Dynamic causal decoupling network based on intervention-invariant features for industrial fault diagnosis
Measurement Science and Technology · 27.03.2026 · 37 (14) · 146201
DOI: 10.1088/1361-6501/ae54be ↗Kurz gefasst
Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 08.04.2026. Verfasst von Jianlong Liu und 4 weiteren Autor:innen. Die Arbeit nennt 40 Literaturangaben. Gefördert von: Natural Science Foundation of Shanghai Municipality, Fundamental Research Funds for the Central Universities.
Abstract (Originalsprache)
Most existing fault diagnosis methods for chemical industries primarily rely on the statistical correlation between learned features and labels, without adequately considering robust dynamic causal features under varying operational states. Consequently, non-causal features influenced by sample selection bias are often incorporated into predictions as unstable paths. To address this, we propose the dynamic causal decoupling graph neural network (DCDGNN) based on a customized structural causal model. This framework decomposes the initial graph into a dynamic causal participation graph and a backdoor participation graph, integrating a dynamic recurrent graph convolutional network to extract both global and instantaneous causal features. An intervention-based training strategy combined with intervention-invariant regularization techniques effectively suppresses confusion bias and enhances causal feature stability. Experiments on benchmark three-phase flow platforms and real polyester esterification process demonstrate that DCDGNN significantly outperforms baseline models in diagnostic accuracy and robustness. Its causal participation graph visualization provides interpretable fault propagation pathways for engineering decisions. These results demonstrate that embedding causal theory into graph deep learning enhances the diagnostic capabilities for multivariate processes.
Förderung
- Natural Science Foundation of Shanghai Municipality
- Fundamental Research Funds for the Central Universities
Zitieren
APA
Liu, J., Peng, C., Wang, H., Chen, L., & Hao, K. (2026). Dynamic causal decoupling network based on intervention-invariant features for industrial fault diagnosis. Measurement Science and Technology, 37(14), 146201. https://doi.org/10.1088/1361-6501/ae54be
Vancouver
Liu J, Peng C, Wang H, Chen L, Hao K. Dynamic causal decoupling network based on intervention-invariant features for industrial fault diagnosis. Measurement Science and Technology. 2026;37(14):146201. doi:10.1088/1361-6501/ae54be
BibTeX
@article{Liu2026,
author = {Liu, Jianlong and Peng, Chuang and Wang, Hengqian and Chen, Lei and Hao, Kuangrong},
title = {Dynamic causal decoupling network based on intervention-invariant features for industrial fault diagnosis},
journal = {Measurement Science and Technology},
year = {2026},
volume = {37},
number = {14},
pages = {146201},
doi = {10.1088/1361-6501/ae54be}
}
EUVIMED