Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Advertisement
Scientific Reports volume 16, Article number: 18169 (2026)
1589
1
Metrics details
Aiming at the problems of current solar photovoltaic (PV) panel defect detection methods, this paper proposes a solar PV panel defect detection and identification algorithm based on improved YOLO11. The algorithm introduces three key enhancements: (1) a Deformable Attention module that strengthens the model’s ability to extract and learn defect features, effectively reducing false detections in complex backgrounds. (2) a Multi-scale Screening-feature Fusion Pyramid Network that improves adaptability to tiny defects under multi-scale conditions, significantly enhancing detection precision. (3) Structural Reparameterization that reduces model parameters and computational complexity, facilitating deployment on resource-constrained devices. The experimental results on the PVEL-AD-2021 dataset show that the precision, recall, and mAP@0.5 improved by 1.6%, 3%, and 3.8%, respectively, compared to the classical YOLO11. Meanwhile, the number of parameters and computation complexity of the model are reduced by 32.1%, and 22.2%, respectively. The algorithm provides a reliable and practical method for detecting defects in outdoor solar PV panels, which can effectively reduce energy loss and operation and maintenance costs caused by panel defects, while simultaneously enhancing the safety of solar PV systems.
Environmental pollution is a growing concern in today’s society, and in order to respond positively to the Carbon Peak and Carbon Neutrality Policy, renewable energy (e.g., solar photovoltaic) is undoubtedly the best choice1,2,3,4. According to the International Energy Agency (IEA), solar photovoltaic (PV) systems will be able to provide 11% of global electricity by 2050, at that time, solar energy will become the largest source of electricity in the world5. Defects in PV modules can lead to a reduction in output power and even safety risks. Therefore, it is crucial to inspect and maintain them6,7. However, traditional inspection methods (e.g., image processing-based methods, electrical detection-based methods, and manual inspection) have low accuracy and efficiency, and with the increasing scale of solar PV panels, it is difficult for traditional methods to meet practical needs.
In recent years, methods based on deep learning have become the mainstream solution for detecting defects in solar PV panels due to their powerful feature capture capability and high detection accuracy8,9. Among various deep learning models, the You Only Look Once (YOLO) series has achieved remarkable success in target detection10,11,12. YOLO11, as the latest version of the series, is even more advantageous in terms of detection accuracy as well as the number of parameters13,14. However, in the actual solar PV panel defect detection scenarios, YOLO11 still has limitations such as misidentification in complex backgrounds, missed detection of tiny defects in multi-scale challenges, and excessive number of parameters and computation complexity of the model, which need to be further optimized to meet the practical applications or special needs. In this paper, the Deformable Attention (DA) module, the Multi-scale Screening-feature Fusion Pyramid Network (MSFPN), and the Structural Reparameterization (SR) technique are integrated into the model to improve it, aiming to improve the accuracy of defect detection of solar PV panels as well as reduce the number of parameters and computation complexity of the model, so as to reduce the energy loss and operation and maintenance cost due to defects of solar PV panels, and to enhance the safety of solar PV systems. In conclusion, the main contributions of this paper are as follows:
To tackle the problem of misidentification of solar PV panel defects in complex backgrounds, the DA module is introduced into the C2PSA module to enhance the model’s ability to extract and learn defect features.
For the problem of missed detection of tiny defects under multi-scale challenges, the MSFPN is incorporated into the neck to improve the model’s adaptability to tiny defects under multi-scale challenges.
To address the hardware deployment issue on some resource-constrained devices, the SR technique is applied to C3k2 to reduce the number of parameters and computation complexity of the model.
Traditional defect detection methods for solar PV panels can be mainly categorized into two types: methods based on image processing and methods based on electrical detection.
The methods based on image processing are mainly used to detect surface defects on solar PV panels by visual and thermal imaging techniques. In 2012, Tsai et al. proposed a method based on image reconstruction for the detection of linear or bar defects, but the method needs to assume a specific shape15. In 2014, Anwar and Abdullah proposed an anisotropic diffusion filter and image segmentation technique for detecting microcracks, but the method requires a higher level of image preprocessing16. In 2015, Tseng et al. proposed a binary classification method based on candidate region features for detecting broken grill defects17. In 2019, Su et al. proposed a new feature descriptor that allows for easier differentiation of defect features in a heterogeneous background18.
Conventional image processing methods for solar PV panels usually characterize defects by edge gradient information, but the gradient variation of defect edges is similar to the grain variation in complex backgrounds, making it easy to interfere when distinguishing defects from background grains. In addition, these methods are often only applicable to smaller datasets and do not have strong generalization ability.
Electrical inspection-based methods can diagnose specific faults based on different electrical characteristics. In 2016, Schuss et al. explored and assessed the impact of PV panel defects on output performance through current-voltage and power-voltage curves as well as by using simultaneous thermography measurements19. In 2022, Azkona et al. constructed a thermography and current-voltage curve based on localized defects as well as thermal breakdown detection models for PV panels20. In 2023, Sohail et al. performed power analysis on multiple models based on the severity of cracks in solar PV panels and found that the output power in the case of deep cracks is directly proportional to the size of the cracks21.
Methods based on electrical inspection usually only provide global information about defects in solar PV panels, but cannot pinpoint the exact location of the defects. Meanwhile, there may be no obvious signal changes for small localized defects (e.g., microcracks), making it difficult to detect some tiny defects.
With the continuous development of deep learning, solar PV panel defect detection has gradually become a research hotspot in deep learning. In 2024, Cao et al. proposed an improved YOLOv8 model, which improved the map@0.5 of defect detection of solar PV panels by 4.2% to 92.8% compared to the original YOLOv8 model by introducing DW-Conv, GSConv, and BiFPN structures, but the number of parameters and computation complexity of the model are still large22. In 2024, Pan et al. proposed a YOLOv5-based adaptive complementary fusion (ACF) module for detecting defects in solar PV panels, with reduced parameters and computational power compared to the benchmark model, and an increase of 5.2% in R compared to YOLOv823. In 2025, Liang et al. proposed a multipath feature-weighted fusion method that enhances feature extraction capability. Experiments on two datasets show that mAP@0.5 improves by 3%, and 5.4% to 91%, and 51%, respectively, compared to the benchmark model YOLOv724. In 2025, Tella et al. applied an integrated strategy for defect detection based on the electroluminescence images of PV panels acquired by drones25. In 2025, Xiang et al. proposed a defect detection system for rooftop solar PV panels based on the DJI Mavic Mini drone and YOLO11 model, which can automatically generate annotated maps and efficiently identify defects. In real-world scenarios, the system achieved a mAP@0.5 of 57.7%26. In 2025, Ghahremani et al. proposed using YOLOv9, YOLOv10, and YOLOv11 algorithms to detect solar panel defects. They trained the models on three datasets and compared their performance in accuracy, recall, mean average precision, and F1 score. YOLOv11-X outperformed the others, especially in mean average precision, achieving 92.7%27.
In summary, methods based on deep learning for solar PV panel defect detection and identification can automatically learn defect features and have high identification accuracy after training with a large amount of data. However, the current solar PV panel defect detection and identification methods still have shortcomings such as misidentification in complex backgrounds, missed detection of tiny defects under multi-scale challenges, and difficulty in deploying the model on resource-constrained devices. To solve the above problems, this paper proposes a solar PV panel defect detection and identification algorithm based on improved YOLO11.
This section firstly introduces YOLO11 and the proposed YOLO11-DRP before detailing the specific improvement modules in YOLO11-DRP.
YOLO11, as the latest version of the YOLO series, can quickly detect targets in images or videos with high accuracy, and has significant advantages in target detection. YOLO11 includes several basic models: n/s/m/l/x. Considering various factors, such as accuracy, number of parameters and computation complexity, YOLO11n is selected as the base model finally.
As shown in Fig. 1a, YOLO11 consists of backbone, neck, and head networks. The backbone network firstly performs feature extraction on the input image, extracting low-level to high-level image features through multiple convolutional layers. Then, the neck network fuses and enhances the features, integrating different levels of features to enable the model to detect targets at different scales. Finally, the head network outputs the final prediction based on the features extracted from the backbone and neck.
YOLO11 and YOLO11-DRP general network structure. (a) YOLO11 general network structure, (b) YOLO11-DRP general network structure.
Although YOLO11 shows overall good performance in solar PV panel defect detection, it still suffers from the following drawbacks: (1) When detecting solar PV panels whose defect features are not obvious or are similar to complex backgrounds, the detection accuracy will be greatly limited. (2) When detecting tiny defects under multi-scale challenges, the attention is not well focused on the defect box, and thus the feature information that can be learned is limited. (3) The YOLO11 benchmark model is large and requires high computational resources and large storage space, which makes it difficult to apply on some resource-constrained devices (e.g., Jetson Nano)28,29,30.
To address the above problems of YOLO11, this paper proposes the YOLO11-DRP model, in which the improved modules are highlighted in the light red area as well as in the red dashed box in Fig. 1b. Each of the improvement modules is described in detail below.
Since the features of defects such as fingers, thick lines, and cracks in the defect dataset of solar PV panels are similar to the complex background, the YOLO11 model can only learn limited defect features and therefore is prone to misidentification during detection. In order to make the model focus more attention on defects in complex backgrounds, the DA module is introduced to improve the self-attention mechanism in the C2PSA module, and the improved module is named C2PSD, as shown in Fig. 2, This improvement allows the model to effectively filter out background information for better feature extraction and learning, and the improvement has almost no effect on the number of parameters and computation complexity of the model, which significantly improves the accuracy of the model for the detection of defects in solar PV panels.
Improved C2PSD module structure.
The structure of the DA module is shown in Fig. 3, where a set of reference points on the defect map of solar PV panels (four reference points as an example, there will be more in practice) firstly predicts the offsets from the query (Q) through an offset network and obtains the focus region determined by multiple deformed sets of sampling points to guide the attention computation. Sampled features based on the offsets are then extracted from the feature map via bilinear interpolation, and these sampled features are projected as keys ((:stackrel{sim}{text{K}}))and values ((:stackrel{sim}{text{V}})) respectively via the projection matrix (M) to accurately capture the information of the important regions31:
Where (:stackrel{sim}{text{K}}), (:stackrel{sim}{text{V}}) and (:varphi:)( ; ) represent the deformed key, the deformed value, and the bilinear interpolation sampling function, respectively.
Structure of DA module.
Solar PV panel defects vary in size from tiny cracks to large black cores, so the detection of defects suffers from the problem of missed detection of tiny defects under multi-scale challenges. To address this, the neck network of the model has been improved. The improved MSFPN module structure is shown in Fig. 4. MSFPN is mainly composed of a feature selection module and a feature fusion module, the feature selection module firstly filters out the feature maps at different scales, and the feature fusion module then effectively combines the high-level semantic information with the low-level localization information to produce rich feature information, which helps to improve the detection capability of the defect detection model of solar PV panels.
Improved MSFPN module structure.
In the feature selection module, the structure of the CA module is shown in Fig. 5, which extracts the important information of each channel through global Average Pooling and Maximum Pooling, which largely reduces the information loss. In addition, to ensure the matching of feature maps at different scales, the dimension matching module uses 1 × 1 convolution (1 × 1Conv) to adjust the number of channels and ultimately optimize the feature fusion. The process of CA module can be described as follows:
CA module structure.
In the feature fusion module, the SFF module is utilized as the weights of the high-level features to filter out the semantic information in the low-level features, and the structure of the SFF module is shown in Fig. 6. When feature sampling is performed, dynamic sample (DySample) and bilinear interpolation are combined to recover the scale information of the high-level features, which can flexibly adjust and recover the features at different scales, and at the same time, adaptive sampling is performed according to the task’s needs to improve the performance of the defect detection model for solar PV panels, in addition, the solar PV panel defects can be selectively feature fusion, which effectively reduces the redundant feature information, so that the number of parameters and computation complexity of the mode can be reduced. The process can be illustrated as follows:
SFF module structure.
In the practical application of solar PV panel defect detection, the model’s excessive number of parameters and computation complexity leading to non-deployment is usually the fatal drawback of some small mobile devices. The residual block in the classical C3k2 module uses two 3 × 3 convolutions to integrate the features, which have large feature dimensions. In order to improve the hardware deployment problem caused by the excessive number of parameters and computation complexity of the model, inspired by RepViT Block, as shown in Fig. 7, the residual block in the C3k2 module is replaced by MobileNet Block, which employs two 1 × 1 convolutions as well as one 3 × 3 convolution to integrate the features, with a slightly reduced number of parameters and computation complexity, and then the SR technique is used to improve it, and the improved module is named C3k2S. The comparison of the structure of the C3k2 module and the C3k2S module without and with C3k is shown in Figs. 7 and 8, respectively.
Improved C3k2S without C3k.
Improved C3k2S with C3k.
The principle of the SR technique is shown on the right side of Fig. 7, where the depth separable convolution (3 × 3Conv) is firstly shifted upwards, and the squeezing and excitation (SE) layer relies on the spatial information interactions, so it is moved to the back of the depth separable convolution, and this structural adjustment effectively separates the token mixer from the channel mixer in the MobileNet Block. The SR technique is then further employed in the deep separable convolution layer to enhance the learning capability of the model during training, and the SR technique eliminates the computational and storage costs associated with skipping connections during inference, effectively reducing the number of parameters and computation complexity of the model, which is highly conducive to the deployment of small mobile devices.
Furthermore, in order to better integrate the feature information from the backbone network and pass it to the neck network, the Efficient Multi-scale Attention (EMA) is incorporated into the improved C3k2S in the backbone network.
In this section, firstly, the dataset as well as the experimental setup are introduced, the various indicators of the improved model and the baseline model are compared and visualized to prove the effectiveness of the improved model, and the ablation experiment and the comparison experiment are conducted, and then the results are analyzed and summarized.
This paper uses two datasets for the experiments. The PVEL-AD-2021 dataset and the panel-2 dataset. The PVEL-AD-2021 dataset, which is a public dataset jointly released by Hebei University of Technology and Beijing University of Aeronautics and Asmetronautics for testing defect detection methods for solar PV panels8. The dataset contains a total of 3979 images of various types of defects (e.g., black core, line crack, finger, star crack, and thick line) as well as a class of normal images, of which the normal images are labeled using Labelimg software. In order to ensure the fairness and impartiality of the method, the dataset is randomly divided into training, validation, and test sets according to 7:2:1, with the training set containing 2785 images, the validation set containing 796 images, and the test set containing 398 images. The detailed defect information of the PVEL-AD-2021 dataset is shown in Fig. 9.
Number of defects in each of the PVEL-AD-2021 dataset.
The panel-2 dataset is a publicly available dataset used for detecting defects in solar PV panels. It is composed of 2400 images of solar PV panels and includes three types of defects: crack, grid, and spot. This dataset is mainly used to validate the effectiveness of the model presented in this paper for detecting defects such as spot on PV panels deployed in the field.
The experiments in this paper were conducted in the environment shown in Table 1, with the following parameter settings during training: the Stochastic Gradient Descent (SGD) optimizer was selected, the momentum parameter was set to 0.937, the initial and final learning rates were 0.01 and 0.0001 respectively, the batch size was 32, and the total number of training rounds was 300 epochs.
The main evaluation metrics used in this paper are precision (P), recall (R), mean average precision (mAP), number of parameters (Paras), and computation complexity (FLOPS). The mAP@0.5 denotes the average of all the category precision mean when the IoU threshold is 0.5. The mAP@0.5–0.95 denotes the value obtained when the IoU threshold is increased from 0.5 to 0.95 in steps of 0.05 and then averaged over the average precision means at different IoU thresholds.
Where true positives (TP) are the number of positive cases correctly predicted as positive, false positives (FP) are the number of negative cases incorrectly predicted as positive, and false negatives (FN) are the number of positive cases incorrectly predicted as negative.
In order to verify the advantages of the YOLO11-DRP model proposed in this paper in terms of confusion matrix, precision (P), recall (R), mean average precision (mAP), speed, number of parameters (Paras) and computation complexity (FLOPS), a comparison with the baseline model is made.
As shown in Fig. 10, the number of true positives of the YOLO11-DRP model for the detection results of star cracks, thick lines, and normal pictures account for 90%, 89%, and 97% of the total number of results respectively, which is much higher than that of the YOLO11 model, which is 70%, 85%, and 91%, only the number of true positives of the detection results of the broken grating defects is slightly lower than those of the baseline model YOLO11, which indicates that the improved YOLO11-DRP model is able to learn the features of different defects more effectively and thus identify various solar PV panel defects more accurately.
Comparison of confusion matrices. (a) Confusion matrix for YOLO11, (b) Confusion matrix for YOLO11-DRP.
As shown in Table 2, compared with the benchmark model YOLO11, the YOLO11-DRP model improves in terms of precision, recall, mAP@0.5 and mAP@0.5–0.95. Meanwhile, the number of parameters and the computation complexity decrease significantly, with only a slight slowdown in detection speed. This indicates that YOLO11-DRP is able to identify defects in solar PV panels more accurately and has outstanding advantages in terms of hardware deployment for practical applications.
In addition, to validate the model’s performance in detecting and recognizing images with multiple defects, this paper randomly selected 500 images containing multiple defects from the dataset for experimentation. The results are shown in Table 3. As can be seen from Table 3, while the model’s performance slightly decreases when an image contains multiple different defects, the precision, recall, mAP@0.5, and mAP@0.5–0.95 still outperform the baseline model by 3.4%, 2.3%, 5.6%, and 0.3%, reaching 83.7%, 80.8%, 86.5%, and 48.7%, respectively. The detection speed is only slightly slower than that of the baseline model and the detection speed shown in Table 2, this is because, on one hand, the recognition speed of a feature map containing a large single defect (such as a black core) is slower than that of a feature map containing multiple small defects. On the other hand, when identifying multiple defects, all defect information is extracted from the same feature map, and the model processes multiple regions and defects in parallel. Therefore, increasing the number of defect types does not significantly increase computational complexity and thus does not noticeably reduce the detection speed, and it remains sufficient to meet practical requirements. These results demonstrate that the model presented in this paper continues to deliver outstanding performance even in the presence of multiple defects, making it more suitable for solar PV panel defect detection applications compared to the baseline model.
In order to show the differences between the YOLO11 model and the improved model YOLO11-DRP more intuitively, the detection results of the two models are visualized and compared below. As shown in Fig. 11, where the upper three graphs show the detection results of the YOLO11 model and the lower three graphs show the detection results of the improved YOLO11-DRP model, it can be clearly seen from Fig. 11 that compared with the improved YOLO11-DRP model, there is an obvious omission of the YOLO11 model for the thick line and crack defects.
Comparison of YOLO11 model and YOLO11-DRP model detection results visualization.
In order to further explore the advantages of the improved model YOLO11-DRP, the improved YOLO11-DRP model as well as the attention graphs of the YOLO11 model are also visualized in this section. As shown in Fig. 12, where the upper three graphs show the detection results of the YOLO11 model and the lower three graphs show the detection results of the improved YOLO11-DRP model, from Fig. 12, it can be seen that the improved YOLO11-DRP model concentrates significantly more attention within the detection bounding box than the YOLO11 model, which allows for better extraction and learning of the defective features, which further explains the improved YOLO11-DRP model shows better detection results compared to the YOLO11 model.
Comparison between YOLO11 model and YOLO11-DRP model attention visualization.
In addition, this paper also explores the advantages of the improved model YOLO11-DRP from the perspective of model sensory field, in the same layer 8 of the model backbone, the ratio of the sensory field of the two models at different stages to the whole feature map is shown in Table 4 (t represents the detection period), from which it can be seen that the sensory field of the improved YOLO11-DRP model at the stages of t = 20%, t = 30%, and t = 50% is basically no difference compared to the YOLO11 model, but at the stage of t = 99%, the ratio of the model’s receptive field reaches 94.76%, which has an obvious advantage over the 32.17% of the YOLO11 model, and it can obtain more information about the defective features of the solar PV panels and therefore will have a higher accuracy rate in the detection of defects in solar PV panels.
In this section, YOLO11n is firstly used as a benchmark model to experimentally verify the effectiveness of the proposed modules in the defect detection model of solar PV panels and to deeply analyze the contribution of each module in improving the overall detection performance of the model as well as realizing model lightweight. The results of the ablation experiments are shown in Table 5, from which it can be seen that:
After the introduction of the DA module, the precision, mAP@0.5 and mAP@0.5–0.95 of the model are improved by 2.7%, 1.7% and 0.5%, respectively, which proves that the DA module can effectively reduce the interference of the complex background information on the model, and thus be able to extract and learn the defective features in a better way.
On the basis of (1), after introducing the MSFPN module, the precision and mAP@0.5 of the model are improved by 0.9% and 0.6%, respectively, and the number of parameters and computation complexity are decreased by 26% and 15.9%, respectively, which indicates that the MSFPN module can significantly reduce the miss detection rate of the model for small defects, and effectively improve the detection accuracy. In addition, by selectively performing feature fusion, the redundant feature information is effectively reduced, so that the number of parameters and computation complexity of the model can be reduced.
On the basis of (2), after the SR technique is applied to C3k2, the recall and mAP@0.5 of the model are improved by 5.5% and 0.5%, respectively, and the number of parameters and computation complexity of the model are reduced by 9.7% and 7.5%, respectively, which proves that the SR technique enhances the model’s ability to learn defective features. Furthermore, the SR technique removes the computational and storage costs associated with skipping connections during the reasoning period, which effectively reduces the complexity of the model.
The experimental results show that the detection performance of this paper’s model on the PVEL-AD-2021 dataset is significantly better than that of the benchmark model YOLO11n, and the number of parameters and computation complexity of this paper’s model decreases significantly while maintaining a higher detection accuracy, which is a significant advantage in the application of defect detection in real solar PV panels.
Secondly, this paper conducts ablation experiments on the positions to which the EMA module can be incorporated to better enhance the model performance, and the results of the ablation experiments are shown in Table 6, which show that when the EMA module is incorporated into the 6th and 8th layers of the model at the same time, the enhancement of the model performance is more obvious, which is due to the fact that these two layers are able to extract more feature information, so that the model is able to learn the feature information more fully, and eventually improve the accuracy of the model for defect identification.
In this section, based on the PVEL-AD-2021 dataset of this paper, the improved YOLO11-DRP model is compared with existing solar PV panel defect detection methods, including other mainstream single-stage as well as two-stage target detection algorithms that pay more attention to the accuracy, and the advantages of this paper’s model are fully proved by comprehensively analyzing the experimental results, as shown in Table 7.
The mAP@0.5 of the model in this paper is 90.4%, which outperforms other single-stage and two-stage target detection methods and other models of the YOLO series, specifically, the mAP@0.5 of this model improves by 5.9%, 5.9%, 3.5%, 3%, 2.8%, and 2% compared to Faster-RCNN, SSD, RTMDet, RetinaNet, DDQ-DETR, and GFL, respectively. Moreover, the mAP@0.5–0.95 of this model improves by 3.8%, 8.4%, 2%, 0.6%, and 1.3% compared to Faster-RCNN, SSD, RTMDet, RetinaNet, and DDQ-DETR, reaching 64.7%, which is only 0.1% lower than GFL’s 64.8%. These results demonstrate the model’s excellent detection performance and fully validate the effectiveness of combining DA module, MSFPN module, and SR technique to improve the model performance.
The experimental data in the table show that the model in this paper not only has higher detection accuracy, but also the number of parameters and the computation complexity have decreased by 32.1% and 22.2%, respectively, compared with the benchmark model, which is also an outstanding advantage compared with other models, therefore it is more suitable for the defect detection of solar PV panels in real situations.
This chapter primarily focuses on designing a software system for the detection and identification of defects in solar PV panels, aiming to provide users with accurate and convenient services for defect detection and identification. The system includes functions such as user login and registration, uploading defect images of solar PV panels, and displaying detection and identification results. The fully deployed solar PV panel defect detection and identification system is shown in Fig. 13.
Figure 14 is the detection and identification results display interface. In the result display area, the image locates the approximate position of the defect, and shows the defect type and confidence level. The detailed defect information is also displayed in the lower-left corner of the interface.
After detection is complete, users can choose to save the images of the detection and identification results according to their needs. The images will be saved in the corresponding target folders, which is highly beneficial for subsequent maintenance work.
The deployed solar photovoltaic panel defect detection and identification system.
Detection and identification results display interface.
In order to validate the effectiveness of the model presented in this paper for detecting defects such as spots on PV panels deployed in the field, further validation was conducted on the panel-2 dataset. The comparison of various metrics is shown in Table 8. As can be seen from Table 8, the improved YOLO11-DRP model not only has advantages in terms of the number of parameters and the computation complexity but also shows significant improvements in precision, mAP@0.5, and mAP@0.5–0.95. Specifically, the precision and mAP@0.5–0.95 were improved by 4.2% and 2.2%, respectively, compared to YOLO11, increasing from 84.4% to 43.7% to 88.6% and 45.9%. These results demonstrate that the improved YOLO11-DRP model achieves excellent defect detection and recognition performance for deployed solar PV panels, and can be effectively applied to real-world solar PV panel defect detection scenarios.
From the test results of this paper on the PVEL-AD-2021 dataset and the panel-2 dataset, it can be found that:
The model in this paper significantly improves the detection accuracy of solar PV panel defects, especially the detection accuracy of star cracks, thick lines, and normal pictures that are difficult to recognize in the baseline model has a significant improvement, and this improvement can make the solar PV system run more stably and reduce unnecessary energy loss.
The model in this paper maintains the real-time detection capability while significantly improving the detection accuracy of solar PV panel defects, which enables timely detection and repair of defects in the production line and solar PV system, largely reducing the energy loss and operation and maintenance costs caused by solar PV panel defects, and at the same time reduces the safety risk of solar PV systems.
The model in this paper adopts the SR technique and optimizes the structure of the neck pyramid network, which significantly reduces the number of parameters and the computation complexity of the model, and reduces the model’s demand for hardware resources, making it easier to deploy and apply on resource-constrained mobile devices.
First of all, aiming at the characteristics of solar PV panels and the practical deployment problems of the model, this paper proposes a solar PV panel defect detection and identification algorithm based on the improved YOLO11, which improves the backbone’s C2PSA module, the neck pyramid network, and the C3k2 module. Experimental results on the PVEL-AD-2021 dataset show that the method proposed in this paper solves the current problems of defect detection in solar PV panels. Quantitatively, our YOLO11-DRP model exhibits improvements of 1.6%, 3%, 3.8%, and 0.5% in precision, recall, mAP@0.5, and mAP@0.5–0.95, respectively. Meanwhile, the number of parameters and computation complexity of the model are reduced by 32.1%, and 22.2%, respectively. In addition, this paper completes the defect detection and recognition on solar PV panels by deploying the improved model on the embedded device Jetson Orin NX, further verifying the model’s feasibility in practical applications. What’s more, this paper validates the improved YOLO11-DRP model for detecting defects, such as spots, on PV panels deployed in the field with the panel-2 dataset, quantitatively, our YOLO11-DRP model exhibits improvements of 4.2%, 1.6% and 2.2% in precision, mAP@0.5, and mAP@0.5–0.95, respectively, which not only verifies the model’s generalization performance but also demonstrates that the improved YOLO11-DRP model can be directly and effectively applied to real-world solar PV panel defect detection scenarios, significantly reducing operational and maintenance costs as well as energy losses caused by defects in solar PV panels.
Future research directions can explore new techniques for rare solar PV panel defects with a small number of labels to further improve the detection accuracy, detection speed, and robustness in practical application scenarios, as well as to reduce the deployment problems caused by the excessive number of parameters and computation complexity in practical applications, in order to make a sustained contribution to the solar PV system and to advance the world towards a low-carbon and environmental-friendly path of sustainable development.
The datasets used in this study have been uploaded to Zenodo and archived with a DOI: https://doi.org/10.5281/zenodo.19384937.
Han, S. et al. Harnessing technological innovation and renewable energy and their impact on environmental pollution in G-20 countries. Sci. Rep-UK. 15 (1), 2236 (2025).
Article ADS CAS Google Scholar
Sun, Y. & Liu, M. A study on the path planning and optimization of carbon peaking and carbon neutrality in the highway service area. Build. Environ. 267, 112187 (2025).
Article Google Scholar
Gado, M. G. et al. Sustainable e-fuels production, liquefaction and transport technoeconomic feasibility using renewable energy resources. Energ Convers. Manage. 345, 120373 (2025).
Article CAS Google Scholar
Wiredu, J. et al. Delving into environmental pollution mitigation: does green finance, economic development, renewable energy resource, life expectancy, and urbanization matter? Environ Dev. Sustain 1–30 (2025).
Dincer, F. The analysis on photovoltaic electricity generation status, potential and policies of the leading countries in solar energy. Renew. Sust Energ. Rev. 15 (1), 713–720 (2011).
Article Google Scholar
Akram, M. W., Li, G., Jin, Y. & Chen, X. Failures of Photovoltaic modules and their Detection: A Review. APPL. ENERG. 313, 118822 (2022).
Article Google Scholar
Polymeropoulos, I. et al. Enhancing solar plant efficiency: A review of vision-based monitoring and fault detection techniques. Technologies 12 (10), 175 (2024).
Article Google Scholar
Su, B., Zhou, Z. & Chen, H. PVEL-AD: A large-scale open-world dataset for photovoltaic cell anomaly detection. IEEE T Ind. Inf. 19 (1), 404–413 (2022).
Article Google Scholar
Su, B. et al. Deep learning-based solar-cell manufacturing defect detection with complementary attention network. IEEE T Ind. Inf. 17 (6), 4084–4095 (2020).
Article Google Scholar
Redmon, J., Divvala, S., Girshick, R. & Farhadi, A. You only look once: Unified, real-time object detection. // P IEEE / CVPR 779–788 (2016).
Wang, C. Y., Bochkovskiy, A. & Liao, H. Y. M. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. // P IEEE / CVF CVPR. 7464–7475 (2023).
Alkhammash, E. H. Multi-Classification Using YOLOv11 and Hybrid YOLO11n-MobileNet Models: A Fire Classes Case Study. Fire 8 (1), 17 (2025).
Article Google Scholar
Wei, J. et al. GFS-YOLO11: A Maturity Detection Model for Multi-Variety Tomato. Agronomy 14 (11), 2644 (2024).
Article Google Scholar
Wang, X. et al. Linear pattern detection of building groups by integrating dynamic snake convolution with YOLO11. Geocarto Int. 40 (1), 2471914 (2025).
Article ADS Google Scholar
Tsai, D. M., Wu, S. C. & Chiu, W. Y. Defect detection in solar modules using ICA basis images. IEEE T IND. Inf. 9, 122–131 (2012).
Article Google Scholar
Anwar, S. A. & Abdullah, M. Z. Micro-crack detection of multicrystalline solar cells featuring an improved anisotropic diffusion filter and image segmentation technique. EURASIP J. Image Vide. 2014(15), 1–17 (2014).
Google Scholar
Tseng, D. C., Liu, Y. S. & Chou, C. M. Automatic finger interruption detection in electroluminescence images of multicrystalline solar cells. Math. Probl. Eng. 2015 (1), 879675 (2015).
Google Scholar
Su, B., Chen, H., Zhu, Y., Liu, W. & Liu, K. Classification of manufacturing defects in multicrystalline solar cells with novel feature descriptor. IEEE T Instrum. Meas. 68 (12), 4675–4688 (2019).
Article ADS Google Scholar
Schuss, C. et al. Detecting defects in photovoltaic cells and panels and evaluating the impact on output performances. IEEE T Instrum. Meas. 65 (5), 1108–1119 (2016).
Article ADS CAS Google Scholar
Azkona, N., Llaria, A., Curea, O., Recart, F. & Detection Characterization and Modeling of Localized Defects and Thermal Breakdown in Photovoltaic Panels from Thermal Images and IV Curves. Electron. Mater. 3 (2), 154–172 (2022).
Article Google Scholar
Sohail, A. et al. Fault detection and computation of power in PV cells under faulty conditions using deep-learning. Energy Rep. 9, 4325–4336 (2023).
Article Google Scholar
Cao, Y. et al. Improved YOLOv8-GD deep learning model for defect detection in electroluminescence images of solar photovoltaic modules. Eng. Appl. Artif. Intel. 131, 107866 (2024).
Article Google Scholar
Pan, W. et al. Enhanced photovoltaic panel defect detection via adaptive complementary fusion in YOLO-ACF. SCI. REP-UK. 14 (1), 26425 (2024).
Article ADS CAS Google Scholar
Liang, Q. & Zhang, X. Research on attention weighted enhanced defect recognition algorithm for photovoltaic panel. Measurement 251, 117334 (2025).
Article Google Scholar
Tella, H. et al. Solar photovoltaic panel cells defects classification using deep learning ensemble methods. Case Stud. Therm. Eng. 66, 105749 (2025).
Article Google Scholar
Xiang, G. Y., I-Ling, A. L., Bingi, K. & Ibrahim, R. Automated Inspection and Fault Detection in Rooftop Solar PV System Using YOLO-Driven Tello Quadrotor Drone. // IEEE / IEEE International Conference on Robotics and Technologies for Industrial Automation (ROBOTHIA) 1–6 (2025). 1–6 (2025). (2025).
Ghahremani, A., Adams, S. D., Norton, M., Khoo, S. Y. & Kouzani, A. Z. Detecting defects in solar panels using the YOLO v10 and v11 algorithms. Electronics 14 (2), 344 (2025).
Article Google Scholar
Liu, Y., Wu, J. & Chen, L. MEIS-YOLO: Improving YOLOv11 for efficient aerial object detection with lightweight design. Intell. Converged Networks. 6 (2), 151–163 (2025).
Article Google Scholar
Haj Hassine, A. & Hamed, K. A Comparative Analysis of YOLOv11 Models: Impact of Object Size on Detection Performance. DiVA Portal (2025).
Luo, Y., Wu, A. & Fu, Q. MAS-YOLOv11: An improved underwater object detection algorithm based on YOLOv11. Sensors 25 (11), 3433 (2025).
Article ADS PubMed PubMed Central Google Scholar
Dong, Y. et al. An arrhythmia classification model based on vision transformer with deformable attention. Micromachines 14 (6), 1155 (2023).
Article PubMed PubMed Central Google Scholar
Ren, S., He, K., Girshick, R., Sun, J. & Faster, R-C-N-N. Towards real-time object detection with region proposal networks. IEEE T Pattern Anal. 39 (6), 1137–1149 (2016).
Article Google Scholar
Lyu, C. et al. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint (2022). arXiv:2212.07784.
Zhang, S. et al. Dense distinct query for end-to-end object detection. // P IEEE / CVF CVPR 7329–7338 (2023).
Download references
This work was supported by Sichuan Science and Technology Program under grant no. 2021JDJQ0027, and the Natural Science Foundation of China under grant no. 61875166.
School of Electrical Engineering and Electronic Information, Xihua University, Chengdu, 610039, China
Wenbo Jiang & Wang Liu
Sichuan Provincial Key Laboratory of Signal and Information Processing, Xihua University, Chengdu, 610039, China
Wenbo Jiang & Wang Liu
PubMed Google Scholar
PubMed Google Scholar
Wenbo Jiang: Conceptualization, Validation, Data curation, Writing—review and editing, Supervision, Funding acquisition. Wang Liu: Conceptualization, Methodology, Validation, Data curation, Investigation, Writing—original draft preparation.
Correspondence to Wenbo Jiang.
The authors declare no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Reprints and permissions
Jiang, W., Liu, W. Study on defect detection and identification of solar photovoltaic (PV) panels based on improved YOLO11. Sci Rep 16, 18169 (2026). https://doi.org/10.1038/s41598-026-48824-5
Download citation
Received:
Accepted:
Published:
Version of record:
DOI: https://doi.org/10.1038/s41598-026-48824-5
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.
Provided by the Springer Nature SharedIt content-sharing initiative
Advertisement
Scientific Reports (Sci Rep)
ISSN 2045-2322 (online)
© 2026 Springer Nature Limited
Sign up for the Nature Briefing: AI and Robotics newsletter — what matters in AI and robotics research, free to your inbox weekly.
