Large-scale chemical process causal discovery from big data with transformer-based deep learning

Abstract

Fault diagnosis is critical for ensuring safe and stable chemical production. Correct identification of causal relationships among variables in large-scale chemical processes is a prerequisite for analyzing the root causes and propagation paths of faults. However, chemical process big data often exhibit nonlinearity and nonstationarity, and contain various forms of noise, rendering conventional causal discovery methods vulnerable. In this paper, a novel causal discovery method based on the causality-gated time series Transformer (CGTST) is proposed to address this challenge. By performing time series prediction using the Transformer-based model on the target variable, CGTST measures the causal strength by assessing the contribution of each variable to the prediction through the causality gate structure. Furthermore, a causal validation method based on permutation feature importance is proposed to eliminate spurious causal relationships and ensure robust results. To enhance the performance of causal discovery on nonlinear and nonstationary chemical process data, ensemble empirical mode decomposition is employed to reduce noise. The CGTST-based method is validated on three case studies: a continuous stirred-tank reactor, the Tennessee Eastman process, and a real-world continuous catalytic reforming process. Our findings demonstrate that the proposed method outperforms conventional causal discovery methods and holds promising prospects for industrial applications.