Part Surface Inspection Through Semi-Supervised Learning to Overcome Limited Data Challenge

Abstract

Abstract Surface inspection plays an important role in manufacturing of parts with complicated geometry. Specifically, the timely detection of minor defects, e.g., dent and scratches, in small-sized parts such as engine blades before acceptance is usually the final step of quality assurance. Current practice, however, mainly relies on human decision which is subjective and labor intense. In this research we propose an automated, image-based inspection system that utilizes robotic automation to acquire images of a part being inspected and employs machine learning to facilitate decision making. The data acquisition aims at high-resolution images that can thoroughly reflect the surface features of parts. While deep learning is leveraged, a series of advancements are designed to overcome the challenges of limited data and especially limited number of labels (i.e., parts with defects) during the initial training process. We synthesize a semi-supervised learning framework, building upon the residual neural network (ResNet) as well as the deep convolutional generative adversarial network (DCGAN), to extract features from the ground truth data and synthetic data. The performance boost of synthetic samples is investigated with different dataset sizes and validated.