An Adversarial Training Framework Based on Unsupervised Feature Reconstruction Constraints for Crystalline Silicon Solar Cells Anomaly Detection

Photovoltaic (PV) power generation has risen prominently across the globe. However, the functioning of PV systems can be substantially diminished by defective solar cells. Electroluminescence (EL) imaging has emerged as an effective tool for identifying such defects. In this article, inspired by the manifold hypothesis, we propose a novel unsupervised adversarial training framework based on feature reconstruction constraints in different spaces for crystalline silicon solar cell anomaly detection. Adversarial constraints are specially designed for both original high-dimensional image space and latent low-dimensional representational space. The adversarial training strategy aims to seek the optimal latent representation of normal data. By minimizing the distance between images and between latent representations, it can not only generate similar images but achieve good latent reconstruction performance. We also employ Gaussian Context Transformer (GCT) to enhance better long-range semantic dependency with less computational complexity. During testing, poor feature reconstruction of either a high-dimension image or a low-dimension latent vector indicates a potential anomaly. Although the proposed framework does not use labeled data for training, it can distinguish out-of-domain defect samples from in-domain normal samples and identify small defects that may not be discernible using other methods. Experimental results over real-life EL datasets demonstrate that the framework achieves satisfactory performance by outperforming other machine learning or deep-learning-driven methods. It offers a resilient solution for defect detection in crystalline silicon solar cells within the industry, bridging the gap of unsupervised approaches in this industrial domain.