Detecting faults in process systems with singular spectrum analysis

Abstract

In this study, process monitoring based on signal decomposition by use of singular spectrum analysis (SSA) is considered. SSA makes use of adaptive basis functions to decompose a time series into multiple components that may be periodic, aperiodic or random. Two variants of SSA are considered in this investigation. In the first, the conventional approach is used based on latent variables extracted from the covariances of the lagged trajectory matrix of the process variables. The second approach is identical to the first approach, except that the covariances of the lagged trajectory matrices are replaced by Euclidean distance dissimilarities to decompose the variables into additive components. These components are subsequently monitored and the merits of the two approaches are considered on the basis of two case studies using simulated nonlinear data and data from the benchmark Tennessee Eastman process.