Date of Award

Summer 8-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Engineering Management & Systems Engineering

Program/Concentration

Engineering Management

Committee Director

Holly A. H. Handley

Committee Member

Resit Unal

Committee Member

Faryaneh Poursardar

Abstract

Industrial fault diagnosis increasingly relies on data-driven methods, yet most approaches assume process stationarity and treat association-based rankings as sufficient evidence for root-cause identification. Neither assumption holds when operating phases and fault progression create distinct behavioral regimes with shifting relationships among variables, and neither answers which feasible corrective action would alter a fault trajectory. The absence of a procedure addressing non-stationarity, causal identifiability, and decision-relevant effect estimation constitutes the gap this dissertation targets. In this context, a regime is a time period in which the process operates under a relatively consistent setting, so the relationships among variables are approximately stable within that period.

This research develops a three-stage, regime-aware causal diagnostic protocol grounded in a Time-Varying Dynamic Bayesian Network (TV-DBN). Stage I detects regime boundaries and learns regime-specific causal skeletons via MDL optimization with Gaussian Process node models. Stage II constructs a Markov equivalence class representation using kernel conditional independence tests and Meek's orientation rules, preserving directional ambiguity. Stage III estimates do-operator intervention effects within MDL-weighted DAG extensions, gated by identifiability and positivity checks. When assumptions are violated, the protocol returns explicit non-estimability outputs, converting validity failures into auditable records.

The protocol is designed and evaluated using Design Science Research Methodology, with stages traceable to the stated requirements. Evaluation uses the Tennessee Eastman Process benchmark under two fault scenarios: Fault 6 represents an actuator-level feed loss with a directed causal path to the monitored output, while Fault 1 represents a compositional disturbance in which no actuator achieves directed reachability to the target variable (fewer than 5% of sampled DAG extensions contain a directed path in any regime).

Regime-aware segmentation, equivalence-class learning, and identifiability-gated effect estimation produce outputs calibrated to what observational data can defensibly support. For Fault 6, the protocol identifies xmv(3) as the sole structurally supported actuator-level candidate with robust directed reachability across all regimes, while signaling that the quantitative verification condition is not met: the estimated effect is directionally consistent but statistically inconclusive at the 95% level, satisfying the structural but not the quantitative condition. For Fault 1, the protocol withholds Regime 2 and 3 estimates due to ancestral support gate failures and refrains from promoting any actuator to root-cause status. The protocol advances practice by replacing association-driven rankings with intervention-grounded assessments and converting validity failures into actionable audit records.

Rights

In Copyright. URI: http://rightsstatements.org/vocab/InC/1.0/ This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s).

DOI

10.25777/43en-fj02

ISBN

9798193214328

ORCID

0009-0000-0836-8481

Download

Share

COinS