Date of Award

Winter 3-1-2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Pradeep Chowriappa

Abstract

This dissertation presents a comprehensive and scalable framework for real-time fault detection and event triage in industrial systems, addressing critical challenges such as class imbalance, ambiguous feature boundaries, and the prioritization of complex, high-dimensional event data. The proposed framework integrates advanced methodologies, including micro-batch processing, retrospective divergence-based event detection (DB-RED), association rule mining (ARM), clustering, and Dempster-Shafer Theory (DST) for conflict resolution. Together, these components enable the systematic stratification of events into actionable priority levels, ensuring robust and interpretable decision-making in real-time environments. DB-RED forms the cornerstone of the framework, leveraging KL-divergence and PE-divergence metrics to detect subtle and transient faults within high-frequency data streams. ARM techniques, including Apriori and FP-Growth, translate these detected events into structured relationships, providing the contextual basis for effective event triage. Clustering methods, such as K-Means and Hierarchical Clustering, further organize events into priority-based groups, while the integration of DST enhances classification precision by resolving ambiguities in boundary and transitional cases. Extensive experiments validated the framework’s efficacy across multiple test-tofailure datasets. Classifiers such as Random Forest (RF) and Support Vector Machine (SVM) consistently achieved high accuracy, precision, and F1-scores, demonstrating the framework's adaptability to diverse fault scenarios. The inclusion of DST-based belief scores dynamically adjusted clustering behaviors, reducing false positive and false negative rates while preserving critical-event hierarchies. Results highlighted the framework’s ability to capture rare but impactful faults, enabling timely interventions and minimizing operational downtime. This research bridges gaps in real-time event monitoring by offering a novel combination of scalable algorithms and interpretable methodologies. Designed for industrial applications, the framework supports systems such as Supervisory Control and Data Acquisition (SCADA) by ensuring reliable event detection and actionable prioritization with human-in-the-loop oversight. Future work will explore the extension of these methodologies to broader industrial contexts, emphasizing scalability, automation, and integration with emerging predictive maintenance technologies. This dissertation contributes a significant step forward in advancing operational reliability and efficiency in high-stakes industrial environments.

Download

Share

COinS