Automatic quality inspection in additive manufacturing using semi-supervised deep learning

Ensuring quality production is essential in additive manufacturing processes such as selective laser sintering (SLS). Deep learning based automated systems may provide a better way for real time monitoring to ensure quality production. These systems are usually trained in a fully supervised manner which require large amount of labeled data. Obtaining labeled training data in large quantities is a tedious and time consuming process. To overcome this, a novel semi-supervised deep learning approach is proposed in this work, which can be trained using both labeled and unlabeled data, and hence, reducing the manual labeling efforts needed to train the system. Experimental results on a SLS powder bed defect detection dataset show that the proposed approach is the new state-of-the-art, and shows its potential as a standalone real time monitoring system for SLS. In this dataset the proposed approach beats the state-of-the-art accuracy of $$98\%$$ 98 % with only $$25\%$$ 25 % of the labeled training data compared to other approaches. In addition, an extensive set of experiments were conducted on three additional public defect inspection datasets (NEU steel surface defects, KolektorSDD surface images of plastic electronics commutators, and surface textures) to show the applicability of the proposed approach on other computer integrated manufacturing environments for quality inspection. In all of these datasets the proposed approach beats the state-of-the-art results with relatively small amounts of training data compared to other approaches, which shows the effectiveness of the proposed approach for real time, automated, and accurate quality inspection.