A Deep Learning Approach for the Identification of Small Process Shifts in Additive Manufacturing using 3D Point Clouds

Abstract Additive manufacturing (AM) refers to a family of manufacturing technologies that fabricate parts by joining materials layer by layer. It has a high level of flexibility in design and manufacturing, which provides a unique opportunity for producing parts with complex geometries that are not feasible using conventional subtractive manufacturing. Due to the sensitivity of AM to machine settings and process conditions, process shifts are oftentimes incurred in AM, which introduce defects and impact the quality and reliability of AM products. As such, it is critical to identify AM process shifts, especially at the incipient stage, for quality assurance. Most existing approaches, however, are limited in their ability to detect small AM process shifts. In this study, a structured light scanner is used to capture 3D point clouds from printed surfaces. A deep learning framework is introduced to extract useful information from point cloud data to delineate geometric variations of the printed surface and detect process shifts. The research methodology is evaluated and validated using both simulation studies and real-world applications. Experimental results have shown that the deep learning approach is with remarkable ability in detecting small process shifts and it outperforms convolutional neural network models when large amounts of training samples are not available. The proposed framework has a strong potential to be used for in-situ layer-wise monitoring of AM processes for quality control and the detection of cyber-physical attacks.