Marine photovoltaic panel defect evaluation method based on multi-modal fusion and spatial reasoning

Abstract

The application discloses a kind of offshore photovoltaic panel defect evaluation method and related device based on multi-modal fusion and space reasoning, wherein the method comprises: the visible light image sequence and infrared thermal imaging sequence of offshore photovoltaic array are collected;Store the GPS coordinates and unmanned aerial vehicle attitude angle of acquisition time;Multiple modal image pair is input to double-flow shared coding network, and the hole and crack class defect candidate region and thermal spot abnormal region are output;According to defect candidate region and thermal spot abnormal region, the temperature gradient statistics of each visible light defect region and its corresponding infrared region are calculated;According to defect propagation probability model, the comprehensive health index of each photovoltaic panel is calculated, and according to the comprehensive health index of each photovoltaic panel, all photovoltaic panels are graded and defect thermodynamic diagram and priority repair list are output.The application establishes defect propagation probability model through photovoltaic array topological graph, simulates the diffusion risk of defect in space and time, realizes from single board detection to overall evaluation.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic panel defect detection technology, specifically to a method, apparatus, and computing device for evaluating defects in marine photovoltaic panels based on multimodal fusion and spatial reasoning. Background Technology

[0002] Offshore photovoltaic systems operate in harsh marine environments characterized by high humidity, high salt spray, strong winds and waves, and intense ultraviolet radiation. This makes the photovoltaic panels susceptible to various defects such as cracks, corrosion, and hot spots, severely impacting power generation efficiency and system lifespan. Failure to detect and repair these defects in a timely manner can lead to the spread of localized faults, causing large-scale system failures and resulting in significant economic losses and energy waste.

[0003] Currently, the main technology for detecting defects in photovoltaic panels utilizes drones equipped with visible light cameras to acquire images of the panels. These images are then processed using image processing or deep learning algorithms (such as YOLO and Faster R-CNN) to identify surface cracks, stains, and damage. However, this approach is sensitive to lighting conditions; strong reflections at sea and fog interference can easily lead to false positives or false negatives. It cannot detect internal defects (such as cell cracks or solder joint failures) or thermal anomalies (such as hot spots). Furthermore, the lack of quantitative assessment of defect severity makes it difficult to guide maintenance prioritization. Typically, photovoltaic panels are treated as independent units, neglecting spatial adjacency, electrical series / parallel relationships, and the coupling effects of corrosion and vibration in marine environments. This results in fragmented assessments that are difficult to support systematic maintenance decisions.

[0004] To address the aforementioned issues, this invention proposes a defect assessment method for marine photovoltaic panels based on multimodal fusion and spatial reasoning. By establishing a defect propagation probability model through a photovoltaic array topology diagram, the method simulates the spatial and temporal spread risk of defects, thereby achieving a comprehensive assessment from single-panel detection. Summary of the Invention

[0005] In view of the above problems, the present invention provides a method, apparatus and computing device for defect assessment of marine photovoltaic panels based on multimodal fusion and spatial reasoning.

[0006] According to one aspect of the present invention, a method for defect assessment of marine photovoltaic panels based on multimodal fusion and spatial reasoning is provided, comprising: The visible light camera and infrared thermal imager installed on the inspection drone simultaneously acquire visible light image sequences and infrared thermal imaging sequences of the marine photovoltaic array; and store the GPS coordinates and drone attitude angles at the time of acquisition. Based on the GPS coordinates and the UAV attitude angle, the visible light image sequence and infrared thermal imaging sequence acquired at the same time are registered to obtain a multimodal image pair; the multimodal image pair is input into a dual-stream shared coding network to output candidate regions for holes and cracks, as well as hot spot anomaly regions; Based on the candidate defect regions and hot spot anomaly regions, calculate the temperature gradient statistics of each visible light defect region and its corresponding infrared region; establish the mapping relationship between the local maximum temperature difference and the hot spot area ratio based on the prior constraint regression head to obtain the preliminary defect score of each photovoltaic panel unit; wherein, the temperature gradient statistics include the local maximum temperature difference and the hot spot area ratio. A defect propagation probability model is constructed based on the preliminary defect score of each photovoltaic panel unit and the photovoltaic array topology diagram. The comprehensive health index of each photovoltaic panel is calculated based on the defect propagation probability model. All photovoltaic panels are classified according to the comprehensive health index of each photovoltaic panel and a defect heat map and priority repair list are output.

[0007] In an optional approach, calculating the temperature gradient statistics of each visible light defect region and its corresponding infrared region based on the defect candidate regions and hot spot anomaly regions further includes: For any visible light defect candidate region, an infrared matching sub-region with the visible light defect region is obtained on the corresponding infrared thermal image using an optical flow tracing algorithm; an adjacent annular buffer region without hot spots is selected at the outer edge of the infrared matching sub-region, and the weighted average temperature of all pixels in the annular buffer region is calculated as the local background reference temperature. The absolute temperature difference value of each pixel in the infrared matching sub-region is calculated based on the local background reference temperature to obtain the local absolute temperature difference field matrix; temperature gradient statistics are extracted based on the local absolute temperature difference field matrix; wherein, the temperature gradient statistics include the local maximum temperature difference, the hot spot area ratio, and the temperature difference gradient distribution entropy.

[0008] In one alternative approach, the formula for calculating the preliminary defect score is: ; in: ; ; ; in, This refers to the number of the photovoltaic panel unit; These represent the local maximum temperature difference and the percentage of hot spot area, respectively. This represents the threshold for the proportion of hot spot area. This is the threshold for the local maximum temperature difference; , , These are the shape parameters determined by the material fatigue model, respectively; and are the shape parameters determined based on the material fatigue model, respectively. The entropy of the temperature gradient distribution; , , To adjust the parameters; The comprehensive environmental aging coefficient; This is the equivalent corrosion activation energy; It is the gas constant; Standard reference temperature; The ambient temperature; The coupling coefficient; This represents the local maximum temperature difference corresponding to the i-th photovoltaic panel unit; This represents the percentage of the hot spot area corresponding to the i-th photovoltaic panel unit; The entropy of the temperature gradient distribution corresponding to the i-th photovoltaic panel unit; The confidence score modulation function; This is the Gaussian error function.

[0009] In one alternative approach, constructing a defect propagation probability model based on the preliminary defect score of each photovoltaic panel unit and the photovoltaic array topology further includes: Based on the preliminary defect score and photovoltaic array topology of each photovoltaic panel unit, the K-hop electrical / physical neighborhood set of each photovoltaic panel unit is obtained; the propagation risk factor of each target photovoltaic panel unit based on the K-hop electrical / physical neighborhood state is calculated; The dynamic defect propagation probability of each target photovoltaic panel unit within the time window is calculated based on the propagation risk factor. The comprehensive health index of each target photovoltaic panel unit is obtained by calculating the state transition probability based on the dynamic defect propagation probability and the preliminary defect score using a Markov chain based on the propagation probability of the defect.

[0010] In one alternative approach, the formula for calculating the comprehensive health index is: ; in: ; ; ; ; ; in, Let be the rate intensity of the i-th photovoltaic panel unit transitioning from a healthy state to a latent defect state; It is a vector-valued function composed of RBF kernels; For RBF center; For bandwidth; are basis vectors; For degenerate-activation coupled functionals; For repair-recovery co-functionality; The correlation coefficient; For Minkowski norm parameters; The virtual energy function on the state transition diagram; It is an inverse temperature parameter; This refers to the number of the photovoltaic panel unit; The comprehensive health index of the i-th photovoltaic panel unit; This represents the total number of radial basis functions; For gradient operators; Let be the rate intensity of the i-th photovoltaic panel unit transitioning from a faulty state to a healthy state; Let be the rate intensity of the i-th photovoltaic panel unit transitioning from a latent defect state to a healthy state; is the activation resonance factor of the i-th photovoltaic panel unit.

[0011] In one alternative approach, classifying all photovoltaic panels based on a comprehensive health index for each panel and outputting a defect heatmap further includes: A statistical distribution fit is performed on the set of comprehensive health indices of all photovoltaic panels, and its kernel density estimate is calculated. Multiple grading thresholds are generated based on the inflection point and standard deviation of the kernel density distribution. The comprehensive health index of each photovoltaic panel is mapped to the corresponding health level according to the grading thresholds. At each photovoltaic panel unit location in the topology map, a static thermal layer is generated by mapping its health level to a preset pseudo-color gradient color band, which represents the severity of defects with color depth. Based on the propagation risk factors of each photovoltaic panel unit obtained from the defect propagation probability model, a dynamic overlay layer is generated to characterize the potential propagation risk of defects using dynamic flicker frequency or halo diffusion effect. The static thermal layer is then fused with the dynamic overlay layer to output an interactive defect heat map, which displays the current spatial distribution severity of defects and future risk propagation trends.

[0012] In one alternative approach, the dual-stream shared coding network comprises a shared backbone encoder, a visible light branch decoder, an infrared branch decoder, and a cross-modal attention fusion module. The shared backbone encoder is a Vision Transformer feature extraction network, which contains multiple cascaded Transformer coding layers. Each Transformer coding layer interacts with features through a multi-head self-attention module. Both the visible light branch decoder and the infrared branch decoder are encoder-decoder structures with skip connections; the encoder-decoder structure makes skip connections with shared features from the corresponding level of the encoder through the upsampling layer.

[0013] In one alternative approach, the photovoltaic array topology is a multi-layer graph structure consisting of node layers and edge layers; The node layer's attribute vector includes basic physical attributes, dynamic detection attributes, spatiotemporal location attributes, and historical health status. The basic physical attributes include model, rated power, area, installation tilt angle, and azimuth angle. The spatiotemporal location attributes include geospatial coordinates obtained from GPS coordinates and 3D reconstruction, as well as 2D grid coordinates in the array's logical layout. The historical health status is a historical sequence of the comprehensive health index calculated from each inspection. The edge layer includes physical proximity edges, electrical connection edges, and maintenance reachable edges. The physical proximity edges are automatically generated based on Delaunay triangulation between nodes, used to model the propagation risks of physical defects caused by structural vibration transmission, salt spray water accumulation, and shading. The electrical connection edges are constructed based on the electrical wiring diagram of the photovoltaic array, used to model the propagation risks of electrical defects such as current mismatch and voltage reversal in series branches and fault current backflow in parallel groups. The maintenance reachable edges are generated based on the platform layout, maintenance channels, and obstacle location information of the photovoltaic array, used to model the cost and path of maintenance operations.

[0014] According to another aspect of the present invention, a defect assessment device for marine photovoltaic panels based on multimodal fusion and spatial reasoning is provided, comprising: The data synchronization acquisition module is used to simultaneously acquire visible light image sequences and infrared thermal imaging sequences of the marine photovoltaic array through the visible light camera and infrared thermal imager installed on the inspection drone; and to store the GPS coordinates and drone attitude angle at the time of acquisition. The image registration and dual-stream defect detection module is used to register the visible light image sequence and the infrared thermal imaging sequence acquired at the same time according to the GPS coordinates and the UAV attitude angle to obtain a multimodal image pair; the multimodal image pair is input into the dual-stream shared coding network, and the output is a candidate region for hole and crack type defects and a hot spot anomaly region; The temperature gradient statistics and preliminary scoring module is used to calculate the temperature gradient statistics of each visible light defect region and its corresponding infrared region based on the defect candidate region and the hot spot abnormal region; and to establish a mapping relationship between the local maximum temperature difference and the hot spot area ratio based on the prior constraint regression head to obtain the preliminary defect score of each photovoltaic panel unit; wherein, the temperature gradient statistics include the local maximum temperature difference and the hot spot area ratio. The comprehensive evaluation module is used to construct a defect propagation probability model based on the preliminary defect score of each photovoltaic panel unit and the photovoltaic array topology map, calculate the comprehensive health index of each photovoltaic panel based on the defect propagation probability model, classify all photovoltaic panels based on the comprehensive health index of each photovoltaic panel, and output a defect heat map and priority repair list.

[0015] According to another aspect of the present invention, a computing device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described method for evaluating defects in marine photovoltaic panels based on multimodal fusion and spatial reasoning.

[0016] According to the solution provided by the present invention, a visible light image sequence and an infrared thermal image sequence of a marine photovoltaic array are simultaneously acquired by a visible light camera and an infrared thermal imager installed on an inspection drone; the GPS coordinates and drone attitude angle at the acquisition time are stored; the visible light image sequence and infrared thermal image sequence acquired at the same time are registered according to the GPS coordinates and drone attitude angle to obtain a multimodal image pair; the multimodal image pair is input into a dual-stream shared coding network to output candidate regions for defects such as holes and cracks, as well as hot spot anomaly regions; the temperature gradient statistics of each visible light defect region and its corresponding infrared region are calculated according to the defect candidate regions and hot spot anomaly regions; a mapping relationship between the local maximum temperature difference and the hot spot area ratio is established according to the prior constraint regression head to obtain a preliminary defect score for each photovoltaic panel unit; wherein, the temperature gradient statistics include the local maximum temperature difference and the hot spot area ratio; a defect propagation probability model is constructed according to the preliminary defect score of each photovoltaic panel unit and the photovoltaic array topology map; the comprehensive health index of each photovoltaic panel is calculated according to the defect propagation probability model; all photovoltaic panels are classified according to the comprehensive health index of each photovoltaic panel and a defect heat map and priority maintenance list are output. This invention establishes a defect propagation probability model through a photovoltaic array topology diagram, simulates the risk of defect diffusion in space and time, and realizes the overall evaluation from single-board detection.

[0017] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating the defect assessment method for marine photovoltaic panels based on multimodal fusion and spatial reasoning according to an embodiment of the present invention is shown. Figure 2 A schematic diagram of a defect heat map according to an embodiment of the present invention is shown; Figure 3 A schematic diagram of the photovoltaic array topology according to an embodiment of the present invention is shown; Figure 4 A schematic diagram of the framework of the marine photovoltaic panel defect assessment device based on multimodal fusion and spatial reasoning according to an embodiment of the present invention is shown; Figure 5 A schematic diagram of the structure of a computing device according to an embodiment of the present invention is shown. Detailed Implementation

[0019] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.

[0020] Figure 1 A flowchart illustrating a method for assessing defects in offshore photovoltaic panels based on multimodal fusion and spatial reasoning, according to an embodiment of the present invention, is shown. Specifically, as... Figure 1 As shown, it includes the following steps: Step S101: The visible light camera and infrared thermal imager installed on the inspection drone simultaneously acquire visible light image sequences and infrared thermal imaging sequences of the marine photovoltaic array; and store the GPS coordinates and drone attitude angles at the acquisition time.

[0021] In this embodiment, synchronous acquisition ensures that the visible light image and infrared thermal image of the same physical point are completely consistent in time, avoiding registration errors introduced by changes in lighting, cloud cover, and slight equipment displacement caused by asynchronous shooting. GPS coordinates and UAV attitude angles (pitch, roll, yaw) provide accurate exterior orientation elements for multimodal image pairs, allowing for high-precision image registration based on geographic coordinates and spatial geometry in complex 3D marine scenes. This overcomes the problem that relying solely on image feature matching is prone to failure in low-texture (such as large photovoltaic panels) or repetitive texture areas, ensuring accurate spatial correlation between infrared hot spots and visible light defect areas. Acquired data with spatiotemporal stamps (GPS coordinates, timestamps) can be correlated with historical inspection data. By comparing the health status of the same geographical location (photovoltaic panel unit) at different times, the evolution trend of defects (such as crack propagation and hot spot aggravation) can be analyzed, providing data support for predictive maintenance.

[0022] Step S102: Register the visible light image sequence and infrared thermal imaging sequence acquired at the same time according to the GPS coordinates and the UAV attitude angle to obtain a multimodal image pair; input the multimodal image pair into a dual-stream shared coding network to output candidate regions for holes and cracks and hot spot anomaly regions.

[0023] In this embodiment, as Figure 2 As shown, the visible light branch accurately identifies surface damage such as cracks, holes, and spots, while the infrared branch detects hot spot anomalies caused by electrical faults or physical damage. This can effectively reduce false alarms (such as misjudging shadows as cracks) and missed alarms (such as internal damage not being manifested on the surface) in single-mode detection.

[0024] In one alternative approach, the dual-stream shared coding network comprises a shared backbone encoder, a visible light branch decoder, an infrared branch decoder, and a cross-modal attention fusion module. The shared backbone encoder is a Vision Transformer feature extraction network, which contains multiple cascaded Transformer coding layers. Each Transformer coding layer interacts with features through a multi-head self-attention module. Both the visible light branch decoder and the infrared branch decoder are encoder-decoder structures with skip connections; the encoder-decoder structure makes skip connections with shared features from the corresponding level of the encoder through the upsampling layer.

[0025] In this embodiment, the shared backbone encoder extracts common low-level features (such as edges and structural information) from visible and infrared images in a unified manner, improving model inference efficiency. Independent branch decoders ensure the complete preservation of high-level, specific features (texture details in visible light, thermal distribution patterns in infrared) for each modality. The Vision Transformer's self-attention mechanism can model long-distance dependencies and handle complex scenes such as sea surface illumination reflection and fog occlusion.

[0026] Step S103: Calculate the temperature gradient statistics of each visible light defect region and its corresponding infrared region based on the defect candidate region and the hot spot abnormal region; establish the mapping relationship between the local maximum temperature difference and the hot spot area ratio based on the prior constraint regression head to obtain the preliminary defect score of each photovoltaic panel unit; wherein, the temperature gradient statistics include the local maximum temperature difference and the hot spot area ratio.

[0027] In this embodiment, the local maximum temperature difference reflects the intensity of the hot spot, which is directly related to the severity of electrical faults and power loss. The proportion of the hot spot area characterizes the influence range of the defect; the larger the area, the wider the impact on the photovoltaic panel's power generation performance. The prior constraint regression head encodes material and thermal knowledge into the scoring model, establishing a mapping relationship that conforms to the failure mechanism of photovoltaic modules, making the scoring results more engineering interpretable and avoiding scoring results that violate physical laws that may be generated by purely data-driven models. The scoring model considers the impact of high humidity and high salt spray environments at sea on the aging rate of photovoltaic modules. Through parameter adjustment, it can adapt to changes in environmental conditions in different sea areas and seasons, solving the problem of poor robustness of traditional threshold methods under changes in sea surface reflection and temperature and humidity.

[0028] In an optional approach, calculating the temperature gradient statistics of each visible light defect region and its corresponding infrared region based on the defect candidate regions and hot spot anomaly regions further includes: For any visible light defect candidate region, an infrared matching sub-region with the visible light defect region is obtained on the corresponding infrared thermal image using an optical flow tracing algorithm; an adjacent annular buffer region without hot spots is selected at the outer edge of the infrared matching sub-region, and the weighted average temperature of all pixels in the annular buffer region is calculated as the local background reference temperature. The absolute temperature difference value of each pixel in the infrared matching sub-region is calculated based on the local background reference temperature to obtain the local absolute temperature difference field matrix; temperature gradient statistics are extracted based on the local absolute temperature difference field matrix; wherein, the temperature gradient statistics include the local maximum temperature difference, the hot spot area ratio, and the temperature difference gradient distribution entropy.

[0029] In this embodiment, the optical flow tracing algorithm can accurately capture the pixel-level correspondence between visible light defect areas and infrared images, overcoming the matching errors of traditional geometric transformation-based registration under local deformation and viewing angle changes. It is particularly suitable for the minute deformations and displacements that may occur in offshore photovoltaic panels in wind and waves. The temperature gradient distribution entropy quantifies the complexity and diffusion trend of hot spot boundaries. It can distinguish different types of hot spots: low entropy values ​​indicate well-defined focal hot spots (such as cell ruptures), while high entropy values ​​indicate diffuse hot spots with blurred boundaries (such as poor solder joints).

[0030] In one alternative approach, the formula for calculating the preliminary defect score is: ; in: ; ; ; in, This refers to the number of the photovoltaic panel unit; These represent the local maximum temperature difference and the percentage of hot spot area, respectively. This represents the threshold for the proportion of hot spot area. This is the threshold for the local maximum temperature difference; , , These are the shape parameters determined by the material fatigue model, respectively; and are the shape parameters determined based on the material fatigue model, respectively. The entropy of the temperature gradient distribution; , , To adjust the parameters; The comprehensive environmental aging coefficient; This is the equivalent corrosion activation energy; It is the gas constant; Standard reference temperature; The ambient temperature; The coupling coefficient; This represents the local maximum temperature difference corresponding to the i-th photovoltaic panel unit; This represents the percentage of the hot spot area corresponding to the i-th photovoltaic panel unit; The entropy of the temperature gradient distribution corresponding to the i-th photovoltaic panel unit; The confidence score modulation function; This is the Gaussian error function.

[0031] In this embodiment, the physical interpretability of the model is enhanced by using material fatigue and environmental aging parameters, such as... , , Derived from material fatigue models, The accelerated aging effects of salt spray, humidity, and temperature cycling in marine environments on components were explicitly modeled (using an Arrhenius-type corrosion kinetic model), so that the score not only reflects the current state but also incorporates the service life degradation law. The exponential-error function composite form is adopted, which provides a smooth response to small temperature difference / small area defects (avoiding false alarms) and a steep response to large temperature difference / large area hot spots (highlighting serious defects). By incorporating the degree of disorder (entropy) in the temperature difference distribution into the score using a convolutional integral operator, high-risk hotspots with blurred edges and strong diffusion are identified; finally, a confidence modulation function is used. The reliability of the test results is weighted to suppress abnormal scores caused by low-quality images or registration errors. Ambient temperature and comprehensive aging coefficient make the same temperature rise comparable under different seasons and service years, solving the assessment drift problem caused by drastic temperature changes in the marine environment.

[0032] Step S104: Construct a defect propagation probability model based on the preliminary defect score of each photovoltaic panel unit and the photovoltaic array topology map; calculate the comprehensive health index of each photovoltaic panel based on the defect propagation probability model; classify all photovoltaic panels based on the comprehensive health index of each photovoltaic panel and output a defect heat map and priority repair list.

[0033] In this embodiment, single-board defects are analyzed within the entire system structure using a photovoltaic array topology diagram. This identifies cascading risks (such as hot spot propagation, current mismatch, and corrosion spread) caused by neighboring board failures, significantly improving the globality and foresight of the assessment. The topology diagram includes three types of associations: physical proximity edges, electrical connection edges, and maintenance reachable edges, corresponding to real-world factors such as salt spray / water accumulation / vibration conduction, series and parallel circuit fault propagation, and maintenance intervention difficulty. The defect propagation probability model differentiates the risk weights of different propagation channels accordingly, ensuring that the health index not only reflects the current state but also predicts future degradation trends. Employing K-hop neighborhood propagation risk factors and Markov chain state transitions, the model can dynamically update the health status of each board within a time window, adapting to the high variability of the marine environment (such as rapid spread of localized damage after a typhoon). The comprehensive health index is statistically graded to generate an interactive defect heatmap, combining static severity and dynamic propagation risk to intuitively present where the board is damaged and where it is likely to fail next. Simultaneously, a priority maintenance list is automatically generated, sorted by risk level, optimizing the scheduling efficiency of limited marine maintenance resources.

[0034] In one alternative approach, constructing a defect propagation probability model based on the preliminary defect score of each photovoltaic panel unit and the photovoltaic array topology further includes: Based on the preliminary defect score and photovoltaic array topology of each photovoltaic panel unit, the K-hop electrical / physical neighborhood set of each photovoltaic panel unit is obtained; the propagation risk factor of each target photovoltaic panel unit based on the K-hop electrical / physical neighborhood state is calculated; The dynamic defect propagation probability of each target photovoltaic panel unit within the time window is calculated based on the propagation risk factor. The comprehensive health index of each target photovoltaic panel unit is obtained by calculating the state transition probability based on the dynamic defect propagation probability and the preliminary defect score using a Markov chain based on the propagation probability of the defect.

[0035] In this embodiment, by setting the dynamic propagation probability within a time window, the potential fault evolution path over the next few hours to days can be simulated, reserving response time for operation and maintenance. The initial defect score (current state) and neighborhood propagation risk (external driver) are used as inputs. Markov chain state transitions are used to mathematically model the transition probability of the photovoltaic panel between "health → latent defect → fault," ensuring that the comprehensive health index not only reflects the current state but also implies future degradation trends, possessing strong interpretability and predictability. For example, the topology graph shows that there is an electrical connection edge (same string) and a physical proximity edge (spacing < 1m) between A and B; assuming K=1, then B's 1-hop neighborhood includes A; the propagation risk factor R of B is calculated. B Because A has a high score and high edge weight, R B A significant increase (e.g., 0.65); within the next 48-hour time window, the model predicts that B will experience a localized temperature rise due to the hotspot of A, accelerating EVA aging, with a dynamic propagation probability of 0.4; Markov chain synthesis S B =0.15 and P prop,B =0.4, thus B's comprehensive health index H is calculated. B =0.68, falling into the warning level; in the final maintenance list, A is for emergency maintenance and B is for preventive inspection. They should be handled simultaneously to prevent B from failing prematurely due to the long-term influence of A.

[0036] In one alternative approach, the formula for calculating the comprehensive health index is: ; in: ; ; ; ; ; in, Let be the rate intensity of the i-th photovoltaic panel unit transitioning from a healthy state to a latent defect state; It is a vector-valued function composed of RBF kernels; For RBF center; For bandwidth; are basis vectors; For degenerate-activation coupled functionals; For repair-recovery co-functionality; The correlation coefficient; For Minkowski norm parameters; The virtual energy function on the state transition diagram; It is an inverse temperature parameter; This refers to the number of the photovoltaic panel unit; The comprehensive health index of the i-th photovoltaic panel unit; This represents the total number of radial basis functions; For gradient operators; Let be the rate intensity of the i-th photovoltaic panel unit transitioning from a faulty state to a healthy state; Let be the rate intensity of the i-th photovoltaic panel unit transitioning from a latent defect state to a healthy state; is the activation resonance factor of the i-th photovoltaic panel unit.

[0037] In this embodiment, a degradation potential field in three-dimensional space is constructed using a radial basis function (RBF) network, and its gradient norm is used as the local degradation rate intensity. This accurately captures the accelerated degradation effect of high-risk locations such as edge panels (susceptible to wind and wave erosion) and low-lying areas (water accumulation corrosion), which is superior to the globally uniform degradation assumption. Activating the resonance factor can be understood as the resonance sensitivity of the photovoltaic panel under specific environmental excitations (such as periodic wave vibration and temperature cycling), improving the early warning capability for sudden failures. For example, in a marine photovoltaic array, there are two panels: panel X is located at the edge of the platform and is subjected to long-term wave splashing, with an initial score of S... X =0.65; Plate Y: Located in the central area, recently covered by bird droppings, S Y =0.70. Although Y has a more obvious defect, X is judged to be of higher risk due to its high degradation rate, high environmental stimulation, and low maintainability. X is prioritized in the maintenance list to prevent it from failing completely in the short term.

[0038] In one alternative approach, classifying all photovoltaic panels based on a comprehensive health index for each panel and outputting a defect heatmap further includes: A statistical distribution fit is performed on the set of comprehensive health indices of all photovoltaic panels, and its kernel density estimate is calculated. Multiple grading thresholds are generated based on the inflection point and standard deviation of the kernel density distribution. The comprehensive health index of each photovoltaic panel is mapped to the corresponding health level according to the grading thresholds. At each photovoltaic panel unit location in the topology map, a static thermal layer is generated by mapping its health level to a preset pseudo-color gradient color band, which represents the severity of defects with color depth. Based on the propagation risk factors of each photovoltaic panel unit obtained from the defect propagation probability model, a dynamic overlay layer is generated to characterize the potential propagation risk of defects using dynamic flicker frequency or halo diffusion effect. The static thermal layer is then fused with the dynamic overlay layer to output an interactive defect heat map, which displays the current spatial distribution severity of defects and future risk propagation trends.

[0039] In this embodiment, by statistically fitting the comprehensive health index of all photovoltaic panels and determining the grading threshold based on kernel density estimation, the health status of each photovoltaic panel is accurately quantified. This identifies which components are in a critical state but have not yet reached the emergency repair standard, allowing for preventative measures to avoid potential failures. A pseudo-color gradient band is used to map the health level to each photovoltaic panel unit location on the topology map, making defects of different severity levels readily apparent. This not only facilitates quick location of problem areas by maintenance personnel but also helps management make more rational resource allocation decisions. Dynamic overlay layers based on propagation risk factors, using flicker frequency or halo diffusion effects to characterize potential future risk propagation trends, not only alert the maintenance team to areas with high propagation risk even without obvious defects but also support the development of long-term maintenance plans. The interactive defect heatmap integrates static severity information and dynamic propagation trend analysis, allowing for in-depth exploration of historical data for specific photovoltaic panels and the impact of their surrounding environment.

[0040] In one alternative approach, the photovoltaic array topology is a multi-layer graph structure consisting of node layers and edge layers; The node layer's attribute vector includes basic physical attributes, dynamic detection attributes, spatiotemporal location attributes, and historical health status. The basic physical attributes include model, rated power, area, installation tilt angle, and azimuth angle. The spatiotemporal location attributes include geospatial coordinates obtained from GPS coordinates and 3D reconstruction, as well as 2D grid coordinates in the array's logical layout. The historical health status is a historical sequence of the comprehensive health index calculated from each inspection. The edge layer includes physical proximity edges, electrical connection edges, and maintenance reachable edges. The physical proximity edges are automatically generated based on Delaunay triangulation between nodes, used to model the propagation risks of physical defects caused by structural vibration transmission, salt spray water accumulation, and shading. The electrical connection edges are constructed based on the electrical wiring diagram of the photovoltaic array, used to model the propagation risks of electrical defects such as current mismatch and voltage reversal in series branches and fault current backflow in parallel groups. The maintenance reachable edges are generated based on the platform layout, maintenance channels, and obstacle location information of the photovoltaic array, used to model the cost and path of maintenance operations.

[0041] In this embodiment, physical proximity edges capture environmental coupling effects such as salt spray corrosion, water spread, and shading; electrical connection edges reflect electrical risks in series and parallel systems such as current mismatch, hot spot cascading, and reverse voltage breakdown; and maintenance reachability edges quantify maintenance difficulty and cost, enabling health assessment results to be directly linked to actual maintenance scheduling, significantly improving the realism and predictive ability of defect propagation simulation. Physical proximity edges are automatically generated using Delaunay triangulation, automatically adapting to irregular array arrangements (such as avoiding wave impact zones and misaligned installations at platform edges), avoiding the inefficiency of manually defining adjacency relationships, and ensuring a high degree of consistency between the topology and the actual physical layout. Nodes contain a comprehensive health index sequence from each inspection, giving the topology a time dimension, which can be used to analyze defect development trajectories, identify accelerated aging areas, and predict remaining lifespan. Maintenance reachability edges incorporate engineering constraints such as platform structure, channels, and obstacles into the graph model, ensuring that the priority maintenance list is based not only on risk level but also on accessibility and operating costs, avoiding the recommendation of theoretically high-risk but practically inaccessible maintenance tasks, and improving the feasibility of the solution.

[0042] Specifically, such as Figure 3 As shown, each photovoltaic panel is treated as a graph node, and its attribute vector is filled. The basic physical attributes are obtained from the equipment ledger, including model, power, area, tilt angle, and azimuth angle. The spatiotemporal location attributes are calculated by combining GPS coordinates collected by the UAV with the 3D reconstructed point cloud to obtain the geospatial coordinates (latitude and longitude + elevation) and the raster row and column number in the logical array. The dynamic detection attributes are the preliminary defect score and temperature gradient statistics obtained from this inspection. The historical health status is obtained by retrieving the comprehensive health index sequence generated from previous inspections from the database to form a time series feature. Physical proximity edges perform Delaunay triangulation on the geospatial coordinates of all nodes, automatically connecting the nearest neighbor plates to form a non-intersecting, fully covered adjacency network. Electrical connection edges connect plates within series branches sequentially according to the electrical wiring diagram in the power plant design drawings, marking the edge direction and current flow direction, and establishing bidirectional connections within the group. Maintenance reachability edges extract obstacle information such as maintenance passages, guardrails, and equipment compartments based on the platform's BIM model or laser scanning map, and use path planning algorithms (such as A*) to calculate the reachability distance or passage difficulty between any two plates, which serves as the edge weight. Node attributes and the three types of edges are stored as node tables and edge tables in the graph database, respectively. After each inspection, only the dynamic detection attributes and historical health status are updated, while the remaining static attributes remain unchanged.

[0043] According to the solution provided by the present invention, a visible light image sequence and an infrared thermal image sequence of a marine photovoltaic array are simultaneously acquired by a visible light camera and an infrared thermal imager installed on an inspection drone; the GPS coordinates and drone attitude angle at the acquisition time are stored; the visible light image sequence and infrared thermal image sequence acquired at the same time are registered according to the GPS coordinates and drone attitude angle to obtain a multimodal image pair; the multimodal image pair is input into a dual-stream shared coding network to output candidate regions for defects such as holes and cracks, as well as hot spot anomaly regions; the temperature gradient statistics of each visible light defect region and its corresponding infrared region are calculated according to the defect candidate regions and hot spot anomaly regions; a mapping relationship between the local maximum temperature difference and the hot spot area ratio is established according to the prior constraint regression head to obtain a preliminary defect score for each photovoltaic panel unit; wherein, the temperature gradient statistics include the local maximum temperature difference and the hot spot area ratio; a defect propagation probability model is constructed according to the preliminary defect score of each photovoltaic panel unit and the photovoltaic array topology map; the comprehensive health index of each photovoltaic panel is calculated according to the defect propagation probability model; all photovoltaic panels are classified according to the comprehensive health index of each photovoltaic panel and a defect heat map and priority maintenance list are output. This invention establishes a defect propagation probability model through a photovoltaic array topology diagram, simulates the risk of defect diffusion in space and time, and realizes the overall evaluation from single-board detection.

[0044] Figure 4 A schematic diagram of the framework of a marine photovoltaic panel defect assessment device based on multimodal fusion and spatial reasoning according to an embodiment of the present invention is shown. The marine photovoltaic panel defect assessment device based on multimodal fusion and spatial reasoning includes: The data synchronization acquisition module 410 is used to simultaneously acquire visible light image sequences and infrared thermal imaging sequences of the marine photovoltaic array through the visible light camera and infrared thermal imager installed on the inspection drone; and to store the GPS coordinates and drone attitude angle at the acquisition time. The image registration and dual-stream defect detection module 420 is used to register the visible light image sequence and the infrared thermal imaging sequence acquired at the same time according to the GPS coordinates and the UAV attitude angle to obtain a multimodal image pair; input the multimodal image pair into the dual-stream shared coding network, and output candidate regions for holes and cracks and hot spot anomaly regions; The temperature gradient statistics and preliminary scoring module 430 is used to calculate the temperature gradient statistics of each visible light defect region and its corresponding infrared region based on the defect candidate region and the hot spot abnormal region; and to establish a mapping relationship between the local maximum temperature difference and the hot spot area ratio based on the prior constraint regression head to obtain the preliminary defect score of each photovoltaic panel unit; wherein, the temperature gradient statistics include the local maximum temperature difference and the hot spot area ratio. The comprehensive evaluation module 440 is used to construct a defect propagation probability model based on the preliminary defect score of each photovoltaic panel unit and the photovoltaic array topology map, calculate the comprehensive health index of each photovoltaic panel based on the defect propagation probability model, classify all photovoltaic panels based on the comprehensive health index of each photovoltaic panel, and output a defect heat map and priority repair list.

[0045] Figure 5 The diagram shows a structural schematic of an embodiment of the computing device of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the computing device.

[0046] like Figure 5 As shown, the computing device may include: a processor 502, a communications interface 504, a memory 506, and a communications bus 508.

[0047] The processor 502, communication interface 504, and memory 506 communicate with each other via communication bus 508. Communication interface 504 is used to communicate with other network elements, such as clients or other servers. Processor 502 executes program 510, specifically performing the relevant steps in the above-described embodiment of the marine photovoltaic panel defect assessment method based on multimodal fusion and spatial reasoning.

[0048] Specifically, program 510 may include program code that includes computer operation instructions.

[0049] Processor 502 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The computing device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0050] Memory 506 is used to store program 510. Memory 506 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0051] According to the solution provided by the present invention, a visible light image sequence and an infrared thermal image sequence of a marine photovoltaic array are simultaneously acquired by a visible light camera and an infrared thermal imager installed on an inspection drone; the GPS coordinates and drone attitude angle at the acquisition time are stored; the visible light image sequence and infrared thermal image sequence acquired at the same time are registered according to the GPS coordinates and drone attitude angle to obtain a multimodal image pair; the multimodal image pair is input into a dual-stream shared coding network to output candidate regions for defects such as holes and cracks, as well as hot spot anomaly regions; the temperature gradient statistics of each visible light defect region and its corresponding infrared region are calculated according to the defect candidate regions and hot spot anomaly regions; a mapping relationship between the local maximum temperature difference and the hot spot area ratio is established according to the prior constraint regression head to obtain a preliminary defect score for each photovoltaic panel unit; wherein, the temperature gradient statistics include the local maximum temperature difference and the hot spot area ratio; a defect propagation probability model is constructed according to the preliminary defect score of each photovoltaic panel unit and the photovoltaic array topology map; the comprehensive health index of each photovoltaic panel is calculated according to the defect propagation probability model; all photovoltaic panels are classified according to the comprehensive health index of each photovoltaic panel and a defect heat map and priority maintenance list are output. This invention establishes a defect propagation probability model through a photovoltaic array topology diagram, simulates the risk of defect diffusion in space and time, and realizes the overall evaluation from single-board detection.

[0052] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination of all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed can be employed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose. Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, in the following claims, any of the claimed embodiments can be used in any combination. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. Unless otherwise specified, the steps in the above embodiments should not be construed as limiting the order of execution.

Claims

1. A method for defect assessment of marine photovoltaic panels based on multimodal fusion and spatial reasoning, characterized in that, include: The visible light camera and infrared thermal imager installed on the inspection drone simultaneously acquire visible light image sequences and infrared thermal imaging sequences of the marine photovoltaic array; and store the GPS coordinates and drone attitude angles at the time of acquisition. Based on the GPS coordinates and the UAV attitude angle, the visible light image sequence and infrared thermal imaging sequence acquired at the same time are registered to obtain a multimodal image pair; the multimodal image pair is input into a dual-stream shared coding network to output candidate regions for holes and cracks, as well as hot spot anomaly regions; Based on the candidate defect regions and hot spot anomaly regions, calculate the temperature gradient statistics of each visible light defect region and its corresponding infrared region; establish the mapping relationship between the local maximum temperature difference and the hot spot area ratio based on the prior constraint regression head to obtain the preliminary defect score of each photovoltaic panel unit; wherein, the temperature gradient statistics include the local maximum temperature difference and the hot spot area ratio. A defect propagation probability model is constructed based on the preliminary defect score of each photovoltaic panel unit and the photovoltaic array topology diagram. The comprehensive health index of each photovoltaic panel is calculated based on the defect propagation probability model. All photovoltaic panels are classified according to the comprehensive health index of each photovoltaic panel and a defect heat map and priority repair list are output.

2. The method for defect assessment of marine photovoltaic panels based on multimodal fusion and spatial reasoning according to claim 1, characterized in that, The calculation of the temperature gradient statistics between each visible light defect region and its corresponding infrared region based on the aforementioned defect candidate regions and hot spot anomaly regions further includes: For any visible light defect candidate region, an infrared matching sub-region with the visible light defect region is obtained on the corresponding infrared thermal image using an optical flow tracing algorithm; an adjacent annular buffer region without hot spots is selected at the outer edge of the infrared matching sub-region, and the weighted average temperature of all pixels in the annular buffer region is calculated as the local background reference temperature. The absolute temperature difference value of each pixel in the infrared matching sub-region is calculated based on the local background reference temperature to obtain the local absolute temperature difference field matrix; temperature gradient statistics are extracted based on the local absolute temperature difference field matrix; wherein, the temperature gradient statistics include the local maximum temperature difference, the hot spot area ratio, and the temperature difference gradient distribution entropy.

3. The method for defect assessment of marine photovoltaic panels based on multimodal fusion and spatial reasoning according to claim 1, characterized in that, The formula for calculating the preliminary defect score is as follows: ； in: ； ； ； in, This refers to the number of the photovoltaic panel unit; These represent the local maximum temperature difference and the percentage of hot spot area, respectively. This represents the threshold for the proportion of hot spot area. This is the threshold for the local maximum temperature difference; , , These are the shape parameters determined by the material fatigue model, respectively; and are the shape parameters determined based on the material fatigue model, respectively. The entropy of the temperature gradient distribution; , , To adjust the parameters; The comprehensive environmental aging coefficient; This is the equivalent corrosion activation energy; It is the gas constant; Standard reference temperature; The ambient temperature; The coupling coefficient; This represents the local maximum temperature difference corresponding to the i-th photovoltaic panel unit; This represents the percentage of the hot spot area corresponding to the i-th photovoltaic panel unit; The entropy of the temperature gradient distribution corresponding to the i-th photovoltaic panel unit; The confidence score modulation function; This is the Gaussian error function.

4. The method for defect assessment of marine photovoltaic panels based on multimodal fusion and spatial reasoning according to claim 1, characterized in that, Based on the preliminary defect score of each photovoltaic panel unit and the photovoltaic array topology diagram, a defect propagation probability model is constructed, which further includes: Based on the preliminary defect score and photovoltaic array topology of each photovoltaic panel unit, the K-hop electrical / physical neighborhood set of each photovoltaic panel unit is obtained; the propagation risk factor of each target photovoltaic panel unit based on the K-hop electrical / physical neighborhood state is calculated; The dynamic defect propagation probability of each target photovoltaic panel unit within the time window is calculated based on the propagation risk factor. The comprehensive health index of each target photovoltaic panel unit is obtained by calculating the state transition probability based on the dynamic defect propagation probability and the preliminary defect score using a Markov chain based on the propagation probability of the defect.

5. The method for defect assessment of marine photovoltaic panels based on multimodal fusion and spatial reasoning according to claim 4, characterized in that, The formula for calculating the comprehensive health index is: ； in: ； ； ； ； ； in, Let be the rate intensity of the i-th photovoltaic panel unit transitioning from a healthy state to a latent defect state; It is a vector-valued function composed of RBF kernels; For RBF center; For bandwidth; are basis vectors; For degenerate-activation coupled functionals; For repair-recovery co-functionality; The correlation coefficient; For Minkowski norm parameters; The virtual energy function on the state transition diagram; It is an inverse temperature parameter; This refers to the number of the photovoltaic panel unit; The comprehensive health index of the i-th photovoltaic panel unit; This represents the total number of radial basis functions; For gradient operators; Let be the rate intensity of the i-th photovoltaic panel unit transitioning from a faulty state to a healthy state; Let be the rate intensity of the i-th photovoltaic panel unit transitioning from a latent defect state to a healthy state; is the activation resonance factor of the i-th photovoltaic panel unit.

6. The method for defect assessment of marine photovoltaic panels based on multimodal fusion and spatial reasoning according to claim 1, characterized in that, All photovoltaic panels are graded based on their comprehensive health index, and a defect heatmap is generated, which further includes: A statistical distribution fit is performed on the set of comprehensive health indices of all photovoltaic panels, and its kernel density estimate is calculated. Multiple grading thresholds are generated based on the inflection point and standard deviation of the kernel density distribution. The comprehensive health index of each photovoltaic panel is mapped to the corresponding health level according to the grading thresholds. At each photovoltaic panel unit location in the topology map, a static thermal layer is generated by mapping its health level to a preset pseudo-color gradient color band, which represents the severity of defects with color depth. Based on the propagation risk factors of each photovoltaic panel unit obtained from the defect propagation probability model, a dynamic overlay layer is generated to characterize the potential propagation risk of defects using dynamic flicker frequency or halo diffusion effect. The static thermal layer is then fused with the dynamic overlay layer to output an interactive defect heat map, which displays the current spatial distribution severity of defects and future risk propagation trends.

7. The method for defect assessment of marine photovoltaic panels based on multimodal fusion and spatial reasoning according to claim 1, characterized in that, The structure of the dual-stream shared coding network includes a shared backbone encoder, a visible light branch decoder, an infrared branch decoder, and a cross-modal attention fusion module; The shared backbone encoder is a Vision Transformer feature extraction network, which contains multiple cascaded Transformer coding layers. Each Transformer coding layer interacts with features through a multi-head self-attention module. Both the visible light branch decoder and the infrared branch decoder are encoder-decoder structures with skip connections; the encoder-decoder structure makes skip connections with shared features from the corresponding level of the encoder through the upsampling layer.

8. The method for defect assessment of marine photovoltaic panels based on multimodal fusion and spatial reasoning according to claim 4, characterized in that, The photovoltaic array topology is a multi-layer graph structure composed of node layers and edge layers; The node layer's attribute vector includes basic physical attributes, dynamic detection attributes, spatiotemporal location attributes, and historical health status. The basic physical attributes include model, rated power, area, installation tilt angle, and azimuth angle. The spatiotemporal location attributes include geospatial coordinates obtained from GPS coordinates and 3D reconstruction, as well as 2D grid coordinates in the array's logical layout. The historical health status is a historical sequence of the comprehensive health index calculated from each inspection. The edge layer includes physical proximity edges, electrical connection edges, and maintenance reachable edges. The physical proximity edges are automatically generated based on Delaunay triangulation between nodes, used to model the propagation risks of physical defects caused by structural vibration transmission, salt spray water accumulation, and shading. The electrical connection edges are constructed based on the electrical wiring diagram of the photovoltaic array, used to model the propagation risks of electrical defects such as current mismatch and voltage reversal in series branches and fault current backflow in parallel groups. The maintenance reachable edges are generated based on the platform layout, maintenance channels, and obstacle location information of the photovoltaic array, used to model the cost and path of maintenance operations.

9. A defect assessment device for marine photovoltaic panels based on multimodal fusion and spatial reasoning, characterized in that, include: The data synchronization acquisition module is used to simultaneously acquire visible light image sequences and infrared thermal imaging sequences of the marine photovoltaic array through the visible light camera and infrared thermal imager installed on the inspection drone; and to store the GPS coordinates and drone attitude angle at the time of acquisition. The image registration and dual-stream defect detection module is used to register the visible light image sequence and the infrared thermal imaging sequence acquired at the same time according to the GPS coordinates and the UAV attitude angle to obtain a multimodal image pair; the multimodal image pair is input into the dual-stream shared coding network, and the output is a candidate region for hole and crack type defects and a hot spot anomaly region; The temperature gradient statistics and preliminary scoring module is used to calculate the temperature gradient statistics of each visible light defect region and its corresponding infrared region based on the defect candidate region and the hot spot abnormal region; and to establish a mapping relationship between the local maximum temperature difference and the hot spot area ratio based on the prior constraint regression head to obtain the preliminary defect score of each photovoltaic panel unit; wherein, the temperature gradient statistics include the local maximum temperature difference and the hot spot area ratio. The comprehensive evaluation module is used to construct a defect propagation probability model based on the preliminary defect score of each photovoltaic panel unit and the photovoltaic array topology map, calculate the comprehensive health index of each photovoltaic panel based on the defect propagation probability model, classify all photovoltaic panels based on the comprehensive health index of each photovoltaic panel, and output a defect heat map and priority repair list.

10. A computing device, comprising: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described method for evaluating defects in marine photovoltaic panels based on multimodal fusion and spatial reasoning.

Images