Accelerating solder joint classification using generative artificial intelligence for data augmentation

Abstract

Despite advancements in computer vision, deploying deep learning algorithms for automated optical inspection (AOI) in printed circuit board (PCB) manufacturing remains challenging due to the need for large, diverse, and high-quality training datasets. AOI programs must be developed quickly, often as soon as the first PCB is assembled, to meet tight production timelines. However, deep learning models require extensive datasets of defect images, which are both scarce and time-consuming to collect. As a result, AOI software developers frequently resort to traditional rule-based methods. This study introduces a novel framework that leverages generative AI and discriminative AI to address dataset limitations. By applying a diffusion model to systematically add and remove Gaussian noise, the framework generates realistic defect images, expanding the available training data. This data augmentation accelerates the learning process of deep learning models, enhancing their robustness and generalizability. Experimental results demonstrate that this approach improves AOI system performance by producing balanced datasets across various defect classes. The framework shortens training times while maintaining high inspection accuracy, facilitating faster deployment of AOI systems in manufacturing. This advancement enhances quality control processes, contributing to more efficient, and reliable mass production of PCBs.