Root cause diagnosis in multivariate time series based on modified temporal convolution and multi-head self-attention. (September 2022)

• Record Type: Journal Article
• Authors: Zhou, Yujie; Xu, Ke; He, Fei
• Is Part Of: Journal of process control. Volume 117(2022)
• Page Start: 14
• Page End: 25
• Publication Date: 2022-09
• Subjects: Causal discovery -- Chemical process -- Deep learning -- Root cause diagnosis -- Self-attention; Process control -- Periodicals; Fabrication -- Contrôle -- Périodiques; Process control; Periodicals; Electronic journals; 660.281
• DOI: 10.1016/j.jprocont.2022.06.014
• Languages: English
• ISSNs: 0959-1524

Abstract

Abstract: Accurate causal discovery is significant for the data-driven root cause diagnosis. A novel framework based on modified temporal convolution and multi-head self-attention (MTCMS) is proposed for causal discovery and root cause diagnosis in multivariate time series. The temporal convolutional network is modified through feature reconstruction and skip connection to heighten the ability of feature extraction, which are further used to mine causalities. The multi-head self-attention is modified via threshold normalization for quantifiable causal inference, elevating the accuracy and generalization performance. Additionally, the calculation rule of root node score based on contrastive causal graph is proposed for root cause diagnosis. The MTCMS network outperforms all ablation structures and comparison methods in the causal discovery experiment based on synthetic data, manifesting its effectiveness and superiority. In the root cause diagnosis experiment based on the Tennessee Eastman benchmark process, the diagnosis result is consistent with the mechanism analysis, which further demonstrates its effectiveness. Highlights: MTCMS is proposed for causal discovery between multivariable time series. Root cause of the fault is diagnosed via contrastive causal graph. The different delay times in causality are fully explored by modified TCN. Causality between variables is quantified by modified self-attention. A joint loss function is designed to train deep learning networks.