Transfer Learning Evaluation of Convolutional Neural Network Architectures for Thermal Image Classification of Solar Modules

Accurate anomaly classification in solar modules is crucial for efficient solar power system monitoring and maintenance. This study explores deep learning, specifically Convolutional Neural Networks (CNNs), for automated defect classification in thermal images. We employed transfer learning with pre-trained AlexNet and GoogleNet architectures, comparing static feature extraction and fine-tuning. Using a dataset of thermal images categorized as no anomaly, cracked, or soiled modules, the fully fine-tuned GoogleNet model achieved the highest classification accuracy 95.5%, outperforming AlexNet and traditional feature-based methods. GoogleNet's superior performance stems from its Inception modules, which learn more discriminative features by employing multiple convolutional filter sizes within the same layer, capturing complex thermal image patterns. Fine-tuning further boosts performance by adapting pre-trained knowledge to the specific task. These results demonstrate deep learning's potential for enhancing solar system maintenance reliability and efficiency.