Evaluating the robustness of signal unmixing methods for multivariate optical spectroscopic process monitoring data

Abstract

The growing application of multivariate spectroscopic methods in process monitoring has created a critical need for robust signal unmixing techniques to resolve complex, overlapping datasets. This dissertation investigates the applicability and limitations of several signal unmixing frameworks for multicomponent optical spectroscopic process monitoring. The first project examines the comparative performance of Multivariate Curve Resolution–Alternating Least Squares (MCR-ALS) and Nonnegative Matrix Factorization using Multiplicative Update (NMF-MU) on simulated excitation–emission matrices (EEMs) under varying conditions of overlap and noise levels. The second project evaluates the limits of tensorization in combination with PARAFAC decomposition for resolving components from simulated emission decay matrices (EDMs), identifying noise thresholds and structural constraints beyond which the tensor model fails to deliver interpretable results. The third project applies constrained matrix factorization methods, including semi-NMF and convex-NMF, to Fourier-transformed EDMs collected during the photodegradation of Rose Bengal, a system known to involve excited-state reactions and overlapping emissive species. Each approach is benchmarked using simulated or experimental datasets, allowing for systematic evaluation of unmixing performance in realistic conditions. Collectively, the findings provide practical guidelines on the selection and application of signal unmixing methods for PAT, with a focus on model interpretability, resolution performance, and robustness under challenging conditions. Across all three projects, findings highlight both algorithmic strengths and contextual limitations, offering practical guidance for selecting appropriate signal unmixing techniques in multivariate spectroscopic monitoring.