Process Anomaly Detection Within IN-SITU Optical Tomography of Additive Manufacturing

Author

Merrill Dennison, Southern Illinois University CarbondaleFollow

Date of Award

12-1-2022

Degree Name

Master of Science

Department

Mechanical Engineering

First Advisor

Chu, Tsuchin

Abstract

The development of relationships between process parameters, process signatures, and final part quality within additive manufacturing (AM) processes has been deemed an essential component of current efforts to characterize and control AM processes towards standardization and qualification of AM parts for end-use. Statistical indicators within nondestructive evaluation data (NDE) obtained during the production process might serve as signatures for certain realities within the manufacturing process that influence the final part quality. End users of AM products and AM system manufacturers alike require quick, efficient, and readily accessible methods for connecting statistical indicators present within in-situ monitoring data to the final quality or characteristics of the part. Machine Learning (ML) approaches for processing and evaluating nondestructive evaluation data have become increasingly relevant in recent years with ever expanding inspection and data requirements. AM systems equipped with in-situ monitoring sensors will produce significant amounts of data, relevant to the quality assurance process, over their entire lifetimes. Failing to leverage this data to improve our understanding of AM processes would constitute a failure to utilize the available data to its full potential. The data utilized within this thesis work is comprised of several images depicting the plan-type view of a fixed band of the solidified regions Near-IR (NIR) light emissions. Samples designed with seeded defects, induced via variations in laser power about the nominal power level, at known locations within the part served to ensure reference defect (process anomaly) signatures presented within the in-situ data. An off-axis optical tomography (OT) imaging sensor captured the NIR melt pool emissions of each fabricated layer. The captured OT image data was cropped, labeled, and augmented to produce datasets for supervised learning. Supervised and unsupervised methods of distinguishing seeded defect layers were explored towards the classification of objects containing seeded defects from in-situ OT images. This research aims to explore the capability of supervised and unsupervised methods to detect the presence of seeded defects within individual layers of OT images. Results showed that effective convolutional neural network (CNN) architectures were obtainable within the small scope of hyperparameters chosen. The mean of the element-wise squared difference plotted over all layers for a particular reference is shown to have potential application to distinguishing anomalous layers within the time-series of images generated by OT. Two methods for defining reference arrays used in the element-wise squared difference operation are explored within this work. This work presents the hyperparameter combinations distinguished by the Keras-Tuner HYPERBAND algorithm for two distinct datasets. Models emphasized by the HYPERBAND algorithm obtained categorical accuracies greater than 90% during training and testing.