Joint active search and neuromorphic computing for efficient data exploitation and monitoring in additive manufacturing

Abstract The recent integration of imaging technology with additive manufacturing (AM) leads to the plethora of in-process and high-dimensional data. Machine learning (ML) methods have been implemented to improve understanding of defect formation in AM-built parts and controlling process variability in real-time. However, modern ML methods, in particular deep neural networks, are empowered by massive high-quality labeled data, which are limited in AM due to the following reasons: First, large data labeling is often tedious, costly, and requires substantial human efforts with considerable expertise. Second, the performance of the learning methods depends to a great extent on the presence of positive data instances (i.e., defective) as they are more informative for monitoring. Third, the rare positives result in a severe imbalanced dataset poses critical challenges in training ML methods designed with the assumption that the input contains an equal number of instances from each class. In this research, we propose novel annotation and learning with limited number of data through the integration of active search and hyperdimensional computing (HDC). The active search is developed to benefit from a single bandit model to learn about the data distribution (exploration) while sampling from the regions potentially containing more positives (exploitation). HDC is introduced as an alternative computing method that mimics important brain functionalities and encodes data with high-dimensional vectors, thereby enabling single-pass learning with just a few samples. Experimental results on a real-world case study of drag link joint build show the proposed model locates the rare positives thoroughly and detects lack of fusion defects with the accuracy of 89.58%, in 3.221 ± 0.029 second training time and with only 66 sample data. The joint active search and neuromorphic computing framework is shown to have strong potentials for general applications in a diverse set of domains with in-situ imaging data.