Publication

Abstract : Reliable photovoltaic (PV) module defect detection is essential for maintaining long term energy efficiency and lowering solar power system maintenance costs. The deep learning model presented in this research is based on a hybridized YOLOv8n architecture and is lightweight and high performing. It is designed for multi scale defect identification in a variety of imaging modalities, such as RGB, grayscale, and infrared datasets. The proposed approach combines a BiFPN based neck, Ghost Bottlenecks, and Efficient Channel Attention (ECA) to improve multi scale representation, decrease redundant computation, and increase feature extraction. The model performs better in terms of detection accuracy and efficiency, as shown by experimental findings on three benchmark datasets: PVEL-AD, PV-Multi-Defect, Solar Panel Anomalies. The model’s respective mAP@50 values are 96.5%, 94.6%, and 97.6%. At a steady inference time of only 1.9 ms and 8.1 GFLOPs, it also achieves near-perfect recall (up to 99.0%) and high precision (up to 98.4%). With just 3M parameters, the proposed hybrid model provides a much better accuracy-latency trade-off 61 current state-of-the-art models, which makes it perfect for real-time solar inspection applications, such as edge deployment in drones and embedded systems. The outcomes confirm that reliable PV fault localization under a range of operating situations may be achieved by combining deep feature fusion, lightweight attention, and efficient convolution.