Mahesh Veershetty

https://doi.org/10.64823/ijter.2507012

The increasing global adoption of Photovoltaic (PV) systems highlights the need for efficient maintenance, as defects such as hotspots, microcracks, and delamination significantly reduce energy output and system lifespan. Manual thermal inspections are slow, subjective, and unsuitable for large solar installations. This work presents an automated, real-time defect detection system using thermal imaging and a lightweight YOLOv9-nano deep-learning model optimized for embedded deployment. The model was trained on a multi-class thermal dataset from Roboflow containing eight types of solar-panel anomalies, following a structured pipeline of preprocessing, augmentation, 50-epoch training, and inference evaluation. The system achieved approximately 94.5% mAP and an inference speed of around 28 FPS in CPU-based simulation, indicating strong suitability for Raspberry Pi 4 Model B deployment after optimization. The results demonstrate the system’s potential as a scalable, low-cost predictive-maintenance tool capable of early fault detection, improved operational reliability, and enhanced energy yield in PV installations.