TabNet-based Self-supervised Fault Diagnosis in Multivariate Time-series Process Data without Labels

Fault diagnosis is an essential field for the safe operation of chemical processes. In this paper, a self-supervised fault diagnosis method employing a tree-based deep learning model is proposed. The temporal information of multivariate time-series data is compressed with a Long Short-Term Memory structure, and the proposed method is demonstrated by performing the classification of fault types in the Tennessee Eastman process. It showed substantial performance enhancement compared to supervised learning, leveraging the feature representation obtained from unlabeled data. Notably, the tree-based characteristic of the proposed method provides interpretability of model results, illuminating the salient features of each fault type.