Author: Dai-Yan Ji
Date/Time: November 10th, 2025 at 11:00AM EST
Location: EGR-2164, Glenn Martin Hall
Committee Members:
Dr. Jay Lee, Chair
Dr. Nii O. Attoh-Okine, Dean’s Representative
Dr. Balakumar Balachandran
Dr. Davis McGregor
Dr. Peter Sandborn
Title of Dissertation: Methodology for Automated Trace Segmentation to Improve Feature Extraction of Time Series Data in Prognostics and Health Management
Abstract: In modern industrial systems, time-series data serve as the foundation for data-driven prognostics and health management (PHM). However, traditional feature extraction approaches, such as computing summary statistics over the whole signal, relying on domain knowledge to manually select informative regions, or using deep learning models to learn representations directly, often fail to capture localized and transient dynamics, leading to reduced interpretability and degraded diagnostic accuracy under nonstationary operating conditions. This dissertation presents a novel methodology for automated trace segmentation to enhance feature extraction of time-series data in PHM applications.
The proposed framework introduces a segmentation-driven statistical feature methodology that decomposes complex temporal signals into physically meaningful segments. A generalized turning-point detection algorithm is developed to identify representative transition points without requiring predefined segment numbers, enabling adaptive segmentation across multivariate signals and diverse datasets. Building upon the segmented structure, localized statistical features are extracted within each segment to preserve discriminative information while reducing noise sensitivity.
Experimental studies were conducted using two representative PHM scenarios: a chemical gas sensor classification dataset and a semiconductor plasma etching health assessment process. In the gas sensor classification task, the proposed method consistently improved classification accuracy and reduced performance variance across multiple machine-learning models, confirming its robustness and generalizability. In the semiconductor case, the segmentation-based approach achieved more stable and interpretable health assessment through segment-wise statistical process monitoring, outperforming the traditional whole-trace method in sensitivity and trend clarity.
Overall, this dissertation establishes a unified and automated segmentation framework that bridges process understanding and data analytics, providing a physically interpretable and scalable solution for feature engineering of industrial time-series data. The methodology offers a fundamental step toward more reliable and explainable PHM modeling across complex and dynamic manufacturing environments. In future research, this framework can be extended toward LLM-based tokenization for time-series modeling and agentic AI PHM systems.