Integrated Statistical and Machine Learning Frameworks for Prediction, Process Monitoring, and Multi-Objective Optimization in Manufacturing Settings

Abstract

This dissertation develops data-driven methodologies that integrate statistical analysis, machine learning (ML), and evolutionary multi-objective optimization to address key challenges in modern manufacturing systems. The research focuses on three connected areas under the umbrella of manufacturing. The first contribution investigates the flowability of Inconel 718 powders, which are widely used in the laser powder bed fusion (L-PBF) processes. Using powder morphological and rheological, and bulk data collected by researchers at NASA, this research develops a ML framework to identify the key powder features governing flow behavior and to predict two widely used flowability metrics: the angle of repose (AOR) and the flow function coefficient (FFC). ML models including LASSO regression, random forest regression (RFR), and support vector regression (SVR) were implemented to predict powder flowability and the ML models achieved prediction errors below 5% for AOR and between 7% and 9% for FFC. The second contribution introduces a novel convex hull (CH)–based control chart for monitoring bivariate manufacturing processes. The proposed CH–based monitoring approach utilizes the geometric structure of multivariate observations to detect deviations from the in-control process. Extensive simulation experiments were conducted using bivariate 𝑡-distributed data with varying degrees of freedom and correlation structures and the CH chart outperformed both MEWMA and MCUSUM control charts in terms of out-of-control average run length. In several scenarios, the convex hull–based chart achieved improvements of up to approximately 60% in detecting moderate mean shifts, particularly under heavy-tailed data distributions. The third contribution develops an integrated ML–based surrogate assisted multi-objective optimization framework in turning Ti-6Al-4V. ML models including SVR, RFR, and Gaussian process regression (GPR) were implemented to predict surface roughness, cutting force, and tool wear. The ML model giving smallest error was further interpreted using SHAP analysis to quantify the influence of machining parameters on the responses. The selected surrogate model was subsequently integrated with the NSGA-II to simultaneously maximize material removal rate while minimizing surface roughness, cutting force, and tool wear. The resulting Pareto front provided a set of optimal trade-off solutions, and a best compromise solution was identified using a minimum-distance-to-ideal-point approach.