R2I: A Deep Fault Diagnosis Semantic Segmentation System Based on Knowledge Distillation in Photovoltaic Panels

With the growing adoption of photovoltaic power systems, ensuring the reliable and efficient operation of photovoltaic panels is critical. However, defects such as cracks, dirt, and failures are common during long-term use. Manual inspection is subjective and resource intensive. Recent computer vision and machine learning techniques present new opportunities for automated diagnosis. This paper reviews photovoltaic panel defect detection using deep learning. Two main approaches exist: visible light and infrared imaging. Earlier works employed object detection with lower precision. Newer methods use semantic segmentation for pixel-level accuracy. However, most focus on single modalities and datasets. Transfer learning across diverse data could improve generalization. Few works estimate defect areas, needed for automated analysis. This paper proposes a dual-modality semantic segmentation approach with knowledge transfer learning. It shares representations across visible and infrared datasets via distillation. A lightweight network reduces parameters and computations for deployment. Soft targets provide latent features transferable to new data. An algorithm estimates defect area proportions to assess hazard levels. The multi-modal transferable learning system aims to advance automated, accurate photovoltaic panel defect diagnosis. Key innovations include tackling insufficient data via shared knowledge, reducing model complexity, improving generalization, and pioneering defect severity assessment. This can facilitate intelligent photovoltaic inspection to ensure system health, safety and longevity.