Wavelet-Based Normalized Flow for Anomaly Detection in Photovoltaic Electroluminescent With Nonstationary Textures

Anomaly detection is vital for maintaining the functionality and productivity of photovoltaic (PV) systems. Previous unsupervised anomaly detection (UAD) methods face challenges when dealing with nonstationary textures in PV electroluminescence (PVEL) images, as they rely solely on spatial-domain information, making it difficult to differentiate between defects and background patterns accurately. In this article, we propose Wave-Flow, a novel anomaly detection model that integrates wavelet-based analysis to overcome these challenges. Wave-Flow can explore information across diverse wavelet domains, significantly improving the differentiation between normal and abnormal states in PVEL images. In addition, the incorporation of the cross-frequency (CF) module enhances the model’s ability to capture correlations between different frequency components, refining defect detection accuracy further. Extensive experiments on benchmark datasets demonstrate Wave-Flow’s superior performance, it achieves 94.4% AUROC and 98.9% area under the precision-recall (AUPR) on the PVEL-AD dataset, outperforming the current state-of-the-art by 5.8 percentage points. Wave-Flow also achieves an AUROC of 98.0% on the crack segmentation dataset (CSD). Our results demonstrate that Wave-Flow is a promising approach for anomaly detection in PVEL images with nonstationary textures.