Layerwise Anomaly Detection in Laser Powder-Bed Fusion Metal Additive Manufacturing

Abstract

A growing research trend in additive manufacturing (AM) calls for layerwise anomaly detection as a step toward enabling real-time process control, in contrast to ex situ or postprocess testing and characterization. We propose a method for layerwise anomaly detection during laser powder-bed fusion (L-PBF) metal AM. The method uses high-speed thermal imaging to capture melt pool temperature and is composed of the following four-step anomaly detection procedure: (1) using the captured thermal images, a process signature of a just-fabricated layer is generated. Next, a signature difference is obtained by subtracting the process signature of that particular layer from a prespecified reference signature, (2) a screening step selects potential regions of interests (ROIs) within the layer that are likely to contain process anomalies, hence reducing the computational burden associated with analyzing the full layer data, (3) the spatial dependence of these ROIs is modeled using a Gaussian process model, and then pixels with statistically significant deviations are flagged, and (4) using the quantity and the spatial pattern of the flagged pixels as predictors, a classifier is trained and implemented to determine whether the process is in- or out-of-control. We validate the proposed method using a case study on a commercial L-PBF system custom-instrumented with a dual-wavelength imaging pyrometer for capturing the thermal images during fabrication.