Using Machine Learning to Analyze Image Data from Advanced Manufacturing Processes

Abstract

Additive manufacturing (AM) - also known as 3D printing - promises a new approach to creating parts in a manufacturing environment; the process allows more design freedom and the production of parts with more complex features, compared to traditional manufacturing processes. The laser powder bed fusion (L-PBF) printer operates by building a part layer by layer in an iterative process of spreading metal powder and melting the desired shape. One particular feature is an overhang (material being melted onto the part over loose un-melted parts). However, some of the un-melted powder from the process could become melted to the overhanging feature - which is known as dross. Overhangs tend to form dross, but the extent of dross created at these features is not fully understood. Due to this unpredictable nature of dross formation, the build process exhibits variability in build quality, deterring industry-wide adoption. The conducted research aims to develop a system that analyzes cross-sectional image data captured from each layer of the print in order to identify dross with a certain level of confidence. Using machine learning techniques, images are used in a model that identifies pixels as a region that contains dross. These images are first labeled with bounding boxes (a coordinate system that identifies features/objects as existing within its boundaries) to train a neural network. The result is an adaptive model that autonomously detects dross in image scans of the part, pointing out these impurities to the printers' users, especially in regions difficult to inspect like interior surfaces of parts. The model aims to further understand L-PBF processing by location regions of excessive dross to relate dross formation with specific design features.