Hybrid Multiscale Contextual Framework for Enhanced Fault Detection in Photovoltaic Electroluminescence Imaging

Abstract

Electroluminescence (EL) imaging has demonstrated efficacy in identifying cracks, inactive areas, and other concealed faults that are frequently undetectable in visible-spectrum examinations. Nonetheless, conventional deep learning models, such as standalone convolutional neural networks (CNNs), encounter limitations in generalization, sensitivity to complex features, and robustness at diverse fault sizes. This study introduces an innovative Hybrid Multiscale Contextual Framework (HMCF) architecture that combines two powerful networks, EfficientNetB0 and ResNet50 to extract diverse features. Proposed model also introduces a Multiscale Feature Fusion Block (MFFB) to handle diverse fault sizes. The findings validate the effectiveness of integrating hybrid CNN architectures with multiscale feature fusion for precise, scalable, and dependable fault classification in photovoltaic (PV) modules, facilitating the development of intelligent and automated solar farm inspection systems. The proposed architecture demonstrates a detection accuracy of 93.18%, significantly outperforming leading deep neural network approaches.