AI identification and vectorization method for photovoltaic power station based on sub-meter image

Abstract

The application relates to a photovoltaic power station AI identification and vectorization method based on sub-millimeter images, which comprises the following steps: acquiring and preprocessing multi-modal image data of a target area; fusing the preprocessed data in multiple modes, identifying photovoltaic panel abnormalities and faults based on a fusion feature map, and grading to generate a pixel-level segmentation mask, feedback error, and obtain an abnormality identification result; combining the abnormality identification result, the fusion feature map and the preprocessed multi-modal data to complete implicit fault confirmation, grading, tracing and trend judgment; inputting the segmentation mask, fault contour and fusion feature map to a coordinate calibration module, calibrating the coordinates combined with the spatial features of the fusion data to obtain calibrated coordinates; and based on the calibrated coordinates, the abnormality identification result and the fault contour, extracting the contour, associating the attributes and converting them into a standard vector format. The application improves the identification accuracy under complex terrain and severe weather, simultaneously realizes accurate grading and judgment of abnormal areas such as photovoltaic panel damage and shielding, and provides direct data support for power station operation and maintenance.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of satellite remote sensing technology, artificial intelligence recognition technology and geographic information data processing technology, specifically to an AI recognition and vectorization method for photovoltaic power plants based on sub-meter level images. Background Technology

[0002] With the rapid development of the global clean energy industry, the construction scale of photovoltaic power plants continues to expand, making the demand for accurate surveys, dynamic monitoring, and standardized management of photovoltaic power plants increasingly urgent. Currently, photovoltaic power plant identification mainly relies on two technical approaches: one is manual interpretation, where professionals visually analyze satellite imagery or aerial photographs to mark the location and extent of photovoltaic power plants; the other is traditional image recognition technology, based on methods such as threshold segmentation and texture feature matching. In recent years, sub-meter-level satellite imagery technology has developed rapidly, with its 0.5m to 1m resolution clearly presenting the texture features, arrangement patterns, and boundary details of photovoltaic panels, providing a data foundation for high-precision identification; while breakthroughs in image semantic segmentation and target detection technologies in artificial intelligence possess powerful feature extraction and complex scene adaptation capabilities.

[0003] However, existing technologies have significant drawbacks: 1. Given the urgent need for large-scale construction and refined management of photovoltaic power plants, manual interpretation technology faces significant bottlenecks. This method struggles to process single-frame 10... 1. Image processing takes several hours, resulting in extremely low efficiency. Furthermore, recognition results are susceptible to subjective interference, with boundary accuracy errors exceeding 1 meter, making it difficult to support large-scale batch operations. 2. Traditional image recognition technologies lack adaptability. In uneven terrain such as mountains and forests, or under complex weather conditions like cloudy or shaded areas, the recognition accuracy is less than 60%. Simultaneously, the technology cannot directly correlate with precise geographic coordinates, making it difficult to integrate output data into GIS systems, significantly reducing its practicality. 3. Existing technologies do not integrate the detail advantages of sub-meter-level satellite imagery with the intelligent recognition capabilities of AI models. They lack model optimization schemes tailored to the unique textures of photovoltaic power plants and have not formed a fully automated system for recognition, coordinate matching, and vector output, thus failing to fully realize the application potential of sub-meter-level imagery.

[0004] Therefore, this invention proposes an AI recognition and vectorization method for photovoltaic power plants based on sub-meter level images. Summary of the Invention

[0005] To overcome the shortcomings of the existing technology, this application provides an AI recognition and vectorization method for photovoltaic power plants based on sub-meter level images, specifically adopting the following technical solution.

[0006] A photovoltaic power plant AI recognition and vectorization system based on sub-meter level images includes:

[0007] Data acquisition module: Acquires raw sub-meter level satellite imagery data, including optical satellite imagery, multispectral imagery, and thermal infrared imagery.

[0008] Image preprocessing module: Performs fully automated preprocessing on raw sub-meter level satellite images to eliminate data noise and distortion. It adopts scene-adaptive preprocessing algorithms, cross-module feedback mechanisms, and latent fault feature enhancement design.

[0009] AI Recognition Model Module: Based on the improved U-Net++ network, a scene-adaptive recognition model is built, which integrates Transformer and attention mechanisms to achieve accurate recognition, graded judgment of explicit anomalies and detection of hidden faults.

[0010] Multimodal data fusion module: Integrates optical images, multispectral images, thermal infrared images, BeiDou positioning data, topographic and meteorological data, and operation and maintenance historical data to achieve multi-dimensional feature fusion, providing comprehensive data support for identification, coordinate matching, and health assessment.

[0011] Coordinate matching module: Combining BeiDou and terrain data from the multimodal data fusion module, a multi-source coordinate fusion calibration system is constructed, and a scenario-differentiated calibration strategy is adopted to achieve high-precision coordinate matching.

[0012] Vector data generation module: Associates the results of hidden fault identification with operation and maintenance data to achieve integrated refinement, standardization and operation and maintenance adaptation of vector data.

[0013] Latent Fault Assessment Module: Based on the characteristics of latent faults obtained from multimodal data fusion and AI recognition models, a system for accurate assessment and tracing of latent faults is constructed, realizing integrated processing of latent faults from detection and classification to tracing.

[0014] Accuracy verification module: Construct a full-process dual accuracy verification and error tracing system, and adopt multi-modal fusion accuracy verification and latent fault judgment accuracy verification.

[0015] Cross-module collaborative control module: enables dynamic collaboration among modules, error propagation control, scenario adaptive adaptation, closed-loop operation and maintenance iteration, and time-series early warning.

[0016] Operations and Maintenance Data Iteration Module: As the support for the closed-loop iteration of operations and maintenance, it is responsible for the collection, organization, analysis and feedback of operations and maintenance data.

[0017] This invention provides an AI recognition and vectorization method for photovoltaic power plants based on sub-meter level images, including the following steps.

[0018] Step 1: Multimodal Data Acquisition: Obtain sub-meter level satellite imagery of the target area through satellite data service providers, including optical, multispectral, and thermal infrared data. Simultaneously collect BeiDou positioning data, topographic and meteorological data, and historical operation and maintenance data of photovoltaic power plants. Verify the completeness and correlation of each modal data; automatically supplement data collection or trigger data repair if data quality is substandard.

[0019] Step 2, Image Preprocessing: Input the selected qualified raw images and multimodal data into the preprocessing module. Through fully automated processing such as radiometric correction, geometric correction, noise removal, and extreme scene adaptation, data noise and distortion are eliminated, and preprocessing parameters are optimized through cross-module feedback.

[0020] Step 3, Multimodal Data Fusion: Input the preprocessed data into the fusion module to complete pixel-level spatiotemporal alignment and multi-dimensional feature deep fusion. Dynamically allocate the fusion weights of each modality according to the scene type, verify the fusion quality, and ensure that high-quality data support is provided for subsequent steps.

[0021] Step 4: AI Identification and Anomaly Classification Assessment: Input the multimodal fusion feature map into the AI ​​identification model, adaptively call the corresponding inference branch, complete the batch identification of explicit anomalies and latent faults of photovoltaic panels, classify and assess explicit anomalies and latent faults, mark relevant information and feed back error optimization parameters.

[0022] Step 5: Accurate assessment and tracing of latent faults: Combining AI identification results, multimodal fusion data and historical operation and maintenance data, latent faults are identified and their classifications are refined. The core causes of the faults are located and the sources are traced. The development trend of the faults is analyzed, the upgrade risks are predicted, and operation and maintenance warnings are triggered.

[0023] Step 6, Multi-source coordinate fusion calibration: Input pixel-level segmentation mask, fault area outline and fusion data, and through coordinate transformation, triple calibration, multi-source fusion and other operations, correct coordinate deviations in various scenarios, trigger deviation warning and closed-loop calibration, and ensure that coordinate accuracy meets the GIS Level 1 application standard.

[0024] Step 7, Refined Vectorization and Operation and Maintenance Adaptation: Based on the calibrated coordinates and judgment results, an adaptation algorithm is used to extract the contours of various regions and optimize their smoothness. Relevant attribute information is associated and converted into a standardized vector format. The vector format is then imported into the operation and maintenance system through a standardized interface to adapt to the vector optimization needs of extreme scenarios.

[0025] Step 8: Full-process accuracy verification and error tracing: Obtain benchmark verification data such as GPS RTK and UAV aerial photography, statistically analyze various error indicators, determine the data qualification and generate a verification report, quantify the error contribution of each module, and optimize the parameters of each relevant module to form a closed-loop verification.

[0026] Step 9, Cross-module collaborative control and time-series dynamic monitoring: Throughout the entire process from Steps 1 to 8, the collaborative control module is invoked to achieve real-time data sharing among modules, dynamic adjustment of core parameters, error propagation blocking, and scenario adaptive adaptation. Combined with time-series images, dynamic monitoring and closed-loop operation and maintenance of the power plant throughout its entire life cycle are completed.

[0027] This invention provides an AI recognition and vectorization method for photovoltaic power plants based on sub-meter level images, including the following steps.

[0028] The technical solution of this application has achieved the following beneficial effects.

[0029] 1. Break through the limitations of isolated modules and build a closed-loop system for the whole process: Add a cross-module collaborative control module to realize data sharing, dynamic control and error blocking among modules, and solve the problems of error accumulation and unstable accuracy in existing technologies; establish a closed-loop system for the whole process of "preprocessing-identification-judgment-verification-optimization", so that the errors of each module can be traced and controlled, the error accumulation is reduced by more than 40% in extreme scenarios, and the accuracy and stability of the whole process is improved by 35%.

[0030] 2. Improve the comprehensiveness and accuracy of recognition: A new multimodal data fusion module has been added, which integrates 6 types of modal data and designs a dynamic weight allocation model to solve the limitation of incomplete feature extraction from a single data source in existing technologies. After fusion, the recognition accuracy in normal scenarios is ≥95%, in extreme scenarios it is ≥88%, and in hidden fault recognition it is ≥90%, all of which are higher than the 85% accuracy limit of existing patents. The IoU value of photovoltaic panel boundary recognition has been improved to over 0.92.

[0031] 3. Upgrade health assessment: Construct a dedicated grading and traceability model that can accurately detect latent faults such as hot spots and component aging, overcoming the limitation of existing technologies that can only identify obvious anomalies; achieve integrated "detection-grading-traceability-trend early warning" for latent faults, with a detection accuracy of ≥90%, providing core support for preventive operation and maintenance of photovoltaic power plants, and reducing operation and maintenance costs by more than 30%.

[0032] 4. Enhance extreme scenario adaptability and break through existing application boundaries: Design exclusive adaptation algorithms and parameter fusion rules for five types of extreme scenarios, including thick fog, snow accumulation, and dust, as well as multiple extreme scenarios, to fill the gap in the adaptation of existing technologies for snow accumulation and dust scenarios; the geometric correction error is ≤0.35m and the vector contour integrity is ≥99.5% under extreme scenarios, which can cover photovoltaic power station scenarios that existing technologies cannot adapt to, such as snowy northern regions and industrial areas, expanding the application scope by more than 60%.

[0033] 5. Improve coordinate accuracy in multiple scenarios: Construct a multi-source coordinate fusion calibration system, combine BeiDou data and terrain data, and design scenario-differentiated coordinate transformation algorithms to solve the limitations of existing technologies that rely heavily on GCP control points and experience sharp drops in accuracy in remote areas; the coordinate error in normal scenarios is ≤0.3m, and in extreme scenarios it is ≤0.35m, meeting the first-level application standard of GIS. The accuracy can still be stably met in scenarios with insufficient GCP, and high-precision coordinate matching can be achieved without manual intervention.

[0034] 6. Achieve linkage between vectorization and operation and maintenance to enhance the practical value of data: Break through the limitations of existing technologies where vectorization and operation and maintenance are disconnected, associate information such as hidden faults and operation and maintenance priorities with vector data, design standardized operation and maintenance interfaces, and directly import data into the operation and maintenance system; the vector data format supports adaptive conversion to adapt to different brands of operation and maintenance systems, achieving seamless connection between "recognition-vectorization-operation and maintenance", and transforming the sub-meter level image detail advantage into practical value.

[0035] 7. Enables precise detection, quantitative analysis, and early warning of changes in photovoltaic power plants; tracks fault development trends and triggers early warnings for operation and maintenance in advance, with an early warning accuracy rate of ≥92%, supporting dynamic monitoring and standardized management of photovoltaic power plants throughout their entire life cycle. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the AI ​​recognition and vectorization method for photovoltaic power plants based on sub-meter level images in the embodiments of this application.

[0037] Figure 2 This is the original sub-meter-level satellite image from the embodiments of this application.

[0038] Figure 3 for Figure 2 Enlarged view of section A.

[0039] Figure 4 for Figure 3 Enlarged view of section B in the middle. Detailed Implementation

[0040] The present application will now be further described with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present application and should not be construed as limiting the scope of protection of the present application. It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present application.

[0041] like Figure 1 , Figure 2 , Figure 3 and Figure 4 As shown, this invention discloses an AI recognition and vectorization system for photovoltaic power plants based on sub-meter level images, including...

[0042] Data Acquisition Module: Collects raw sub-meter level satellite imagery data, including optical satellite imagery, multispectral imagery, and thermal infrared imagery. Thermal infrared imagery is used to capture latent fault features such as hot spots on photovoltaic panels, while multispectral imagery is used to enhance the spectral difference between the photovoltaic panels and the background. Historical data on the target area's topography, meteorology, and basic attributes of the photovoltaic power station are collected. BeiDou satellite positioning data is collected as a supplementary data source for coordinate matching. Historical operation and maintenance data of the photovoltaic power station, including fault records, maintenance records, and power generation data, are collected for model iteration and health assessment. Secondary fine-grained annotation is performed on the raw data, covering photovoltaic power station boundaries, photovoltaic panel arrays, abnormal areas, and material information. Geographic, meteorological, and operation and maintenance data are associated as reference samples, and a sample library is constructed simultaneously to avoid duplicate annotations. The total number of samples reaches over 15,000, significantly improving the sample generalization ability. The sample areas cover major photovoltaic production areas in China, including the Northwest desert, the Eastern plains, the Southwest plateau, and the South China tropics. Module types cover mainstream power from 255W to 550W and service life from 1 to 15 years, ensuring a balanced sample proportion under different scenarios and operating conditions. The annotation boundary error is ≤0.1m. A dual-annotation + third-party review model is adopted. Only annotation consistency Kappa values ​​≥0.85 are included in the sample library. Regular training for annotators is conducted to reduce subjective errors. BeiDou data supplementation scenarios: Automatically activated in scenarios with <10 GCPs / image, steep mountain slopes ≥30°, and snow / dust accumulation. The fusion logic uses a BeiDou data weight of 0.4 + RPC model data weight of 0.6 to improve coordinate accuracy. A dynamic update mechanism for the sample library is established: updated quarterly, with no fewer than 1000 new samples added, focusing on supplementing extreme scenarios, false positives / false negatives, and new regional / component type samples; outdated samples with an accuracy rate <85% are discarded to ensure the timeliness of the sample library.

[0043] Image preprocessing module: Performs fully automated preprocessing on raw sub-meter level satellite images to eliminate data noise and distortion. It adopts scene-adaptive preprocessing algorithms, cross-module feedback mechanisms, and latent fault feature enhancement designs, specifically including...

[0044] Radiation correction unit: Employing a scene-adaptive 6S model, it automatically matches atmospheric patterns and aerosol types based on terrain and meteorological data. It sets interval-specific reflectivity normalization thresholds according to photovoltaic panel material, and includes a fallback rule for missing data. The entire process is automated, eliminating the influence of differences in illumination and material composition. A new thermal infrared image radiation correction subunit uses an atmospheric window correction algorithm to eliminate the impact of atmospheric attenuation on thermal infrared signals, ensuring accurate extraction of photovoltaic panel hotspot features. The reflectivity thresholds are 0.3-0.6 for monocrystalline silicon, 0.25-0.55 for polycrystalline silicon, and 0.2-0.5 for thin films. The thermal infrared atmospheric window correction algorithm adapts to the 8-14μm wavelength range, closely matching the thermal radiation characteristics of photovoltaic panels. For single-type data missing ≤20%, linear interpolation is used for supplementation; for 20%-50%, BeiDou positioning and historical data from the same period are used for replacement; and for >50%, re-acquisition is triggered to ensure uninterrupted preprocessing.

[0045] Geometric correction unit: Based on the RPC parameter and quadratic polynomial fitting fusion algorithm, the elevation compensation term weight is optimized for mountainous steep slope ≥30° scenarios, and the terrain slope factor is introduced to correct the projection difference; a new geometric correction adaptation strategy for snow cover and industrial dust scenarios is added, and the image distortion is corrected by combining terrain and meteorological data to ensure that the geometric correction error is ≤0.5m / pixel in various scenarios, which breaks through the limitation of existing technologies that do not consider the projection difference of steep slopes, snow cover and dust scene distortion.

[0046] Noise Removal Unit: Employs a combined adaptive median filtering and wavelet thresholding algorithm, dynamically adjusting parameters based on noise intensity to remove salt-and-pepper noise and stripe noise while preserving the details of the photovoltaic panel grid texture. A combined Gaussian filtering and morphological filtering algorithm removes thermal noise from the thermal infrared image, enhancing the recognizability of hot spot features. Noise Removal Parameters: Adaptive median filtering window size is 3×3-7×7; a 3×3 window is used when noise intensity is ≤0.1, and a 5×5-7×7 window is used when noise intensity is >0.1. Wavelet thresholding denoising threshold ranges from 0.02-0.08, dynamically adjusted according to image noise intensity.

[0047] Extreme Scene Preprocessing Adaptation Unit: Dedicated adaptation algorithms are designed for five extreme scene types: thick fog, strong shadows, steep mountain slopes, snow cover, and industrial dust. Parameter fusion rules for multiple extreme scene overlays, such as steep slopes + strong shadows + snow cover, are adopted. Customized processing solutions are generated according to weight and priority to address the inability of existing technologies to adapt to multiple extreme scene overlays and snow / dust scenes. Algorithms such as spectral enhancement and temperature threshold segmentation are used to enhance the characteristics of latent faults such as hot spots on photovoltaic panels and module aging, laying the foundation for subsequent latent fault identification. For scenes near water surfaces, a reflection suppression algorithm is used to weaken water surface reflection interference, and multispectral imagery is combined to distinguish photovoltaic panels from the water surface. For temporary construction obstruction scenes, a target recognition auxiliary algorithm is used to distinguish between construction equipment and damaged photovoltaic panels, avoiding misjudgments.

[0048] Cross-module collaborative feedback unit: This unit provides real-time feedback of preprocessing quality indicators to the AI ​​recognition module and the latent fault assessment module. These indicators include the clarity of hotspot features in thermal infrared images. It receives feedback on recognition and fault assessment errors and automatically optimizes preprocessing parameters, forming a closed loop of preprocessing-recognition-fault assessment-feedback-optimization, unlike existing technologies where each module operates in isolation. Preprocessing quality qualification thresholds: thermal infrared image hotspot feature clarity ≥ 0.85, optical image texture clarity ≥ 0.8, and reflectivity deviation after radiometric correction ≤ 5%. If the values ​​are below these thresholds, the preprocessing parameters are automatically optimized.

[0049] AI Recognition Model Module: Based on the improved U-Net++ network, a scene-adaptive recognition model is built, which integrates Transformer and attention mechanisms to overcome the limitation of existing technologies that can only identify obvious anomalies. It achieves integrated accurate recognition, hierarchical judgment of obvious anomalies, and detection of hidden faults, specifically including...

[0050] Sample library construction unit: Construct a sample library of 15,000+ multi-scene sub-meter level images, with complex scenes, extreme scenes, and hidden fault scenes accounting for ≥40%, ≥20%, and ≥15% of the samples, respectively. Associate it with multi-source auxiliary data, including terrain, meteorology, operation and maintenance, and thermal infrared features, to improve the model's generalization ability and hidden fault identification ability.

[0051] Model architecture design unit: Introduces Transformer attention mechanism and ASPP dilated convolution pooling module, including scene classification branch, extreme scene-specific inference branch and hidden fault identification branch; The hidden fault identification branch adopts thermal infrared and optical image feature fusion algorithm to capture the correlation features of photovoltaic panel temperature anomaly and texture anomaly, and realize the accurate detection of hidden faults such as hot spots and component aging; Design boundary refinement branch to solve the problems of adjacent panel adhesion and blurred boundary segmentation in existing technologies; Integrate multispectral, optical and thermal infrared image features to improve the comprehensiveness of recognition.

[0052] Model training optimization unit: Employs a weighted fusion algorithm of Dice loss, cross-entropy loss, and temperature loss to address class imbalance and the lack of obvious latent fault features. Shortens the training cycle through transfer learning and introduces a reverse iteration mechanism based on maintenance data, feeding back confirmed fault data from the maintenance process to the model for continuous parameter optimization. Ensures test set recognition accuracy ≥95%, extreme scenario recognition accuracy ≥88%, and latent fault recognition accuracy ≥90%. Latent fault recognition recall ≥92%, precision ≥93%, and explicit anomaly recognition recall ≥96%, precision ≥95%. Accuracy calculation criteria are defined as follows: recognition results with boundary deviation ≤0.2m are considered acceptable; separately labeled areas with blurred boundaries are not included in accuracy statistics. Based on a photovoltaic-specific pre-trained U-Net model, layers 3-5 of the encoding layer are selected as transfer layers. Small-batch, low-learning-rate fine-tuning is employed, with the transfer layers frozen for the first 5 rounds and unfrozen and trained synchronously in subsequent rounds, shortening the training cycle by more than 30%.

[0053] The photovoltaic power plant health status preliminary assessment unit: Based on multiple indicators such as reflectivity distribution, array integrity, and temperature distribution, it can classify obvious anomalies such as photovoltaic panel damage and shading into mild / moderate-severe levels, and at the same time classify latent faults such as hot spots and module aging into minor, moderate, and severe levels, marking the coordinates of abnormal areas and fault types, providing accurate basis for operation and maintenance.

[0054] Error feedback linkage unit: It feeds back the identification error and the hidden fault judgment error to the preprocessing module and the multimodal data fusion module, triggers parameter optimization, and forms a closed loop, which is different from the design of existing technology without error linkage.

[0055] Multimodal data fusion module: Integrates optical imagery, multispectral imagery, thermal infrared imagery, BeiDou positioning data, topographic and meteorological data, and historical operation and maintenance data to achieve multi-dimensional feature fusion, providing comprehensive data support for identification, coordinate matching, and health assessment, specifically including...

[0056] Data Alignment Unit: Employing a spatiotemporal alignment algorithm, multispectral and thermal infrared images are aligned pixel-level with optical images, achieving an alignment error ≤0.1 pixels. This unit spatiotemporally correlates BeiDou positioning data, topographic and meteorological data, and operational data with image data to ensure consistency and correlation of multimodal data. Time Synchronization Accuracy Requirements: The time difference between multimodal image captures must be ≤1 hour. If this range is exceeded, a temporal interpolation correction algorithm is used, combined with meteorological data, to correct temperature and spectral characteristic deviations, ensuring fusion accuracy.

[0057] Feature Fusion Unit: Employing a feature fusion algorithm that combines convolutional neural networks and Transformer, this unit extracts core features from various modalities, including texture features from optical images, spectral features from multispectral images, temperature features from thermal infrared images, spatial features from BeiDou data, and fault features from maintenance data. This achieves deep fusion of multi-dimensional features, improving recognition accuracy and fault assessment precision, and addressing the issue of incomplete feature extraction from single data sources. Feature Fusion Layers: The convolutional neural network extracts shallow texture features (layers 1-2) and mid-level spectral features (layers 3-4), while the Transformer extracts deep semantic features (layers 5-6). Feature fusion is performed at layer 4, balancing feature details with semantic information.

[0058] The data weight dynamic allocation unit dynamically allocates the fusion weights of each modality of data based on scene type and data quality. For example, in extreme scenes, the weights of thermal infrared and BeiDou data are increased, while in conventional scenes, the weights of optical and multispectral data are increased. When a data source is missing, the weight allocation rules are automatically adjusted to ensure that high processing accuracy can be maintained under any combination of data sources, further reducing dependence on a single data source. Weight adjustment trigger conditions: When snow cover ≥30%, dust impact ≥20%, or steep slope ≥30°, the weight of thermal infrared data is automatically increased to 0.35, and the weight of BeiDou data is increased to 0.25; in conventional scenes, the weights of optical imagery (0.4) and multispectral imagery (0.3) are maintained.

[0059] Fusion Quality Verification Unit: Sets fusion quality evaluation indicators, including feature consistency, data correlation and accuracy contribution, to verify the multimodal fusion effect in real time. When the fusion quality fails to meet the standards, it automatically adjusts the fusion algorithm and weight allocation rules to ensure the reliability of the fused data and provide high-quality data support for subsequent modules.

[0060] Coordinate matching module: Combining BeiDou and terrain data from the multimodal data fusion module, a multi-source coordinate fusion calibration system is constructed. A scenario-differentiated calibration strategy is adopted to achieve high-precision coordinate matching, specifically including...

[0061] Coordinate transformation algorithm unit: It adopts the RPC model and quadratic polynomial fitting fusion algorithm with a fusion weight of 0.7:0.3, and combines the terrain slope factor to optimize the elevation compensation term to correct the coordinate offset of steep mountain slopes; it adopts the coordinate transformation adaptation algorithm for snow and dust scenes, and combines meteorological data to correct the coordinate deviation caused by atmospheric refraction.

[0062] Triple coordinate calibration unit: Employing a triple mechanism of transformation calibration, overlapping area cross-calibration, and GCP sampling calibration, it uses BeiDou positioning data and terrain data to perform secondary corrections on the calibration results, ensuring coordinate errors ≤0.3m in normal scenarios and ≤0.35m in extreme scenarios, meeting the GIS Level 1 application standard. GCP sampling rules: 8-12 GCPs are sampled from each image within 10km², and 12-16 GCPs are sampled from images within 10-20km², evenly distributed in the four corners and central area of ​​the image, with a positional deviation of ≤0.05m for the sampled GCPs.

[0063] Multi-source coordinate fusion calibration unit: When the number of ground control points is less than 10 per frame, BeiDou satellite positioning data and terrain data are called to supplement the calibration. The weight allocation rules for missing data sources are adopted, and the weights are dynamically adjusted according to the accuracy of the data sources. For example, when the accuracy of BeiDou data is 0.1m, the weight is 0.4, and the weight of terrain data is 0.2, to ensure that the accuracy does not decrease under any combination of data sources.

[0064] Coordinate Deviation Early Warning Unit: Sets a deviation threshold; when the threshold is exceeded, it automatically returns to the geometric correction stage, forming a closed-loop calibration to prevent error propagation. It locates the source of the deviation, such as atmospheric refraction or terrain distortion, and feeds it back to the preprocessing module to optimize parameters. Coordinate Deviation Early Warning Threshold: An early warning is triggered when the coordinate deviation is >0.3m in normal scenarios and >0.35m in extreme scenarios, automatically returning to the geometric correction stage for reprocessing to prevent error propagation.

[0065] Vector data generation module: Associates the results of hidden fault identification with operation and maintenance data to achieve integrated refinement, standardization and operation and maintenance adaptation of vector data, specifically including...

[0066] Vectorization processing unit: Employing the Douglas-Peucker algorithm, the threshold is dynamically adjusted according to the scene, with a standard threshold of 0.1m and an extreme threshold of 0.08m. This extracts the contours of photovoltaic power plants, visible anomaly areas, and hidden fault areas, preserving key inflection points. Thermal infrared features are combined to mark the boundaries of fault areas, ensuring accurate contours. Threshold adjustment criteria: A standard threshold of 0.1m and an extreme threshold of 0.08m are used when the image resolution is 0.5m; when the resolution is 1m, the standard threshold is adjusted to 0.12m and the extreme threshold to 0.1m; when the terrain slope is ≥30°, the threshold is further reduced by 0.02m to improve contour accuracy.

[0067] Boundary smoothing optimization unit: Gaussian filtering eliminates jagged boundaries, ensuring the deviation between the contour and the actual boundary is ≤0.2m, and ≤0.15m in extreme scenarios. For contour breakage issues in snow and dusty scenes, multimodal data is used for repair, ensuring contour integrity ≥99.5%. Contour integrity evaluation method: Contour breakage length ≤0.15m and number of breaks ≤2 per image are considered complete. A combination of manual sampling and automatic algorithm detection is used, with a sampling ratio ≥10%, ensuring reliable evaluation results.

[0068] Attribute information association unit: Automatically calculates and classifies the power plant's footprint and array quantity, and associates multiple attributes such as health status, coordinate accuracy, material, fault type, and fault level; according to the fault level and impact range, it marks the priority of operation and maintenance, such as emergency, routine, and postponement, to improve the usability of data.

[0069] Operation and Maintenance Interface Adaptation Unit: Utilizing standardized HTTP / HTTPS interfaces, this unit directly imports vector data, anomaly information, and operation and maintenance priorities into the photovoltaic power plant operation and maintenance management system, achieving seamless integration of identification, vectorization, and operation and maintenance. It adapts to the interface specifications of different brands of operation and maintenance systems, improving compatibility. Compatibility Scope: Supports mainstream operation and maintenance systems such as Huawei, Sungrow Power, and Goldwind Technology. The interface protocol uses HTTP / HTTPS version 1.1 and supports the RESTful API interface specification, ensuring compatibility.

[0070] Extreme Scene Vector Optimization Unit: Performs contour integrity verification and repair on vector data in extreme scenes to ensure integrity ≥99.5%; in snow-covered scenes, a photovoltaic panel contour extraction optimization algorithm is used to distinguish between snow and photovoltaic panels to avoid incorrect contour extraction.

[0071] Latent Fault Assessment Module: Based on the characteristics of latent faults from multimodal data fusion and AI recognition models, a system for accurate assessment and tracing of latent faults is constructed, realizing integrated processing of latent faults from detection and classification to tracing. Specifically, it includes...

[0072] Latent Fault Detection Unit: Based on the temperature characteristics of thermal infrared images, the spectral characteristics of multispectral images, and the texture characteristics of optical images, combined with the latent fault branch output of the AI ​​recognition model, it accurately detects latent faults such as hot spots, component aging, poor circuit contact, and minor leakage, eliminating environmental interference such as uneven illumination and localized temperature anomalies, with a detection accuracy of ≥90%. It employs a dual exclusion logic of temperature threshold + texture features: the normal temperature range for photovoltaic panels is set at 25-45℃; those exceeding this range and accompanied by texture damage / abnormalities are judged as genuine hot spots; temperature anomalies caused by uneven illumination are verified using the spectral characteristics of multispectral images to eliminate false detection results.

[0073] Fault Classification Unit: Combining photovoltaic panel temperature deviation, power generation data, and operation and maintenance history, a latent fault classification model is constructed, classifying latent faults into three levels: minor, moderate, and severe, providing a priority basis for operation and maintenance handling. Prioritizing Fault Source Tracing: First, determine directly observable causes, such as shading > module damage, then determine latent causes, such as poor line contact > module aging. Based on historical operation and maintenance data, prioritize locating high-frequency causes to shorten the source tracing time.

[0074] Fault tracing unit: By combining multimodal data, it analyzes the causes of hidden faults, locates the root cause of the fault, generates a fault tracing report, and helps maintenance personnel to handle the faults accurately and reduce maintenance costs.

[0075] Fault Development Trend Analysis Unit: Combining time-series images and operation and maintenance data, it analyzes the development trend of hidden faults, predicts the risk of fault escalation, triggers operation and maintenance early warnings in advance, realizes preventive operation and maintenance, and fills the gap in existing technologies that cannot achieve hidden fault trend analysis.

[0076] Accuracy Verification Module: Constructs a full-process dual accuracy verification and error tracing system, adopting multi-modal fusion accuracy verification and latent fault judgment accuracy verification, specifically including.

[0077] Verification Data Acquisition Unit: Verification baseline data is acquired through a fusion of GPS RTK and UAV aerial photography, with enhanced aerial and field testing verification in extreme scenarios. Thermal infrared verification data, measured by an infrared thermometer with an accuracy of ±0.1℃, is used to verify the accuracy of latent fault diagnosis. Verification Data Accuracy: GPS RTK measured accuracy ±0.05m, UAV aerial photography resolution ≥0.1m, and infrared thermometer measured accuracy ±0.1℃, ensuring that the verification data accuracy is higher than the system output accuracy, and the verification results are reliable.

[0078] Dual error statistics unit: This unit statistically analyzes boundary errors, area errors, anomaly identification accuracy, and latent fault assessment accuracy, establishing quantitative correlation rules between errors and parameter optimization to provide a basis for subsequent optimization; it also statistically analyzes the accuracy contribution under different fusion weights to optimize the fusion strategy. Correlation logic: When the boundary error is ≥0.2m, the weight of the geometric correction elevation compensation term is increased by 0.1; when the latent fault false detection rate is ≥5%, the weight of the thermal infrared image fusion is increased by 0.05, establishing a quantitative correspondence between errors and parameter adjustments.

[0079] Cross-module verification feedback unit: The decision tree algorithm is used to quantify the error contribution of each module, determine the responsible module and trigger automatic optimization; when the fusion quality does not meet the standard, the fusion algorithm optimization is triggered to form a closed-loop verification process.

[0080] Cross-module collaborative control module: Enables dynamic collaboration among modules, error propagation control, scenario adaptive adaptation, closed-loop operation and maintenance iteration, and time-series early warning, specifically including...

[0081] Data Collaboration Unit: Establish a unified data collaboration platform to integrate data from all modules, including multimodal data and operational data, enabling real-time sharing and visualization, and solving the problem of fragmented data in existing technologies; ensuring real-time consistency of data across modules and improving collaboration efficiency. Specifically, this includes: Platform Specifications: Data storage uses JSON and GeoTIFF formats, and the transmission protocol uses TCP / IP, supporting real-time data synchronization with a latency of ≤100ms, ensuring efficient data sharing across modules.

[0082] Dynamic adjustment unit: Based on data from each module, it automatically adjusts core parameters such as preprocessing thresholds, model attention weights, and multimodal fusion weights to ensure a balance between accuracy and efficiency throughout the entire process. A new scenario-differentiated adjustment subunit is added to formulate differentiated adjustment strategies for different scenarios. Parameter adjustment step size: The fusion weight adjustment step size is 0.05, ranging from 0.05 to 0.4; the preprocessing threshold adjustment step size is 0.02, ranging from 0.02 to 0.1, adopting a small step size and multiple iteration mode to avoid accuracy fluctuations.

[0083] Error propagation control unit: Sets error thresholds for each module. When the threshold is exceeded, it automatically blocks error propagation and triggers module optimization to avoid error accumulation. Different optimization strategies are adopted according to the severity of the error, such as local optimization for minor errors and full-process optimization for severe errors.

[0084] Scene adaptive adaptation unit: automatically determines the scene type, calls the exclusive adaptation parameters of each module, and generates customized solutions by weight fusion rules for multiple extreme superposition scenarios; continuously optimizes the scene determination accuracy by combining operation and maintenance data.

[0085] Time-series image comparison and analysis unit: Integrates multimodal time-series images with operation and maintenance data to achieve dynamic monitoring of the entire lifecycle of photovoltaic power plants, including time-series image registration, accurate difference detection, quantitative change analysis, operation and maintenance early warning, and data closed-loop iteration, linking with the operation and maintenance interface to achieve early warning push; tracks the development trend of hidden faults and provides early warning of fault escalation risks. Time-series image acquisition interval: once per quarter for normal scenarios, once per month for extreme scenarios, and once per week for fault areas to ensure accurate tracking of fault development trends.

[0086] Operation and Maintenance Closed-Loop Iteration Unit: This unit feeds back maintenance records and fault confirmation results from the operation and maintenance system to the sample library and AI recognition model, continuously optimizing the parameters of the sample library and model. It also feeds back accuracy issues discovered during operation and maintenance to each module, triggering parameter optimization and achieving a closed-loop iteration of recognition-operation and maintenance-optimization-re-recognition, continuously improving system performance. Closed-Loop Collaboration Rules: Priority is set for the verification feedback closed loop > preprocessing closed loop > coordinate calibration closed loop. After high-priority closed loop parameters are optimized, low-priority closed loops automatically adjust their parameters accordingly. A new coordination unit is added to monitor the parameters of each closed loop in real time, avoiding conflicts.

[0087] System Integration and Control Unit: Achieves integrated control over the entire process, supports unified parameter configuration and automated scheduling, with a single 10-square-kilometer image processing time of ≤30 minutes; supports parallel processing of multiple regions and batches of images, improving processing efficiency by more than 60% compared to existing technologies, and adapts to the needs of large-scale photovoltaic power plant surveys. Technical Specifications: Revised to a single 10-square-kilometer sub-meter resolution image processing time of ≤30 minutes, while supplementing the processing time for a single 0.5m resolution, 1 square kilometer image of ≤5 minutes, ensuring that processing efficiency is assessable and verifiable.

[0088] Operations and Maintenance Data Iteration Module: As the support for the closed-loop iteration of operations and maintenance, it is responsible for the collection, organization, analysis and feedback of operations and maintenance data, specifically including...

[0089] Operation and Maintenance Data Acquisition Unit: Connects to the photovoltaic power station operation and maintenance management system via standardized interfaces to collect real-time operation and maintenance data such as maintenance records, fault confirmation results, power generation data, and component replacement records, ensuring data timeliness and completeness. Acquisition Frequency: Fault records and maintenance records are collected in real-time; power generation data is collected hourly; component replacement records are synchronized daily; and meteorological data is collected every 30 minutes, ensuring data timeliness.

[0090] Data Processing and Analysis Unit: This unit cleans, classifies, and analyzes operational data, extracting key information such as fault characteristics and accuracy optimization needs, and establishing the correlation between operational data and identification accuracy and fault assessment accuracy. Correlation Analysis Method: Pearson correlation analysis is used to quantify the degree of correlation between operational data and identification accuracy; a correlation coefficient ≥ 0.7 is considered a strong correlation. Combined with linear regression analysis, a predictive model for operational data and fault assessment accuracy is established to achieve targeted optimization.

[0091] Data Feedback Iteration Unit: The organized and analyzed operation and maintenance data is fed back to the sample library, AI recognition model, cross-module collaborative control module, and hidden fault judgment module to achieve iterative upgrades of the entire system and continuously improve technical performance and practicality.

[0092] This invention provides an AI recognition and vectorization method for photovoltaic power plants based on sub-meter level images, including the following steps.

[0093] Step 1: Multimodal Data Acquisition: Acquire sub-meter level satellite imagery of the target area (0.5m-1m) through satellite data service providers, including optical, multispectral, and thermal infrared data. Simultaneously collect BeiDou positioning data, topographic and meteorological data, and historical operation and maintenance data of photovoltaic power plants. Call the image quality pre-inspection unit to screen qualified images based on four indicators: grayscale mean 120-200, texture clarity ≥0.8, thermal infrared signal intensity ≥0.7, and spectral consistency. Remove invalid images if cloud cover ≤20%, snow cover ≤10%, and dust impact ≤15%. Automatically determine scene type, including conventional, single extreme, and multi-extreme overlay scenes, and push them to the cross-module collaborative control module to provide a basis for subsequent parameter adaptation. Verify the integrity and correlation of each modal data. If the data quality does not meet the standards, automatically supplement the collection or trigger data repair.

[0094] Step 2, Image Preprocessing: Input the selected qualified raw images and multimodal data into the preprocessing module. Through fully automated processing such as radiometric correction, geometric correction, noise removal, and extreme scene adaptation, data noise and distortion are eliminated, and preprocessing parameters are optimized through cross-module feedback.

[0095] Step 3, Multimodal Data Fusion: Input the preprocessed data into the fusion module to complete pixel-level spatiotemporal alignment and multi-dimensional feature deep fusion. Dynamically allocate the fusion weights of each modality according to the scene type, verify the fusion quality, and ensure that high-quality data support is provided for subsequent steps.

[0096] Step 4: AI Identification and Anomaly Classification Assessment: Input the multimodal fusion feature map into the AI ​​identification model, adaptively call the corresponding inference branch, complete the batch identification of explicit anomalies and latent faults of photovoltaic panels, classify and assess explicit anomalies and latent faults, mark relevant information and feed back error optimization parameters.

[0097] Step 5: Accurate assessment and tracing of latent faults: Combining AI identification results, multimodal fusion data and historical operation and maintenance data, latent faults are identified and their classifications are refined. The core causes of the faults are located and the sources are traced. The development trend of the faults is analyzed, the upgrade risks are predicted, and operation and maintenance warnings are triggered.

[0098] Step 6, Multi-source coordinate fusion calibration: Input pixel-level segmentation mask, fault area outline and fusion data, and through coordinate transformation, triple calibration, multi-source fusion and other operations, correct coordinate deviations in various scenarios, trigger deviation warning and closed-loop calibration, and ensure that coordinate accuracy meets the GIS Level 1 application standard.

[0099] Step 7, Refined Vectorization and Operation and Maintenance Adaptation: Based on the calibrated coordinates and judgment results, an adaptation algorithm is used to extract the contours of various regions and optimize their smoothness. Relevant attribute information is associated and converted into a standardized vector format. The vector format is then imported into the operation and maintenance system through a standardized interface to adapt to the vector optimization needs of extreme scenarios.

[0100] Step 8: Full-process accuracy verification and error tracing: Obtain benchmark verification data such as GPS RTK and UAV aerial photography, statistically analyze various error indicators, determine the data qualification and generate a verification report, quantify the error contribution of each module, and optimize the parameters of each relevant module to form a closed-loop verification.

[0101] Step 9, Cross-module collaborative control and time-series dynamic monitoring: Throughout the entire process from Steps 1 to 8, the collaborative control module is invoked to achieve real-time data sharing among modules, dynamic adjustment of core parameters, error propagation blocking, and scenario adaptive adaptation. Combined with time-series images, dynamic monitoring and closed-loop operation and maintenance of the power plant throughout its entire life cycle are completed.

[0102] Example 1.

[0103] This embodiment 1 discloses an AI recognition and vectorization method for photovoltaic power plants based on sub-meter level images, wherein step 2 includes the following steps.

[0104] 2.1 Radiation Correction: The scene-adaptive 6S model is adopted, and the atmospheric mode and aerosol type are automatically matched in combination with terrain and meteorological data. The reflectance is normalized according to the photovoltaic panel material, ranging from 0.2 to 0.7. The thermal infrared image is simultaneously radiometrically corrected, and the atmospheric window correction algorithm is used to eliminate the influence of atmospheric attenuation. The normalized temperature range is 25-60℃. When data is missing, a fallback rule is executed. The whole process is automated.

[0105] The photovoltaic panel reflectivity correction model is as follows.

[0106] .

[0107] In the formula, The corrected reflectivity of the photovoltaic panel is used for subsequent photovoltaic panel feature extraction and identification. The original reflectance of the photovoltaic panel is the initial reflectance data directly extracted from sub-meter level optical images; Atmospheric path radiation is calculated by the scene-adaptive 6S model in combination with meteorological data, representing the attenuation interference of the atmosphere on radiation; τ is atmospheric transmittance, output by the 6S model after matching the current atmospheric model and aerosol type. The solar zenith angle is the angle parameter at which the photovoltaic panel receives sunlight, obtained in conjunction with terrain data. To observe the zenith angle, the satellite observes the angle parameters of the photovoltaic panel, which are taken from satellite imagery metadata; Atmospheric hemispherical albedo represents the hemispherical reflectivity of the atmosphere to solar radiation, and its value is optimized for photovoltaic scenarios. These are correction factors for photovoltaic panel materials: 0.92 for monocrystalline silicon, 0.88 for polycrystalline silicon, and 0.85 for thin-film photovoltaic panels. These factors are used to eliminate differences in reflectivity between photovoltaic panels made of different materials. This is a latent fault characteristic reflectivity correction term for the hot spot region. =0.03, normal range =0, used to enhance the difference in reflectivity characteristics in areas with latent faults.

[0108] The thermal infrared image radiation correction model is as follows.

[0109] .

[0110] In the formula, This represents the actual temperature of the photovoltaic panel, used to detect latent fault characteristics such as hot spots. The initial temperature data is directly extracted from sub-meter level thermal infrared images for satellite observation brightness temperature; ε is the emissivity of the photovoltaic panel, dynamically determined according to the photovoltaic panel material: 0.91 for monocrystalline silicon, 0.90 for polycrystalline silicon, and 0.89 for thin film. Atmospheric emissivity, calculated in conjunction with real-time meteorological data, characterizes the atmosphere's own thermal radiation capacity; This is a hotspot characteristic correction coefficient, dynamically ranging from 0.95 to 1.02 based on the intensity of the thermal infrared signal. The stronger the hotspot signal, the larger the value, used to enhance the temperature characteristics of the hotspot region. This is an environmental temperature interference correction term, calculated in conjunction with real-time meteorological data, used to eliminate the interference of environmental temperature on the photovoltaic panel temperature inversion, ensuring accurate extraction of photovoltaic panel hot spot characteristics.

[0111] 2.2 Geometric Correction: Based on the fusion algorithm of RPC parameters and quadratic polynomial fitting, geometric distortion is corrected; for mountainous steep slopes ≥30°, the weight of the elevation compensation term is optimized; for snow and dust scenes, meteorological data is used to correct deviations caused by atmospheric refraction, ensuring an error ≤0.5m / pixel. The geometric correction model is as follows.

[0112] .

[0113] .

[0114] In the formula, X and Y are the corrected geographic coordinates of the photovoltaic panel, adapted to the CGCS2000 coordinate system, used for subsequent coordinate matching and vectorization, which meets the needs of GIS applications; x and y are the pixel coordinates of the photovoltaic panel in the original sub-meter level image, which are directly extracted from the original image and are the initial input parameters for coordinate transformation. _ The X-direction coordinate transformation coefficient is calculated by the fusion algorithm of RPC parameters and quadratic polynomial fitting, and is dynamically adjusted according to the scene type. In extreme scenes, the coefficient iteration step size is optimized. _ This represents the Y-axis coordinate transformation coefficient, corresponding to the transformation relationship from pixel coordinates in the Y-axis to geographic coordinates; For extreme scene distortion correction factors, snow scene =0.08, Dust Scene =0.06, Steep Slope Scene =0.12, typical scenario =0 is used to correct image distortion in different extreme scenarios. α is the slope of the target area, which is calculated from the terrain data. When α≥30°, the secondary slope correction logic is enabled. It is different from the general slope correction and is used to correct the projection difference in mountainous steep slope scenarios to ensure the correction accuracy.

[0115] 2.3 Noise Removal: An adaptive median filtering and wavelet threshold denoising algorithm is adopted to dynamically adjust parameters according to noise intensity and preserve the texture details of the photovoltaic panel; simultaneously, noise removal is performed on the thermal infrared image using a Gaussian filtering and morphological filtering algorithm to remove thermal noise and enhance hot spot features.

[0116] 2.4 Extreme Scene Adaptation: Dedicated algorithms are executed for single extreme scenes, and weighted fusion parameters are used to process multiple extreme scenes; a snow cover scene adaptation algorithm is adopted to distinguish snow from photovoltaic panels and avoid texture feature confusion; for industrial dust scenes, a haze removal algorithm is used to weaken the impact of dust; through algorithms such as spectral enhancement and temperature threshold segmentation, the hidden fault characteristics such as photovoltaic panel hot spots and component aging are enhanced.

[0117] 2.5 Cross-module feedback: Preprocessing indicators are fed back to the AI ​​recognition module and the latent fault assessment module. The module receives feedback on recognition errors and fault assessment errors and automatically optimizes parameters, forming a closed loop. Preprocessing indicators include the clarity of hotspot features in thermal infrared images.

[0118] Example 2.

[0119] This embodiment 2 discloses an AI recognition and vectorization method for photovoltaic power plants based on sub-meter level images, wherein step 3 includes the following steps.

[0120] 3.1 Data Alignment: A spatiotemporal alignment algorithm is adopted to align multispectral and thermal infrared images with optical images at the pixel level, and to spatiotemporally correlate BeiDou positioning data, topographic and meteorological data, operation and maintenance data with image data to ensure the consistency of multimodal data, with an alignment error of ≤0.1 pixels.

[0121] 3.2 Feature Fusion: A feature fusion algorithm combining convolutional neural networks and Transformer is adopted to extract the core features of each modality data, and to achieve deep fusion of optical texture features, multispectral spectral features, thermal infrared temperature features, BeiDou spatial features, and operation and maintenance fault features, generating a multimodal fusion feature map.

[0122] 3.3 Dynamic Weight Allocation: Based on scene type and data quality, the fusion weights of each modality of data are dynamically allocated. For example, in extreme scenarios, the weights of thermal infrared and BeiDou data are increased to 0.35 and 0.25, respectively; in normal scenarios, the weights of optical and multispectral data are increased to 0.4 and 0.3, respectively. When data sources are missing, the weight allocation rules are automatically adjusted to ensure that the fusion accuracy is not reduced. The dynamic weight allocation model is as follows.

[0123] .

[0124] In the formula, The fusion weight for the i-th modal data is a core output parameter that characterizes the contribution of this modal data to the multimodal fusion result. The sum of the weights of all modal data is 1 to ensure the fusion logic is rigorous. This is the quality evaluation index for the i-th modal data, with a value ranging from 0 to 1. It is determined based on a comprehensive assessment of data completeness, accuracy, and sharpness. For example, if the optical image sharpness is ≥ 0.8... ≥0.9, when the BeiDou positioning data accuracy is 0.1m =1.0; Let be the adaptation coefficient of the i-th modal data in the current scene, representing the optical image in a typical scene. =0.4, multispectral image =0.3, thermal infrared imagery in extreme scenarios =0.35, Beidou positioning data =0.25, dynamically matching scenario requirements; n is the number of modal data types. In this application, n=6, corresponding to six major categories of modal data: optical images, multispectral images, thermal infrared images, BeiDou positioning data, topographic and meteorological data, and operation and maintenance historical data; where j is the summation index, corresponding to the j-th modal data, and is only used to distinguish the summation variable from the target variable. Let j be the quality evaluation index for the j-th modal data. Let be the adaptation coefficient of the j-th modal data in the current scenario. For the modal cooperative gain term of the j-th modal data, Let be the latent fault correlation coefficient of the j-th modal data. This is the modal synergy gain term, reflecting the synergistic effect of multimodal fusion, such as when thermal infrared data is synergistic with operational data. =0.3, when a single modality participates in fusion independently =0, enhancing the feature contribution of collaborative modes. The correlation coefficient for latent faults aligns with the core requirement of latent fault identification in this application. Different modal data exhibit varying degrees of correlation with latent faults, and thermal infrared images... =1.2, Operation and Maintenance Data =1.1, optical image =1.0, the higher the correlation, the larger the coefficient, and the higher the fusion weight of latent fault features; for example, in extreme scenarios, the weight of thermal infrared and Beidou data is increased to 0.35 and 0.25, respectively, and in normal scenarios, the weight of optical and multispectral data is increased to 0.4 and 0.3, respectively.

[0125] 3.4 Fusion Quality Verification: Set fusion quality evaluation indicators to verify the multimodal fusion effect in real time. When the fusion quality fails to meet the standards, automatically adjust the fusion algorithm and weight allocation rules. After the verification is qualified, push the fusion feature map to the AI ​​recognition model module and the hidden fault judgment module.

[0126] Example 3.

[0127] This embodiment 3 discloses an AI recognition and vectorization method for photovoltaic power plants based on sub-meter level images, wherein step 4 includes the following steps.

[0128] 4.1 Model Adaptation: Based on the scene type, the model automatically calls the scene classification branch, the extreme scene-specific reasoning branch, and the hidden fault identification branch to achieve adaptive reasoning; the hidden fault identification branch focuses on capturing the correlation features between thermal infrared temperature anomalies and texture anomalies to improve the accuracy of hidden fault detection.

[0129] 4.2 Batch Inference: The batch size is automatically adjusted according to the image resolution, and the inference time for a single 10 images is ≤30 minutes, improving processing efficiency; a new parallel inference sub-step is added to support simultaneous inference of multiple images, further improving the efficiency of large-scale census.

[0130] 4.3 Anomaly Classification and Judgment: Based on multimodal fusion characteristics, the system identifies explicit and implicit anomalies in photovoltaic panels. Explicit anomalies are classified as mild / moderate-severe, and implicit faults are classified as slight / moderate / severe. Coordinates and fault types are marked, and detailed descriptions are generated. Explicit anomalies include damage and shading; implicit faults include hot spots and module aging.

[0131] 4.4 Error Feedback: The boundary segmentation IoU value, anomaly identification accuracy, and latent fault judgment accuracy are fed back to the preprocessing module and the multimodal data fusion module. When the error exceeds the standard, the preprocessing parameters and fusion weights are optimized to form a closed loop.

[0132] Example 4.

[0133] This embodiment 4 discloses an AI recognition and vectorization method for photovoltaic power plants based on sub-meter level images, wherein step 5 includes the following steps.

[0134] 5.1 Fault Confirmation: Combining thermal infrared measurement data with maintenance history records, the hidden faults identified by AI are confirmed to eliminate false detections caused by environmental interference and ensure that the fault detection accuracy rate is ≥90%.

[0135] 5.2 Fault Classification Refinement: Based on photovoltaic panel temperature deviation, power generation data, and module service life, refine the latent fault classification standards, clarify the handling requirements for each level of fault, and mark the priority of operation and maintenance handling, such as emergency, routine, and postponement.

[0136] 5.3 Fault Root Cause Analysis: Combining terrain, meteorological, and historical operation and maintenance data, analyze the causes of hidden faults, such as hot spots caused by component damage, obstruction, or line faults, locate the root cause of the fault, and generate a fault root cause report.

[0137] 5.4 Trend Analysis: By combining time-series image data, analyze the development trend of hidden faults, predict the risk of fault escalation, and if there is an escalation risk, trigger an operation and maintenance early warning and push the early warning information to the cross-module collaborative control module.

[0138] Example 5.

[0139] This embodiment discloses an AI recognition and vectorization method for photovoltaic power plants based on sub-meter level images, wherein step 6 includes the following steps.

[0140] 6.1 Coordinate Extraction and Preprocessing: Extract pixel coordinates of photovoltaic power stations, visible abnormal areas, and hidden fault areas, remove duplicates and smooth out noise; set up a special sub-step for extracting coordinates of hidden fault areas to ensure accurate coordinates of fault areas.

[0141] 6.2 Coordinate Transformation: An RPC and quadratic polynomial fusion algorithm is used to construct a transformation model from pixel coordinates to the CGCS2000 coordinate system. The elevation compensation term and atmospheric refraction correction parameters are optimized for mountainous steep slopes, snow accumulation, and dust scenes to reduce the deviation caused by scene distortion.

[0142] The coordinate transformation model is as follows.

[0143] .

[0144] .

[0145] In the formula, (X,Y) are the final corrected geographic coordinates of the photovoltaic panel, the core output parameters, adapted to the CGCS2000 coordinate system, used for subsequent vectorization, fault area location and GIS applications, to ensure that the coordinate accuracy meets the first-level application standard; 0.7 and 0.3 are the fusion weights of the RPC model and the quadratic polynomial fitting algorithm, which take into account the global accuracy of the RPC model and the local correction advantages of the quadratic polynomial, and optimize the coordinate deviation in the mountainous steep slope scene; , The coordinates for the RPC rational polynomial coefficient model are calculated from the RPC parameters in the satellite imagery metadata and serve as the basic input for coordinate transformation, representing the initial geographic coordinates of the photovoltaic panel; , The coordinates are transformed by a quadratic polynomial, calculated using the quadratic polynomial formula of the geometric correction unit mentioned earlier, and are used to correct the coordinate deviation of the RPC model in local areas; H is the terrain elevation value, provided by the terrain data from the multimodal data fusion module, representing the altitude of the area where the photovoltaic panel is located, and is used to optimize the elevation compensation term; α is the terrain slope of the target area, calculated from the terrain data, and is mainly adapted to mountainous steep slopes ≥30°. This is a slope-elevation co-correction factor, dynamically determined based on the slope α. When α ≥ 30°... =0.15, α<30° =0.08, used to collaboratively correct coordinate offsets caused by both slope and elevation. This is a multi-source data collaborative calibration item, combining BeiDou positioning data and terrain data. The calibration deviation for BeiDou data is calculated from the difference between the BeiDou positioning data and the initial coordinates, representing the amount of correction the BeiDou data makes to the coordinates. The calibration deviation for terrain data is calculated from the elevation correlation deviation between terrain data and preliminary coordinates, representing the correction amount of terrain data to coordinates. 0.4 and 0.2 are the calibration weights of BeiDou data and terrain data, adapting to the high precision advantage of BeiDou data.

[0146] A coordinate transformation adaptation algorithm for snow and dust scenarios is adopted, and the coordinate deviation caused by atmospheric refraction is corrected by combining meteorological data. The optimized atmospheric refraction correction amount is shown in the model.

[0147] .

[0148] In the formula, the meanings of each symbol are as follows: The value represents the correction for the X-direction coordinate deviation caused by atmospheric refraction, used to offset the interference of atmospheric refraction on the X-direction coordinate in snow and dust scenes. The correction effect is directly reflected in the final (X,Y) coordinates. 0.00128 is the atmospheric refraction correction coefficient, adapted to the coordinate accuracy requirements of sub-meter level imagery. P is the atmospheric pressure, provided by meteorological data from the multimodal data fusion module, representing the atmospheric pressure state of the current scene. T is the atmospheric temperature, provided by meteorological data from the multimodal data fusion module, representing the atmospheric temperature state of the current scene. X is the X-direction geographic coordinate of the photovoltaic panel after preliminary correction, i.e., the X value before adding ΔX correction in the fused coordinate transformation model, serving as the calculation benchmark for atmospheric refraction deviation. For meteorological environment correction factors, snow scene =1.1, Dust Scene =1.05, typical scenario =1.0, used for differential correction of atmospheric refraction deviation under different meteorological conditions.

[0149] 6.3 Triple calibration: Perform conversion calibration, overlapping area cross calibration, and GCP sampling calibration. Set up a multimodal data-assisted calibration sub-step and use BeiDou positioning data and terrain data to perform secondary correction on the calibration results to ensure that the coordinate accuracy meets the standards.

[0150] 6.4 Multi-source fusion: When GCP is insufficient, Beidou data and terrain data are called to supplement calibration, and weights are allocated according to the accuracy of the data source. When the data source is missing, the weights are dynamically adjusted.

[0151] 6.5 Deviation Warning and Source Tracing: When the coordinate deviation exceeds the standard, it automatically returns to the geometric correction stage for reprocessing, forming a closed loop; the source of the deviation is located and fed back to the preprocessing module to optimize the parameters.

[0152] Example 6.

[0153] This embodiment discloses an AI recognition and vectorization method for photovoltaic power plants based on sub-meter level images, wherein step 7 includes the following steps.

[0154] 7.1 Vectorization Processing: The Douglas-Peucker algorithm is adopted to dynamically adjust the threshold according to the scene, extract the contours of photovoltaic power plants, obvious abnormal areas, and hidden fault areas, and retain key inflection points; a special optimization sub-step for the contour of hidden fault areas is added, which combines hot spot features to ensure the accuracy of fault area contours.

[0155] 7.2 Boundary Smoothing: Gaussian filtering is used to eliminate jagged boundaries and control contour deviation within the standard range; contour breakage problems in extreme scenarios are repaired to ensure contour integrity ≥99.5%.

[0156] 7.3 Attribute Association: Automatically calculate and classify power plant parameters, and associate multiple attributes such as health status, coordinate accuracy, material, fault type, fault level, and maintenance priority with vector contours.

[0157] 7.4 Format Conversion and Batch Output: Converts to mainstream formats such as SHP and GeoJSON, supports batch export and unified naming, and is suitable for large-scale censuses; a new format adaptive conversion sub-step is added to automatically adjust the vector format according to the needs of the operation and maintenance system, improving compatibility.

[0158] 7.5 Operation and Maintenance Adaptation and Extreme Optimization: Import into the operation and maintenance system through standardized interfaces to fix the vector contour breakage problem in extreme scenarios; add special markers for vector data of hidden fault areas to facilitate operation and maintenance personnel to quickly locate fault locations.

[0159] Example 7.

[0160] This embodiment discloses an AI recognition and vectorization method for photovoltaic power plants based on sub-meter level images, wherein step 8 includes the following steps.

[0161] 8.1 Verification Data Acquisition: GPS RTK and UAV aerial photography are fused to acquire verification baseline data, and aerial photography verification and field measurement verification are enhanced in extreme scenarios; thermal infrared verification data (measured by an infrared thermometer) is used to verify the accuracy of hidden fault assessment; multimodal fusion accuracy verification data is used to verify the fusion effect.

[0162] 8.2 Dual Error Statistics: Sampling statistics are performed on boundary error, floor area error, anomaly identification accuracy and hidden fault judgment accuracy by scenario to establish a quantitative correlation between error and parameter optimization.

[0163] 8.3. Conformity Assessment and Report Generation: Automatically assesses data conformity and generates special verification reports for routine / extreme scenarios / latent faults, including error details, conformity conclusions, and fault verification results.

[0164] 8.4 Error Source Tracing and Feedback: Quantify the error contribution of each module, determine the responsible module and trigger automatic optimization; through the accuracy feedback of the multimodal fusion module, when the fusion quality does not meet the standard, trigger the fusion algorithm optimization to form a closed loop verification.

[0165] Example 8.

[0166] This embodiment discloses an AI recognition and vectorization method for photovoltaic power plants based on sub-meter level images, wherein step 9 includes the following steps.

[0167] 9.1 Dynamic Collaboration: Integrates data from all modules, automatically adjusts core parameters of each module, including multimodal fusion weights, to ensure a balance between accuracy and efficiency throughout the process; enables real-time synchronization of data from each module, improving collaborative efficiency.

[0168] 9.2 Error Blocking: Monitor the errors of each module, block the propagation of errors and trigger optimization when they exceed the standard to avoid accumulation; adopt differentiated optimization strategies according to the severity of the error.

[0169] 9.3 Scene Adaptation: Based on the scene type, automatically call the adaptation parameters of each module, and execute the weight fusion scheme for multiple extreme superimposed scenes; continuously optimize the scene judgment accuracy through scene self-learning.

[0170] 9.4 Time-series monitoring: Combining multi-period and multi-modal images, the system can accurately detect, quantitatively analyze, and provide early warnings for changes in photovoltaic power plants, push early warning information to the operation and maintenance system, and update the historical database to form a closed loop; it can also perform time-series change analysis of latent faults and track the development trend of faults.

[0171] 9.5 Closed-loop Iteration of Operation and Maintenance: Feedback the maintenance records and fault confirmation results of the operation and maintenance system to the sample library and AI recognition model to optimize the parameters of the sample library and model; Feedback the accuracy problems found in the operation and maintenance process to each module to trigger parameter optimization, realize the closed-loop iteration of recognition-operation and maintenance-optimization-re-recognition, and continuously improve the system performance.

[0172] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for AI recognition and vectorization of photovoltaic power plants based on sub-meter level images, characterized in that, include: Acquire multimodal image data of the target area and perform preprocessing; Multimodal fusion is performed on the preprocessed data to complete spatiotemporal alignment, feature fusion, and quality verification, resulting in a fused feature map. Based on the fused feature map, photovoltaic panel anomalies and faults are identified and classified, pixel-level segmentation masks and feedback errors are generated, and anomaly identification results are obtained. By combining the anomaly identification results, fused feature maps, and preprocessed multimodal data, the system completes the confirmation, classification, source tracing, and trend analysis of latent faults, marks the basic contours and fault contours, and issues early warnings. Based on the segmentation mask, fault contour, and fused feature map, coordinate calibration is performed in conjunction with the spatial features in the fused data to obtain calibration coordinates; based on the calibration coordinates, anomaly identification results, basic contours, and fault contours, corresponding contours and associated attributes are extracted and converted into standard vector format.

2. The method for AI recognition and vectorization of photovoltaic power plants based on sub-meter level images according to claim 1, characterized in that... Image data preprocessing includes the following steps: Radiation correction: A scene-adaptive 6S model is used for radiation correction. The atmospheric model and aerosol type are matched with terrain and meteorological data. The reflectivity is normalized according to the photovoltaic panel material, and the temperature is normalized for thermal infrared images. Automatic fallback rules are executed when data is missing. Geometric correction: A rational polynomial coefficient model and a quadratic polynomial fitting fusion algorithm are used for geometric correction, and distortion optimization correction is performed for extreme scenes; Noise Removal: A filtering algorithm is used for noise removal, preserving the texture details of the photovoltaic panel and enhancing the hot spot characteristics; Extreme scenario adaptation: In snow-covered scenarios, snow and photovoltaic panels are distinguished; in industrial dust scenarios, dust interference is weakened; in scenarios with multiple extremes, a weighted fusion strategy is adopted; and hidden fault characteristics are enhanced through spectral enhancement and temperature threshold segmentation. The preprocessing quality indicators are fed back to the subsequent identification and fault assessment steps, and the preprocessing parameters are automatically optimized based on the identification error and fault assessment error to form a closed loop.

3. The method for AI recognition and vectorization of photovoltaic power plants based on sub-meter level images according to claim 2, characterized in that... Multimodal fusion includes the following steps: Data alignment: A spatiotemporal alignment algorithm is used to achieve pixel-level alignment of multiple types of images and spatiotemporal correlation of multimodal data; Feature fusion: A convolutional neural network and a converter fusion algorithm are used to extract features from each modality and fuse them deeply to generate a multimodal fusion feature map; Dynamic weight allocation: The fusion weights of each modality are dynamically allocated according to the scenario type and data quality, and the weight allocation standards for extreme and normal scenarios are clearly defined. When the data source is missing, the rules are automatically adjusted to ensure fusion accuracy. Fusion quality verification: Set fusion quality evaluation indicators and verify them in real time. If the indicators are not met, the fusion algorithm and weights will be automatically adjusted. If the indicators are qualified, the subsequent identification and fault judgment steps will be carried out.

4. The method for AI recognition and vectorization of photovoltaic power plants based on sub-meter level images according to claim 1, characterized in that... Identifying photovoltaic panel anomalies includes the following steps: Model adaptation: Automatically call the corresponding inference branch according to the scenario type, and improve the detection accuracy of hidden faults through the hidden fault identification branch; Batch inference: Adjust batch size based on image resolution; Anomaly classification and judgment: Identify latent faults based on fused features, classify them according to corresponding standards, and mark relevant information; Error feedback: The segmentation threshold, recognition accuracy, and judgment accuracy are fed back to the preprocessing step and the multimodal fusion step. When the error exceeds the standard, the parameters are optimized to form a closed loop.

5. The method for AI recognition and vectorization of photovoltaic power plants based on sub-meter level images according to claim 4, characterized in that... The identification, classification, source tracing, and trend analysis of latent faults include the following steps: Fault confirmation: By combining thermal infrared measurement data with operation and maintenance history records, the identified hidden faults are confirmed to eliminate false detections caused by environmental interference; Fault classification is refined: Based on the temperature deviation of photovoltaic panels, power generation data and the service life of the components, the classification standards for latent faults are refined, the handling requirements for each level of fault are clarified, and the priority of operation and maintenance is marked. Fault tracing: By combining terrain, meteorological, and historical operation and maintenance data, analyze the causes of hidden faults and generate a fault tracing report; Trend analysis: By combining time-series image data, we analyze the development trend of latent faults and predict the risk of fault escalation.

6. The method for AI recognition and vectorization of photovoltaic power plants based on sub-meter level images according to claim 2, characterized in that... The calibration of coordinates includes the following steps: Coordinate extraction and preprocessing: Extract pixel coordinates of photovoltaic power plants, obvious abnormal areas, and hidden fault areas, and remove noise after deduplication and smoothing; Coordinate transformation: A transformation model from pixel coordinates to the standard coordinate system is constructed by using a rational polynomial coefficient and quadratic polynomial fusion algorithm. Triple calibration: Perform conversion calibration, overlapping area cross calibration, and sampling calibration, and use BeiDou positioning data and terrain data to make secondary corrections to the calibration results; Multi-source fusion: When there are insufficient ground control points, BeiDou data and terrain data are called up to supplement calibration, and weights are allocated according to the accuracy of the data source. When the data source is missing, the weights are dynamically adjusted. Deviation warning and source tracing: When the coordinate deviation exceeds the standard, it automatically returns to the geometric correction stage for reprocessing, forming a closed loop; and locates the source of the deviation.

7. The method for AI recognition and vectorization of photovoltaic power plants based on sub-meter level images according to claim 2, characterized in that... The photovoltaic panel reflectivity correction model is as follows: ； In the formula, The reflectivity of the photovoltaic panel after correction; The original reflectivity of the photovoltaic panel; τ represents atmospheric path radiation; τ represents atmospheric transmittance. The solar zenith angle; To observe the zenith angle; The atmospheric hemispherical albedo; This is a correction factor for the photovoltaic panel material. This is a correction term for the reflectivity of latent fault characteristics.

8. The method for AI recognition and vectorization of photovoltaic power plants based on sub-meter level images according to claim 2, characterized in that... The geometric correction model is as follows: ； ； In the formula, X and Y are the corrected geographic coordinates of the photovoltaic panel; x and y are the pixel coordinates of the photovoltaic panel in the original sub-meter level image. , , , , , The X-axis coordinate transformation coefficient; , , , , , This refers to the Y-axis coordinate transformation coefficient. α is the distortion correction factor for extreme scenes, and α is the slope of the target area terrain.

9. The method for AI recognition and vectorization of photovoltaic power plants based on sub-meter level images according to claim 1, characterized in that... Converting to a standard vector format involves the following steps: Vectorization processing: Dynamically adjust thresholds according to the scenario to extract the outlines of photovoltaic power stations, abnormal areas and fault areas, retain key inflection points and optimize the outline accuracy of hidden fault areas; Boundary smoothing: Eliminates jagged boundaries and repairs broken contours, ensuring contour integrity; Attribute association: Calculate and classify power plant parameters, and bind various related attributes to vector profiles; Convert the vector outlines with associated attributes to a standard vector format and perform batch output; Operation and maintenance adaptation and extreme optimization: Import into the operation and maintenance system through standardized interfaces to fix contour problems in extreme scenarios and specifically mark hidden fault areas.

10. A photovoltaic power plant AI recognition and vectorization system employing the sub-meter level image-based AI recognition and vectorization method for photovoltaic power plants as described in any one of claims 1-9, comprising a data acquisition module, an image preprocessing module, an AI recognition model module, a multi-modal data fusion module, a coordinate matching module, a vector data generation module, a latent fault assessment module, an accuracy verification module, a cross-module collaborative control module, and an operation and maintenance data iteration module, characterized in that... : Data acquisition module: Acquires raw sub-meter level satellite imagery data, including optical satellite imagery, multispectral imagery, and thermal infrared imagery; Image preprocessing module: Performs fully automated preprocessing on raw sub-meter resolution satellite imagery; AI Recognition Model Module: Constructs a scene-adaptive recognition model, integrates a converter and an attention mechanism, and achieves integrated recognition, graded judgment of explicit anomalies, and detection of hidden faults; Multimodal data fusion module: integrates optical imagery, multispectral imagery, thermal infrared imagery, BeiDou positioning data, topographic and meteorological data, and historical operation and maintenance data to achieve multi-dimensional feature fusion; Coordinate matching module: Combines BeiDou and terrain data from the multimodal data fusion module to construct a multi-source coordinate fusion calibration system; Vector data generation module: Associates the results of hidden fault identification with operation and maintenance data to achieve integrated refinement, standardization and operation and maintenance adaptation of vector data; Latent Fault Assessment Module: Based on the results of multimodal data fusion and the characteristics of latent faults, a system for accurate assessment and tracing of latent faults is constructed to realize the integrated processing of latent faults from detection and classification to tracing. Accuracy verification module: Constructs a full-process dual accuracy verification and error tracing system, and adopts multi-modal fusion accuracy verification and latent fault judgment accuracy verification; Cross-module collaborative control module: enables dynamic collaboration among modules, error propagation control, scenario adaptive adaptation, closed-loop operation and maintenance iteration, and time-series early warning; Operations and maintenance data iteration module: collects, organizes, analyzes, and provides feedback on operations and maintenance data.

Images