Defect detection in photovoltaic modules based on image-to-image generation and deep learning

Abstract

The autonomous monitoring of photovoltaic modules is emerging as an integral approach to maximize performance and reliability of photovoltaic systems, primarily in large-scale applications. However, it suffers with data acquisition, volume, diversity and performance constraints. This study proposed a method to detect multi-defects at module level in electroluminescence images of photovoltaic panels using limited data with synergistic integration of image-to-image generation and transfer deep learning from specific data and knowledge. Therein, StyleGAN3-t based image-to-image generation is firstly used to augment training data. Subsequently, the real-synthetic mix data is used to train YOLOv9 GELAN-e network using develop-model transfer learning from a pre-trained custom cell level model. To validate the effectiveness of proposed method, multiple iterations of image-to-image and object detection networks are studied using real and real-synthetic mix data with different formations, mixed training, and pre-trained general and specific weights. This method achieves FID score of 15.01 for images generation and 7% higher [email protected] for detection of seven classes compared to real data model trained without transfer learning, indicating the synergy and effectiveness of integrating image-to-image generation and transfer learning from specific data. The non-deterministic training for multiple runs also demonstrates the accuracy, stability and reliability of the method. This study contributed a module-level dataset and not only deals with enhanced detection of multi-defects at module and outdoor level but also addresses limited and diverse data constraints, leading to enhanced performance, operation and management of photovoltaic systems.