ANNOUNCEMENT OF DEFENSE OF DISSERTATION RESEARCH
The faculty of the Engineering Management and Systems Engineering department is pleased to issue an invitation to Ms. Cansu Yalim鈥檚 defense of the research conducted for his dissertation.
The defense is open to the public.
This defense is via Zoom ONLY.
Date: Thursday, March 26th, 2026,
Time: 11:00 AM to 12:30 PM
Location: Online only
Online Access: Via Zoom Meeting ID: 943 3695 8487 Passcode: 609342
Doctoral Candidate: Cansu Yalim
B.S. 2020, Izmir 91短视频 of Economics
M.S. 2022, Celal Bayar 91短视频
A CAUSAL INFERENCE METHODOLOGY FOR ROOT-CAUSE DIAGNOSIS IN NONSTATIONARY INDUSTRIAL TIME SERIES
Director: Holly A. H. Handley, PhD, PE
Abstract:
This research addresses a gap in industrial fault diagnosis: the lack of a procedure that learns time-varying causal structure from observational time series, while representing what is and is not identifiable, and then uses intervention-based reasoning to support root-cause assessment under representative industrial data constraints. Predictive maintenance can reduce downtime, yet diagnosis often remains driven by expert rules or association-based pattern recognition. As a result, pipelines may elevate downstream symptoms alongside true drivers and provide limited guidance on the decision that maintenance teams actually face, namely, which feasible action would change a fault trajectory under the current operating conditions.
Furthermore, industrial systems rarely follow a single, stable mechanism in practice, as operating phases and fault progression can create distinct operating regimes. To address this gap, the study develops and evaluates a three-stage, regime-aware causal diagnostic protocol based on a time-varying Dynamic Bayesian Network. 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. Stage I identifies regime boundaries and learns a regime-specific dependency skeleton, defined as a map of conditional dependencies among variables after accounting for other measured variables. This skeleton describes which variables are directly related to the data within a regime, but it purposefully does not claim causal direction. Stage II then learns a regime-specific causal representation up to Markov equivalence, meaning it preserves multiple directional explanations when observational data cannot uniquely determine edge orientations. Stage III estimates forward horizon what-if effects for candidate control levers by simulating do interventions within the learned model, while enforcing identifiability and empirical support checks, and returning explicit non-estimability outputs when the assumptions required for quantitative effect estimation are not satisfied.
Developed through Design Science Research Methodology, the protocol is demonstrated on the Tennessee Eastman Process benchmark using representative fault scenarios, including Fault 1 and Fault 6. The results indicate that combining regime-aware segmentation, cautious causal structure learning, and intervention-based effect estimation can provide a structured, interpretable approach to exploit non-stationary industrial time series for decision-relevant diagnosis. The protocol supports regime-aware explanations that help distinguish plausible root causes and differentiate between alternative control actions, and it also provides auditable refusal outcomes when identifiability or support conditions fail, reducing the risk of overconfident maintenance recommendations driven by association rather than defensible intervention statements.
Short Biography:
Cansu Yalim is pursuing her Ph.D. in Engineering Management and Systems Engineering at Old Dominion 91短视频, where she also serves as a graduate research and teaching assistant. Before academia, she worked in the thermotechnology and automotive industries. Yalim鈥檚 research reframes industrial root-cause diagnosis as a causal inference problem, overcoming the correlational limits of traditional ML by integrating time, causal structure, and system dynamics to deliver trustworthy fault attribution in complex, changing environments.