Photovoltaic module dust accumulation and hot spot defect ai intelligent identification detection method

Abstract

This application discloses an AI-powered intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules, comprising the following steps: S1, synchronously acquiring visible light images and infrared thermal images of photovoltaic modules, and performing spatiotemporal registration of the visible light images and infrared thermal images. Thus, using an AI-powered intelligent identification mode, there is no need for staff to observe with their naked eyes or use handheld devices to inspect each module individually, significantly reducing the investment of manpower and resources. Relying on a dual-stream collaborative deep network model, comprehensive and efficient detection of defects in photovoltaic modules is achieved. The visible light feature extraction branch captures visual information related to dust accumulation, and the infrared feature extraction branch captures thermal information related to hot spots. Combined with a cross-modal attention fusion module, bidirectional interaction of dual features is achieved, effectively identifying hidden hot spots. Through the causal inference module for inducing risks, it can be determined whether hot spot defects are induced by dust accumulation and the degree of causal coupling between the two can be quantified. It can predict the hot spot risks that dust accumulation may cause in advance and extend the service life of the modules.

Description

Technical Field

[0001] This application relates to the technical field of photovoltaic equipment testing, and in particular to an AI-based intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules. Background Technology

[0002] As the core energy conversion unit of a photovoltaic power generation system, photovoltaic modules are exposed to complex outdoor environments for extended periods. They are susceptible to factors such as dust accumulation, stains, uneven lighting, and module aging, leading to two typical faults: dust accumulation and hot spot defects. Dust accumulation blocks the light incident on the surface of the photovoltaic module, reducing photoelectric conversion efficiency. Hot spot defects are abnormally high temperatures generated in localized areas of the photovoltaic module. Long-term hot spots can accelerate cell aging, degrade encapsulation materials, and even cause safety hazards such as module burnout and fires. Therefore, accurate identification and risk prediction of dust accumulation and hot spot defects in photovoltaic modules are core requirements for the intelligent operation and maintenance of photovoltaic power plants.

[0003] In traditional operation and maintenance, staff need to inspect each component by visual inspection or handheld testing equipment. This not only consumes a lot of manpower and resources, but the testing efficiency is also difficult to meet the operation and maintenance needs of large-scale photovoltaic power plants. In addition, manual judgment is affected by subjective experience and ambient light, resulting in low accuracy in identifying minor dust accumulation and hidden hot spots, which can easily lead to missed or false detections. Furthermore, working at height poses safety hazards and makes it impossible to achieve comprehensive defect detection.

[0004] Secondly, existing machine vision inspection methods only use visible light images for dust detection, which cannot capture abnormal internal temperatures of components and make it difficult to identify hidden hot spots. Furthermore, under complex working conditions such as insufficient lighting and shadow occlusion, the distinction between dust accumulation areas and component backgrounds is reduced, making misjudgments more likely.

[0005] In addition, most existing detection methods can only independently identify dust accumulation and hot spot defects and output detection results separately. This makes it impossible for maintenance personnel to predict the hot spot risk that dust accumulation may cause in advance. They can only carry out passive repairs after the hot spot defects appear. This can only stop the damage from worsening and cannot undo the damage that has already occurred. In the long run, this will significantly shorten the lifespan of photovoltaic modules and increase the cost of module replacement.

[0006] Application content

[0007] This application aims to address, at least to some extent, the technical problems in the related art.

[0008] To achieve the above objectives, this application proposes an AI-based intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules, comprising the following steps:

[0009] S1. Simultaneously acquire visible light images and infrared thermal images of photovoltaic modules, perform spatiotemporal registration on the visible light images and infrared thermal images, and obtain aligned multimodal image pairs.

[0010] S2. Construct a dual-stream collaborative deep network model, including a visible light feature extraction branch and an infrared feature extraction branch, to extract visual and thermal features from aligned visible light and infrared thermal images, respectively, and establish a bidirectional interactive mapping between visual and thermal features through a cross-modal attention fusion module.

[0011] S3. Using the dual-stream collaborative deep network model, identify and segment the dust accumulation area on the surface of the photovoltaic module in the visible light image space, output the dust accumulation mask and calculate the dust accumulation coverage level. At the same time, detect and locate the hot spot defect area in the infrared thermal image space, output the hot spot mask and calculate the hot spot level.

[0012] S4. The causal deduction module for inducing risk maps the dust mask to the infrared thermal image coordinate system, determines whether the dust area induces a hot spot effect, and if the dust area and the hot spot defect area have spatial overlap and the temperature in the overlapping area exceeds a preset threshold, a dust-induced hot spot warning is triggered, and a causal coupling quantitative index between dust and hot spot is output.

[0013] S5. Based on the dust coverage level, hot spot level, and causal coupling quantitative index, generate hierarchical operation and maintenance decision recommendations, including cleaning priority score and hot spot risk level assessment.

[0014] In addition, the application may also include the following additional technical features:

[0015] Specifically, the spatiotemporal registration in step S1 includes: extracting common feature points in visible light images and infrared thermal images based on scale-invariant feature transformation or accelerated robust feature extraction, and achieving spatial alignment through homography transformation matrix; at the same time, achieving time synchronization by using the synchronous trigger signal of the image acquisition device or timestamp interpolation method.

[0016] Specifically, the visible light feature extraction branch and the infrared feature extraction branch both use a deep residual network or an efficient visual Transformer as the backbone network, and output multi-scale visual feature maps and multi-scale thermal feature maps respectively.

[0017] Specifically, the cross-modal attention fusion module includes: calculating a first cross-attention using visual features as a query and thermal features as keys and values ​​to obtain thermally enhanced visual features; calculating a second cross-attention using thermal features as a query and visual features as keys and values ​​to obtain visually enhanced thermal features; and adaptively aggregating the two enhanced features with the original features through learnable fusion weights.

[0018] Specifically, the identification and segmentation of the dust accumulation area in step S3 includes: using a depth convolution-based encoder-decoder structure to classify dust, glass, and battery cell categories pixel by pixel on the visible light feature map, and outputting a dust accumulation mask; the dust accumulation coverage level is divided into four levels: clean, light dust accumulation, moderate dust accumulation, and heavy dust accumulation, according to the proportion of the dust accumulation pixel area to the total pixel area of ​​the photovoltaic module.

[0019] Specifically, the detection and location of hot spot defect areas in step S3 includes: using an instance segmentation network based on anchor frames or without anchor frames on the infrared feature map to extract temperature abnormal connected regions, filtering non-defect hot spots through a preset temperature threshold, and outputting a hot spot mask. The hot spot level is divided into four levels: slight, moderate, severe, and critical, based on the temperature difference between the hot spot area and the normal component background.

[0020] Specifically, the process of determining whether the dust accumulation area induces a hot spot effect in step S4 by the risk causal deduction module is as follows: after mapping the dust accumulation mask to the infrared thermal image coordinate system, the spatial overlap rate between the dust accumulation area and the hot spot defect area is calculated. When the overlap rate is greater than the preset overlap threshold and the average temperature in the overlap area exceeds the sum of the normal average temperature of the component and the preset temperature difference threshold, it is determined to be a dust accumulation-induced hot spot. The causal coupling quantification index includes the overlap rate, the temperature anomaly multiple of the overlap area, and the Pearson correlation coefficient between dust density and temperature rise.

[0021] Specifically, in step S4, when determining whether the dust accumulation area induces a hot spot effect, both spatial and temperature conditions must be met simultaneously: the spatial overlap rate between the dust accumulation area and the hot spot area is greater than a threshold, and the average temperature in the overlapping area exceeds the sum of the normal temperature of the component and the temperature difference threshold. If the dust accumulation area shows a trend of temperature rise rate exceeding a preset threshold during continuous detection, it will be used as a basis for strengthening the determination of dust accumulation-induced hot spot.

[0022] Specifically, the generation of hierarchical operation and maintenance decision suggestions in step S5 includes: outputting a cleaning plan based on the cleaning priority score, wherein the cleaning priority score is positively correlated with the dust coverage level and the causal coupling quantitative index; outputting maintenance or replacement suggestions based on the hot spot risk level assessment; and setting the operation and maintenance priority of dust-induced hot spots to be higher than that of non-dust-induced hot spots with the same temperature difference.

[0023] In summary, the beneficial effects of the AI-based intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules presented in this application are as follows:

[0024] 1. Adopting an AI intelligent recognition mode, it eliminates the need for staff to observe with their naked eyes or use handheld devices to inspect each module, greatly reducing the investment of manpower and resources. Relying on a dual-flow collaborative deep network model, it can accurately identify minor dust accumulation and hidden hot spots, reducing the rate of missed detection and false detection. At the same time, it eliminates the need for high-altitude operations, thus eliminating safety hazards and achieving comprehensive and efficient detection of defects in photovoltaic modules.

[0025] 2. By capturing visual information related to dust accumulation through the visible light feature extraction branch and capturing thermal information related to hot spots through the infrared feature extraction branch, and combining the cross-modal attention fusion module to achieve bidirectional interaction of dual features, it can accurately capture internal temperature anomalies of components, effectively identify hidden hot spots, and solve the problem of low differentiation between dust accumulation areas and background under complex working conditions, thus avoiding misjudgment.

[0026] 3. The risk causal deduction module can determine whether hot spot defects are induced by dust accumulation and quantify the degree of causal coupling between the two. It can predict the hot spot risk that dust accumulation may cause in advance, and can deal with the dust accumulation cause in a timely and targeted manner to avoid the appearance of hot spots or curb the further deterioration of hot spots, recover potential power generation losses, and at the same time prevent irreversible damage to the components and extend the service life of the components. Attached Figure Description

[0027] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0028] Figure 1 This is a flowchart of the AI-based intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules according to this application.

[0029] Figure 2 This is a flowchart of the cross-modal attention fusion module of the AI ​​intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules in this application.

[0030] Figure 3 This is a flowchart illustrating the computational space overlap rate of the AI-based intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules in this application. Detailed Implementation

[0031] To make the technical means, inventive features, objectives, and effects of this application easier to understand, the application is further described below with reference to specific illustrations. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0032] The present application will now be described in further detail with reference to the accompanying drawings.

[0033] like Figures 1-3 As shown in the figure, the AI-powered intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules according to an embodiment of this application includes the following steps:

[0034] S1. Simultaneously acquire visible light and infrared thermal images of photovoltaic modules, perform spatiotemporal registration on the visible light and infrared thermal images, and obtain aligned multimodal image pairs.

[0035] It should be noted that visible light images rely on the reflection of natural ambient light to fully present visual information such as the surface texture, stains, dust accumulation, and surface shading of photovoltaic modules. Infrared thermal images are based on the principle of infrared radiation of objects and can intuitively reflect the surface temperature field distribution of photovoltaic modules, capture local overheating and abnormal temperature areas, and are data for identifying hot spot defects.

[0036] Since the two images are acquired by different imaging devices, there are hardware differences in shooting field of view, installation position, focal length and angle, and imaging frame rate. At the same time, real-time changes in external light, wind speed and ambient temperature will cause asynchronous time dimension. Directly mixing them will result in pixel misalignment and area mismatch at the same component position.

[0037] Therefore, it is necessary to first complete spatiotemporal dual registration. Temporal registration is achieved by synchronizing timestamps and aligning frame sequences to ensure that the two types of images were acquired at the same time and under the same conditions. Spatial registration is achieved by correcting image distortion, matching feature points, transforming homography matrix, and correcting pixel-level mapping to eliminate spatial misalignment caused by equipment installation deviations.

[0038] S2. Construct a dual-stream collaborative deep network model, including a visible light feature extraction branch and an infrared feature extraction branch, to extract visual and thermal features from aligned visible light and infrared thermal images, respectively, and establish a bidirectional interactive mapping between visual and thermal features through a cross-modal attention fusion module.

[0039] Specifically, visible light alone cannot detect internal temperature anomalies in components, making it difficult to identify hidden hot spots. Infrared images alone lack surface details, making it impossible to distinguish whether hot spots are caused by dust accumulation, battery aging, or microcracks. Based on the theory of multimodal feature complementarity, a dual-stream collaborative deep network model is constructed to achieve parallel extraction and deep fusion of dual features.

[0040] The model adopts a dual-branch independent coding structure. The visible light feature extraction branch focuses on the visual low-level features such as shallow texture, edge, and color, as well as the high-level semantic features such as dust distribution and surface contamination. The infrared feature extraction branch focuses on mining the thermal low-level features such as temperature gradient, thermal radiation difference, and local temperature rise, as well as the semantic features of hot spot defects and abnormal heating areas. The two branches are adapted to their respective modal data characteristics to form a dedicated feature extraction backbone structure, ensuring the relevance and effectiveness of feature extraction.

[0041] A cross-modal attention fusion module is introduced to solve the problems of feature fragmentation and information redundancy in traditional simple splicing and fusion. Through the attention weight allocation mechanism, a two-way interactive mapping between visual features and thermal features is established: the visible light feature branch adaptively focuses on key areas of high temperature anomalies in the infrared image, and the infrared feature branch reinforces the learning of visual semantic information of dust occlusion and surface contamination. The two-way complementarity enhances the feature expression capability.

[0042] S3. Using a dual-stream collaborative deep network model, the dust accumulation area on the surface of the photovoltaic module is identified and segmented in the visible light image space, the dust accumulation mask is output and the dust accumulation coverage level is calculated. At the same time, the hot spot defect area is detected and located in the infrared thermal image space, the hot spot mask is output and the hot spot level is calculated.

[0043] It is understandable that dust accumulation is a visible defect on the surface of the photovoltaic module. In visible light images, it has the highest feature recognition, clear boundaries, and strong regional continuity, making it suitable for pixel-level semantic segmentation algorithms. The model accurately classifies the foreground dust-covered pixels and the background clean module pixels, outputting a binary dust mask to accurately outline the dust coverage contour. Then, by statistically analyzing the ratio of the dust-covered pixel area to the effective power generation area of ​​the photovoltaic module, and combining it with the industry operation and maintenance standard threshold gradient, the dust coverage level is quantitatively classified, thus quantitatively representing the degree of module contamination.

[0044] Hot spots are essentially localized temperature anomalies in solar cells, with their core characteristics manifested in the infrared temperature dimension. In infrared thermal images, the temperature difference between abnormal temperature rise areas and normal areas is significant, and the feature distinction is high. Therefore, defect target detection and region segmentation are carried out in the infrared image space to accurately locate the position and range of hot spots in single or multiple solar cells and output hot spot masks. At the same time, by combining multi-dimensional parameters such as the temperature rise amplitude of the hot spot region, defect area, and temperature dispersion, the severity level of hot spots is classified.

[0045] S4. The causal deduction module for inducing risk maps the dust mask to the infrared thermal image coordinate system to determine whether the dust area induces a hot spot effect. If the dust area and the hot spot defect area have spatial overlap and the temperature in the overlapping area exceeds the preset threshold, a dust-induced hot spot warning is triggered, and a causal coupling quantitative index between dust and hot spot is output.

[0046] It should be noted that dust accumulation on the surface of photovoltaic modules can block the incident light. The cells in the blocked areas cannot perform normal photoelectric conversion, resulting in a significant decrease in power generation. The cells become a load in the series working circuit of the module, and are powered in reverse by the surrounding normally generating cells, generating additional power consumption and continuous heat generation, which can induce local high temperatures and form hot spots.

[0047] Because the coordinate systems and projection rules of visible light images and infrared images are different, direct comparison of the two masks will result in spatial misalignment. Therefore, by using coordinate geometric mapping and image transformation algorithms, the dust accumulation mask in the visible light domain is uniformly mapped to the infrared thermal image coordinate system, so as to achieve the same scale and same coordinate spatial comparison between the dust accumulation area and the hot spot defect area.

[0048] The system employs a dual-condition judgment logic combining spatial overlay determination and temperature threshold constraint: the region correlation is determined by the pixel overlap ratio, and the validity of the heating anomaly is determined by the infrared temperature threshold. When both conditions are met simultaneously, the hot spot can be determined to be a dust-induced hot spot, which is different from spontaneous hot spots caused by battery aging, internal microcracks, or circuit faults.

[0049] S5. Based on dust coverage level, hot spot level and causal coupling quantitative indicators, generate hierarchical operation and maintenance decision suggestions, including cleaning priority score and hot spot risk level assessment.

[0050] It should be noted that the dust coverage level determines the degree of light transmission attenuation of the module and the basic value of power generation loss, and is the core basis for cleaning operations; the hot spot level is directly related to the degree of cell damage, fire safety hazards, and the risk of accelerated module degradation, and is a key indicator for equipment safety management. The causal coupling quantitative indicator clarifies the root cause of hot spot failure.

[0051] Based on the weighted scoring model, the degree of pollution, hot spot risk coefficient and causal correlation strength are comprehensively calculated to generate a quantitative component cleaning priority score. Priority is given to cleaning operations for components with severe dust accumulation, easy to induce hot spots and high power generation loss. At the same time, the overall hot spot risk level is assessed by combining hot spot temperature rise risk, fault coupling type and component operating years.

[0052] In one embodiment of this application, the spatiotemporal registration in step S1 includes: extracting common feature points in visible light images and infrared thermal images based on scale-invariant feature transformation or accelerated robust feature extraction, and achieving spatial alignment through homography transformation matrix; at the same time, time synchronization is achieved by using the synchronous trigger signal of the image acquisition device or timestamp interpolation.

[0053] It should be noted that, firstly, common feature points in the two types of images are extracted based on Scale Invariant Feature Transform (SIFT) or Accelerated Robust Feature Transform (SURF) algorithms. The SIFT algorithm can effectively resist the effects of image scaling, rotation, and illumination changes, and can stably extract unique feature points in the two types of images. The SURF algorithm accelerates the feature extraction process through integral images, improves registration efficiency, and is suitable for the rapid processing of large-scale photovoltaic module images.

[0054] After extracting common feature points, the pixel coordinate correspondence between the two types of images is established by solving the homography transformation matrix, achieving precise spatial alignment and eliminating image offset problems caused by differences in shooting angle and device position. Simultaneously, to ensure the synchronization of the two types of images in the temporal dimension and avoid feature misalignment due to acquisition time differences, time synchronization is achieved using a synchronization trigger signal from the image acquisition device or timestamp interpolation: the synchronization trigger signal controls the visible light camera and infrared thermal imager to start acquisition simultaneously, ensuring temporal consistency from the source.

[0055] In one embodiment of this application, both the visible light feature extraction branch and the infrared feature extraction branch use a deep residual network or an efficient visual Transformer as the backbone network to output multi-scale visual feature maps and multi-scale thermal feature maps, respectively.

[0056] It should be noted that deep residual networks effectively solve the gradient vanishing and gradient exploding problems in the training process of deep networks by introducing residual connection structures. They can deeply mine the deep semantic features in images. Their convolutional layers at different levels can output feature maps of different scales. Shallow feature maps focus on the texture details of photovoltaic modules (such as scratches on the glass surface and the outline of dust particles), mid-level feature maps capture the structural features of the modules (such as the arrangement of battery cells and the shape of the frame), and deep feature maps correspond to the global semantic features of the modules (such as the integrity of the overall modules).

[0057] The high-efficiency visual Transformer is based on a self-attention mechanism, which can break the local receptive field limitation of convolutional networks and quickly capture global correlation features in images. It is especially suitable for extracting global contrast features between hot spot areas and normal areas in infrared thermal images. Through block embedding and multi-head attention computation, it can output thermal feature maps of different scales, accurately depicting the temperature distribution differences on the surface of components, as well as the range and contour of hot spot areas.

[0058] In one embodiment of this application, the cross-modal attention fusion module includes: calculating a first cross-attention using visual features as a query and thermal features as keys and values ​​to obtain thermally enhanced visual features; calculating a second cross-attention using thermal features as a query and visual features as keys and values ​​to obtain visually enhanced thermal features; and adaptively aggregating the two enhanced features with the original features through learnable fusion weights.

[0059] It should be noted that the multi-scale visual feature map output from the visible light feature extraction branch is used as the query, and the multi-scale thermal feature map output from the infrared feature extraction branch is used as the key and value. The first cross-attention operation is achieved through attention weight calculation, focusing on the enhancement effect of thermal features on visual features to obtain thermally enhanced visual features. That is, by using the cues of temperature anomaly areas in the thermal features, the texture details of the corresponding areas in the visual features are enhanced, which facilitates the subsequent accurate identification of the associated areas of dust accumulation and hot spots.

[0060] Subsequently, the thermal feature map is used as the query and the visual feature map as the key and value, and a second cross-attention operation is performed. By utilizing the texture structure information in the visual features, the positioning accuracy of the temperature anomaly area in the thermal features is optimized, and visually enhanced thermal features are obtained, avoiding false hot spot misjudgments caused by infrared image noise.

[0061] Finally, learnable fusion weight parameters are introduced to adaptively aggregate thermally enhanced visual features, visually enhanced thermal features, and the two original features. The weight parameters are dynamically adjusted during the model training process, and the optimal weights are automatically assigned according to the feature importance of different defect types, ensuring that the fused features can both retain visual texture details and accurately reflect temperature distribution differences.

[0062] In one embodiment of this application, the identification and segmentation of the dust accumulation area in step S3 specifically includes: using a depth convolution-based encoder-decoder structure to classify dust, glass and battery cell categories pixel by pixel on the visible light feature map, and outputting a dust accumulation mask; the dust accumulation coverage level is divided into four levels: clean, light dust accumulation, moderate dust accumulation and heavy dust accumulation, according to the proportion of the dust accumulation pixel area to the total pixel area of ​​the photovoltaic module.

[0063] It should be noted that the pixel features of the dust accumulation area are characterized by low gray values, coarse texture and no fixed shape, while the glass area is characterized by uniform gray values ​​and smooth texture, and the battery cell area presents a regular rectangular texture and a fixed gray range. The classifier outputs the class probability of each pixel, and finally generates a binarized dust accumulation mask. In the mask, the white area corresponds to the dust accumulation area and the black area corresponds to the non-dust accumulation area.

[0064] A cleanliness level corresponds to a dust coverage rate of <5%, at which point dust accumulation has minimal impact on module power generation efficiency; light dust accumulation corresponds to a dust coverage rate of 5%-20%, resulting in a slight decrease in module power generation efficiency; moderate dust accumulation corresponds to a dust coverage rate of 20%-50%, resulting in a significant decrease in power generation efficiency; and heavy dust accumulation corresponds to a dust coverage rate of >50%, resulting in a substantial reduction in module power generation efficiency, requiring immediate cleaning.

[0065] In one embodiment of this application, the detection and location of hot spot defect areas in step S3 specifically includes: using an instance segmentation network based on anchor frames or without anchor frames on the infrared feature map to extract temperature abnormal connected regions, filtering non-defect hot spots through a preset temperature threshold, and outputting a hot spot mask. The hot spot level is divided into four levels: slight, moderate, severe, and critical, based on the temperature difference between the hot spot area and the normal component background.

[0066] It should be noted that anchor-frame instance segmentation networks use preset anchor frames of different sizes and proportions to traverse infrared feature maps, filter out candidate regions that may have hot spot defects, and then accurately locate the hot spot region and generate a hot spot mask through feature matching and bounding box regression. Anchor-frame-less instance segmentation networks, on the other hand, do not require preset anchor frames. They directly predict the center coordinates and size of the hot spot region to achieve rapid detection and segmentation of hot spots. This is suitable for scenarios with large differences in hot spot size, improving detection efficiency and flexibility.

[0067] After extracting the connected regions with abnormal temperatures, to avoid misjudging environmental interference as hot spot defects, a preset temperature threshold is used for filtering. The temperature threshold is set according to the normal operating temperature range of the photovoltaic module. For example, if the normal operating average temperature of the module is 25-35℃ and the preset temperature difference threshold is 5℃, then connected regions with temperatures higher than 40℃ are identified as suspected hot spots, and regions with temperatures lower than this threshold are considered as non-defect regions, thus outputting an accurate hot spot mask.

[0068] Hot spot levels are classified into four categories—minor, moderate, severe, and critical—based on the temperature difference between the hot spot area and the normal module background. Specifically: Minor hot spots correspond to a temperature difference of 5-10℃, with a small hot spot area and minimal impact on the module; moderate hot spots correspond to a temperature difference of 10-15℃, with a medium-sized hot spot area and a localized decrease in module power generation efficiency; severe hot spots correspond to a temperature difference of 15-20℃, with a large hot spot area and a risk of localized module damage; critical hot spots correspond to a temperature difference >20℃, with the hot spot area spreading rapidly and easily leading to module burnout, requiring immediate shutdown and repair.

[0069] In one embodiment of this application, the specific process of determining whether the dust accumulation area induces a hot spot effect in step S4 by the risk causal inference module is as follows: after mapping the dust accumulation mask to the infrared thermal image coordinate system, the spatial overlap rate between the dust accumulation area and the hot spot defect area is calculated. When the overlap rate is greater than the preset overlap threshold and the average temperature in the overlap area exceeds the sum of the normal average temperature of the component and the preset temperature difference threshold, it is determined to be a dust accumulation-induced hot spot. The causal coupling quantification index includes the overlap rate, the temperature anomaly multiple of the overlap area, and the Pearson correlation coefficient between dust density and temperature rise.

[0070] It should be noted that the normal average temperature of the module is obtained by statistically analyzing the infrared temperatures of modules from the same batch, operating under the same conditions, without dust accumulation or defects. The preset temperature difference threshold is set according to the module performance parameters to ensure the accuracy of the judgment results. To quantify the causal coupling between dust accumulation and hot spots, causal coupling quantification indicators are introduced: first, the overlap rate, which reflects the spatial correlation strength between dust accumulation and hot spots; second, the temperature anomaly multiple of the overlapping area, i.e., the ratio of the average temperature of the overlapping area to the normal average temperature of the module, reflecting the degree of influence of dust accumulation on the temperature rise of the module; and third, the Pearson correlation coefficient between dust density and temperature rise, which quantifies the linear correlation between the two by calculating the correlation coefficient between the density of the dust accumulation area and the temperature rise in that area. The closer the correlation coefficient is to 1, the higher the dust density, the more obvious the temperature rise of the module, and the stronger the causal relationship.

[0071] In one embodiment of this application, in step S4, when determining whether the dust accumulation area induces a hot spot effect, both spatial and temperature conditions must be met simultaneously: the spatial overlap rate between the dust accumulation area and the hot spot area is greater than a threshold, and the average temperature in the overlapping area exceeds the sum of the normal temperature of the component and the temperature difference threshold. If the dust accumulation area shows a trend of temperature rise rate exceeding a preset threshold during continuous detection, it will be used as a basis for strengthening the determination of a dust accumulation-induced hot spot.

[0072] It should be noted that if the dust-accumulated area shows a temperature rise rate exceeding a preset threshold in multiple consecutive tests (e.g., three consecutive tests with a one-hour interval between each test), this trend will be used as the basis for further categorizing it as a dust-induced hotspot. This is because non-dust-induced hotspots typically have relatively stable temperatures, while dust-induced hotspots exhibit a continuous temperature rise trend as environmental conditions change with dust accumulation (e.g., increased light intensity). By continuously monitoring the temperature rise rate, it is possible to effectively distinguish between accidental correlations and genuine causal relationships, reducing the false positive rate.

[0073] In one embodiment of this application, the step S5 of generating hierarchical operation and maintenance decision recommendations includes: outputting a cleaning plan based on a cleaning priority score, wherein the cleaning priority score is positively correlated with the dust coverage level and the causal coupling quantitative index; outputting maintenance or replacement recommendations based on the hot spot risk level assessment; and setting the operation and maintenance priority of dust-induced hot spots to be higher than that of non-dust-induced hot spots with the same temperature difference.

[0074] It should be noted that modules with heavy dust accumulation and severe hot spots require the highest cleaning priority and must be cleaned immediately; modules with light dust accumulation and no associated hot spots have a lower cleaning priority and can be cleaned according to the regular schedule. Secondly, based on the hot spot risk level assessment, targeted repair or replacement recommendations should be provided: minor hot spots do not require immediate repair, only periodic monitoring of temperature changes; general hot spots require timely inspection of module wiring, encapsulation, etc., and targeted repairs; severe hot spots require shutdown for inspection and replacement of damaged cells or parts of the module; critical hot spots require immediate replacement of the entire module to prevent the fault from spreading.

[0075] Considering the preventability of dust-induced hot spots, the maintenance priority of dust-induced hot spots is set higher than that of non-dust-induced hot spots with the same temperature difference. For example, for severe hot spots with the same temperature difference, dust-induced hot spots should be prioritized for cleaning and maintenance, while non-dust-induced hot spots can be handled according to the regular maintenance procedures.

[0076] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0077] The present application and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present application. The actual structure is not limited to this. In conclusion, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present application, such design should fall within the protection scope of the present application.

Claims

1. An AI-powered intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules, characterized in that, Includes the following steps: S1. Simultaneously acquire visible light images and infrared thermal images of photovoltaic modules, perform spatiotemporal registration on the visible light images and infrared thermal images, and obtain aligned multimodal image pairs; S2. Construct a dual-stream collaborative deep network model, including a visible light feature extraction branch and an infrared feature extraction branch, extracting visual features and thermal features from aligned visible light images and infrared thermal images, respectively, and establishing a bidirectional interactive mapping between visual features and thermal features through a cross-modal attention fusion module; S3. Using the dual-stream collaborative deep network model, identify and segment the dust accumulation area on the surface of the photovoltaic module in the visible light image space, output the dust accumulation mask and calculate the dust accumulation coverage level. At the same time, detect and locate the hot spot defect area in the infrared thermal image space, output the hot spot mask and calculate the hot spot level. S4. The causal deduction module for inducing risk maps the dust mask to the infrared thermal image coordinate system to determine whether the dust area induces a hot spot effect. If the dust area and the hot spot defect area have spatial overlap and the temperature in the overlapping area exceeds a preset threshold, a dust-induced hot spot warning is triggered, and a causal coupling quantitative index between dust and hot spot is output. S5. Based on the dust coverage level, hot spot level, and causal coupling quantitative index, generate hierarchical operation and maintenance decision recommendations, including cleaning priority score and hot spot risk level assessment.

2. The AI-powered intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules according to claim 1, characterized in that, The spatiotemporal registration in step S1 includes: extracting common feature points in visible light images and infrared thermal images based on scale-invariant feature transformation or accelerated robust feature extraction, and achieving spatial alignment through homography transformation matrix; at the same time, time synchronization is achieved by using the synchronous trigger signal of the image acquisition device or timestamp interpolation.

3. The AI-powered intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules according to claim 1, characterized in that, Both the visible light feature extraction branch and the infrared feature extraction branch use a deep residual network or an efficient visual Transformer as the backbone network, and output multi-scale visual feature maps and multi-scale thermal feature maps, respectively.

4. The AI-powered intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules according to claim 1, characterized in that, The cross-modal attention fusion module includes: calculating a first cross-attention using visual features as a query and thermal features as keys and values ​​to obtain thermally enhanced visual features; calculating a second cross-attention using thermal features as a query and visual features as keys and values ​​to obtain visually enhanced thermal features; and adaptively aggregating the two enhanced features with the original features through learnable fusion weights.

5. The AI-powered intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules according to claim 1, characterized in that, The specific steps of identifying and segmenting the dust accumulation area in step S3 include: using a depth convolution-based encoder-decoder structure to classify dust, glass, and solar cell categories pixel by pixel on the visible light feature map, and outputting a dust accumulation mask; the dust accumulation coverage level is divided into four levels: clean, light dust accumulation, moderate dust accumulation, and heavy dust accumulation, according to the proportion of dust accumulation pixel area to the total pixel area of ​​the photovoltaic module.

6. The AI-powered intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules according to claim 1, characterized in that, The detection and location of hot spot defect areas in step S3 specifically includes: using an instance segmentation network based on anchor frames or without anchor frames on the infrared feature map to extract temperature abnormal connected regions, filtering non-defect hot spots through a preset temperature threshold, and outputting a hot spot mask. The hot spot level is divided into four levels: slight, moderate, severe, and critical, based on the temperature difference between the hot spot area and the normal component background.

7. The AI-powered intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules according to claim 1, characterized in that, The specific process of determining whether the dust accumulation area induces a hot spot effect in step S4 by the causal inference module is as follows: After mapping the dust accumulation mask to the infrared thermal image coordinate system, the spatial overlap rate between the dust accumulation area and the hot spot defect area is calculated. When the overlap rate is greater than the preset overlap threshold and the average temperature in the overlap area exceeds the sum of the normal average temperature of the component and the preset temperature difference threshold, it is determined to be a dust accumulation-induced hot spot. The causal coupling quantification index includes the overlap rate, the temperature anomaly multiple of the overlap area, and the Pearson correlation coefficient between dust density and temperature rise.

8. The AI-powered intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules according to claim 1, characterized in that, In step S4, when determining whether the dust accumulation area induces a hot spot effect, both spatial and temperature conditions must be met simultaneously: the spatial overlap rate between the dust accumulation area and the hot spot area is greater than a threshold, and the average temperature in the overlapping area exceeds the sum of the normal temperature of the component and the temperature difference threshold. If the dust accumulation area shows a trend of temperature rise rate exceeding a preset threshold during continuous detection, it will be used as a basis for strengthening the determination of dust accumulation-induced hot spot.

9. The AI-powered intelligent identification and detection method for dust accumulation and hot spot defects in photovoltaic modules according to claim 1, characterized in that, The generation of hierarchical operation and maintenance decision suggestions in step S5 includes: outputting a cleaning plan based on the cleaning priority score, which is positively correlated with the dust coverage level and the causal coupling quantitative index; outputting maintenance or replacement suggestions based on the hot spot risk level assessment; and setting the operation and maintenance priority of dust-induced hot spots to be higher than that of non-dust-induced hot spots with the same temperature difference.

Images