Reduced Stationary Subspace Analysis for Anomaly Detection in Nonstationary Industrial Processes

Abstract

Anomaly detection is significant to ensure production efficiency and system safety of industrial processes. Complex processes generally show nonstationary properties, so nonstationary process monitoring has become a research hotspot recently. As two major methods, adaptive methods are difficult to model significant nonstationary trends, and cointegration analysis-based approaches are insensitive to the faults that are orthogonal to cointegration vectors. For monitoring nonstationary industrial processes, reduced stationary subspace analysis (SSA) is proposed in this paper. Different from traditional SSA, reduced SSA only considers the first-order stationarity. The simplified form facilitates the estimation of the stationary projection matrix, whose optimal solution is derived analytically by solving a generalized eigenvalue problem. After projection, stationary components can be obtained, which are monitored by the Mahalanobis distance. At last, the effective performance of reduced SSA is verified through the simulation on a nonstationary continuous stirred tank reactor.