Unsupervised and few-shot segmentation in photovoltaic electroluminescence images for defect detection via a novel enhanced iterative autoencoder with simple implementation

Abstract

Photovoltaic electroluminescence (PVEL) imaging captures material-level degradation in PV modules and offers high-resolution input for machine learning (ML) models to perform automated fault detection and health evaluation, reducing reliance on manual inspection. It is expected to have a simple and efficient defect detection ML model to achieve accurate segmentation for the fine-featured identification of defects in fabricated PV modules. This study proposes a novel enhanced iterative autoencoder (EI-AE), a completely new model that differs fundamentally from existing approaches which rely directly on classical ML models for defect detection. The proposed EI-AE, which for the first time introduces an iterative mechanism into the traditional AE framework, features a simple yet effective architecture and achieves accurate unsupervised pixel-level segmentation of all defect types using only normal PVEL images. In addition, few-shot learning can be realized by extending the unsupervised EI-AE with a small number of annotated masks, allowing more detailed functional defect detection while mitigating background interference. Theoretical proof demonstrates the benefits of the proposed EI-AE in improving defect detection compared to the conventional AE. Experimental results further validate its superiority, showing consistently better performance across multiple pixel-level metrics and outperforming both widely used unsupervised and few-shot baseline approaches.