An Attention Minimal Gated Unit-Based Causality Analysis Framework for Root Cause Diagnosis of Faults in Nonstationary Industrial Processes

Abstract

Root cause diagnosis is an important part of the fault diagnosis framework, which is often used to locate the root causes and identify the propagation paths. Most of the traditional root cause diagnosis methods consider the time series of industrial processes to be stationary or nearly stationary after faults occur. Since fault information is often propagated according to the causalities between process variables, and the pseudo-regression caused by nonstationary characteristics is not conducive to correct causality analysis, further affects the root cause diagnosis performance. Associated with those trends, in this article, a new causality analysis framework is proposed for root cause diagnosis of faults in nonstationary industrial processes. Specifically, the augmented Dickey-Fuller (ADF) test is first used to determine the stationarity of the time series, and the combination method of cointegration analysis (CA) and higher order difference is used for extracting the stationarity factors from nonstationary time series. Then, an attention minimal gated unit (AMGU)-based nonlinear dynamic causality analysis method is developed for causal topology construction and root cause diagnosis. Finally, industrial verifications on two datasets from actual hot rolling processes (HRPs) show that the proposed scheme is feasible, and is superior to competitive methods in terms of solving the issues of root cause diagnosis of faults in nonstationary industrial processes.