Automated Defect Detection in Photovoltaic Modules Using Deep Learning on Electroluminescence Imaging

Abstract

Accurate defect detection in photovoltaic (PV) modules is essential to ensure energy efficiency and system reliability. Although large-scale electroluminescence (EL) datasets such as PVEL-AD have enabled significant progress in deep learning–based inspection, rare defect types remain difficult to detect due to severe class imbalance. Moreover, the extremely limited number of samples available for certain defect categories restricts the diversity of defect appearances, preventing models from learning the full variability in terms of shape, size, orientation, and spatial location, which further degrades generalization performance.This paper proposes a defect-aware data preparation and evaluation framework to improve detection performance for underrepresented PV defects. Rare defect classes are enhanced using a blending-based augmentation pipeline that combines geometric transformations with Laplacian (Poisson) blending and adaptive alpha blending to generate realistic and structurally consistent synthetic samples. To ensure fair and unbiased evaluation, this paper proposed a strict 7-fold cross-validation strategy , guaranteeing complete separation between training and test data and preventing any data leakage.Multiple detection models—including YOLO variants, Faster R-CNN, and DETR—are evaluated using AP50 and AP50:95 metrics computed exclusively on real test images. Experimental results show consistent performance improvements for rare defects, achieving an aggregated mAP50 of 86.28%