This paper discusses a deep learning approach for detecting defects in photovoltaic (PV) modules using electroluminescence (EL) images. The method addresses key challenges in two practical areas: Creating high-quality EL images to overcome imbalance issues in existing datasets. This is accomplished by employing generative adversarial network (GAN) properties to generate new images. Enhancing training efficiency and performance through a one-cycle policy with optimized learning rate settings, designed to overcome hardware limitations. The research highlights that while automatic defect classification in PV modules is gaining attention as an alternative to visual/manual inspection, the process remains challenging due to the inhomogeneous nature of cell cracks and complex backgrounds in crystalline solar cells. A comparison was made between popular deep learning models (Densenet169, Densenet201, Resnet101, Resnet152, Senet154, Vgg16, and Vgg19) to assess the effectiveness of our approaches on multiple variants of our dataset. We also observe a shift in the phenomenon of moving the threshold in regression estimates because of employing a policy that uses a dynamic threshold instead of a standard threshold (0.5). We have employed two different categorizations that use binary numbers; the first employs four classes (0%, 33%, 67%, and 100%), while the second employs eight classes that are identical to four classes. However, each class has two varieties (monocrystalline and polycrystalline) and a boundary beyond which results will be obtained. Based on the performance results, it was found that the pre-trained Resnet152 model achieved the highest classification accuracy (90.13% for Datasets) of all approaches. Additionally, we have demonstrated that approaches that utilize over-sampling have the greatest performance. These findings emphasize the strength and innovation of our approach, combining advanced data augmentation, adaptive thresholding, and optimized learning strategies. The proposed system not only achieved a peak classification accuracy of 90.13% using ResNet152 but also demonstrated high robustness, reduced training time, and superior generalization across defect types and cell categories. This positions our framework as a scalable and deployment-ready solution for real-world photovoltaic quality inspection systems.
Ebied et al. (Wed,) studied this question.