Multi-modal sensor fusion with machine learning for data-driven process monitoring for additive manufacturing

This paper presents a complete concept and validation scheme for potential inter-layer flaw detection from in-situ process monitoring for powder bed fusion additive manufacturing (PBFAM) using supervised machine learning. Specifically, the presented work establishes a meaningful statistical correlation between (i) the multi-modal sensor footprint acquired during the build process, and (ii) the existence of flaws as indicated by post-build X-ray Computed Tomography (CT) scans. Multiple sensor modalities, such as layerwise imagery (both pre and post laser scan), acoustic and multi-spectral emissions, and information derived from the scan vector trajectories, contribute to the process footprint. Data registration techniques to properly merge spatial and temporal information are presented in detail. As a proof-of-concept, a neural network is used to fuse all available modalities, and discriminate flaws from nominal build conditions using only in-situ data. Experimental validation was carried out using a PBFAM sensor testbed available at PSU/ARL. Using four-fold cross-validation on a voxel-by-voxel basis, the null hypothesis, i.e. absence of a defect, was rejected at a rate corresponding to 98.5% accuracy for binary classification. Additionally, a sensitivity study was conducted to assess the information content contributed by the individual sensor modalities. Information content was assessed by evaluating the resulting correlation as classification performance when using only a single modality or a subset of modalities. Although optical imagery contains the highest amount of information for flaw detection, additional information content observed in other modalities significantly improved classification performance.