Distance-Based Multivariate Anomaly Detection in Wire Arc Additive Manufacturing

Abstract

Wire Arc Additive Manufacturing (WAAM) offers the possibility to build up large-scale metal parts. Data which is obtained from a multivariate sensor system in-situ must be analyzed automatically to ensure an early and reliable detection of defects to reduce the costs due to production scrap. For that reason, a modular anomaly detector for multivariate time series in WAAM was investigated in this paper. The approach adressed major topics in real-life data sets of industrial applications such as miscellaneous signal sample rates, lack of synchronization and concept drift. A reference data set based on an anomaly-dependently splitted time horizon was defined to reduce the sensitivity loss of the detector after an anomaly. To avoid the need for labeled data, an unsupervised anomaly detection method based on neural networks was used. Hence, no time and costs for artificial defect creation on the machine tool are required when implementing the approach in industrial applications.