<![CDATA[Automatic defect detection in photovoltaic (PV) panels is a crucial challenge for maintaining energy efficiency and reliability in renewable power systems. However, the limited availability of labeled datasets and high environmental variability often lead deep learning models to overfit and lose generalization capability. This study investigates the combined effects of data augmentation and dropout regularization on improving the generalization performance of transfer learning-based models for multi-class PV defect classification. Two pretrained architectures, VGG16 and InceptionV3, were fine-tuned using the Faulty Solar Panel dataset comprising six defect categories. Experiments were conducted under three main configurations: (1) baseline without regularization, (2) augmentation only, and (3) combined augmentation–dropout with dropout rates of 0.2, 0.3, and 0.5. Performance evaluation employed accuracy, precision, recall, macro-F1, and confusion matrix analysis for each defect class. The results demonstrate that the combination of data augmentation and moderate dropout (0.3) significantly enhances generalization, achieving 92.10% accuracy and 0.90 macro-F1 with the InceptionV3 architecture. Higher dropout values (0.5) caused slower convergence and decreased accuracy. These findings confirm that balanced integration of augmentation and dropout effectively mitigates overfitting and strengthens model robustness under limited and imbalanced data conditions. The proposed approach provides practical implications for implementing reliable, lightweight, and deployable deep learning-based inspection systems in real-world PV monitoring using edge computing devices.]]>
Copyrights © 2025
