Novel Multi-Scale Attention Generative Adversarial Network for Photovoltaic Solar Cell Defect Inspection Using Electroluminescence Images

In the pursuit of promoting green energy, efficient defect inspection in solar cell manufacturing is crucial in enhancing the reliability of solar energy systems. However, traditional deep learning models for automatic defect inspection in photovoltaic (PV) cell electroluminescence (EL) images encounter challenges in industrial settings due to difficulties associated with data acquisition, imbalance, and variability of defects. This paper presents a novel Multi-Scale Attention Generative Adversarial Network (MAGAN), an innovative GAN-based framework specifically designed for data augmentation in the context of solar cell defect detection. When integrated with automated detection techniques, MAGAN markedly improves the accuracy and efficiency of current models. A method for augmenting image datasets of EL was developed to generate a sufficient quantity of images for training machine learning models, addressing sample scarcity and bolstering CNN-based defect classification accuracy. The core of this approach lies in the application of the MCA (Multi-channel Spatial Attention Mechanism) and GLSA (Gate-like Spatial Attention Mechanism) modules, which enhance feature extraction by leveraging channel attention and spatial attention, respectively, thereby reflecting the most recent advancements in attention mechanism technology. The MCA dissects channels into sub-features across various scales, ensuring detailed attention mapping, whereas the GLSA refines spatial cues with a gating mechanism, shedding computational inefficiencies. The effectiveness of this approach is validated by comprehensive experiments against state-of-the-art deep learning models. The experiments demonstrate the exceptional performance of MAGAN, achieving a low FID score of 141.98 and KID score of 0.106 on complex EL images, surpassing previous models and emphasizing data augmentation’s importance in defect detection. With an industry-leading detection accuracy of 87.3%, this study makes a substantial contribution to mitigating data imbalance. This method enhances quality control in solar cell manufacturing. Additionally, it advances defect inspection in the industrial semiconductor sector.