Nonlinear Process Fault Detection Method using Slow Feature Analysis

Abstract

Slow feature analysis (SFA) is a method that extracts the invariant or slowly varying features from an input signal based on a nonlinear expansion of it. This paper introduces SFA into industrial process monitoring. It overcomes the innate drawback of principal component analysis (PCA) that it fails to draw the more complex features or underlying nonlinear structure of the industrial process signals. Moreover, the invariance and slowness indicate the intrinsic properties of data. Thus the extracted information is interesting for data analysis. For the purpose of fault detection, two statistics are constructed: the T2 statistic and the SPE statistic. Then, these two statistics are applied to perform process monitoring. Simulations are run on the Tennessee Eastman (TE) process and the results illustrate the effectiveness of the proposed method.