Abstract/Description/Artist Statement

Industrial fault diagnosis lacks a procedure that learns time-varying causal structure from observational time series, makes identifiability limits explicit, and uses intervention-based reasoning to support root-cause assessment under industrial constraints. Although predictive maintenance can reduce downtime, diagnosis is often expert-rule-based or association-driven; pipelines may elevate downstream symptoms alongside true drivers and provide limited guidance on which feasible action would change a fault trajectory under current operating conditions. Because operating phases and fault progression create regime shifts, industrial systems rarely follow a single stable mechanism. This study develops and evaluates a three-stage, regime-aware causal diagnostic protocol based on a time-varying Dynamic Bayesian Network. A regime is a period in which process settings are relatively consistent, so relationships among variables are approximately stable. Stage I identifies regime boundaries and learns a regime-specific dependency skeleton (conditional dependencies given other measured variables) without asserting causal direction. Stage II learns a regime-specific causal representation up to Markov equivalence, preserving multiple edge-orientation explanations when directions are not identifiable from observational data. Stage III estimates forward-horizon what-if effects for candidate control levers via simulated do-interventions, enforcing identifiability and empirical support checks, and returning explicit non-estimability outputs when required assumptions fail. Developed through Design Science Research Methodology, the protocol is demonstrated on the Tennessee Eastman Process benchmark for representative faults, including Fault 1 and Fault 6. Results indicate that combining regime segmentation, cautious causal learning, and intervention-based effect estimation enables interpretable, decision-relevant diagnosis with auditable refusal outcomes that reduce overconfident association-based maintenance recommendations.

Presenting Author Name/s

Cansu Yalim

Faculty Advisor/Mentor

Holly A.H. Handley

Faculty Advisor/Mentor Email

hhandley@odu.edu

Faculty Advisor/Mentor Department

Department of Engineering Management and Systems Engineering

College/School Affiliation

Batten College of Engineering & Technology

Student Level Group

Graduate/Professional

Presentation Type

Poster

Share

COinS
 

A Causal Inference Methodology for Root-Cause Diagnosis in Nonstationary Industrial Time Series

Industrial fault diagnosis lacks a procedure that learns time-varying causal structure from observational time series, makes identifiability limits explicit, and uses intervention-based reasoning to support root-cause assessment under industrial constraints. Although predictive maintenance can reduce downtime, diagnosis is often expert-rule-based or association-driven; pipelines may elevate downstream symptoms alongside true drivers and provide limited guidance on which feasible action would change a fault trajectory under current operating conditions. Because operating phases and fault progression create regime shifts, industrial systems rarely follow a single stable mechanism. This study develops and evaluates a three-stage, regime-aware causal diagnostic protocol based on a time-varying Dynamic Bayesian Network. A regime is a period in which process settings are relatively consistent, so relationships among variables are approximately stable. Stage I identifies regime boundaries and learns a regime-specific dependency skeleton (conditional dependencies given other measured variables) without asserting causal direction. Stage II learns a regime-specific causal representation up to Markov equivalence, preserving multiple edge-orientation explanations when directions are not identifiable from observational data. Stage III estimates forward-horizon what-if effects for candidate control levers via simulated do-interventions, enforcing identifiability and empirical support checks, and returning explicit non-estimability outputs when required assumptions fail. Developed through Design Science Research Methodology, the protocol is demonstrated on the Tennessee Eastman Process benchmark for representative faults, including Fault 1 and Fault 6. Results indicate that combining regime segmentation, cautious causal learning, and intervention-based effect estimation enables interpretable, decision-relevant diagnosis with auditable refusal outcomes that reduce overconfident association-based maintenance recommendations.