Theses and Dissertations

Date of Award

5-1-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Statistics

First Advisor

Farid Ahmed

Second Advisor

Tamer Oraby

Third Advisor

Jianzhi Li

Abstract

The Laser Powder Bed Fusion Process (LPBF) has been one of the main processes of additive manufacturing, enabling the manufacturing of complex geometries, customization, and lightweight parts. Modern LPBF processes have integrated monitoring systems that capture the light emissions per layer for quality assurance. However, standard defect detection algorithms have not yet achieved the high precision required due to the inherently variable nature of the signal, insufficient data for model training, and the confounding effects of the print.

The processes still have some challenges, such as characterizing the roughness from the build parameters alone, improving the pore detection using the 1D light emission from the print, the monitoring system cannot quantify the size of the pore using the light emissions, and the simulation of light emission given build characteristics such as roughness and defect size. This dissertation addresses these challenges by developing machine learning (ML) methodologies that leverage machine learning models to (a) characterize surface roughness, (b) detect pores using the average 1D light emissions, (c) estimate pore size using the 1D light emission, and (d) use the build characteristics to simulate 1D light emissions to be used to train defect detection modules.

A set of deep learning architectures, including a novel signal-to-image convolutional neural network for 1D functional data, was trained on experimental printed IN 718 samples. The roughness estimation model’s variability was trained using bootstrapping on a single set, while the defect detection variability was measured across several runs. Light emission signals were used for defect detection using classification models and pore characteristic dimensions using the ML regression models. Additionally, an ML model linked the roughness and characteristic dimensions to produce synthetic 1D light signals to expand the training dataset for the defect size estimation modules.

Neural networks accurately characterize the surface using build angles, agreeing with experimental data. Light emissions-based defect detection reached 87% having the limitation in confounding signals due to the overlap in the distribution. Pore characteristic dimension estimation was achieved, having cylindrical or asymmetrical shapes, having the best performance with a correlation coefficient of 0.8906 and a Root Mean Square Error (RMSE) of 40.8788. Synthetic data augmentation improved the correlation coefficient of the characteristic dimension estimation of the pore up to 12% and reduced the RMSE error up to 16%.

This dissertation introduces a novel deep learning approach to analyze functional data and a framework for building quality assurance for LPBF prints. This work showcases the use of ML methodologies for defect detection and quantification, advancing the understanding of the link between build characteristics, defects, and light emissions. Establishes quantification of variability of the estimations for ML monitoring and demonstrates the effectiveness of simulated light emissions to enhance the size estimation of the models. The framework proposed can be applied to other LPBF process machines and potentially can be extended to other manufacturing processes where 1D data is monitored.

Comments

Copyright 2026 Jose Galarza. All Rights Reserved. https://proquest.com/docview/3371103981

Download

Included in

Mathematics Commons

Share

COinS