Unmanned aerial vehicle visual angle large-scale photovoltaic module infrared image defect detection algorithm model

Abstract

The invention discloses an unmanned aerial vehicle visual angle large-scale photovoltaic module infrared image defect detection algorithm model, belongs to the technical field of photovoltaic module defect detection, is named as GVT-YOLO, and is used for photovoltaic module infrared image defect detection. The GVT-YOLO mainly comprises the following three parts: a backbone network (Backkbone), a neck network (Neck) and a head network (Head). Through the mode, the invention discloses an efficient, real-time and accurate infrared thermal imaging detection method which is used for detecting the hot spot effect of a photovoltaic module in a large-scale photovoltaic power station, so that potential safety hazards caused by hot spots can be found and processed in time, and safe operation of the photovoltaic power station and long-term stability of the module are guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic component defect detection, and in particular to an infrared image defect detection algorithm model for large-scale photovoltaic components from the perspective of an unmanned aerial vehicle. Background Art

[0002] As the global demand for sustainable energy grows, the photovoltaic industry has experienced significant development. Solar photovoltaic technology has become an important means to reduce dependence on fossil fuels and mitigate climate change. However, the rapid development of photovoltaic technology has also brought challenges, especially in the quality and performance of photovoltaic modules. Photovoltaic modules, as the key to converting solar energy into electrical energy, may have defects in their production process, such as cold solder joints, broken grids and hidden cracks, which seriously affect the performance and safety of photovoltaic panels. In addition, natural environmental conditions, such as wind, sand, rain, ice and snow, may also cause physical damage to photovoltaic panels, further affecting their energy conversion efficiency and power output performance.

[0003] Therefore, defect detection technology is crucial in the production and operation of photovoltaic modules. At present, the widely used defect detection technologies in the industry include manual appearance inspection, volt-ampere (IV) curve analysis, infrared thermal imaging, photoluminescence (PL) imaging, and electroluminescence (EL) imaging. The manual appearance inspection method is simple and fast, and is suitable for the detection of defects such as spacing and foreign matter. The volt-ampere (IV) curve analysis method determines whether the module has defects by observing the attenuation of the IV characteristics. Infrared thermal imaging is a real-time non-destructive defect imaging technology that can accurately locate defects by capturing the thermal radiation generated by local overheating. It is suitable for the detection of "hot spot effect" in the operation of large-scale photovoltaic power stations. The photoluminescence (PL) imaging detection method irradiates the cell or module with a laser of a specific wavelength and receives the infrared light emitted by the recombination of electron-hole pairs. It is mainly used for defect detection in the production of cells. The electroluminescence (EL) imaging detection method is similar to the PL method in principle. It injects minority carriers to recombine and emit light by applying a forward bias, and is widely used for defect detection in the production process of components.

[0004] However, most of these technologies still have certain limitations:

[0005] 1. Limitations of manual inspection: Although manual appearance inspection is simple and quick, it has obvious limitations in detecting internal defects of photovoltaic modules, such as hidden cracks and cold solder joints. Manual inspection relies on the operator's experience, is easily affected by subjective judgment, and has low efficiency, which is not suitable for large-scale production needs.

[0006] 2. Complexity and time-consuming nature of PL and EL imaging techniques: Although PL and EL imaging techniques can provide detailed component images to reveal defects, these techniques are relatively complex in operation, require laser irradiation or applied voltage, and the imaging process is time-consuming, which is not conducive to rapid detection and real-time monitoring.

[0007] 3. Shortcomings of the IV curve method: Although the IV curve method can determine whether a component has defects, it cannot display the specific location and type of the defects, which limits its application in accurately locating defects.

[0008] 4. Low level of automation and intelligence of EL imaging: At present, the EL imaging results on crystalline silicon photovoltaic production lines are mainly judged manually, which is not only costly and inefficient, but also difficult to adapt to the needs of the operation stage of large-scale photovoltaic power stations.

[0009] 5. Challenges in detecting “hot spot effect”: In the operation of large-scale photovoltaic power stations, the detection of “hot spot effect” is particularly important, but existing technologies are insufficient in real-time, accuracy and automation, and it is difficult to meet the needs of fast and accurate detection.

[0010] 6. Equipment requirements and cost issues: Some detection technologies, such as PL imaging, have high requirements on equipment and require a stable laser light source, which increases the cost and complexity of detection.

[0011] 7. The volt-ampere (IV) curve analysis method, PL imaging detection method and EL imaging detection method are all used for defect detection in the production process of battery components. However, a serious problem currently faced is the problem of defect detection of photovoltaic components in the operation stage of large-scale photovoltaic power stations.

[0012] Photovoltaic modules are the most important component of photovoltaic power stations. During the operation of large-scale photovoltaic power stations, the safety issues of photovoltaic modules caused by the "hot spot effect" have always troubled the operators of photovoltaic power stations. The "hot spot effect" can cause problems such as spotting, streaking, and circuit breaking. If these problems are ignored or not discovered in time, they may reduce the output power of photovoltaic modules. If the heating temperature reaches a certain threshold, it will cause local burning of photovoltaic modules to form dark spots, melting of solder joints, aging of packaging materials and other permanent damage, affecting the output power and service life of photovoltaic modules, and may even cause safety hazards. Therefore, it is very important to detect the "hot spot effect" of photovoltaic modules during the operation of large-scale photovoltaic power stations.

[0013] Due to its real-time and non-destructive nature, infrared thermal imaging shows its unique value in the detection of "hot spot effect" during the operation of large-scale photovoltaic power stations.

[0014] Based on this, the present invention designs a large-scale photovoltaic module infrared image defect detection algorithm model from the perspective of a drone to solve the above problems. Summary of the invention

[0015] In view of the above-mentioned shortcomings of the prior art, the present invention provides an algorithm model for defect detection of large-scale photovoltaic modules using infrared images from a drone's perspective.

[0016] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0017] The defect detection algorithm model of large-scale photovoltaic module infrared images from the perspective of drones is named GVT-YOLO, which is used for defect detection of photovoltaic module infrared images;

[0018] The GVT-YOLO mainly includes the following three parts: backbone network (Bacbkbone), neck network (Neck) and head network (Head);

[0019] The process of GVT-YOLO is as follows:

[0020] (1) First, the input image is scaled normalized;

[0021] (2) Next, the input image passes through the backbone network to capture different levels of features in the image;

[0022] Among them, the backbone network includes VoVGSCSPC module, SPPF module and ternary attention mechanism;

[0023] (3) Then, the feature maps of each scale are input into the neck network;

[0024] Among them, the neck network includes GSConv convolution and VoVGSCSP modules;

[0025] (4) Then, it passes through the lightweight head network Gdetect;

[0026] (5) Finally, non-maximum suppression is used to remove redundant low-quality detection frames and output the final detection results.

[0027] Furthermore, in step (1), the resolution is adjusted to 640×640.

[0028] Furthermore, in step (2), different levels of features in the image are captured, including low-level detail features and high-level semantic features.

[0029] Furthermore, the present invention is used to detect defects in infrared images of photovoltaic arrays. The specific method includes the following steps:

[0030] Step 1: Data collection;

[0031] Step 2: Input the infrared image of the photovoltaic module collected by the drone into the GVT-YOLO algorithm of the present invention to detect defects in the infrared image of the photovoltaic array, and output the location and category of the defects.

[0032] Furthermore, the collection process uses the DJI M300RTK drone flight platform and the equipped DJI Zenmuse H20T gimbal camera.

[0033] In order to better achieve the purpose of the present invention, the present invention also provides

[0034] Compared with the prior art, the present invention has the following beneficial effects: the present invention discloses an efficient, real-time and accurate infrared thermal imaging detection method, which is used for detecting the "hot spot effect" of photovoltaic modules in large-scale photovoltaic power stations, so as to timely discover and deal with potential safety hazards caused by hot spots, and ensure the safe operation of photovoltaic power stations and the long-term stability of modules.

[0035] The present invention introduces the VoVGSCSPC module in the Slimneck network into the backbone network of GVT-YOLO, replaces the C2f module in the original model, improves the backbone network's ability to extract feature information and lightweights the model.

[0036] The present invention introduces the GSConv convolution and VoVGSCSP modules in the Slimneck network into the neck network of GVT-YOLO, replaces the Conv convolution and C2f modules in the original model, improves the feature fusion capability of the neck network and lightens the model.

[0037] After a large number of experimental comparisons, the present invention determined the optimal position of the ternary attention mechanism (Triplet Attention), and further improved the detection accuracy of the algorithm without almost increasing the amount of model calculation.

[0038] This paper proposes a lightweight head network: GDetect, which significantly reduces the computational complexity of the model and maintains detection accuracy.

[0039] The GVT-YOLO algorithm proposed in the present invention has high precision, few parameters and high speed. It can acquire large-scale infrared images through an infrared thermal imaging camera carried by a UAV platform, and directly perform infrared image defect detection of photovoltaic modules in the UAV onboard GPU. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0041] Figure 1 This is the structural diagram of GVT-YOLO;

[0042] Figure 2 This is a structural comparison diagram of the C2f module and the VoVGSCSPC module;

[0043] Figure 3 This is the structural diagram of the improved backbone network;

[0044] Figure 4 This is the structural diagram of the Triplet Attention mechanism;

[0045] Figure 5 This is the structural diagram of the improved neck network;

[0046] Figure 6 It is the structural diagram of GSConv module and DWConv module;

[0047] Figure 7 This is the structural diagram of the GDetect head network;

[0048] Figure 8 Schematic diagram of image acquisition for drones;

[0049] Fig. 9 To detect the defects of dot spot (DB), streak spot (TB) and open circuit (DL) in the infrared images of photovoltaic modules in large-scale photovoltaic power stations;

[0050] Fig.10 The evaluation curves of each algorithm on the photovoltaic module infrared image dataset;

[0051] Fig.11 The evaluation curves of each algorithm on the photovoltaic module infrared image enhancement dataset;

[0052] Fig.12 Visual comparison of various algorithms on the photovoltaic module infrared image dataset;

[0053] Fig.13 Evaluation curves of ablation experiments for GSConv, VoVGSCSP, and VoVGSCSPC;

[0054] Fig.14 Comparison of thermal images of ablation experiments. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] Embodiment 1: In some embodiments, please refer to the drawings of the specification Figure 1 , a large-scale photovoltaic module infrared image defect detection algorithm model from the perspective of drones, named GVT-YOLO, is used for photovoltaic module infrared image defect detection;

[0057] The GVT-YOLO mainly includes the following three parts: backbone network (Bacbkbone), neck network (Neck) and head network (Head). The process of GVT-YOLO is as follows:

[0058] (1) First, the input image is scaled and the resolution is adjusted to 640×640;

[0059] (2) Next, the input image passes through the backbone network (Backbone) to capture different levels of features in the image, including low-level detail features and high-level semantic features;

[0060] The VoVGSCSPC module replaces the C2f module in the original backbone network. The VoVGSCSPC module can more effectively fuse feature information of different scales, enabling the backbone network to better capture the contextual information of the target and improve the accuracy and robustness of target detection.

[0061] The SPPF module obtains multi-scale information of objects in the feature map through pooling operations of different scales on multiple branches, and fuses them to enhance the feature expression ability of the network;

[0062] Subsequently, the captured feature information is further extracted through the Triplet Attention mechanism.

[0063] (3) Then, the feature maps of each scale are input into the neck network (Neck);

[0064] Among them, the Conv convolution in the neck network is replaced with a more efficient GSConv convolution, and the C2f module in this part is replaced with a VoVGSCSP module;

[0065] The neck network mainly performs multi-scale feature fusion, and through a series of upsampling, downsampling and cross-scale connection operations, it fuses feature maps of different scales to obtain richer contextual information and more accurate feature representation. This multi-scale feature fusion helps the model better detect objects of different sizes and improves the accuracy and robustness of object detection.

[0066] (4) Subsequently, the lightweight head network GDetect improved the model’s convergence speed and detection accuracy.

[0067] (5) Finally, non-maximum suppression (NMS) is used to remove redundant low-quality detection boxes and output the final detection results.

[0068] The present invention discloses an efficient, real-time and accurate infrared thermal imaging detection method for detecting the "hot spot effect" of photovoltaic modules in large-scale photovoltaic power stations, so as to timely discover and deal with potential safety hazards caused by hot spots, and ensure the safe operation of photovoltaic power stations and the long-term stability of modules.

[0069] The present invention introduces the VoVGSCSPC module in the Slimneck network into the backbone network of GVT-YOLO, replaces the C2f module in the original model, improves the backbone network's ability to extract feature information and lightweights the model.

[0070] The present invention introduces the GSConv convolution and VoVGSCSP modules in the Slimneck network into the neck network of GVT-YOLO, replaces the Conv convolution and C2f modules in the original model, improves the feature fusion capability of the neck network and lightens the model.

[0071] After a large number of experimental comparisons, the present invention determined the optimal position of the ternary attention mechanism (Triplet Attention), and further improved the detection accuracy of the algorithm without almost increasing the amount of model calculation.

[0072] This paper proposes a lightweight head network: GDetect, which significantly reduces the computational complexity of the model and maintains detection accuracy.

[0073] The GVT-YOLO algorithm proposed in the present invention has high precision, few parameters and high speed. It can acquire large-scale infrared images through an infrared thermal imaging camera carried by a UAV platform, and directly perform infrared image defect detection of photovoltaic modules in the UAV onboard GPU.

[0074] Example 2 Backbone Network

[0075] In mainstream target detection algorithms, effectively extracting feature information from images is the key to the entire algorithm. However, due to the dense distribution of defects, small defect sizes, fuzzy defect features, and large environmental influences in infrared images of photovoltaic modules, it is difficult for traditional target detection networks to extract feature information more effectively.

[0076] To solve the above problems, GVT-YOLO introduces the VoVGSCSPC module in the Slimneck network. With its unique design and efficient performance, the VoVGSCSPC module is the key to improving feature extraction capabilities. It can fuse feature information of different scales and capture subtle and fuzzy features in the image, thereby achieving a more comprehensive and accurate extraction of defect features in infrared images of photovoltaic modules.

[0077] from Figure 2-3 It can be seen that in the backbone network of GVT-YOLO, multiple improved modules are applied, including multiple CBS convolution modules, multiple VoVGSCSPC modules inserted between CBS convolution modules, SPPF modules and ternary attention mechanisms; the VoVGSCSPC module includes conventional convolution blocks and GSBottleneck modules;

[0078] The first few layers of the backbone network are composed of multiple CBS convolutional modules and VoVGSCSPC modules alternating with each other. These modules first extract local features of the input image, and then further process these features through VoVGSCSPC, fusing multiple layers of features to enhance the recognition ability of complex targets.

[0079] The VoVGSCSPC module uses the more efficient GSconv convolution in the Bottleneck module and proposes the GSBottleneck module, which transmits information more efficiently and reduces the risk of gradient disappearance through two convolution operations and cross-layer residual connections. In addition, the VoVGSCSPC module also deletes the excessive concatenation operations in the C2f module and only performs one concatenation operation, which greatly simplifies the computational complexity of the module.

[0080] At the end of the backbone network, the feature map is processed by the SPPF module and then by the Triplet Attention mechanism. This design allows the network to better capture global and local information in the image. Through the combination of SPPF and Triplet Attention, the network has stronger robustness and flexibility when processing objects of different scales.

[0081] The design of the GVT-YOLO backbone network effectively combines the efficient feature extraction capabilities of the traditional YOLO model, and improves the network's processing capabilities for multi-scale and complex scenes through multiple innovative modules.

[0082] In the field of deep learning, the attention mechanism has become an indispensable key technology with its excellent performance and plug-and-play flexibility. With its unique advantages, it not only provides solid support for the optimization of neural networks, but also greatly improves the learning efficiency and accuracy of the model, enabling it to efficiently and accurately process various complex data. With the continuous innovation and progress of artificial intelligence technology, the attention mechanism has become an important research direction for improving the learning ability of neural networks and has received widespread attention from many researchers.

[0083] Due to the plug-and-play nature of the attention mechanism, a large number of experiments were conducted to explore the impact of adding the attention mechanism at different locations in the algorithm on the overall performance. First, the attention mechanism was introduced at multiple key locations, including the input layer, intermediate layer, and output layer. Second, multiple attention mechanisms were added at different locations in the algorithm at the same time. After each adjustment, a rigorous performance evaluation and effect comparison were carried out, and detailed data and results were recorded. After in-depth analysis and repeated verification, the best solution was finally determined after applying the attention mechanism to the backbone network. This discovery not only significantly improved the performance of the algorithm, but also fully demonstrated the potential of the attention mechanism. Table 1 shows the performance of the attention mechanism introduced at different locations.

[0084] Table 1 Performance comparison of different positions of attention mechanism in the validation set

[0085]

[0086] It can be found from Table 1 that not all positions are suitable for introducing the attention mechanism. Although the introduction of the attention mechanism in some positions can indeed improve performance, for example, the attention mechanism is introduced after each C2f in the backbone network, but the cost is that too many attention mechanisms will increase the number of layers of the algorithm, resulting in a slower algorithm reasoning speed. This situation shows that the introduction of the attention mechanism needs to find a balance between performance improvement and computational efficiency. Based on this discovery, the present invention finally determined to introduce the attention mechanism after the SPPF module of the backbone network. The choice of this position not only effectively improves the performance of the algorithm, but also avoids the computational burden caused by too many attention mechanisms. Through such optimization, not only the accuracy and stability of the model are improved, but also a high reasoning speed is maintained.

[0087] After a large number of experimental comparisons, the present invention introduces the Triplet Attention mechanism after the backbone network of GVT-YOLO. This strategy aims to further explore and strengthen the feature extraction ability of the model, ensuring that the model can fully capture the global and local information in the image. The Triplet Attention mechanism is a three-dimensional attention mechanism at the spatial, channel and pixel levels, which realizes multi-dimensional and refined attention to the feature map, thereby significantly improving the efficiency and accuracy of feature extraction.

[0088] The ternary attention mechanism focuses on the key areas of the image at the spatial level, and accurately emphasizes the features that are critical to target detection by calculating the spatial attention map, thereby improving the expressiveness of spatial features. At the channel level, it uses the channel attention mechanism to intelligently screen and strengthen useful feature channels, while weakening or eliminating irrelevant or redundant features, thereby enhancing the recognition and robustness of feature representation. At the pixel level, the ternary attention mechanism uses the pixel-level attention mechanism to deeply mine and finely express the details of the feature map, making feature extraction more accurate, detailed and comprehensive. Figure 4 The model structure diagram and network structure diagram of the ternary attention mechanism (MISRA D, NALAMADA T, ARASANIPALAI AU, et al. Rotate to Attend: Convolutional Triplet Attention Module[C] / / 2021IEEE Winter Conference on Applications of Computer Vision(WACV).2021:3138-3147.10.1109 / WACV48630.2021.00318.).

[0089] like Figure 4 As shown in the figure, the ternary attention mechanism (Triplet Attention) mainly consists of three branches. The first branch is the traditional channel attention mechanism, and the remaining two branches are used to calculate the interaction information between the channel C dimension and the spatial W dimension and the channel C dimension and the spatial H dimension respectively.

[0090] The first branch of Triplet Attention uses the traditional channel attention mechanism. The input features are first subjected to the Z-Pool operation, followed by a 7×7 convolution, and finally the attention weights of the spatial dimension are generated through the Sigmoid function.

[0091] The second branch of Triplet Attention mainly captures the interactive information between the channel dimension and the spatial dimension W. The input features are transformed (permute) into the H×C×W format. Then Z-Pool is applied to the spatial H dimension, and after corresponding processing, it is activated by Sigmoid, and finally the dimensions are reshaped into the channel C dimension and the spatial dimension W into the C×H×W format for subsequent element-by-element addition.

[0092] The third branch of Triplet Attention focuses on the interaction between the channel and another spatial dimension H. First, the dimensions of channel C and spatial H are exchanged (through permute), and then the Z-Pool operation is performed on the spatial dimension W. The subsequent steps are similar to the second branch. Finally, the outputs of the three branches are averaged and the final result is output.

[0093] The Z-pool layer is responsible for reducing the C-dimensional Tensor to 2 dimensions by concatenating the average pooled features and the maximum pooled features on that dimension. This allows the layer to retain the rich representation of the actual tensor while reducing its depth to make further computations lighter.

[0094] Example 3 Neck Network

[0095] The neck network connects the backbone network and the head network in the model, and improves the model's ability to detect objects of various scales by fusing and enhancing features of different scales. The neck network improves the effect of feature fusion through structures such as feature pyramid (FPN) and path aggregation network (PAN).

[0096] The present invention introduces the GSConv module in the neck network, which reduces the amount of calculation and parameters of the neck network while maintaining efficient feature extraction capabilities. The GSConv module uses the idea of ​​the Ghost module and generates redundant features using fewer convolution kernels, thereby greatly reducing the complexity and computational cost of the model without affecting the model performance. The GSConv module mainly includes two stages: standard convolution and linear transformation. First, the initial basic features are extracted from the input feature map through standard convolution, and then more feature maps are generated through simple linear transformation, thereby avoiding repeated complex convolution calculations. Thereby achieving efficient feature fusion.

[0097] At the same time, the introduction of the VoVGSCSP module further improves the efficiency and effect of feature fusion by combining spatial convolution and pyramid structure guided by variational overlap. The VoVGSCSP module uses variational methods to guide the overlap and fusion of features, enhances the expressiveness of features, and improves the fusion effect of multi-scale features by combining spatial convolution and pyramid structure. This makes the model perform better in multi-scale target detection, can more effectively capture features of different scales, and improve the accuracy and robustness of detection.

[0098] The combined use of the GSConv module and the VoVGSCSP module can improve the performance of GVT-YOLO in target detection while maintaining efficient calculations. The GSConv module reduces the amount of calculations and parameters, making the model more lightweight, while the VoVGSCSP module improves the accuracy and robustness of detection by enhancing the feature fusion capability. Therefore, the introduction of these two modules in the neck network not only improves the performance of the GVT-YOLO network, but also enables it to run efficiently in a resource-constrained environment, meeting the needs of actual applications.

[0099] The improved neck network structure and the structure of the GSConv module are as follows: Figure 5 and Figure 6 shown.

[0100] from Figure 5It can be found that GVT-YOLO consists of three parts: the backbone network (Backbone), the neck network (Neck) and the head network Gdetect. First, the input image is scaled normalized and the resolution is adjusted to 640×640. Next, the input image passes through the Backbone feature extraction network to capture features at different levels in the image. The backbone network contains CBS convolution module, VoVGSCSPC module, SPPF module and ternary attention mechanism. Among them, the VoVGSCSPC module replaces the C2f module in the original Backbone. The VoVGSCSPC module can more effectively fuse feature information of different scales, so that Backbone can better capture the contextual information of the target and improve the accuracy and robustness of target detection. The SPPF module obtains multi-scale information of objects in the feature map through pooling operations of different scales on multiple branches, and fuses them to enhance the feature expression ability of the network. Subsequently, the captured feature information is further extracted through the ternary attention mechanism. Next, the feature maps of each scale are input into the neck network (Neck), where the CBS convolution module in the Neck is replaced by a more efficient GSConv convolution module, and the C2f module in this part is replaced by a VoVGSCSP module. The Neck part mainly performs multi-scale feature fusion. In the processing of the neck network, first, each input feature map combines feature information of different levels through the Concat operation to enrich the feature expression. Then, some feature maps are improved in resolution through the Upsample operation and aligned with the high-resolution feature map for multi-scale fusion. After multiple Concat and Upsample operations, the feature maps of all levels are fully fused, ensuring that the model can capture targets of different scales.

[0101] The Neck part fuses feature maps of different scales through a series of upsampling, downsampling and cross-scale connection operations to obtain richer contextual information and more accurate feature representation. This multi-scale feature fusion helps the model better detect objects of different sizes and improves the accuracy and robustness of object detection. Subsequently, the lightweight Head detection head improves the convergence speed and detection accuracy of the model. Finally, NMS is used to remove redundant low-quality detection boxes and output the final detection results.

[0102] Example 4 Head Network GDetect

[0103] The head network performs specific object detection and classification. It is responsible for converting the extracted features into specific detection results, including information such as the location, category, and confidence score of the bounding box.

[0104] In the process of further optimizing the object detection algorithm, the author found that the computational cost and memory requirements of traditional convolutional layers increased sharply with the increase of model complexity and the resolution of input images. This situation is particularly evident in the YOLOv8 model, especially in the head network. Through experiments, the author observed that the head network of YOLOv8 accounts for 25.1% of the total parameters of the model. This has become a major bottleneck that restricts the deployment of the model in resource-constrained environments.

[0105] In order to break through this bottleneck, a new head network, GDetect, is proposed. This head network introduces group convolution. Group convolution significantly reduces computational complexity by dividing the input feature map into several small groups and performing convolution operations independently in each group. This group processing method not only retains the spatial information and expression ability of the features, but also speeds up the model's reasoning by reducing the amount of calculation.

[0106] By using group convolution, GVT-YOLO can maintain detection accuracy while greatly improving the model's operating efficiency on resource-limited devices. In addition, the introduction of group convolution further enhances the model flexibility of GVT-YOLO. By adjusting parameters such as the number of groups and convolution kernel size, the complexity and performance of the model can be flexibly controlled to meet the needs of different scenarios. The GDetect head network structure is shown in the figure. Figure 7 shown.

[0107] In the processing of the GDetect detection head, the input features are first processed by the GConv convolution in the GBS module, the feature map is batch normalized (BN) to stabilize the model training, and then the nonlinear feature representation is enhanced by the SiLU activation function. Multiple layers of GBS modules are continuously superimposed in the GDetect detection head to further enhance the extraction and fusion of multi-scale features, which is particularly suitable for capturing target objects of different sizes. Through this improved structure, GDetect not only improves parameter efficiency, but also accelerates model convergence and improves detection accuracy.

[0108] from Figure 7 It can be found that the original Detect head network structure consists of three detection branches, which are used to detect large, medium and small objects. Each branch has the same internal structure and consists of two CBS convolution modules. The CBS convolution module contains conventional convolution (Conv), batch normalization (Batch Normalization, BN) and SiLU activation function.

[0109] In the improved detection head network, the three branches of the original detection head are retained, and the GBS module is used in each branch to replace the original CBS convolution module. The GBS module replaces the conventional convolution operation through group convolution (GConv). Each improved detection branch is still composed of two convolution modules, but both layers of convolution modules use GBS. After the feature is refined through group convolution, the extracted features are finally passed to the detection layer for target recognition.

[0110] Finally, the present invention conducts actual tests on the original head network and the optimized head network of YOLOv8.

[0111] The results show that the number of parameters in the optimized head network is 0.16M, and the total number of model parameters is 1.8M, accounting for 8.8% of the total calculation. In contrast, the number of parameters in the original head network of YOLOv8 is 0.75M, and the total number of model parameters is 3.0M, accounting for 25.1% of the total calculation. Through optimization, the number of parameters in the GDetect head network is reduced by about 78.6% compared with the original head network, and the total number of model parameters is reduced by 39.2%. This means that this optimization significantly reduces the amount of calculation while maintaining a high degree of accuracy.

[0112] Embodiment 5

[0113] Please refer to the instruction manual Figure 8 , the present invention is used to detect defects in the infrared image of a photovoltaic array, and the specific method includes the following steps:

[0114] Step 1: Data collection, involving four main parts: photovoltaic arrays in natural scenes, data collection platforms carried by drones, planning of drone cruise routes, and photovoltaic array images collected by cameras. The collection process mainly relies on the DJI M300RTK drone flight platform and the DJI Zenmuse H20T gimbal camera equipped with it. The M300RTK drone has a single cruise time of 55 minutes and a flight load of 2.7 kg, ensuring the efficiency and reliability of photovoltaic array defect image collection. The H20T gimbal camera integrates a variety of sensors, including 20 million pixel zoom, 12 million pixel wide angle, 1200 meter laser ranging and 640×512 pixel thermal imaging functions, which can fully meet the needs of photovoltaic array infrared image collection.

[0115] Step 2: Input the infrared image of the photovoltaic module collected by the drone into the GVT-YOLO algorithm of the present invention to detect defects in the infrared image of the photovoltaic array, and output the location and category of the defects.

[0116] Fig. 9 To detect the dot spot (DB), streak spot (TB) and open circuit (DL) defects in the infrared images of photovoltaic modules in large-scale photovoltaic power stations.

[0117] The present invention acquires large-scale infrared images by carrying an infrared thermal imaging camera on a drone. Infrared thermal imaging is a real-time, non-destructive defect imaging technology. Infrared thermal imaging captures the thermal radiation generated by local overheating caused by current consumption at the defect site to accurately locate the defect. In addition, the technology can provide real-time image output, accelerate the detection process, and improve work efficiency. In particular, its value is more prominent for application scenarios that require rapid evaluation of a large number of photovoltaic panels. Infrared thermal imaging can also intuitively reveal hot spots and thermal effects caused by defects, which helps to detect potential electrical faults early and avoid further damage caused by thermal effects. Easy integration and automation are another advantage of infrared thermal imaging. The non-contact characteristics of infrared thermal imaging make it more convenient and safer when detecting large photovoltaic power stations. Therefore, infrared thermal imaging shows wider applicability and high efficiency in practical applications, and is particularly suitable for the detection and maintenance of large-scale outdoor photovoltaic power stations.

[0118] Experimental example

[0119] In order to fully demonstrate the performance advantages of GVT-YOLO, this paper compares and analyzes the detection performance of GVT-YOLO with five other advanced target detection algorithms on the photovoltaic module infrared image dataset and its enhanced dataset. The compared detection algorithms include YOLOX-n, YOLOv5n, YOLOv7-tiny, YOLOv8 and the most advanced YOLOv9-t.

[0120] The quantitative analysis results of each target detection algorithm on the photovoltaic module infrared image validation set and its enhanced dataset are shown in Table 2. In the photovoltaic module infrared image dataset, GVT-YOLO achieved the best recall rate and detection accuracy, with a recall rate of 85.9% and a mAP@0.5 of 87.9%. The mAP@0.5 is 1.1% higher than the second-best algorithm YOLOv9-t and 2.7% higher than the second-place YOLOv8n. GVT-YOLO has significantly improved the mAP@0.5 indicator and significantly improved the detection performance. At the same time, GVT-YOLO also achieved the best recall rate. For the photovoltaic module infrared image enhanced dataset, the performance indicators of all mainstream algorithms have been improved to varying degrees, which shows the effectiveness of the enhancement strategy and can achieve better training results with limited data resources. Among them, GVT-YOLO's mAP@0.5 is 90.2%, which is slightly lower than YOLOv9-t's 90.4%, but GVT-YOLO's FPS can reach 249, which is much higher than YOLOv9-t's 72. This is because the lightweight module and simplified head network technology introduced in GVT-YOLO have a relatively simple model structure. In contrast, although YOLOv9-t achieved the best mAP@0.5 with a huge number of 658 convolutional layers, it came at the cost of reducing the inference speed. This further proves that GVT-YOLO maintains the best balance between computing resources and inference speed, rather than blindly pursuing high mAP@0.5. Figure 10-11 The performance index curves of various mainstream algorithms on the initial data set and the enhanced data set are shown in detail. It can be seen that GVT-YOLO performs well in all indicators.

[0121] Table 2 Performance comparison of different target detection algorithms on the photovoltaic module infrared image test set

[0122]

[0123] In terms of algorithm detection performance comparison, visual analysis of detection results is more intuitive than quantitative analysis evaluation indicators. Fig.11 The results of different detection algorithms in infrared images of photovoltaic modules are visualized and analyzed in Figure 1. In order to demonstrate the balance and generalization ability of GVT-YOLO, five representative infrared images of photovoltaic modules are selected for detection.

[0124] from Fig.12It can be found that for small defects such as the first and second images, only the GVT-YOLO algorithm can correctly detect the defects of DB, and the confidence of GVT-YOLO is higher. Other algorithms have over-detection or missed detection. In the remaining three images, it can be seen that the proposed GVT-YOLO has achieved the best detection effect. The above visualization result analysis further verifies the superiority of the proposed algorithm.

[0125] In order to better explain the purpose and effectiveness of each improved strategy in GVT-YOLO, the present invention conducted an ablation experiment, which not only verified the effectiveness of different strategies, but also located the most effective position of each improved strategy. The quantitative analysis results are shown in Table 3 and Fig.13 shown.

[0126] Table 3 Ablation experiment results

[0127]

[0128] It can be seen that the three modules of GSConv, VoVGSCSP and VoVGSCSPC can all be used to lighten the algorithm. Among them, the GSConv module performs poorly in the backbone network part, the VoVGSCSPC performs well in the head network part but poorly in the backbone network part, and the VoVGSCSP performs well in the head network part but poorly in the backbone network part. Combining all experiments, it is finally selected to use the VoVGSCSPC module in the backbone network part, and the VoVGSCSP and GSConv modules in the head network part. Finally, GVT-YOLO achieved the best performance in terms of parameter quantity, recall rate and mAP@0.5.

[0129] In order to more intuitively show the actual effects of different improvement strategies, CAM technology is used to extract and visualize the characteristic heat map of each strategy. Fig.14 From the heat map in Figure 2, we can clearly see that GVT-YOLO has significantly improved its detection performance.

[0130] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. The defect detection algorithm model of large-scale photovoltaic modules infrared images from the perspective of drones is characterized by: Named GVT-YOLO, it is used for infrared image defect detection of photovoltaic modules; The GVT-YOLO mainly includes the following three parts: backbone network (Bacbkbone), neck network (Neck) and head network (Head); The process of GVT-YOLO is as follows: (1) First, the input image is scaled normalized; (2) Next, the input image passes through the backbone network to capture different levels of features in the image; Among them, the backbone network includes VoVGSCSPC module, SPPF module and ternary attention mechanism; (3) Then, the feature maps of each scale are input into the neck network; Among them, the neck network includes GSConv convolution and VoVGSCSP modules; (4) Then, it passes through the lightweight head network Gdetect; (5) Finally, non-maximum suppression is used to remove redundant low-quality detection frames and output the final detection results.

2. The large-scale photovoltaic module infrared image defect detection algorithm model from the perspective of a drone according to claim 1 is characterized in that: In step (1), the resolution is adjusted to 640×640.

3. The large-scale photovoltaic module infrared image defect detection algorithm model from the perspective of a drone according to claim 2 is characterized in that: In step (2), different levels of features in the image are captured, including low-level detail features and high-level semantic features.

4. The large-scale photovoltaic module infrared image defect detection algorithm model from the perspective of a drone according to claim 3 is characterized in that: GVT-YOLO is used to detect defects in infrared images of photovoltaic arrays. The specific method includes the following steps: Step 1: The drone collects data; Step 2: Input the collected infrared image of the photovoltaic module into GVT-YOLO and output the location and category of the defect.

5. The large-scale photovoltaic module infrared image defect detection algorithm model from the perspective of a drone according to claim 4 is characterized in that: In step 1, use the DJI M300RTK drone flight platform and the DJI Zenmuse H20T gimbal camera.

Images