Diagnostic Methodologies for Fault Management and Performance Optimization

Document Type

Dissertation

College

College of Engineering

Department

Industrial and Engineering Management

Degree

PhD in Engineering Management

Dissertation Defense Date

2025-11-25

First Committee Member

Li, Zhaojun Steven

Second Committee Member

Rupe, Jason Wade

Third Committee Member

Salmon, Christian Mauriz

Additional Committee Member(s)

Huang, Jiali

Hóu, Yǔ

Abstract

This research investigates diagnostic approaches applied to fault data to enhance predictive accuracy and support proactive maintenance. Engineering systems degrade over time due to operational stresses, making proactive maintenance essential for reliable performance. Unlike reactive policies, proactive strategies aim to address potential faults early using model-based, data-driven, or hybrid diagnostic methods. Challenges in real-world applications include data imbalance, process variability, uncontrolled operating conditions, and varying levels of diagnostic analysis, from fault detection to root-cause identification. Fault diagnosis encompasses detection, localization, failure mode identification, and severity assessment, with methods tailored to data availability and system knowledge.

To address these challenges, four fault diagnosis methods are proposed. The first uses a hierarchical ensemble learning structure to maximize prediction accuracy on raw data, validated on field data from a US cable operator. To overcome limitations of computational cost and sequential task performance, a generic data-driven method with feature extraction and selection is developed, improving diagnostic performance while addressing model complexity. Multi-objective optimization (MOO) balances accuracy, efficiency, and interpretability, though interpretability challenges remain. To further enhance interpretability, a sparse feature identification approach combines adaptive windowing with PAC-Bayesian learning, selecting minimal features while preserving accuracy.

Finally, multi-class fault scenarios are addressed using a lightweight convolutional neural network with an attention mechanism, achieving high diagnostic accuracy and improved explainability. These methods collectively advance fault diagnostics for engineering systems, providing robust, interpretable, and efficient solutions for proactive maintenance across diverse operational conditions.

This document is available upon request to Western New England University faculty, students, and staff. Please contact D'Amour Library at for access.

Share

COinS