A Study on the Optimization of Fine-tuning and Dataset for Defect Detection Models in Photovoltaic Modules Images

Fault diagnosis of photovoltaic (PV) modules is essential for reliable solar energy generation.Previously, defects in solar modules were identified through manual inspection of PV module images; however, these methods are inefficient and prone to errors.To address this issue, deep-learning-based approaches have been explored to automate defect diagnosis in solar modules.Nevertheless, there remains a lack of systematic analysis on how model structure optimization and dataset composition affect classification performance.This study investigates the effectiveness of layer-wise fine-tuning and dataset integration using image classification models retrained on infrared (IR) and electroluminescence (EL) images.Specifically, we compared three AlexNet-based training strategies: (1) retraining AlexNet from scratch using only IR images, (2) fine-tuning a model pretrained on general image datasets with IR images, and (3) fine-tuning a model pretrained on general image datasets using both IR and EL images.The results indicated that the best performance was achieved by freezing the first three convolutional layers during fine-tuning, with the AlexNet model fine-tuned on the IR dataset achieving the highest classification accuracy.These findings suggest that combining fine-tuned models with selective layer freezing and tailored datasets can effectively enhance the design of defect detection systems for photovoltaic module diagnostics.