Deep Convolutional Neural Networks for Accurate Solar Module Classification in Thermal Images

The rapid and continuous technological advancements in computer and internet technologies, combined with data management techniques, find applications in various domains. For example, they aid in predictive maintenance systems, evaluate the performance of solar modules, and differentiate efficiently performing solar modules from less efficient ones using thermal imagery. The accuracy of solar module classification significantly impacts the subsequent assessment. In this study, we employ advanced techniques, particularly deep learning through Convolutional Neural Networks (CNN), to classify thermal images of solar modules. The practical aspects of applying CNN to the dataset of thermal images are explored, including transfer learning from pretrained CNN architectures and custom CNN architecture training. Experiments are conducted with three prominent types of solar module conditions: normal, cracked, and dusty. The results reveal that the most effective approach involves learning and fine-tuning pre-trained architectures on detailed studies, achieving a classification accuracy of up to 94.5
^%
.