Text-to-Image and Image-to-Image Augmentation and Classification of Defects in Photovoltaic Modules

The autonomous photovoltaic (PV) monitoring is an emerging field with potential application in large-scale PV farms to maximize their output and reliability. However, it experiences several critical challenges like limited data constraints, intricate data collection and insufficient performance on unseen data in unknown working conditions. Addressing these challenges, the present study explores text-to-image and image-to-image synthetic data generation methods. The synthetic electroluminescence (EL) images are generated using Stable Diffusion (text-to-image) and Style Generative adversarial Networks, StyleGAN (image-to-image) variants. The real and synthetic data combinations are then used for solo and mixed training experiments that demonstrated enhanced (highest) classification performance with text-to-image data augmentation followed by image-to-image augmentation and real data. Moreover, a hybrid classification architecture based on MobileNetv3 small and YOLOv11 is proposed that achieved enhanced precision, recall and accuracy of 94.3 %, 87.3 % and 91.3 % respectively compared to baseline and other models. The introduction of inverse and effective sampling along with class weighting is also studied to handle class imbalance. This study explored synthetic data generation methods and proposed a classification architecture to improve the performance of autonomous PV monitoring as well as addresses limited volume and diverse data constraints, simplifying the data acquisition process.