Fault detection in the Tennessee Eastman benchmark process with nonlinear singular spectrum analysis

Abstract

Multivariate statistical process monitoring methods aim at detecting and identifying faults in the performance of processes over time in order to keep the process under control. Singular spectrum analysis (SSA) is a potential tool for multivariate process monitoring. It allows the decomposition of dynamic process variables or time series into additive components that can be monitored separately to identify hidden faults that may otherwise not be detectable. However, SSA is a linear method and can give misleading information when it is applied to dynamic processes with strong nonlinearity. Therefore, in this paper, nonlinear versions of SSA based on the use of auto-associative neural networks or auto-encoders and dissimilarity matrices are considered. This is done based on the benchmark Tennessee Eastman process that is widely used in the evaluation of statistical process monitoring methods.