Automatic detection system for hot spot defect of photovoltaic module

Abstract

The invention relates to the photovoltaic field, and provides a photovoltaic module hot spot intelligent inspection method based on an unmanned aerial vehicle and infrared image processing. Aiming at the problems of large deployment difficulty and relatively high false detection rate and omission rate when the unmanned aerial vehicle detects the hot spots of the photovoltaic module, the invention designs a lightweight network model S-YOLO by taking YOLOv8 as a base line. The method comprises the following steps: firstly, providing a space confusion backbone network, reinforcing an attention mechanism while lightweight design is performed, and reducing model parameter quantity and calculation complexity; secondly, newly adding a small target prediction scale, constructing a multi-scale weighted feature pyramid network (MSBFPN), and improving the detection capability of small target hot spots; and finally, replacing CIoU with an Ang-IoU loss function, solving the problem of penalty failure, and promoting rapid convergence and accurate regression of the model. Experiments show that the weight of the S-YOLO model is reduced by 41.3%, the parameter quantity is reduced by 21.2%, the average precision is improved by 2.3%, the S-YOLO model has the advantages of light weight and high precision, and the S-YOLO model is suitable for deployment of the unmanned aerial vehicle in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of photovoltaics, and in particular to an intelligent inspection method for hot spots in photovoltaic modules based on unmanned aerial vehicles and infrared image processing. Background Technology

[0002] During operation, solar panels are prone to localized hot spot faults due to uneven illumination, internal defects, or surface accumulation of dirt and snow. These faults cause internal energy loss, reducing power generation efficiency and potentially severely damaging the panels. Currently, solar power plants commonly use drones to capture infrared images for rapid detection of hot spot faults. Developing a lightweight, efficient, and accurate hot spot detection model suitable for deployment on drones is crucial for achieving intelligent inspection of photovoltaic power plants.

[0003] With the continuous development of drone performance and artificial intelligence technology, drones combined with detection algorithms are gradually becoming the mainstream method to replace manual inspection. Traditional machine vision technology detects defects by analyzing and extracting image features, but it requires manual parameter adjustment for images with different resolutions and complex backgrounds, resulting in limited generalization ability and automated detection performance. In addition, manually designed feature models are limited and cannot comprehensively and accurately describe defect features, making it difficult to further improve detection accuracy.

[0004] In recent years, with the rise of deep convolutional neural networks (CNNs), their application advantages in the field of target detection have become increasingly significant. Compared with traditional methods, deep convolutional neural networks have stronger feature learning and transfer capabilities. For example, Ai Shangmei et al. built an image super-resolution network based on the SSD model and replaced the backbone network with a feature extraction network that uses a dual-pooling method, achieving the ability to quickly identify hot spots on photovoltaic modules. Zhang Wenbiao et al. improved the Faster R-CNN algorithm and adopted a dual-branch structure with non-shared parameters to more accurately identify defects on the surface of solar panels. Nevertheless, these methods still face the following challenges: (1) In aerial infrared images, the interference background and the infrared characteristics of hot spot faults are extremely similar; (2) Hot spot faults are usually small targets with low detection accuracy; (3) The computing power of UAVs is limited, and the existing detection models are large in size, which is not conducive to the rapid inspection of UAVs.

[0005] Therefore, a lightweight inspection model suitable for drone inspection is needed. While ensuring high inspection speed, it should effectively reduce false positives and false negatives for small targets such as solar panels, thereby meeting the actual needs of intelligent inspection of photovoltaic power plants. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a lightweight network model, S-YOLO, based on YOLOv8, for detecting hot spots on solar panel surfaces. This model effectively reduces the interference of complex backgrounds in infrared images of photovoltaic modules on hot spot detection, improves the performance of detecting small target hot spots, and thus achieves accurate hot spot identification. Experimental results show that compared to the original YOLOv8 model, the model of this invention achieves a 2.3% improvement in average accuracy while reducing weights by 41.3% and parameters by 21.2%. This improvement fully demonstrates the model's lightweight and high-precision advantages, making it highly suitable for deployment on UAVs with limited computing power, especially in complex environments.

[0007] This invention includes the following steps:

[0008] S1. Take infrared images of the photovoltaic modules of the photovoltaic power station using a drone, and preprocess the collected image data;

[0009] S2. Using the YOLOv8 model as the basic framework, a Spatial Confusion Backbone was designed. The backbone network was lightweighted while the attention mechanism was strengthened, reducing the number of model parameters and lowering computational complexity. Secondly, a new small target prediction scale was added, and a Multi-scale Weighted Feature Pyramid Network (MSBFPN) was constructed to enhance the detection capability of small target hot spots on battery cells. Finally, the superior Ang-IoU loss function was used to replace the traditional CIoU loss function, solving the problem of penalty failure and promoting the model's rapid convergence and accurate regression.

[0010] S3. The target detection algorithm model is trained using infrared images of photovoltaic power stations captured by UAVs after processing in S1. The optimal training parameters are obtained by optimizing the target detection algorithm model based on the training sample image set. During training, the input layer performs Mosaic enhancement, adaptive anchor box calculation, and image scaling operations to convert the image into a target tensor input model. Subsequently, image features are extracted through the backbone network, and the feature maps are aggregated through the spatial pyramid pooling module to generate target features. Finally, the output layer predicts the target detection results.

[0011] S4 compares the detection performance of the optimal training model obtained from S3 with the original model.

[0012] Preferably, the solar cell image data in step S1 is an infrared image of the photovoltaic modules of a photovoltaic power station captured by an infrared imager mounted on a drone. The drone flies autonomously according to a preset inspection route and sets the shooting altitude to ensure the accuracy and consistency of image acquisition.

[0013] Preferably, the images captured by the drone are screened, removing defect-free images and retaining those containing hotspots. Since hotspots typically occupy only a small area of ​​the image, the Slicing Aided HyperInference (SAHI) algorithm is used to process the images to improve detection accuracy. Specifically, the image is divided into multiple regions using a sliding window, and each region is predicted separately while inference is performed on the entire image. Subsequently, the prediction results for each region are merged with the prediction results for the entire image to generate segmented image sub-images.

[0014] Preferably, the segmented images are labeled using LabelImg software to generate training and validation sets for model training. To further expand the dataset, data augmentation operations such as rotation, flipping, translation, brightness adjustment, and adding Gaussian noise are applied to the segmented images. Finally, the dataset is divided into training, validation, and test sets in a 6:3:1 ratio for model training and performance validation.

[0015] Preferably, step S2 improves and optimizes the YOLOv8 model, specifically including:

[0016] S21. Since photovoltaic power plants often use drones to detect hot spots, they frequently face challenges related to limited drone power and computing resources. This necessitates that the detection model improve accuracy while minimizing drone power and computing resource consumption. Therefore, this invention designs a Spatial Confusion Backbone, which is lightweight while strengthening the attention mechanism. This invention transforms the backbone network into two main modules: a downsampling module and a channel partitioning module. The downsampling module expands the receptive field, increases network depth, and acquires global information. Feature map downsampling is achieved by changing the stride of the separable convolutions on the two branches. When the stride is 2, the size of the output feature map is halved. Then, the two compressed feature maps are concatenated along the channel dimension, doubling the number of channels. The channel segmentation module divides the channel into two parts. One part is directly connected to the output via a shortcut connection, while the other part undergoes a series of convolutional operations before being fused with the output of the shortcut connection. Finally, the Channel Shuffle function divides the input feature map channels into several groups and shuffles these groups to enhance information exchange between different channels while keeping the feature map size unchanged. The split convolution DWConv applied to the downsampling module and the channel segmentation module further reduces the model size, as shown in equations (1)-(2).

[0017]

[0018] Where: P, F—the number of parameters and computational cost of standard convolution; P DW F DW — Depthwise separable convolution parameter count and computational cost; H — Height of the input feature map; W — Width of the input feature map; K×K — Width and height dimensions of the convolution kernel; C1 — Number of input channels; C2 — Number of output channels. From formulas (1) and (2), it can be seen that the parameter count and computational cost of separable convolution are only a fraction of those of ordinary convolution. This further reduces the number of model parameters and computational cost. Finally, to compensate for the accuracy loss caused by lightweighting, this invention adds a Spatial-to-Depth Attention Module (SPD) to the backbone network. The SPD consists of a spatial-depth transformation layer and a 1×1 non-staggered convolutional layer, which respectively implement segmentation fusion and channel control functions. In this way, the SPD module controls the model parameters, avoids information loss, and retains important information in the channel dimension while expanding the receptive field, thereby reducing information loss.

[0019] S22. This invention constructs a multi-scale weighted feature pyramid network (MSBFPN), which can effectively detect small target hotspots. The YOLOv8 model provides outputs at three different scales of feature layers, corresponding to the middle layer, lower-middle layer, and bottom layer, respectively. The 20×20 prediction scale has the deepest network layers and is suitable for detecting large targets; while the 40×40 and 80×80 prediction scales have moderate receptive fields and are suitable for detecting medium and small targets. However, in drone aerial images, small targets such as hotspots on solar panels are relatively small, with pixel sizes generally between 10×10 and 30×30, which does not match the model's prediction scale. In practical applications, this leads to problems such as feature loss and high false negative rates, affecting the model's performance in complex scenes. Based on the original three prediction scales, this invention adds a 160×160 prediction scale to the P2 layer, with the corresponding detection head being S-Head. This gives the P2 layer a smaller receptive field, enabling it to detect relatively small hot spots. At the same time, a feature fusion structure capable of bidirectional cross-scale connectivity and weighted feature map fusion operations is designed, which helps to quickly capture and fuse multi-scale features.

[0020] S23. Bounding Box Regression Loss: This measure is used to evaluate the difference between the predicted location of the detected bounding box and the actual labeled bounding box. In the YOLOv8 series of algorithms, CIoU is used as the bounding box loss function, and the CIoU calculation formula is shown in Equation (3).

[0021]

[0022] In the formula: IoU—Intersection over Union (IoU); B gt —True frame; B prd —Predicted box; ρ —Distance between the center point of the ground truth box and the center point of the predicted box; c —Diagonal length of the minimum bounding rectangle of the ground truth box and the predicted box; v —Similarity of the aspect ratio between the target box and the predicted box; α —Influence factor of v. It can be seen from formula (5) that when the aspect ratio of the predicted box and the target box are the same, v takes 0. At this time, the aspect ratio penalty term is invalid, CIoU cannot be stably expressed, which will lead to slower model convergence speed and lower efficiency, and may eventually produce a worse training model. This paper studies the actual situation of hot spot detection of photovoltaic panels and introduces a new loss function Ang-IoU. The Ang-IoU loss function further considers the vector angle between the center points of the two boxes when the predicted box regresses the ground truth box on the basis of the intersection-union ratio loss, and redefines the penalty term. Ang-IoU consists of four parts: angle loss (angle_cost), shape loss (Ω), distance loss (Δ), and overlap area (IoU). The calculation of the Ang-IoU loss function is as follows. Figure 5 As shown in the figure. Using Ang-IoU as the bounding box loss function makes the loss function converge more smoothly, which greatly improves the training and validation of the object detection algorithm model. The loss function Ang-IoU is shown in equation (7).

[0023]

[0024] The Ang-IoU loss function incorporates the aspect ratio of the predicted bounding box to the ground truth bounding box, making the loss function focus more on the shape of the bounding box. This avoids the problem of the penalty term failing when the aspect ratios of the predicted and target boxes are the same, as shown in formula (5). Attached Figure Description

[0025] To more clearly illustrate the technical solutions in this invention, the accompanying drawings used in the technical description will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 The attached figure is a flowchart of the YOLOv8 photovoltaic module hot spot defect identification method provided by the present invention;

[0027] Figure 2The attached figure shows the structure of the S-YOLO photovoltaic module hot spot surface defect detection network provided by the present invention: it consists of three parts: Spatial Confusion Backbone, MSBFPN, and 4 inspection heads;

[0028] Figure 3 The attached figure is a schematic diagram of the lightweight backbone network Spatial Confusion Backbone structure in S-YOLO provided by the present invention;

[0029] Figure 4 The attached figure is a schematic diagram of the MSBFPN structure with bidirectional cross-scale connectivity and weighted feature map fusion operation in S-YOLO provided by the present invention;

[0030] Figure 5 The attached figure is a schematic diagram showing the initial YOLOv8 model detection results and the bounding box loss of the model provided by this invention after 50-300 training rounds;

[0031] Figure 6 The attached figures show a comparison of the model detection results: Figure (b) shows the detection results of the initial YOLOv8 model, and Figure (c) shows the detection results of the model provided by this invention. Detailed Implementation

[0032] This invention discloses an automatic detection system for hot spot defects in photovoltaic modules, the flowchart of which is attached to the specification. Figure 1 As shown, the specific implementation includes the following steps:

[0033] S1. Use a drone to collect hot spot images and preprocess the collected image data.

[0034] S11. The dataset used in this embodiment is open-source solar cell images from the internet, totaling 1305 images.

[0035] S12. Perform data augmentation on 1305 images: Expand the dataset to 6314 images by performing data augmentation operations such as rotating, flipping, translating, adjusting brightness, and adding Gaussian noise on the hotspot images.

[0036] S13. From the expanded dataset, randomly allocate the test set, validation set, and training set in a ratio of 1:3:6.

[0037] S14. Using the labelImg image visualization annotation tool, the defective parts in the training and validation sets were manually annotated one by one and output in YOLO format. After annotation, each image was matched with a txt file containing the annotation information to form the final dataset used to train YOLO v8.

[0038] S2. Based on YOLOv8, a Spatial Confusion Backbone was designed. While streamlining the backbone network, the attention mechanism was strengthened, reducing the number of model parameters and computational complexity. A Multi-scale Weighted Feature Pyramid Network (MSBFPN) was constructed, which can effectively detect small target hotspots. S3. Ang-IoU was used as the bounding box loss function to obtain the optimized S-YOLO model, as shown in the appendix of the instruction manual. Figure 2 As shown;

[0039] The implementation process of the S21 Spatial Confusion Backbone feature extraction backbone network is shown in the appendix. Figure 3 As shown, the feature information first passes through the SPD module, where the feature map is segmented into different spatial regions, and feature fusion is performed within each region. This step expands the receptive field of the feature map while preserving important information. After spatial depth transformation, channel control is applied to the features using 1×1 non-staggered convolutional layers. Next, the feature information processed by the SPD module is input into the SpatialConfusion Backbone module, where the channels of the feature map are divided into several groups, each containing a subset of channels. This step reduces intra-group correlation and increases feature diversity. After channel grouping, a channel shuffling operation rearranges the channels within each group, promoting information exchange between groups and enhancing the representational power of the feature map. This structure significantly reduces the model size, making it easier to deploy on UAVs while meeting the need for rapid detection.

[0040] S22. The initial YOLOv8 model provides three prediction scale feature layers: 20×20 is suitable for large targets, and 40×40 and 80×80 are suitable for small and medium-sized targets. However, the hot spots on solar panels in UAV aerial photography are small (between 10×10 and 30×30 pixels), which do not match the existing prediction scales, leading to feature loss and missed detections. To improve the detection performance of small target hot spots, this paper adds a 160×160 prediction scale P2 layer to the original three prediction scales, using a smaller receptive field S-Head to better detect small targets. At the same time, the feature layers in the backbone network are optimized, and a multi-scale weighted feature pyramid network (MSBFPN) is constructed to improve the multi-scale feature fusion capability, thereby enhancing the model's detection performance for small targets.

[0041] S23. The initial YOLOv5 loss function includes bounding box loss (box_loss), confidence loss (obj_loss), and classification loss (cls_loss). CIoU is used as the bounding box loss function. When the aspect ratios of the predicted box and the target box are the same, v is 0. In this case, the aspect ratio penalty term becomes ineffective, CIoU cannot be stably represented, which will lead to slower model convergence and lower efficiency, and may ultimately result in a worse trained model.

[0042] To address the aforementioned problems, this invention introduces a novel loss function, Ang-IoU. The Ang-IoU loss function, based on the Cross-Union Ratio (CIRR) loss, further considers the vector angle between the center points of the predicted and ground truth boxes during regression, and redefines the penalty term. Ang-IoU consists of four parts: angle loss, shape loss, distance loss, and overlap area. A comparison of the loss functions is shown in the appendix to the specification. Figure 5 As shown.

[0043] S3. Using the dataset preprocessed in step S1, train the optimized S-YOLO model and perform comparative verification.

[0044] S31 loads the preprocessed dataset from step S1 and divides it into training and validation sets to ensure a balanced data distribution.

[0045] S32 initializes the optimized S-YOLO model, sets the training parameters (such as learning rate, batch size, optimizer, etc.) as shown in Table 1, and loads the pre-trained weights.

[0046] S33 uses the training set to train the S-YOLO model, employs the Ang-IoU loss function to optimize the weights, and reduces detection errors through multiple iterations.

[0047] S4 compares the detection performance of the optimal training model obtained from S3 with the original model.

[0048] S41 uses a validation set to evaluate the performance of the trained model, recording key metrics such as mAP, number of parameters, and weights, and analyzing the model's performance in small object detection.

[0049] S42 compares the optimized S-YOLO model with the baseline model, analyzing the differences in detection accuracy, model size, number of parameters, and inference speed.

[0050] Table 1 Training parameters

[0051]

[0052] This invention uses several metrics, including mean precision (mAP), weight size, and number of parameters, to evaluate the performance of the algorithm model. The mean precision (mAP) is calculated with an IoU threshold of 0.5 and takes precision and recall into account. Precision (P) refers to the proportion of samples predicted as positive by the model that are actually positive. Recall (R) refers to the proportion of samples correctly predicted as positive by the model out of all positive samples. Average precision (AP) is the area enclosed by the PR curve (P on the vertical axis and R on the horizontal axis) and the x-axis. mAP is the average AP across all categories, as shown in formula (11).

[0053]

[0054] In the formula: T P —The number of positive samples predicted by positive samples; F P —Negative samples predict the number of positive samples; F N — The number of negative samples predicted by positive samples; N — The number of categories of the target detected in this paper, AP C —The average precision of the c-th category.

[0055] Table 2 Comparison of Algorithm Results

[0056]

[0057] The S-YOLO algorithm proposed in this invention exhibits excellent performance in terms of detection accuracy and average precision, while also boasting a fast detection speed. Compared to YOLOv8, S-YOLO improves average precision by 2.3%, while effectively reducing model size.

[0058] S35. Compare the detection performance of the model before and after the improvement.

[0059] As per the instruction manual Figure 6As shown in the figure, the detection results of this paper are compared with those of the original YOLOv8 model. It is clearly observed from the comparison of the detected images that S-YOLO (the model proposed in this paper) exhibits higher accuracy and precision compared to YOLOv8. Specifically, this is reflected in the following aspects: 1) In the YOLOv8n model, overlapping bounding boxes occur in areas where hot spots are concentrated, meaning the same target is detected multiple times; 2) During the detection process, YOLOv8n encounters complex backgrounds due to power line obstruction, resulting in varying hot spot sizes and missed detections, i.e., some small hot spots are ignored; 3) In complex backgrounds, YOLOv8n mistakenly identifies interference hotspots under power lines as hot spots, leading to false detections. In conclusion, the S-YOLO model demonstrates better performance than YOLOv8n in hot spot image detection tasks, reducing issues such as overlapping bounding boxes, missed detections, and false detections, thus improving detection accuracy.

[0060] To address the hotspot defect problem in photovoltaic modules, this implementation uses a Spatial Confusion Backbone as the backbone network, making the model more lightweight. Simultaneously, a Multi-Scale Weighted Feature Pyramid Network (MSBFPN) is introduced, and a dedicated head for small target detection is added, significantly improving the network's ability to perceive small hotspots. Through these optimizations, the model's average detection accuracy reaches 97.0%. Furthermore, the Ang-IoU loss function is used instead of the original CIoU loss function, improving accuracy by 1.7%, accelerating convergence speed, and significantly improving FPS.

[0061] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A photovoltaic module hot spot defect automatic detection system, characterized by, The method comprises the following steps: data collection: the image data set used for training and verification in the application is mainly from a mountain photovoltaic power station, and infrared images of the surface of photovoltaic cells and hot spots thereof are collected by aerial photography with a drone. The data set contains hot spot images under different climate conditions and different light conditions, ensuring the diversity and representativeness of the data; data preprocessing: image size unification, random cropping, color enhancement and contrast adjustment, etc. and data set expansion; model training: the processed data are trained by using a YOLOv8 model to generate an original fault diagnosis model; model improvement: including 1) a spatial confusion backbone network (SpatialConfusion Backbone) is constructed for lightweight design of the backbone network, and the attention mechanism is strengthened to reduce the model parameter quantity and calculation complexity. 2) a multi-scale weighted feature pyramid network (MSBFPN) is used to enhance the detection ability of small target hot spots, and the newly added small target prediction scale helps to improve the perception of small targets. 3) an Ang-IoU loss function is used to replace the traditional CIoU loss function to solve the penalty failure problem and promote the rapid convergence and accurate regression of the model.

2. The method of claim 1, wherein, The hot spot is caused by local circuit failure of a photovoltaic module, shadow shielding, cell damage, wiring problem or surface dirt, etc., resulting in abnormal temperature rise in a specific area, thereby affecting the performance of the module and possibly causing permanent damage, which appears as a distinct red area inconsistent with the surrounding temperature in the infrared image taken by the drone.

3. The method of claim 1, wherein, The data preprocessing further comprises: screening the pictures taken by the drone, eliminating pictures without defects and retaining pictures containing hot spot types. Since the hot spot usually only occupies a very small area of the picture, in order to improve the detection accuracy, the Slicing Aided Hyper Inference (SAHI) algorithm is used to process the picture. The image is divided into multiple regions by a sliding window, and each region is predicted separately, while the whole picture is inferred. Then, the prediction results of each region and the prediction results of the whole picture are combined to improve the accuracy of hot spot detection. The segmented images are labeled by using LabelImg software to generate a training set and a verification set for training the model. In order to further expand the data set, the segmented images are subjected to data enhancement operations such as rotation, flipping, translation, brightness adjustment and addition of Gaussian noise. Finally, the data set is divided into a training set, a verification set and a test set in a ratio of 6:3:1 for model training and performance verification.

4. The method of claim 1, wherein, The YOLOv8 model is optimized by the following steps. The backbone network optimization: a lightweight backbone network such as Spatial Confusion Backbone is adopted, which reduces the network parameter amount and strengthens the attention mechanism, so that the model can maintain high detection ability while reducing the computational complexity, and adapt to resource-constrained devices such as unmanned aerial vehicles. Small target detection ability enhancement: to adapt to the detection needs of small targets, a small target prediction scale is added, and a multi-scale weighted feature pyramid network (MSBFPN) is used to improve the detection ability of small targets such as hot spots. Loss function optimization: the original CIoU loss function is replaced by the Ang-IoU loss function, which solves the problem of punishment failure through more accurate boundary box regression, accelerates the convergence speed, and improves the detection accuracy.

5. The method of claim 1, wherein, The improved lightweight model device is used on the unmanned aerial vehicle to perform real-time hot spot detection of the photovoltaic power station according to the specified inspection route. The unmanned aerial vehicle is equipped with an optimized hot spot detection model, which can automatically capture infrared images of photovoltaic modules during flight and process the images through the model to identify hot spot areas. If hot spots are detected, the system will automatically mark the relevant areas and determine whether they exceed the normal threshold according to the size and severity of the hot spots. Once the hot spots are detected, the information is transmitted to the cloud in real time for further analysis and evaluation by the cloud computing platform. The cloud system can process multiple unmanned aerial vehicle detection data, summarize and visualize, and generate regular inspection reports for maintenance personnel to view. If there are serious hot spots, the system will trigger an alarm mechanism to notify or alert relevant personnel, ensuring timely handling of potential risks that may cause photovoltaic module failure or damage. This automated hot spot detection process greatly improves the efficiency and accuracy of photovoltaic power station inspection, reduces the workload of manual inspection, reduces the rate of missed detection and false detection, and also provides important data support for subsequent equipment maintenance and fault warning.

6. The method of claim 2, wherein, The backbone part, the Spatial Confusion Backbone feature extraction backbone is designed, and the feature fusion includes the following steps: the feature information is first processed by the SPD module, the feature map is divided into multiple spatial regions, and feature fusion is performed in each region to expand the receptive field and retain important information. Then, a 1x1 convolution layer is used to control the channels. The processed feature map is input into the Spatial Confusion Backbone module, the channels are divided into multiple groups to reduce the correlation within the group and improve the diversity. Finally, through the channel shuffling operation, the channels within the group are rearranged to promote information exchange between groups and enhance the performance of the feature map. Through lightweight design and strengthened attention mechanism, efficient processing of small target detection is ensured, which is especially suitable for embedded devices such as unmanned aerial vehicles.

7. The method of claim 2, wherein, The multi-scale weighted feature pyramid network (MSBFPN) improves the ability of small target detection by fusing features from different scales. Its core steps are as follows: First, the model extracts features from different scales (such as P2, P3, P4, P5, etc.). Each scale of feature map has a different receptive field, which can capture targets of different sizes. Second, MSBFPN introduces a weighting mechanism to weight and fuse features from different scales. This allows the network to adjust the weights of different feature maps according to their importance, highlighting important features and suppressing noise. Finally, after weighting, the features are fused through cross-scale connections to ensure that the network can better capture target information at different scales. In this way, the feature information of small and large targets can be effectively utilized. The fused feature map is further processed and optimized to enhance the network's ability to detect small targets (such as hot spots) and targets of different scales. The design of this multi-scale weighted feature pyramid network improves the accuracy of target detection, especially in complex scenarios for small target recognition.

8. The automatic detection system for hot spot defects of a photovoltaic module according to claim 3, wherein The implementation is achieved through the following steps: First, use the input end to complete Mosaic enhancement, adaptive anchor box calculation and image scaling operations, convert the image to a target tensor and input it to the model; then, use the Spatial Confusion Backbone backbone network to extract image features, and aggregate the feature maps through the multi-scale weighted feature pyramid network (MSBFPN) to generate target features; finally, the output layer optimized by the Ang-IoU loss function is used to predict the target detection result.

9. An operation and maintenance optimization method for photovoltaic module hot spot detection driving, characterized in that, By combining hot spot detection data and historical operation information, a predictive maintenance plan is developed, including: scheduling inspection personnel to prioritize tasks that need to be repaired based on hot spot type, detection confidence, and the impact of hot spots on component performance; dynamically adjusting operation resources based on the recommendations provided by the detection system to improve photovoltaic component power generation efficiency and optimize resource allocation; continuously introducing new hot spot cases through an online learning system to continuously enhance the adaptability and robustness of the detection system.

Images