Document Type

Article

Publication Date

2026

DOI

10.1155/int/3436744

Publication Title

International Journal of Intelligent Systems

Volume

2026

Pages

3436744

Abstract

Predictive maintenance (PdM) systems effectively forecast failures, but they often fail to find root causes, particularly when system dynamics change over time. This limitation arises from applying static causal models or decoupled segmentation to handle nonstationary industrial time series. For regime-aware causal diagnostics and interventional effect estimation, we introduce a three-stage time-varying dynamic Bayesian network (TV-DBN) protocol. Using a minimum description length (MDL) objective that connects segmentation to mechanism changes, Stage I jointly infers change points and regime-specific graph structure. Stage II produces a completed partially directed acyclic graph (DAG) by orienting edges within each regime using a combination of score-based search and conditional independence testing with false-discovery-rate control. Stage III uses MDL-based weights to average over DAG extensions, truncated factorization to calculate h-step interventional effects, and ancestral support gating to prevent risky extrapolation. The protocol localizes causal regime changes, generates more plausible regime-specific structures than static baselines, and generates regime-specific effect estimates that are consistent with known fault mechanisms, according to experiments conducted on the Tennessee Eastman process benchmark. In nonstationary industrial settings, the suggested "auditable contract" of assumptions supports more dependable root-cause diagnosis and maintenance choices by giving practitioners evident standards for verifying causal claims.

Rights

Copyright © 2026 Cansu Yalim et al.

This is an open access article under the terms of the Creative Commons Attribution 4.0 International (CC BY 4.0) License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.

Data Availability

Article states: "The Tennessee Eastman Process dataset used in this study is publicly available. Code implementation that supports the findings of this study is available from the corresponding author upon reasonable request."

ORCID

0009-0000-0836-8481 (Yalim), 0000-0002-4798-003X (Handley)

Original Publication Citation

Yalim, C., Unal, R., & Handley, H. A. H. (2026). A three-stage causal root-cause diagnostic protocol for nonstationary industrial time series data. International Journal of Intelligent Systems, 2026, Article 3436744. https://doi.org/10.1155/int/3436744

Download

Share

COinS