Engineering Applications of Artificial Intelligence 131 (2024) 107866 Contents lists available at ScienceDirect Engineering Applications of Artificial Intelligence journal homepage: www.elsevier.com/locate/engappai Improved YOLOv8-GD deep learning model for defect detection in electroluminescence images of solar photovoltaic modules Yukang Cao a ,1 , Dandan Pang a ,1 , Qianchuan Zhao b , Yi Yan a , Yongqing Jiang c , Chongyi Tian a ,∗, Fan Wang a , Julin Li a a Shandong Key Laboratory of Intelligent Buildings Technology, School of Information and Electrical Engineering, Shandong Jianzhu University, Jinan 250101, Shandong province, China b Department of Automation, Tsinghua University, Beijing 100018, China c MOE Key Laboratory of Deep Earth Science and Engineering and Department of Civil Engineering, College of Architecture and Environment, Sichuan University, Chengdu 610065, Sichuan province, China ARTICLE INFO ABSTRACT Keywords: Photovoltaic systems Health monitoring Defect detection Deep learning YOLOv8 Photovoltaic defect detection is an essential aspect of research on building-distributed photovoltaic systems. Existing photovoltaic defect detection models based on deep learning, such as YOLOv5 and YOLOv8, have significantly improved the accuracy of photovoltaic defect detection. However, these models are too large, and their feature extraction ability is insufficient, leading to low detection efficiency and inability to cope with the continuous evolution of defects. Therefore, this study proposes an accurate and lightweight YOLOv8 (You Only Look Once v8) GD algorithm. The algorithm is an improved version of YOLOv8, wherein DW-Conv (DepthWise-Conv) is applied to the YOLOv8 backbone network. Moreover, convolution is replaced with the GSConv (Group-shuffle Conv) and the BiFPN (bidirectional feature pyramid network) structure is added to the architecture. Several electroluminescent photovoltaic defect datasets are used to verify the effectiveness of the proposed method. The final experimental results show that the map@0.5 and map@0.5∼0.95 of YOLOv8-GD are 92.8% and 63.1%, respectively, which are 4.2% and 5.7% higher than those of the original algorithm, respectively, and the model volume is reduced by 16.7%. Thus, the proposed algorithm shows considerable potential in the field of photovoltaic defect detection. 1. Introduction can easily lead to fires and cause significant losses (Addabbo et al., 2017). Therefore, PV defect detection should be an essential part of safety inspection and maintenance. Currently,PV defect detection methods can be classified into traditional object detection methods (Juan and Kim, 2020) and object detection algorithms based on deep learning (Tang et al., 2022a, 2020). The traditional object detection algorithms utilize a sliding window approach to traverse the entire image and generate a certain number of candidate frames. The features of the candidate frames are then extracted and classified using a classification method such as a support vector machine (SVM) to obtain the results. The primary detection methods include the Viola–Jones detector (Viola and Jones, 2001), the HOG detector (Dalal and Triggs, 2005), and the part-based deformable model (DPM) (Felzenszwalb et al., 2008). Owing to the lack of an effective image representation, traditional object detection methods can only use limited computational resources to design complex feature representations for detection. Building distributed photovoltaic power generation systems (Zhu et al., 2015; Brown and Green, 2004) are the primary tools for resolving building energy resource shortage, energy conservation, and environmental protection issues; however, they are directly related to the effectiveness and safety of building electricity consumption (Gorge et al., 2013). The compact configuration of distributed photovoltaic power generation systems also places higher demands on the quality of power generation, and fault detection (Mansouri et al., 2012). An accurate assessment of the health of solar panels is of great important to ensure the smooth operation of the power system (Kun et al., 2019). However, in the actual operation of building distributed photovoltaic power generation systems, solar cells are prone to defects such as hot spots, hidden cracks, ash accumulation, and shading, and many other types of defects (Wang et al., 2019). This is particularly significant in hot spot areas where the temperature is relatively high; such defects ∗ Corresponding author. 1 E-mail address: tianchongyi86@163.com (C. Tian). Yukang Cao and Dandan Pang are co-first authors of the article. https://doi.org/10.1016/j.engappai.2024.107866 Received 7 June 2023; Received in revised form 19 October 2023; Accepted 6 January 2024 Available online 13 January 2024 0952-1976/© 2024 Elsevier Ltd. All rights reserved. Engineering Applications of Artificial Intelligence 131 (2024) 107866 Y. Cao et al. Deep learning-based defect detection methods primarily involve defect imaging and feature extraction. They rely on infrared and electroluminescent technology (Demirci et al., 2021) to realize defect imaging and use a K-means color quantization and clustering algorithm to segment and extract defect regions (Sun et al., 2013). In 2014, the R-CNN algorithm (Girshick et al., 2014) was introduced as a deep learning method for object detection. However, with the extensive development of computer vision technology, the mainstream object detection algorithms are now divided into one-stage and twostage object detection methods. The one-stage object detection algorithms usually produce candidate frames and simultaneously performs classification and boundary frame regression which is characterized by high speed but poor accuracy. Examples include the YOLO (Redmon et al., 2016) series and the SSD algorithms (Berg et al., 2015), such as the classic YOLOv3 (Redmon and Farhadi, 2018), and YOLOv4 (Bochkovskiy et al., 2020) algorithms, and then the subsequent emergence of YOLOv5, YOLOv6 (Li et al., 2022a), YOLOv7 (Wang et al., 2022) and YOLOv8 algorithms. These methods characterized by high speed but poor accuracy. The two-stage object detection algorithms first generate regional candidate frames and extract the features of each candidate frame; they finally generate position frames and predict the corresponding categories. Examples include Fast R-CNN (Girshick, 2015), SPPNet (He et al., 2014), Faster R-CNN (Ren et al., 2017), and FPN (Lin et al., 2017) algorithms. They are characterized by high accuracy but low speed. Therefore, improving the accuracy and speed of the algorithms and their deployment ability for the sparse and regional characteristics of building distributed PV defects is necessary. To identify abnormal regions, Li et al. (2020) proposed using the Kirsch operator that could reduce computational complexity and error rate. They then used a deep convolution neural network to extract defect features and a multi-classification support vector machine to assist defect detection decisions. Su et al. (2021) proposed an attention-based top-down and bottom-up architectures to accomplish multiscale feature fusion and solve the feature disappearance problem by deepening the network. This architecture known as the bidirectional attention feature pyramid network (BAFPN), enables all the pyramid layers to share similar semantic features. Zhang et al. (2013) indicated that defect images were reconstructed based on independent component analysis (ICA) to enhance local anomalies to detect defects. Fioresi et al. (2022) proposed a segmented Deeplabv3 model using the ResNet-50 backbone to identify and segment defects in silicon photovoltaic cells’ electroluminescence (EL) images. Doll et al. (2021) used photoluminescence (PL) as an outdoor characterization method for full-size module imaging to provide economical, fast, and non-destructive detection of PV defects. Akram et al. (2020) proposed an automated detection method for PV module defects using deep learning methods based on infrared thermal imaging technology. Ding et al. (2018) used migration learning to make a PV defect detection algorithm more robust. Wang et al. (2021) proposed using unsupervised learning to automatically detect defects in EL images and to statistically classify them based on electrical sources. Pratt et al. (2021) identified and segmented PV defects in EL images based on the U-Net architecture. Demirci et al. (2021) used a lightweight convolutional neural network (L-CNN) architecture to reduce the number of parameters in the model and thus enhancing its deployment capability. In general, algorithm innovation for the detection of building distributed PV defects needs to be based on accuracy, speed, and deployment ability, not only to achieve a balance between accuracy and speed but also to ensure that it can meet the requirements for deployment, particularly on unmanned aircraft and embedded devices for large-scale building-distributed PV detection and maintenance. This study makes the following contributions in the areas of accuracy, speed balance, and deployment ability. The main contributions of this study are as follows: 2. Related methods Research on solar panel defect detection can be broadly divided into two categories. One approach is to use devices such as unmanned aerial vehicles (UAVs), to acquire large-scale aerial images acquisition for defect detection. For example, Li et al. (2019) combined deep learning, data argumentation, and migration learning to propose a nondestructive vision inspection system based on UAV and edge computing to solve the problem of a lack of training samples on edge devices. Grimaccia et al. (2017) used a light UAV to collect the image of the surface conditions of the photovoltaic panels. Image post-processing technology was then utilized for rapid fault detection to minimize the impact of faults on the photovoltaic power generation performance. Tang et al. (2022b) considered the balance between defect detection performance and computational complexity and proposed the distribution of computational tasks among edge devices, edge servers, and cloud servers to perform defect detection using deep learning methods. Li et al. (2017) indicated that an image-processing algorithm based on the first-order derivative of the Gaussian function and feature matching can detect defects in aerial parts. Salimzadeh and Hammad (2017) used building information modeling (BIM) and ranging (LiDAR) techniques to build distributed PV for potential estimation. Quan et al. (2017) compressed the defect mapping image by eliminating the information redundancy in the sensing stage of defect location; that is, in the process of defect location, the photocurrent of test points was measured using an LCD to illuminate at the same time instead of point-by-point; hence, a small number of photocurrent measurements were used to solve the convex problem of positioning of defects. Xiaoxia et al. (2019) proposed using images captured by drones for defect pattern recognition. Convolution neural networks are used during machine learning to classify various forms of modular defects. The second category of methods include the use of deep learning algorithms for defect detection. For example, Zhang (2023) proposed a virtual convolution method based on BottleneckCSP and YOLOv5 (GBH-YOLOv5) with tiny object prediction heads for PV panel defect detection, which increases the inference speed and reduces the number of parameters by adding prediction heads and replacing convolutions. • This paper proposes an improved method based on YOLOv8 for building distributed PV defects, named YOLOv8-GD. • YOLOv8-GD includes improved methods such as GSConv, BiFPN and DW-Conv, which shows improved accuracy and speed. • YOLOv8-GD can be used for defect detection and health monitoring in building distributed PV, which can provide a reference for future research in this field. 3. YOLOv8-GD defect detection algorithm for building distributed PV In this study, we proposed an improved defect detection algorithm based on YOLOv8 for the sparse characteristics of building distributed PV defects. After verifying the performance of YOLOv8 using the PV defect datasets, the best-performing YOLOv8s were selected for improvement and named YOLOv8-GD. The main improvement strategies include the following: (1) Application of DW-Conv to the YOLO v8 backbone network. This is used to reduce the number of parameters by 14% and improve the inference speed. (2) Utilization of the GSConv to replace the standard convolution in the YOLOv8 neck network. This better balances the accuracy and speed of the model. (3) Utilization of the BiFPN structure in the neck network. This enriches the contextual information (information around the object pixel). The specific network structure is illustrated in Fig. 1. 2 Engineering Applications of Artificial Intelligence 131 (2024) 107866 Y. Cao et al. Fig. 1. The network architecture of YOLOv8-GD. Fig. 2. The DW-Conv structure diagram. 3.1. DW-Conv (DepthWise-Conv) and computational effort involved in extracting defect information from the model backbone network. In the DW-Conv, the number of DW-Conv (Chollet, 2017) channels is equal to that of the channels (Fig. 2). Each channel is convolved separately as a group, and the results of the convolution of each group are spliced for the output. Here, output channels are input to DWConv, divided into groups with an equal number of input channels, and there is no information flow between channels. Specifically, for several convolution kernels, each convolution kernel only completes the convolution of the input single channel to acquire the feature map and then stacks a plurality of feature maps to get the output with the same input depth. However, the difference is that the spatial dimension is compressed. Briefly, the number of channels in the convolution kernel of a conventional convolution operation is the same as the number of channels in the input image. Conversely, deep convolution can be regarded as a special case of grouping convolution. By grouping the number of channels in the input picture, a convolution operation is performed, thus reducing the parameter quantity of the model and the convolution calculation cost and improving the inference speed. This module is primarily designed to solve the problems of the parameters 3.2. GSConv (Group-shuffle Conv) In the CNN structure, image transmission must undergo a conversion process in which spatial feature information is gradually transferred to channel information. The feature map must be compressed in height and width during the conversion process; however, the compression process tends to lose the semantic information of defective features. If a dense convolution is used, the connections between channels must be preserved to the maximum extent possible; however, this will significantly increase the computational cost. With sparse convolution, these connections are completely severed, and critical information about defective features is easily lost. Conversely, the GSConv approach (Li et al., 2022b) (Fig. 3) preserves as many of these connections as possible. However, if used in both the backbone and neck stages, it would increase the number of network layers while exacerbating the data flow resistance and inference time. Therefore, because no transformation of the feature map is required in the neck stage, the channel dimensions reach their 3 Engineering Applications of Artificial Intelligence 131 (2024) 107866 Y. Cao et al. Fig. 3. The GSConv structure diagram. maximum. GSConv is used to process the spliced feature map obtained from the backbone and reduce redundant repetitive information. It increases the inference speed while ensuring feature extraction accuracy; moreover, it reduces the number of parameters and computational costs. 3.3. BiFPN (bidirectional feature pyramid network) The BiFPN proposed by EfficientDet (Tan et al., 2020) is a weighted bidirectional feature pyramid network. In contrast to previous feature fusion networks, the BiFPN does not treat input features equally during the fusion process. It allocates additional weights to each input feature, enabling the network to learn the importance of each input feature. The BiFPN has branches from top to bottom and from bottom to top. Moreover, it increases the direct jump connection between the input and output nodes; thus, the feature information (context information) around the target pixel can be fused through various branches and jump connections to extract features. The BiFPN structure used in this study was provided by YOLOv5 officials, with learnable weight parameters set to handle the input features. Unlike previous feature fusion networks, the BiFPN no longer treats input features equally during the fusion process, but allocates additional weights to each input feature, allowing the network to learn the importance of each input feature. Fig. 4. The defect data set after data enhancement. Defects types such as black core sheet, black spot sheet, and hot spot can similarly affect the cell’s performance; hence, these types were uniformly labeled as spots for easy detection (Cao et al., 2023). The total defect types were divided into cracks and spots (Fig. 5). The samples were labeled using labelimg1.8.3 dataset labeling software. Finally, there is no fixed basis for the division between training and validation or test sets. Generally, it is divided according to the number of features in the dataset; however, if the conditions permit, more images can be used for training. In this study, the division ratio of 9:1 was selected for the experiment. And the experimental results in Section 4.3 also showed that the training results were not significantly different from the test results, which fully meets the relevant requirements of object detection algorithms. 4. Experiments 4.1. Data set and preprocessing The data set (Buerhop et al., 2018) used in this study was presented at the 35th European Photovoltaic Solar Energy Conference and Exhibition. Images of solar cells extracted from high-resolution electroluminescence images of photovoltaic modules were thus provided at the conference. The data set comprised 44 PV modules, including 18 single crystals and 26 polymorphic modules. The images could be split into 2624 EL images of 300 × 300-pixel cells. These images contained common internal and external defects such as black core sheets, black spot sheets, short circuit black sheets, over-soldered sheets, broken grids, bright and dark sheets, hidden cracks, etc. A total of 1550 defect images were selected from this dataset (Fig. 4). The dataset was expanded using data enhancement methods, such as random flip, rotation, translation transformation, and scale transformation, resulting in a final dataset of 4530 defect images. 4.2. Experimental parameter setting and model training To make the YOLOv8-GD algorithm more effective, all of the training and testing tasks described in this paper were performed on a server with a single GPU (CPU: Intel(R) Core(TM) i7-9700F CPU: @3.00 GHz, RAM: 16 GB GPU: NVIDIA GeForce GTX 3060) and Python 3.8. etc. Fig. 6 shows that 16 samples were used to train the model. Each sample consisted of four images augmented by stitched data, effectively solving the small-object problem during model training. The algorithms used for comparison were YOLOv3, YOLOv5s, YOLOv6s, YOLOXS, YOLOv8n, YOLOv8s, and YOLOv8s-GS-BiFPN. The specific parameter settings are listed in Table 1. The batch size was set to 32, the input size of image was 416 × 416, and an SGD optimizer was used, after which training was performed. At the end of 4 Engineering Applications of Artificial Intelligence 131 (2024) 107866 Y. Cao et al. Fig. 5. Labeling process of data set. Fig. 6. Training visualization. Table 1 Hyperparameter setting. Parameter Settings lr0(initial learning rate) lrf(final learning rate) momentum weight_decay warmup_epochs warmup_momentum warmup_bias_lr epochs optimizer Mosaic hsv_h(HSV-Hue augmentation) hsv_s(HSV-Saturation augmentation) hsv_v(HSV-Value augmentation) 0.01 0.01 0.937 0.0005 3.0 0.8 0.1 300 SGD 1.0 0.015 0.7 0.4 score, and PR curve. The area under the PR curve (Davis, 2006) was used to evaluate the model’s accuracy, where P represents precision and R represents recall. The area enclosed by the PR curve is the AP, and the average AP of all categories is the map, as shown in Eq. (1). 𝑇𝑃 𝑇𝑃 + 𝐹𝑃 𝑇𝑃 𝑅𝑒𝑐𝑎𝑙𝑙 = 𝑇𝑃 + 𝐹𝑁 𝑃 𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = (1) where TP represents the number of correctly detected ground-truth boxes, including the number of defects in the ground-truth boxes; FN represents the number of undetected ground-truth boxes and FP represents the number of incorrect bounding boxes. This experiment uses the YOLOv8s algorithm as a baseline for improvement. The comparison algorithms used include the classic YOLOv3-tinyu (the simplified version of YOLOv3), YOLOv5s (the smallest version of YOLOv5 in size), YOLOv6s (the lightest version of YOLOv6), YOLOv8n (the smallest version of YOLOv8), YOLOv8s-DCNv2 (which replaces the Conv module with DCNv2), YOLOv8s-ConvNext (which uses ConvNext as a backbone network), YOLOv8s-GS-SPD (wherein SPDConv is inserted into the backbone and GSConv replaces the Conv module), YOLOv8s-GS-SimAM (wherein GSConv replaces convolution and the SimAM attention mechanism is added). Other algorithms are obtained after training. The PR curves of training process are shown in Fig. 7. The parameters floating point operations (FLOPs), mean of average precision (MAP), and frames per second (FPS) were taken as the model performance metrics of this experiment, in which the parameters refer to the total number of parameters to model training, metrics such as model layers, number of parameters, computation complexity (FLOPs), map@0.5, and map@0.5∼0.95 were estimated to demonstrate the model’s performance during the training process. 4.3. Model validation and results At the end of model training, the weights of the last epoch of the training process and the best training weights were obtained. Generally, the best training weights are used as the output weights of the model for validation. The validation process generated a confusion matrix, F1 5 Engineering Applications of Artificial Intelligence 131 (2024) 107866 Y. Cao et al. Fig. 7. Comparison of PR curves for different models. Table 2 Test results comparison of different models. Method FLOPs (G) Params (M) Layers Map@0.5 Map@0.5 ∼ 0.95 YOLOv3-tinyu YOLOv5s YOLOv6s YOLOv8n YOLOv8s YOLOv8s-DCNv2 YOLOv8s-ConvNext YOLOv8s-GS-SPD YOLOv8s-GS-SimAM YOLOv8s-GS-ConvNext YOLOv8s-GS-BiFPN YOLOv8s-GS-Repvgg-SPD YOLOv8s-GS-DW YOLOv8s-GS-slim YOLOv8s-GS-DW-slim YOLOv8s-GS-DW-SPD YOLOv8-GD 19.9 15.8 57.3 8.1 28.4 25.4 23.7 7.3 26.8 21.6 26.4 8.4 24.3 23.0 19.3 6.5 24.5 12.13 7.02 18.75 3.00 11.13 11.24 9.06 10.13 10.19 7.99 10.06 11.08 9.20 8.88 7.32 9.35 9.27 63 213 446 168 168 168 198 190 191 219 189 210 174 266 266 175 174 0.834 0.865 0.906 0.896 0.886 0.875 0.821 0.862 0.901 0.886 0.928 0.796 0.916 0.896 0.894 0.8 0.928 0.486 0.487 0.558 0.558 0.574 0.55 0.467 0.509 0.589 0.563 0.64 0.437 0.613 0.575 0.578 0.437 0.631 be trained in model training, which is used to describe the size of the model. FLOPs refer to the floating-point operands and the calculation amount of the model; this is used to measure the complexity of the algorithm or model. MAP refers to averaging the AP values of all classes to describe the accuracy of the model prediction. FPS refers to the number of frames transmitted per second and is used to measure the model’s speed (Zhang et al., 2015). Fig. 7(a) shows a comparison of the algorithms in the YOLO series, where the YOLOv8s algorithm has the best training performance on this data set, followed by YOLOv6s, YOLOv3-tinyu, and YOLOv5s. This indicates that the YOLOv8s algorithm performs well in distinguishing between positive and negative training samples while also aggregating positive samples to a high confidence level to increase coverage, that is, the area enclosed by the PR curve and coordinate axis. In Fig. 7(b), the curve coverage of YOLOv8-GD is slightly higher than that of YOLOv8s, whereas the curve coverage of YOLOv8s is higher than that of YOLOv8n, which indicates that the training performance of the proposed YOLOv8-GD algorithm is better than that of the comparison algorithm mentioned in Table 2. The experiment validated the classical YOLOv3-tinyu and YOLOv5s algorithms (Table 2), both of which effectively reduced the computational complexity and number of parameters while achieving the accuracy of 83.4% and 86.5% on the data set, respectively. This confirmed the generality of the simplified classical algorithms in the field of object detection. YOLOv6s also shows good potential owing to its improved network structure, achieving an accuracy of 90.6% on this data set; however, its computational complexity and number of parameters are considerable, which is not conducive to later quantitative deployment. By contrast, YOLOv8n and YOLOv8s achieved high accuracy and lightweight. Therefore, this study mainly focused on the improvement of YOLOv8s. Comprehensive experimental validation of various improvement methods was conducted by replacing the modules, replacing backbone, and Neck, including the more popular ConvNext, DCNv2, GSConv, SPD-Conv, and other modules. Although these improvements showed promising results in terms of accuracy, the number of parameters, and complexity, they could still not balance model accuracy and speed. The best results were achieved by embedding GSConv and DWConv into the network and using the BiFPN structure on the experimental data set;this network was named YOLOv8-GD. The map@0.5 and map@0.5∼0.95 of YOLOv8-GD achieved 92.8% and 63.1%, respectively, a 4.2% and 5.7% improvement compared to the original algorithm. The FLOPs and parameters were reduced by 13.7% and 16.7% respectively, thus achieving the ideal results. The final detection results are shown in Fig. 8. The prediction box represents the actual application results of the model, the red and the pink boxes represent the positions of the two types of defects detected by the model, and the numbers represent the probability of such defects. The higher the probability, the more accurate is the defect detection. The inference speed is also an integral aspect of the algorithm performance quantification, This experiment used the FPS to characterize the inference speed of the model. The formula for calculating FPS is as shown in Eq. (2): 𝐹𝑃𝑆 = 1000 𝑝𝑟𝑒 + 𝑖𝑛𝑓 𝑒𝑟𝑒𝑛𝑐𝑒 + 𝑝𝑜𝑠𝑡 (2) where pre is the preprocessing time of the model, including the time consumed by image scaling, padding, and channel transformation. Inference is the inference time of the model, which refers to the time from the preprocessed image input to the model output result. Post is the post-processing time of the model, which refers to the time spent on the line conversion of the model output results. FPS is the inference speed of the model. 6 Engineering Applications of Artificial Intelligence 131 (2024) 107866 Y. Cao et al. Fig. 8. Visualization of the detection process. Fig. 9. PVEL-AD data set. Table 3 Comparison of inference speed of different models. Model Pre-process (ms) Inference (ms) Post-process (ms) FPS YOLOv3-tinyu YOLOv5s YOLOv6s YOLOv8n YOLOv8s YOLOv8s-DCNv2 YOLOv8s-ConvNext YOLOv8s-GS-SPD YOLOv8s-GS-SimAM YOLOv8s-GS-ConvNext YOLOv8s-GS-BiFPN YOLOv8s-GS-Repvgg-SPD YOLOv8s-GS-DW YOLOv8s-GS-slim YOLOv8s-GS-DW-slim YOLOv8s-GS-DW-SPD YOLOv8-GD 0.5 0.5 0.4 0.5 0.7 0.4 0.4 0.5 0.5 0.5 0.5 0.4 0.3 0.4 0.5 0.6 0.5 8.3 11 18.4 9.7 10.3 10.7 10.6 10.3 10.6 11.1 10.3 11.4 10.5 10.3 10.4 10.1 10.1 1.8 1.4 1.0 1.7 1.8 1.5 1.6 1.7 1.6 1.7 1.6 1.7 1.6 1.4 1.8 1.7 1.6 94.34 77.52 50.51 84.03 78.13 79.37 79.37 80 78.74 75.19 80.65 74.07 80.65 84.64 78.74 80.65 81.97 Table 3 shows a comparison of the inference speeds of different models. With approximately the same model size, YOLOv8-GD improved the inference speed to 81.97 FPS, surpassing the benchmark YOLOv5s, YOLOv6s, YOLOv8s, and various improved methods. This also shows that YOLOv8-GD ensures a balanced speed with improved model accuracy, and provides feasibility for subsequent deployment. 4.4. Performance of the proposed model on other data sets In this study, the PVEL-AD data set (Su et al., 2023) was used to demonstrate the generalizability of YOLOv8-GD further. The dataset includes crack, finger, black_core, thick_line, star_crack, corner, fragment, scratch, horizontal_dislocation, vertical_dislocation, printing_error, 7 Engineering Applications of Artificial Intelligence 131 (2024) 107866 Y. Cao et al. Table 4 Comparison of model test results. Method FLOPs (G) Params (M) Map@0.5 Map@0.5 ∼ 0.95 FPS YOLOv5lu YOLOv5mu YOLOv5nu YOLOv5su YOLOv8l YOLOv8m YOLOv8n YOLOv8s YOLOv8-GD 134.7 64 7.1 23.9 164.9 78.7 8.1 28.5 24.1 50.68 23.89 2.39 8.69 41.60 24.65 2.87 10.61 9.27 0.925 0.908 0.901 0.907 0.917 0.923 0.916 0.93 0.926 0.557 0.587 0.559 0.578 0.581 0.589 0.573 0.58 0.585 54.64 68.03 90.09 80 53.48 68.97 80.65 72.99 75.19 GSConv BiFPN Map@0.5 Map@0.5 ∼ 0.95 FPS ✓ 0.886 0.927 0.923 0.903 0.916 0.928 0.906 0.574 0.628 0.625 0.599 0.613 0.64 0.601 78.13 78.13 80 81.3 80.65 80.65 76.34 Table 5 Ablation experiment. DW-Conv ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ short_circuit, and 12 other defects (Fig. 9). In this experiment, 4500 defect images with a resolution of 1024 × 1024 were selected from the original data as datasets. The training and verification sets were divided according to the ratio of 9:1, in which the training set included 4051 defect images and the verification set included 449 defect images. The algorithm’s performance cannot be demonstrated using a single dataset, and the defect category and quantity of the dataset used in this study are insufficient. Therefore, this study evaluated the algorithm’s performance on more larger datasets. As shown in Table 4, the accuracy, model size, FPS, and complexity of the YOLOv8-GD algorithm still show promising results in the same experimental environment, fully demonstrating the proposed algorithm’s generalization ability. 5. Discussion The improved YOLOv8-GD algorithm comprises three main modules: DW-Conv, GSConv, and BiFPN. In this study, the mechanisms of action of these three modules were experimentally investigated to illustrate the advantages of YOLOv8-GD. The results are summarized in Table 5. After adding DW-Conv, the model’s accuracy improved by 4.1% and 5.4%, respectively, but the inference speed was the same as that of the original model. This indicates that DW-Conv improves the feature extraction capability of the backbone network while ensuring convolution performance, resulting in a more defect-informed feature representation. The experimental results show that embedding GSConv into the neck network with less redundant information maximizes its effectiveness, with a more significant increase in accuracy and speed. In particular, the connections between channels retained by GSConv ensure the model’s accuracy while reducing the number of parameters, enabling the network to obtain richer information on defect features. The addition of the BiFPN structure improved the model accuracy by 1.7% and 2.5%, respectively, and the inference speed with improved feature fusion capability, demonstrating that BiFPN enables neck networks to focus on the contextual information of defective features and ensure a balance between model accuracy and inference speed. Fig. 10 shows a plot of the PR curve of YOLOv8s after inserting the three modules. The accuracy of the improved method is higher than that of the original YOLOv8s algorithm after the insertion of the three modules separately. Fig. 11 shows a heat map visualization after adding the three modules. The insertion of DW-Conv and GSConv improve the model’s accuracy more because the convolution operation is used more in the Fig. 10. PR comparison graph for the three modules. algorithm, which has a more far-reaching influence. The BiFPN structure, conversely, focuses on improving the feature fusion capability. Overall, DW-Conv, when used to replace the standard convolution, improves the extraction capability of the backbone network, thus increasing the focus on defective feature information. In contrast, the neck network uses the GSConv and BiFPN structures to fuse further the multiscale feature information extracted from the backbone network and balance the accuracy and speed of the model while preserving the connectivity between channels. The advantages of YOLOv8-GD in model size, accuracy, and speed are illustrated. Additionally, the interpretability (Duval, 2019) of a model is an important aspect that requires attention. Establishing an interpretable photovoltaic defect detection model (Nauck, 2019) while maintaining a high level of learning performance (such as prediction accuracy) is the next direction of study. In this process, balancing the model interpretability and performance according to the principles of artificial intelligence, fairness, privacy, and data fusion is necessary. 6. Conclusions This study proposes a defect detection algorithm based on the improved YOLOv8 for detecting sparse and hard to distinguish distributed PV defects; the algorithm mainly included DW-Conv, GSConv, and BiFPN structures. The improved strategy proposed in this study were 8 Engineering Applications of Artificial Intelligence 131 (2024) 107866 Y. Cao et al. Fig. 11. The Grad-CAM heat map of three modules. Yongqing Jiang: Supervision. Chongyi Tian: Writing – review & editing. Fan Wang: Software. Julin Li: Software. validated by creating a dataset containing 4530 defect images. The effectiveness of the proposed method was confirmed through model training and validation. PVEL-AD datasets with more types and abundant defects were used to verify the robustness of the YOLOv8-GD algorithm to reflect the generalization ability of the model. The accuracy and speed of YOLOv8GD were still improved, demonstrating the potential and popularization value of YOLOv8-GD in photovoltaic defect detection. The final experimental results show an improvement in detection accuracy and speed. To our knowledge, most current practical applications in this field sacrifice accuracy to improve detection speed. This algorithm’s improved accuracy and detection speed can meet the actual deployment requirements. In the next step, we will deploy the algorithm to unmanned aerial vehicles and other carriers to monitor the distributed photovoltaics of buildings. This will facilitate the reduction of energy consumption of buildings and improve the quality of power generation. Declaration of competing interest We wish to confirm that there are no known conflicts of interest associated with this publication and there has been no significant financial support for this work that could have influenced its outcome. We confirmed that we had reviewed the final manuscript and approved its publication. To the best of our knowledge and belief, this manuscript has not been published in whole or in part, nor has it been considered for publication elsewhere. We confirm that the manuscript has been read and approved by all named authors and that there are no other persons who satisfied the criteria for authorship but are not listed. We further confirm that the order of authors listed in the manuscript has been approved by all of us. We confirm that we have given due consideration to the protection of intellectual property associated with this work and that there are no impediments to publication, including the timing of publication, with respect to intellectual property. In so doing we confirm that we have followed the regulations of our institutions concerning intellectual property. CRediT authorship contribution statement Yukang Cao: Conceptualization, Methodology, Writing – original draft. Dandan Pang: Data curation, Software, Validation. Qianchuan Zhao: Supervision, Validation. Yi Yan: Visualization, Investigation. 9 Engineering Applications of Artificial Intelligence 131 (2024) 107866 Y. Cao et al. We understand that the Corresponding Author is the sole contact for the Editorial process (including Editorial Manager and direct communications with the office). He/she is responsible for communicating with the other authors about progress, submissions of revisions and final approval of proofs. Grimaccia, F., Leva, S., Niccolai, A., 2017. PV plant digital mapping for modules’ defects detection by unmanned aerial vehicles. Iet Renew. Power Gener. 11 (10), 1221–1228. http://dx.doi.org/10.1049/iet-rpg.2016.1041. He, K., Zhang, X., Ren, S., Sun, J., 2014. Spatial pyramid pooling in deep convolutional networks for visual recognition. In: IEEE Transactions on Pattern Analysis & Machine Intelligence. pp. 1904–1916. http://dx.doi.org/10.1007/978-3-319-105789_23. Juan, R., Kim, J., 2020. Photovoltaic cell defect detection model based-on extracted electroluminescence images using SVM classifier. http://dx.doi.org/10. 1109/ICAIIC48513.2020.9065065. Kun, D., Fudong, C., Yuanliang, L., 2019. Research on the health status of photovoltaic system based on Hilbert-Huang transform and grey relational degree. Renew. Energy Resour. Li, C., Li, L., Jiang, H., Weng, K., Geng, Y., Li, L., Ke, Z., Li, Q., Cheng, M., Nie, W., Li, Y., Zhang, B., Liang, Y., Zhou, L., Xu, X., Chu, X., Wei, X., Wei, X., 2022a. YOLOv6: a single-stage object detection framework for industrial applications. arXiv:2209.02976. Li, H., Li, J., Wei, H., Liu, Z., Zhan, Z., Ren, Q., 2022b. Slim-neck by GSConv: A better design paradigm of detector architectures for autonomous vehicles. arXiv preprint arXiv:2206.02424. Li, X., Li, W., Yang, Q., Yan, W., Zomaya, A.Y., 2019. Building an online defect detection system for large-scale photovoltaic plants. In: The 6th ACM International Conference. Li, X., Li, W., Yang, Q., Yan, W., Zomaya, A.Y., 2020. An unmanned inspection system for multiple defects detection in photovoltaic plants. IEEE J. Photovolt. 10 (2), 568–576. http://dx.doi.org/10.1109/JPHOTOV.2019.2955183. Li, X., Qiang, Y., Chen, Z., Luo, X., Yan, W., 2017. Visible defects detection based on UAV-based inspection in large-scale photovoltaic systems. IET Renew. Power Gener. 11 (10), 1234–1244. http://dx.doi.org/10.1049/iet-rpg.2017.0001. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S., 2017. Feature pyramid networks for object detection. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition. CVPR. Mansouri, A., Zettl, M., Mayer, O., Lynass, M., Burhenne, R., 2012. Defect detection in photovoltaic modules using electroluminescence imaging. Nauck, D., 2019. Responsible AI. J. Inst. Telecommun. Prof. (Apr./Jun. Pt.2), 13. Pratt, L., Govender, D., Klein, R., 2021. Defect detection and quantification in electroluminescence images of solar PV modules using U-net semantic segmentation. Renew. Energy 178, http://dx.doi.org/10.1016/j.renene.2021.06.086. Quan, L., Xie, K., Xi, R., Liu, Y., 2017. Compressive light beam induced current sensing for fast defect detection in photovoltaic cells. Sol. Energy 150 (jul.), 345–352. http://dx.doi.org/10.1016/j.solener.2017.04.059. Redmon, J., Divvala, S., Girshick, R., Farhadi, A., 2016. You only look once: unified, real-time object detection. Comput. Vis. Pattern Recognit. Redmon, J., Farhadi, A., 2018. YOLOv3: an incremental improvement. http://dx.doi. org/10.48550/arXiv.1804.02767, arXiv e-prints. Ren, S., He, K., Girshick, R., Sun, J., 2017. Faster R-CNN: towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39 (6), 1137–1149. http://dx.doi.org/10.1109/TPAMI.2016.2577031. Salimzadeh, N., Hammad, A., 2017. High-level framework for GIS-based optimization of building photovoltaic potential at urban scale using BIM and LiDAR. pp. 123–134. http://dx.doi.org/10.1061/9780784481196.012. Su, B., Chen, H., Zhou, Z., 2021. BAF-detector: an efficient CNN-based detector for photovoltaic cell defect detection. IEEE Trans. Ind. Electron. PP (99), 1. http: //dx.doi.org/10.1109/TIE.2021.3070507. Su, B., Zhou, Z., Chen, H., 2023. PVEL-AD: a large-scale open-world dataset for photovoltaic cell anomaly detection. IEEE Trans. Ind. Inform. 19 (1), 404–413. http://dx.doi.org/10.1109/TII.2022.3162846. Sun, Z.J., Xue, L., Xu, Y.M., 2013. Marginal Fisher feature extraction algorithm based on deep learning. Dianzi Yu Xinxi Xuebao/J. Electron. Inform. Technol. 35 (4), 805–811. http://dx.doi.org/10.3724/SP.J.1146.2012.00949. Tan, M., Pang, R., Le, Q.V., 2020. EfficientDet: scalable and efficient object detection. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. CVPR. Tang, W., Yang, Q., Hu, X., Yan, W., 2022a. Convolution neural network based polycrystalline silicon photovoltaic cell linear defect diagnosis using electroluminescence images. Expert Syst. Appl. (Sep.), 202. Tang, W., Yang, Q., Hu, X., Yan, W., 2022b. Deep learning-based linear defects detection system for large-scale photovoltaic plants based on an edge-cloud computing infrastructure. Sol. Energy 231, 527–535. http://dx.doi.org/10.3390/ electronics12030561. Tang, W., Yang, Q., Xiong, K., Yan, W., 2020. Deep learning based automatic defect identification of photovoltaic module using electroluminescence images. Sol. Energy 201, 453–460. http://dx.doi.org/10.1016/j.solener.2020.03.049. Viola, P.A., Jones, M.J., 2001. Rapid object detection using a boosted cascade of simple features. In: Computer Vision and Pattern Recognition, 2001. CVPR 2001. Proceedings of the 2001 IEEE Computer Society Conference on. Wang, C.Y., Bochkovskiy, A., Liao, H., 2022. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. http://dx.doi.org/10.48550/ arXiv.2207.02696, arXiv e-prints. Wang, Y., Li, J., Chen, Y., Zhou, J., Chu, J., 2019. Effects of working pressure and power onphotovoltaic and defect properties of MagnetronSputtered Sb2Se3 thin-film solar cells. Appl. Opt. 59 (4), http://dx.doi.org/10.1364/AO.382805. Data availability The authors do not have permission to share data. Acknowledgments The authors are grateful for the National Natural Science Foundation of China (No. 52007109, No. 62076150), the Natural Science Foundation of Shandong Province, China (No. ZR2022MF302) and all the reviewers for their constructive comments and valuable suggestions to improve the quality of the article. References Addabbo, P., Angrisano, A., Bernardi, M.L., Gagliarde, G., Ullo, S., 2017. A UAV infrared measurement approach for defect detection in photovoltaic plants. In: 2017 IEEE International Workshop on Metrology for AeroSpace (MetroAeroSpace). Akram, M.W., Li, G., Jin, Y., Chen, X., Ahmad, A., 2020. Automatic detection of photovoltaic module defects in infrared images with isolated and develop-model transfer deep learning. Sol. Energy 198, 175–186. http://dx.doi.org/10.1016/j. solener.2020.01.055. Berg, A.C., Fu, C.Y., Szegedy, C., Anguelov, D., Erhan, D., Reed, S., Liu, W., 2015. SSD: single shot MultiBox detector. Bochkovskiy, A., Wang, C.Y., Liao, H., 2020. YOLOv4: optimal speed and accuracy of object detection. http://dx.doi.org/10.48550/arXiv.2004.10934. Brown, A.S., Green, M.A., 2004. Impurity photovoltaic effect with defect relaxation: Implications for low band gap semiconductors such as silicon. J. Appl. Phys. 96 (5), 2603–2609. http://dx.doi.org/10.1063/1.1777394. Buerhop, C., Deitsch, S., Maier, A., Gallwitz, F., Brabec, C.J., 2018. A benchmark for visual identification of defective solar cells in electroluminescence imagery. In: 35th European Photovoltaic Solar Energy Conference and Exhibition. Cao, Y., Pang, D., Yan, Y., Jiang, Y., Tian, C., 2023. A photovoltaic surface defect detection method for building based on deep learning. J. Build. Eng. 70, 106375. http://dx.doi.org/10.1016/j.jobe.2023.106375. Chollet, F., 2017. Xception: Deep learning with depthwise separable convolutions. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition. CVPR. Dalal, N., Triggs, B., 2005. Histograms of oriented gradients for human detection. In: IEEE Computer Society Conference on Computer Vision & Pattern Recognition. Davis, J., 2006. The relationship between precision-recall and ROC curves. In: Proceedings of the 23th International Conference on Machine Learning, 2006. Demirci, M.Y., Beli, N., Gümüü, A., 2021. Efficient deep feature extraction and classification for identifying Defective Photovoltaic Module Cells in Electroluminescence Images. Expert Syst. Appl. 175 (2), 114810. http://dx.doi.org/10.1016/j.eswa.2021. 114810. Ding, S., Yang, Q., Li, X., Yan, W., Wei, R., 2018. Transfer learning based photovoltaic module defect diagnosis using aerial images. In: 2018 International Conference on Power System Technology (POWERCON). Doll, B., Hepp, J., Hoffmann, M., Schuler, R., Buerhop-Lutz, C., Peters, I.M., Hauch, J.A., Maier, A., Brabec, C.J., 2021. Photoluminescence for defect detection on fullsized photovoltaic modules. IEEE J. Photovolt. (11–6), http://dx.doi.org/10.1109/ JPHOTOV.2021.3099739. Duval, A., 2019. Explainable artificial intelligence (XAI). http://dx.doi.org/10.13140/ RG.2.2.24722.09929. Felzenszwalb, P.F., Mcallester, D.A., Ramanan, D., 2008. A discriminatively trained, multiscale, deformable part model. In: 2008 IEEE Conference on Computer Vision and Pattern Recognition. Fioresi, J., Colvin, D.J., Frota, R., Gupta, R., Li, M., Seigneur, H.P., Vyas, S., Oliveira, S., Shah, M., Davis, K.O., 2022. Automated defect detection and localization in photovoltaic cells using semantic segmentation of electroluminescence images. IEEE J. Photovolt. (12–1), http://dx.doi.org/10.1109/JPHOTOV.2021.3131059. Girshick, R., 2015. Fast R-CNN. Comput. Sci. http://dx.doi.org/10.1109/ICCV.2015. 169. Girshick, R., Donahue, J., Darrell, T., Malik, J., 2014. Rich feature hierarchies for accurate object detection and semantic segmentation. IEEE Comput. Soc. http: //dx.doi.org/10.1109/CVPR.2014.81. Gorge, V., Migan-Dubois, A., Djebbour, Z., Pantzas, K., Gautier, S., Moudakir, T., Suresh, S., Ougazzaden, A., 2013. Theoretical analysis of the influence of defect parameters on photovoltaic performances of composition graded InGaN solar cells. Mater. Sci. Eng. B: Solid-State Mater. Adv. Technol. 178 (2), 142–148. http: //dx.doi.org/10.1016/j.mseb.2012.10.033. 10 Engineering Applications of Artificial Intelligence 131 (2024) 107866 Y. Cao et al. Zhang, J., Sclaroff, S., Lin, Z., Shen, X., Mech, R., 2015. Minimum barrier salient object detection at 80 FPS. In: IEEE International Conference on Computer Vision. Zhang, X., Sun, H., Zhou, Y., Xi, J., Li, M., 2013. A novel method for surface defect detection of photovoltaic module based on independent component analysis. Math. Probl. Eng. 2013 (pt.6), 147–160. http://dx.doi.org/10.1155/2013/520568. Zhu, W., Wang, Z., Yuan, X., Fan, R., 2015. Security assessment of distribution system with distributed photovoltaic. J. Power Energy Eng. 03 (04), 152–162. http://dx.doi.org/10.4236/jpee.2015.34034. Wang, Y., Li, L., Sun, Y., Xu, J., Akiyama, H., 2021. Adaptive automatic solar cell defect detection and classification based on absolute electroluminescence imaging. Energy 229, 120606. http://dx.doi.org/10.1016/j.energy.2021.120606. Xiaoxia, L., Yang, Q., Lou, Z., Yan, W., 2019. Deep learning based module defect analysis for large-scale photovoltaic farms. IEEE Trans. Energy Convers. Zhang, T., 2023. GBH-YOLOv5: ghost convolution with BottleneckCSP and tiny target prediction head incorporating YOLOv5 for PV panel defect detection. Electronics 12, http://dx.doi.org/10.3390/electronics12030561. 11
0
You can add this document to your study collection(s)
Sign in Available only to authorized usersYou can add this document to your saved list
Sign in Available only to authorized users(For complaints, use another form )