the infrared imaging plane, sub-image coordinates in the PL preprocessed image, component number and flight line segment group identifier, and records the temperature statistical characteristics of the reference point in the stable region, such as average temperature, extreme temperature, and temperature change amplitude within a certain time window, in conjunction with local environmental monitoring parameters.

[0059] The regional temperature statistics module performs more refined temperature statistics on the regions where each reference point is located, building upon the background reference point screening. The module treats each background reference point's corresponding time-series slice within a stable temperature field region as a record, calculating the equilibrium level, upward or downward trend of temperature readings at different time scales, and their correlation with environmental monitoring parameters. These statistical results are then archived as a time-series temperature feature summary. For multiple background reference points within the same component number and adjacent component regions, the module establishes cross-reference relationships. This involves recording temperature difference statistics between local areas and checking for long-term deviations of a reference point from neighboring reference points. If such deviations exist, the reference point is marked as unreliable and removed from subsequent modeling processes. Finally, the stable temperature field region identification results, noise fluctuation segment classification results, intact component candidate region information, and regional temperature statistics results are combined into structured data, forming a time-series temperature compensation reference point set. The time-series temperature compensation reference point set contains multiple reference point records. Each record is bound to a component number, spatial location index, stable time segment description, and temperature statistical summary. This set serves as the sole input source for the subsequent sub-step S330 of this main step to construct the time-series temperature compensation model structure. At the same time, at the cross-main step level, it provides background reference point parameters that can be directly called for frame-by-frame temperature compensation calculations in S400.

[0060] S330. The time-series temperature compensation reference point set is grouped by time period, temperature trend modeled and compensation coefficient registered to generate a time-series temperature compensation model structure. In this implementation, the time-series temperature compensation modeling subsystem reads each background reference point record from the time-series temperature compensation reference point set output by S320 and groups them according to task number, component array region, and acquisition date, thereby constructing multiple time period groups. Each group corresponds to a batch of reference point temperature sequences acquired under similar environmental and operational conditions. When constructing the groups, the time period grouping unit comprehensively uses acquisition time, estimated solar altitude angle, ambient temperature level, and inspection batch information. Tasks that are adjacent on the time axis and have small differences in environmental conditions are grouped into the same time period group, while tasks with significant seasonal differences or large differences in ambient temperature levels are created into independent group records. In this way, the time-series temperature compensation modeling subsystem forms multi-dimensional partitions in the data structure based on time and environmental conditions, constructing dedicated temperature compensation strategies for different time periods and environmental backgrounds.

[0061] After grouping by time period, the temperature trend modeling unit models the temperature sequences recorded by all background reference points within each time period group. Based on the average level, slow change trend, and occasional fluctuations of the reference point temperature sequences, the unit decomposes temperature changes into three levels: background temperature level, slow change component, and local disturbance component. The background temperature level component synthesizes the temperature statistics of multiple reference points within a time period to form a baseline temperature description for that period. The slow change component generates trend parameters describing the overall slow temperature drift behavior within that time period by analyzing the overall rising or falling trend of reference point temperatures along the time axis. The local disturbance component records small-scale shifts caused by short-term weather changes or changes in operational status. The temperature trend modeling unit employs different modeling granularities for different time period groups. For time period groups with relatively slow environmental changes, a longer time window is used to describe the trend; for time period groups with rapid environmental changes, a shorter time window and finer-grained trend description are used. Simultaneously, it continuously records information such as the number of background reference points used in model construction, their distribution range, and data coverage duration for subsequent auditing and model quality assessment.

[0062] After acquiring the temperature trend modeling results, the compensation coefficient registration unit extracts a set of compensation coefficients for infrared temperature field correction from the background temperature level, slowly changing components, and local disturbance components, based on the needs of signal measurement and model use. The minimum set of compensation coefficients typically includes the temperature difference offset and temperature drift rate for each component array region and time period group, used to correct the original infrared temperature values ​​during S400 frame-by-frame temperature compensation calculations. In a preferred embodiment, the compensation coefficients can also be extended to include extended parameters such as ambient temperature response sensitivity coefficients and wind speed and direction correction factors to improve compensation accuracy in special scenarios. When registering compensation coefficients for each component array region, the compensation coefficient registration unit simultaneously records the corresponding time period group number, background reference point list, modeling version number, and creation timestamp, forming a compensation parameter archive with version management capabilities. This archive is organized in the parameter management module by task number and effective time range, enabling the S400 to automatically query the matching compensation coefficients based on the infrared frame acquisition time during subsequent frame-by-frame temperature compensation calculations, reducing the need for manual intervention.

[0063] The time-series temperature compensation modeling subsystem encapsulates the time-segment grouping structure, temperature trend modeling results, and compensation coefficient registration archive into a unified time-series temperature compensation model structure. This time-series temperature compensation model structure is stored in the model management module using task number and model version number as indexes, and exposes an interface containing component array region-level compensation parameters, time-segment grouping information, and reference point traceability information. Internally, the time-series temperature compensation model structure serves as the sole modeling basis for S400 to acquire the time-series temperature compensation model structure and perform frame-by-frame temperature compensation calculations, temperature field smoothing resampling, and abnormal pixel marker updates. Simultaneously, at the cross-main-step level, it provides a time-series reference for the correlation analysis between subsequent microcrack identification results and environmental background temperature behavior. The technical effects of this step can be summarized as follows: By performing time-segment grouping, temperature trend modeling, and compensation coefficient registration on the time-series temperature compensation reference point set, this invention constructs a well-defined, traceable time-series temperature compensation model structure with version management capabilities. This allows for the use of differentiated compensation parameters for different time periods and environmental conditions in subsequent infrared temperature compensation stages, reducing the interference of environmental temperature field fluctuations and channel anomalies on the microcrack feature analysis process, and forming an auditable record of the compensation parameter evolution.

[0064] Step S400 includes at least steps S410-S430: S410. Obtain the time-series temperature compensation model structure, perform frame-by-frame temperature compensation calculation, temperature field smoothing resampling and abnormal pixel marker update processing, and obtain the compensated infrared temperature distribution map. In this embodiment, the infrared compensation and resampling subsystem reads the corresponding compensation coefficient files from the time-series temperature compensation model structure output from the previous main step S300, grouped by task number, component array region, and time period. Simultaneously, it obtains the raw infrared temperature sequence data and the infrared raw temperature preprocessing result set related to the same task number from the infrared acquisition and preprocessing module. The time-series temperature compensation model structure records the compensation parameter set corresponding to the background temperature level, slow temperature change characteristics, and local disturbance behavior for different time period groups. It also records the version number, effective time range, and reference point traceability information of the compensation parameters. When the system detects a new infrared raw temperature preprocessing result set being written, the task scheduling module automatically triggers the frame-by-frame temperature compensation calculation process according to the task number order, using the time-series temperature compensation model structure corresponding to the current task as the modeling basis for this step, without requiring manual activation. Specifically, when processing each frame of infrared image or each group of regional temperature sampling values, the frame-by-frame temperature compensation calculation subunit queries the temporal and spatial matching combination of compensation coefficients in the temporal temperature compensation model structure based on the acquisition time stamp and component array region identifier attached to that frame. It then performs calculations on the original infrared temperature value along with the corresponding compensation offset and temperature drift characteristics to generate a background-corrected temperature value. During this calculation, for pixel locations marked as suspected frozen or mid-failure channels in the channel health status record, the frame-by-frame temperature compensation calculation subunit uses neighborhood interpolation or reference point substitution strategies to fill in their temperature values. Simultaneously, it retains the original anomaly markers in the intermediate results to avoid overwriting hardware defect information. After the frame-by-frame temperature compensation calculation is completed, the system constructs a preliminary compensated infrared temperature distribution map at the image level and appends the version number of the compensation coefficients used and the corresponding time period group number to each frame record, forming an intermediate compensation result with parameter traceability attributes.

[0065] After completing frame-by-frame temperature compensation calculations, the temperature field smoothing and resampling subunit performs smoothing and resampling operations on the intermediate compensation results in both spatial and temporal dimensions. Spatially, based on the physical resolution of the infrared sensor and the geometric layout of the component array, the subunit constructs a regular sampling grid for each frame of compensated infrared temperature distribution, mapping the original irregular pixel distribution onto the regular grid. For regions with temperature differences within a certain range, a local smoothing strategy is used to reduce the influence of isolated noise points; for regions with significant gradient changes within the neighborhood, an edge protection strategy is used to preserve gradient information. Temporally, the subunit constructs a sliding window along the time axis within each component array region, performing smoothing and pitch reshaping on the compensated temperature sequence within the window. This ensures that the temperature sequences of the same component unit at different time points have a uniform sampling frequency and comparability. For time periods marked as noise fluctuation segments in the channel health status record, the subunit employs a stronger smoothing strategy and records the smoothing intensity level; for stable temperature regions, a lighter smoothing strategy is used to preserve subtle temperature differences to the maximum extent. Through the above process, the obtained temperature sequence not only achieves spatial uniformity and noise suppression at the single-frame level, but also achieves sampling rhythm uniformity across time, providing a foundation for subsequent hotspot region identification and temperature gradient calculation.

[0066] The anomaly pixel labeling update is performed by the anomaly update subunit. This subunit comprehensively utilizes reference point information from the time-series temperature compensation model structure and channel health status records from the infrared raw temperature preprocessing result set to re-examine the compensated temperature field. For pixels already marked as suspected anomalies before compensation, the anomaly update subunit observes whether their temperature values ​​still deviate from the average level of the same component region after compensation. If the deviation significantly converges after compensation, the anomaly labeling level is adjusted, downgraded from severe anomaly to mild anomaly, or the anomaly label is removed. If there is still a significant deviation after compensation, the anomaly labeling level is retained or increased. For pixels that were not previously labeled but exhibit unreasonable temperature abrupt changes after compensation, the anomaly update subunit marks them as newly appearing anomaly pixels based on their persistence on the time axis and their spatial clustering. After the abnormal pixel marker update is completed, the infrared compensation and resampling subsystem organizes each frame of infrared image that has undergone compensation, smooth resampling, and abnormal marker update processing into a compensated infrared temperature distribution map set according to the task number and flight segment group identifier. The compensated infrared temperature distribution map set specifically includes fields such as frame content, compensation parameter version number, time period group number, and abnormal pixel mask. This map set serves as the sole input for S420 to extract hotspot region contours, temperature gradient fields, and component edge contours from the compensated infrared temperature distribution map set. At the same time, at the cross-main step level, the compensated infrared temperature distribution map set also provides the original temperature field basic data for the infrared local temperature gradient calculation in the subsequent S500.

[0067] S420. Extract the hot spot region contour, temperature gradient field and component edge contour from the compensated infrared temperature distribution map set, perform key feature point annotation and geometric constraint generation, and generate PL infrared registration control feature set. In this embodiment, the feature extraction and constraint generation subsystem reads each frame of compensated infrared image from the compensated infrared temperature distribution map set output by S410 in order of component number and flight segment group identifier. Simultaneously, it reads the component-level sub-map in the PL preprocessed image data structure and the field-of-view configuration, component layout, and UAV attitude record in the dual-modal acquisition task configuration structure. After the task scheduling module issues the "Infrared Feature Extraction Task Start" command, the feature extraction and constraint generation subsystem automatically processes all compensated infrared temperature distribution map frames in the task list to be analyzed, without relying on manual frame-by-frame operation. The hotspot region contour extraction unit first constructs candidate hotspot regions on each frame of compensated infrared image based on the temperature value range and abnormal pixel masks. It uses a step-by-step threshold segmentation method to determine the boundaries of high-temperature regions, identifying connected high-temperature connected regions with an area exceeding a certain threshold as hotspot regions, and obtaining the fine contour curve of the region through region growing. Simultaneously, the hotspot region contour extraction unit sets an edge protection strip near the component boundaries to avoid misidentifying slight temperature changes at the edges due to changes in viewing angle as hotspots. After extraction, the hotspot region contour is converted into a set of contour nodes, each node recording its position in the infrared image coordinate space and its corresponding temperature range.

[0068] After extracting the contours of the hotspot region, the temperature gradient field construction unit calculates the temperature gradient changes in the horizontal and vertical directions across the entire image based on the temperature value of each pixel in the compensated infrared temperature distribution map, forming a temperature gradient field description. For the internal region of the component, the temperature gradient field construction unit focuses on the gradient distribution along the long and short sides of the component, extracting the set of gradient abrupt change locations in this direction as possible defect boundary clues. When constructing the temperature gradient field, the system references anomaly pixel masks to exclude pixels already marked as severely anomalous from the gradient calculation, preventing individual outliers from having an excessive impact on the overall gradient distribution. After construction, the temperature gradient field is stored as a set of gradient feature maps under the corresponding component number. Each map records the gradient direction, gradient intensity, and spatial location, providing a basis for subsequent geometric constraint generation. The component edge contour extraction unit combines the information about field coverage and component layout in the dual-modal acquisition task configuration structure to perform edge detection on the compensated infrared temperature distribution map. After clarifying the approximate position of the component row and column distribution and geometric boundaries, the outer contour curve and internal dividing line position of the component are obtained through edge tracking. The edge contour is then structured and encoded using the component number and cell partitioning method.

[0069] After obtaining the hotspot region contour, temperature gradient field, and component edge contour, the key feature point annotation and geometric constraint generation unit begins to construct the PL infrared registration control feature set. This unit first selects several representative feature points at the intersection of the component edge contour and the temperature gradient field, such as corner points of the component's outer contour and intersections of internal structural lines with significant temperature gradient locations. These feature points are then labeled according to the component number and its row and column index in the image. For the hotspot region contour, the key feature point annotation and geometric constraint generation unit samples contour nodes at certain intervals along the hotspot region boundary. These nodes are used as feature points describing the local temperature rise pattern, and a summary of the brightness level at the same spatial location in the corresponding PL preprocessed image sub-image is appended to the node record. This summary serves as the basis for multi-source feature comparison during subsequent dual-modal matching. The geometric constraint generation process constructs a component-level geometric relationship description based on the above feature points, including the parallel and perpendicular relationships between the four sides of the component's outer contour, the relative positional relationships between adjacent components, and the spacing relationships of internal structural lines within the same component. These geometric relationships are recorded in the registration control feature set in the form of a constraint list. After feature point annotation and geometric constraint generation are completed, the system encapsulates the hotspot region contour feature point set, temperature gradient field feature point set, component edge contour feature point set, corresponding PL preprocessing subgraph spatial index, and geometric constraint list into a PL infrared registration control feature set. The PL infrared registration control feature set serves as the direct input for S430 to perform feature matching search, UAV attitude constraint superposition, and pixel mapping relationship calculation. At the same time, at the cross-main step level, the hotspot and gradient information in this feature set can also participate in the construction of spatial overlap and temperature difference indices in S500.

[0070] S430: Perform feature matching search on the PL infrared registration control feature set, obtain the UAV attitude constraint superposition and pixel mapping relationship, and generate a pixel-level aligned dual-modal image data structure. In this embodiment, the dual-modal registration and mapping acquisition subsystem reads the feature point set and geometric constraint description of each component region from the PL infrared registration control feature set output by S420 in order of task number, and obtains the standard-size PL sub-map of the corresponding component from the PL preprocessed image data structure. Simultaneously, it obtains the UAV attitude record and imaging geometric parameters from the dual-modal acquisition task configuration structure. The feature matching search unit first performs edge detection and brightness structure analysis on the PL preprocessed sub-map to obtain the component edge contour and internal dark line distribution on the PL side. It then performs structured encoding of these contour and dark line feature points and performs candidate matching with the component edge contour feature points on the infrared outer side of the PL infrared registration control feature set. Specifically, the feature matching search unit performs initial matching on the four corner points and boundary line segments of the component's outer contour according to positional relationships, using point pairs with similar shapes and close angular relationships as matching seeds. Then, it expands the matching around the corresponding positions of the internal structure lines and hotspot areas, adding point pairs located within the same geometric constraint framework and highly corresponding to the dark line features on the PL side and the temperature gradient features on the infrared outer side to the matching set. During this process, the feature matching search unit continuously refers to the list of geometric constraints, eliminates abnormal matching point pairs that violate the aspect ratio of the component or the spacing relationship of the internal structure, and gradually converges a set of high-confidence feature matching pairs through iterative optimization.

[0071] After obtaining the initial feature matching pairs, the UAV attitude constraint overlay unit further corrects the matching relationship based on the UAV attitude information recorded in the dual-modal acquisition task configuration structure. The UAV attitude information includes the heading angle, pitch angle, and roll angle records at each frame imaging moment, serving as the basic imaging attitude data for the PL imaging module and the infrared thermal imager during flight. The UAV attitude constraint overlay unit first converts the attitude information into a description of the viewpoint changes and projection deformation of the corresponding imaging plane. Based on this, it performs unified viewpoint normalization on the geometric coordinates in the PL preprocessed sub-image and the compensated infrared temperature distribution map, aligning the two modes under a virtual standard viewpoint. Subsequently, the UAV attitude constraint overlay unit performs attitude constraint correction on the feature point pairs output by the feature matching search unit. For slight positional shifts caused by attitude differences, small-range coordinate corrections are used to bring the corresponding point pairs back to a reasonable position range. Point pairs incompatible with both attitude and geometric constraints are discarded. After the attitude constraint overlay operation, the final set of retained feature matching pairs spatially better conforms to the component geometry and actual shooting posture, providing a reliable basis for subsequent pixel mapping relationship determination.

[0072] The pixel mapping relationship acquisition unit constructs a pixel mapping relationship description from the PL image coordinate space to the infrared image coordinate space based on the feature matching pair set after pose correction. This unit obtains a family of spatial mapping functions that maintains geometric stability within the component region based on the positional relationship of component-level feature points in both modes. It selects the most suitable mapping function according to different component row and column positions and different field-of-view coverage, generating a separate mapping parameter set for each component region. The pixel mapping relationship acquisition unit then performs coordinate transformation on all PL sub-image pixels within the component, mapping each pixel position in the PL sub-image to its corresponding position or nearest neighbor position in the infrared image. In scenarios where mapping fails or the position falls into an abnormal pixel mask area, a mapping anomaly marker is recorded. For successfully mapped pixels, the system internally constructs a dual-modal pixel pair data record, which includes information such as PL side brightness value, infrared outer compensation temperature value, component number, sub-image index, and task number. After mapping and pixel pair generation for all component regions and all task frames, the dual-modal registration and mapping subsystem encapsulates the mapped PL sub-image, infrared temperature map, pixel pair records, and corresponding geometric and attitude parameters into a pixel-level aligned dual-modal image data structure. In this invention, this pixel-level aligned dual-modal image data structure serves as the sole input source for the S500 during PL dark line region candidate extraction, infrared local temperature gradient calculation and component unit number binding, geometric length, width, spatial overlap, and temperature difference index extraction, normalization rule mapping, and convolutional neural network inference processing. It also serves as the foundational data for constructing multi-source overlay display layers for various visualization modules.

[0073] In summary, the technical effects of this step are as follows: By introducing frame-by-frame temperature compensation operations and temperature field smoothing resampling processing based on a time-series temperature compensation model structure in S410 to S430, and combining hotspot region contours, temperature gradient fields, component edge contour extraction, and UAV attitude constraint superposition, high-precision spatial registration and pixel-level mapping of PL images and infrared images within the component-level range are achieved. This generates a structured pixel-level aligned dual-modal image data structure, thereby providing a stable and consistent input data foundation for subsequent hidden crack multimodal feature construction and convolutional neural network inference, and reducing the interference of environmental temperature field fluctuations and attitude changes on the hidden crack identification process.

[0074] Step S500 includes at least steps S510-S530: S510. Obtain pixel-level aligned dual-modal image data structure, perform candidate extraction of PL dark line region, infrared local temperature gradient calculation and component unit number binding processing to obtain hidden crack candidate feature vector set; In this embodiment, the hidden crack candidate feature construction subsystem reads each frame of aligned photoluminescence image and infrared temperature image from the pixel-level aligned dual-modal image data structure output from the previous main step S400, according to task number, component array row and column number, and time sequence. The pixel-level aligned dual-modal image data structure records the brightness value of each pixel in the photoluminescence image, the compensated temperature value in the infrared temperature image, and the component number, component unit number, flight line segment number, and shooting time tag to which the pixel belongs. It serves as the basic carrier for realizing multimodal feature fusion. When the system detects a new pixel-level aligned dual-modal image data structure being written to the task queue, the task scheduling module triggers the hidden crack candidate feature construction subsystem to automatically start the processing flow of this step, without manual intervention. Specifically, the PL dark line region candidate extraction unit first locates the corresponding photoluminescence image channel in the pixel-level aligned dual-modal image data structure and divides the image into blocks according to the component unit number, constructing a sub-image set with a single component unit as the smallest processing unit. In each PL sub-image, the PL dark line region candidate extraction unit detects slender regions whose brightness is significantly lower than the background brightness of the same component unit through brightness histogram statistics, local contrast analysis, and connectivity analysis based on neighborhood relationships. Connected regions that meet the constraints of length, width, and slenderness are identified as PL dark line candidate regions, and the pixel set, geometric bounding rectangle, and relative position of the candidate region in the component unit coordinate system are recorded.

[0075] Furthermore, the infrared local temperature gradient calculation unit, based on the aligned infrared temperature image channels in the pixel-level aligned dual-modal image data structure, performs local analysis of the temperature field within a preset neighborhood range around each PL dark line candidate region. Under the same component unit number as the PL dark line candidate region, this unit calculates the temperature variation trend with spatial location along the long and short sides of the component, respectively, to obtain local temperature gradient information describing the strength of temperature changes. Simultaneously, it constructs a comparison window inside and outside the candidate region, comparing the differences between the average temperature, maximum temperature, and background temperature within the window to form a local temperature feature description reflecting the thermal anomaly behavior of the candidate region. During the local temperature gradient calculation process, for pixels marked as severely anomalous in previous steps, the infrared local temperature gradient calculation unit removes the influence of these pixels on the gradient calculation through neighborhood interpolation or masking strategies, and records the relevant processing in the intermediate state field for subsequent auditing and traceability. The component unit number binding is completed by the component mapping and encoding unit. This unit reads the configuration data about the component array topology and component unit division in the dual-modal acquisition task configuration structure, overlays and matches the geometric outer rectangle of the PL dark line candidate area with the component unit boundary, records the component number and component unit number of each candidate area, and records the normalized position index and coverage ratio information of the candidate area in the component unit.

[0076] After completing the candidate extraction of dark line regions in the PL (Plastic Probe) system, the calculation of local infrared temperature gradients, and the binding of component unit numbers, the hidden crack candidate feature construction subsystem constructs a feature vector entry for each candidate region. It encapsulates the brightness statistics of the PL side (e.g., average brightness, minimum brightness, and brightness variation trend along the dark line direction), the local temperature gradient features of the infrared side (e.g., temperature difference between the candidate region and the background on both sides, and the strength of temperature change along the component structure line), geometric shape features (e.g., length, width, and aspect ratio), and metadata such as component unit number, time tag, and task number into a unified structured record. Each structured record corresponds to a hidden crack candidate entity. The hidden crack candidate feature construction subsystem organizes all structured records according to task number and component unit number into an indexable data set, forming a hidden crack candidate feature vector set, and marks the interface field names with subsequent processing steps in this set. In one feasible embodiment, for a photovoltaic power station with a large installed capacity, the pixel-level aligned dual-modal image data structure generated by a single UAV inspection mission can contain several flight lines and tens of thousands of component units. The hidden crack candidate feature construction subsystem can automatically and in parallel traverse all component units on the server side, complete the PL dark line region candidate extraction and infrared local temperature gradient calculation for each frame of data, and write the hidden crack candidate feature vector set into the feature database in batches after the mission is completed for subsequent steps. The hidden crack candidate feature vector set is the output field name of this subsection and is also the only input source when extracting geometric length, width, spatial overlap and temperature difference indicators from the hidden crack candidate feature vector set in the next subsection S520. At the same time, at the cross-main step level, after the final hidden crack identification result structure is generated, this set will also be used backtracking for sample interpretation and incremental learning sample selection.

[0077] S520. Extract geometric length, width, spatial overlap and temperature difference indices from the candidate feature vector set of hidden cracks, perform normalization rule mapping and confidence segmentation, and generate a set of hidden crack confidence classification results. In this embodiment, the hidden crack feature normalization and confidence construction subsystem reads all candidate region entries from the hidden crack candidate feature vector set output in the previous section. Each entry contains at least the PL dark line morphology features, infrared local temperature gradient features, component unit number, and candidate region coordinates. The hidden crack feature normalization and confidence construction subsystem automatically starts after the task scheduling module issues the "candidate feature normalization and classification" instruction, and performs unified processing on all candidate regions within the same task number range. Specifically, the geometric length extraction unit first calculates the extension distance of the dark line candidate region along the direction of the main structure inside the component based on the projection of the geometric bounding rectangle of the candidate region in the component unit coordinate system, and records it as a geometric length index; the geometric width extraction unit then calculates the coverage area of ​​the dark line candidate region in the direction perpendicular to the main structure direction, converts this range into a relative proportion of the component unit size, and records it as a geometric width index. For candidate regions with bifurcations or multiple connected segments, the geometric length and geometric width extraction unit traverses the region skeleton, takes the longest continuous skeleton segment as the length reference, takes the maximum extension range on both sides of the skeleton as the width reference, and records additional information such as the number of segments in the feature vector, providing a basis for distinguishing different types of hidden crack entities.

[0078] The spatial overlap index is calculated by the spatial coverage analysis unit. This unit reads the component unit boundary information and candidate region coverage ratio information generated by the component mapping and encoding unit in the previous section, calculates the coverage area ratio of the candidate region within its component unit and across adjacent component units, and records the cross-unit coverage pattern when multiple component units share coverage. The spatial overlap index describes the spatial relationship of the candidate dark line relative to the standard component unit division and is one of the important features for distinguishing between small defects within a component and large-scale anomalies across components. The temperature difference index is constructed by the temperature comparison analysis unit. Based on the local temperature statistics recorded in the candidate feature vector set of the microcrack, this unit calculates the temperature difference between the average temperature within the same candidate region and the average temperature of the stable region of its component unit. At the same time, multiple background windows are selected in the neighborhood of the candidate region to compare the temperature difference between the candidate region and the background window, and these differences are recorded as the temperature difference index in a centralized expression. Understandably, geometric length, geometric width, spatial overlap, and temperature difference indicators constitute the minimum core feature set that is indispensable for the hidden crack identification in this invention, and directly participate in the subsequent normalization rule mapping and convolutional neural network inference process; on this basis, the system can also select extended features such as texture directionality and local contrast change to further refine the identification performance in specific scenarios.

[0079] After extracting the core indicators, the hidden crack feature normalization and confidence construction subsystem performs normalization rule mapping on the geometric length, geometric width, spatial overlap, and temperature difference indicators of all candidate regions, mapping indicators with different dimensions and value ranges to a unified feature space. The normalization rule mapping unit selects an appropriate mapping strategy for each type of indicator based on the system's pre-configured rule set. For example, by referring to the indicator distribution range obtained from historical inspection tasks, extreme outliers are cropped to a reasonable range, and the remaining indicators are mapped to a unified relative scale according to component type or operating condition type. The normalization rule mapping process records the original indicator value, the normalized indicator value, and the rule number used for each candidate region in the intermediate result field, providing a basis for subsequent auditability. In the normalized feature space, the confidence level segmentation unit maps the position of candidate regions in the multi-dimensional space of core indicators to several confidence level intervals according to a pre-defined segmentation strategy. For example, candidate regions that simultaneously meet the following criteria are assigned to higher confidence levels: larger geometric length, spatial overlap concentrated within a single component unit, and significant temperature differences. Candidate regions that are shorter or have insignificant temperature differences are assigned to lower confidence levels. While generating confidence levels, the confidence level segmentation unit also records the triggering conditions and reference rule numbers for each level, facilitating the interpretation of the model output in subsequent steps.

[0080] After completing the normalization rule mapping and confidence segmentation, the hidden crack feature normalization and confidence construction subsystem encapsulates the core indicator normalization result, confidence level identifier, component unit number, and time label of each candidate region into a new structured record, organizing them into a hidden crack confidence level result set. The hidden crack confidence level result set serves as the output field name for this subsection, where each record can establish a one-to-one correspondence with the hidden crack candidate feature vector set from the previous subsection through the component unit number and candidate region index. The hidden crack confidence level result set is then passed to the next subsection in the workflow, where it is used as input in S530 by the Convolutional Neural Network (CNN) inference and component-level aggregation subsystem for further deep feature fusion inference and component-level hidden crack aggregation.

[0081] S530. Perform convolutional neural network model inference, threshold clustering and component-level numbering on the confidence grading result set of hidden cracks to generate the hidden crack identification result structure. In this embodiment, the convolutional neural network model inference and component-level aggregation subsystem obtains the core indicator normalization results and confidence level information of all candidate regions from the hidden crack confidence grading results set output in the previous section. Simultaneously, it extracts local multimodal image fragments corresponding to these candidate regions from the pixel-level aligned bimodal image data structure as needed. The image data characterizing the PL dark line morphology and infrared temperature distribution around the candidate regions are jointly processed with tabular indicator features. After the task scheduling module issues the "deep inference and aggregation" command, the convolutional neural network model inference and component-level aggregation subsystem automatically traverses all tasks to be processed, constructing each candidate region as a multi-source input sample, including a photoluminescence local sub-image, an infrared temperature local sub-image, and the aforementioned normalized indicators such as geometric length, geometric width, spatial overlap, and temperature difference. The convolutional neural network model inference unit uses a network structure pre-trained on historical inspection data and laboratory calibration samples. It inputs the photoluminescence local sub-image and infrared temperature local sub-image as image channels into the network front end, enabling the network to automatically learn the spatial correspondence between PL dark line morphology and infrared temperature anomalies in the convolution and feature extraction layers. At the same time, a connection layer for fusing tabular indicator features is introduced in the middle and later stages of the network. The normalized geometric length, geometric width, spatial overlap and temperature difference indicators are concatenated with the high-dimensional feature vector output by the convolutional layer, so that the model can give a prediction result of whether the candidate region is a hidden crack and a more detailed confidence score under the condition of comprehensively considering image pattern and structured indicators.

[0082] After the convolutional neural network model inference is completed, the threshold clustering unit performs multi-threshold clustering and spatial merging on the candidate regions based on the confidence score output by the model and the confidence levels already established in the previous section. The threshold clustering unit first divides the candidate regions into several initial categories based on the multiple thresholds configured by the system, such as high-confidence suspected hidden cracks, medium-confidence suspicious regions, and low-confidence background noise regions. Then, it performs spatial proximity clustering on high-confidence and some medium-confidence candidate regions within the component unit coordinate system, merging spatially adjacent or partially overlapping candidate regions into a unified hidden crack entity. During the threshold clustering process, the system combines component edge contours and internal structural line information to perform cross-unit clustering on candidate regions that span multiple component units and geometrically conform to possible extension paths, forming component-level or cross-component-level hidden crack entity descriptions. For candidate regions determined to be isolated and unable to form a stable geometric structure, they are classified into the noise category using thresholds and clustering rules, and the reasons for removal and the decision basis are recorded. Through the above process, the threshold clustering unit transforms the judgment results given by the convolutional neural network model at the candidate region granularity into a set of hidden crack entities that are closer to the application granularity.

[0083] After clustering hidden crack entities, the component-level numbering and aggregation unit categorizes and organizes these entities according to component number, component unit number, and flight segment number. It numbers and sorts one or more hidden crack entities within the same component unit, forming a component-level hidden crack list, and assigns a globally unique hidden crack number to each entity. During the component-level aggregation process, the unit also summarizes information such as the number of hidden crack entities, average geometric length, maximum geometric length, and typical temperature differences within each component unit, and writes this information, along with the confidence score output by the convolutional neural network model, into the component-level record field. For hidden crack entities spanning multiple component units during clustering, the unit establishes a connection between them and multiple component units by recording the cross-unit coverage pattern and the main component unit number, ensuring that subsequent visualization and operational decision-making can be queried and analyzed at the component level. Finally, the convolutional neural network model inference and component-level aggregation subsystem encapsulate the hidden crack entity list under all component numbers, component-level aggregation statistics, and the mapping relationship from candidate regions to hidden crack entities into a unified hidden crack identification result structure.

[0084] In summary, the technical effects of this step are as follows: By constructing a candidate feature vector set for hidden cracks around the pixel-level aligned dual-modal image data structure in S510 to S530, core indicators such as geometric length, geometric width, spatial overlap, and temperature difference are extracted and normalized rule mapping is completed. Then, a convolutional neural network model is used to jointly infer the multimodal images and structured indicators, supplemented by threshold clustering and component-level numbering and summarization processing. This step achieves deep fusion of PL dark line morphology and infrared temperature anomaly at the hidden crack identification level, and orderly convergence from candidate pixels to component-level entities. This reduces the probability of misjudgment caused by single-modal noise or local artifacts, and at the same time prepares a structured hidden crack identification result structure for subsequent visualization and incremental learning.

Claims

1. A method for inspecting microcracks in photovoltaic modules using PL infrared fusion timing compensation, characterized in that, include: Acquire photovoltaic power station flight path planning data, UAV platform parameters, environmental monitoring parameters and incremental learning update model structure, register task numbers, construct module mapping relationships, verify basic safety boundaries and organize and bind synchronous triggering and attitude compensation control parameters to generate a dual-modal acquisition task configuration structure; Based on the dual-modal acquisition task configuration structure, index binding, image integrity verification, format standardization and flat field correction, local contrast enhancement and standard size cropping are performed to generate PL preprocessed image data structure. Based on the PL preprocessed image data structure, timestamp alignment, channel health status detection, preliminary marking of abnormal segments, extraction of stable temperature field regions, background reference point screening, and regional temperature statistical processing are performed to generate a time-series temperature compensation model structure. Based on the temporal temperature compensation model structure, frame-by-frame temperature compensation calculation, temperature field smoothing resampling, abnormal pixel marking update, hot spot region contour, temperature gradient field and component edge contour feature extraction and pixel mapping relationship calculation are performed to generate a pixel-level aligned dual-modal image data structure. Based on a pixel-level aligned dual-modal image data structure, candidate extraction of PL dark line regions, calculation of infrared local temperature gradient and binding of component unit numbers are performed. Geometric length, width, spatial overlap and temperature difference indices are extracted, normalized rule mapping and convolutional neural network inference processing are performed to generate the hidden crack identification result structure.

2. The method according to claim 1, characterized in that, The data on photovoltaic power plant flight path planning, drone platform parameters, environmental monitoring parameters, and incremental learning update model structure include: The photovoltaic power station route planning data includes the photovoltaic field boundary, module array layout, waypoint sequence of each inspection route, inspection altitude, heading angle, overlap rate configuration, and preset inspection time window. The parameters of the UAV platform include UAV model identification, payload capacity, remaining battery capacity estimation method and descent rate, allowable attitude angle range, gimbal control accuracy, installation position and field-of-view geometry of photoluminescence imaging module and infrared thermal imaging module. The environmental monitoring parameters include the current solar irradiance of the site, ambient temperature, array surface temperature and background level, wind speed and direction, air humidity, and weather event labels. The incremental learning update model structure records the version identifier of the currently active hidden crack identification model, the distribution characteristics of the sample data used during model training, the model's requirements for PL image spatial resolution and dynamic range, the model's requirements for infrared temperature resolution and temperature change sensitivity, as well as the model's recommended requirements for acquisition coverage redundancy, flight path spacing, and viewpoint diversity.

3. The method according to claim 1, characterized in that, The process of basic security boundary verification also includes: The basic safety boundary verification process includes reading flight path planning data and UAV platform parameters, comparing the planned flight altitude with the minimum recommended flight altitude in the UAV platform parameters (if it is lower than the minimum recommended flight altitude, it is considered too low); comparing the planned turning radius with the minimum allowed turning radius in the UAV platform parameters (if it is smaller than the minimum allowed turning radius, it is considered too small); comparing the planned flight speed with the maximum cruising speed in the UAV platform parameters (if it exceeds the maximum cruising speed, it is considered too high); and marking risk indicators for flight path sub-tasks with such conditions. Combined with wind speed and direction labels, wind resistance capability verification records are added to the relevant flight path sub-tasks. A set of no-fly polygon areas is generated for geometric detection, and waypoints with intrusion risk are marked as waypoints requiring adjustment.

4. The method according to claim 1, characterized in that, The process of synchronizing and binding control parameters for synchronous triggering and attitude compensation also includes: The processing and binding of synchronous triggering and attitude compensation control parameters includes reading the time resolution requirements for synchronous acquisition of PL and IR, reading the expected speed and length segment information of each route, reading the maximum triggering frequency and load processing capacity, combining and constructing synchronous triggering parameter placeholder fields, generating attitude compensation parameter placeholder fields, and binding the synchronous triggering parameter placeholder fields and attitude compensation parameter placeholder fields to each route subtask record respectively.

5. The method according to claim 1, characterized in that, The process of flat field correction, local contrast enhancement, and standard size cropping also includes: Flat field correction processing includes loading a flat field reference file and performing pixel-level response correction through a lookup table. Local contrast enhancement processing involves dividing the data into several local windows, grouping each local window, and applying different contrast stretching strategies. The standard size cropping process involves cropping the image into several standard size sub-images and assigning a component number, row and column index, and line segment group identifier to each sub-image.

6. The method according to claim 1, characterized in that, The process of extracting stable temperature regions, selecting background reference points, and statistically processing regional temperatures also includes: The temperature field stable region extraction process includes removing candidate segments with hidden anomalies, smoothing the remaining segments and counting the fluctuation amplitude, and registering segments with fluctuation amplitudes below a preset stability threshold as temperature field stable regions. Background reference point screening process includes reading the brightness distribution and dark line feature summary of the component-level sub-image, judging the intact component candidate areas with uniform brightness and no obvious dark lines or local abnormal shadows through simple geometric rules and brightness thresholds, and cross-matching with temperature field stable areas to select background reference point candidates. Regional temperature statistical processing includes calculating the equilibrium level of temperature readings at different time scales, their upward or downward trends, and their correlation with environmental monitoring parameters.

7. The method according to claim 1, characterized in that, The process of extracting hotspot region contours, temperature gradient fields, and component edge contour features, and determining pixel mapping relationships also includes: Hotspot region contour extraction processing includes constructing candidate hotspot regions, determining the boundaries of high-temperature regions, and obtaining fine contour curves; The temperature gradient field construction process involves calculating the temperature gradient changes in the horizontal and vertical directions to form a temperature gradient field description. The component edge contour extraction process includes edge detection and edge tracking to obtain the component's outer contour curve and the position of the internal dividing lines. The pixel mapping relationship retrieval process involves obtaining a family of spatial mapping functions and mapping each pixel position in the PL subgraph to the corresponding position in the infrared image.

8. The method according to claim 1, characterized in that, The process of extracting candidate dark line regions in PL, calculating local infrared temperature gradients, and binding component unit numbers also includes: The PL dark line region candidate extraction process involves dividing the image into blocks, constructing a set of sub-graphs with a single component unit as the smallest processing unit, and detecting elongated regions whose brightness is significantly lower than the background brightness of the same component unit through brightness histogram statistics, local contrast analysis, and connectivity analysis based on neighborhood relationships. Connected regions that satisfy length, width, and elongation constraints are identified as PL dark line candidate regions. The infrared local temperature gradient calculation process includes local analysis of the temperature field within a preset neighborhood range around each candidate region of PL dark line, calculating the temperature variation trend with spatial location along the long and short sides of the component to obtain local temperature gradient information, and constructing a comparison window inside and outside the candidate region to calculate the difference between the average temperature, maximum temperature and background temperature. The component unit number binding process includes reading the configuration data about the component array topology and component unit division in the dual-modal acquisition task configuration structure, overlaying and matching the geometric outer rectangle of the PL dark line candidate area with the component unit boundary, and recording the component number and component unit number.

9. The method according to claim 1, characterized in that, The process of extracting geometric length, width, spatial overlap, and temperature difference indices also includes: The extraction and processing of geometric length and width indicators includes calculating the extension distance of the dark line candidate region along the direction of the main structure inside the component and the coverage area in the vertical direction; The spatial overlap index extraction process includes statistical analysis of the coverage area ratio of candidate regions within their respective component units and across adjacent component units; The temperature difference index extraction process includes calculating the temperature difference between the average temperature within the same candidate region and the average temperature of the stable region of the component unit, and comparing the temperature difference between the candidate region and the background window.

10. The method according to claim 1, characterized in that, The process of normalization rule mapping and convolutional neural network inference also includes: The normalization rule mapping process involves mapping indicators with different dimensions and different value ranges to a unified feature space, and performing cropping and mapping by referring to the indicator distribution range obtained from historical inspection tasks. The convolutional neural network inference processing involves inputting a local sub-image of photoluminescence and a local sub-image of infrared temperature as the image channel input to the network front end. It automatically learns the spatial correspondence between the morphology of PL dark lines and infrared temperature anomalies. It then concatenates the normalized geometric length, width, spatial overlap, and temperature difference indices with the high-dimensional feature vector output by the convolutional layer to provide a prediction result and confidence score for whether the candidate region is a hidden crack.

Images