A Novel Fault Detection Method Based on the Extraction of Slow Features for Dynamic Nonstationary Processes

Abstract

The industrial process often shows nonstationary characteristic, such as time-varying mean and variance, due to the unmeasured disturbances, adjustments of production plans, operator interventions, etc. Because of the overlook of nonstationarity, traditional multivariate statistical methods based on the assumption of stationarity often cause low fault detection rate (FDR) and high false alarm rate (FAR). In this article, a novel fault detection framework is proposed to improve the performance of fault detection when nonstationary and stationary variables coexist. It combines slow feature analysis (SFA) and K-nearest neighbor (KNN) based on Mahalanobis distance (SFA-MDKNN). First, stationary/nonstationary variables are distinguished by the unit root test. For nonstationary variables, the stationary residuals are obtained by Johansen cointegration analysis (CA) to form a combination matrix. Second, slow features in the combination matrix are extracted to achieve feature level fusion, and KNN distance based on Mahalanobis one is introduced as the monitoring index. Finally, considering that the effect of KNN algorithm is susceptible to data distribution, the FDR and FAR are taken into account by setting dynamic control limits. Tennessee Eastman process and hot rolling process (HRP) are used to validate the proposed framework and the computational efficiency of the algorithm.