Shift detection and source identification in multivariate autocorrelated processes

Abstract

Motivated by the challenges of identifying the true source of shift in multivariate processes, we propose a neural-network-based identifier (NNI) for multivariate autocorrelated processes. A rather extensive performance evaluation of the proposed scheme is carried out, benchmarking it against three statistical control charts, namely the Hotelling T 2 control chart, the MEWMA chart, and the Z chart. The comparative study shows the strengths and weaknesses of each control scheme. The proposed NNI is most effective in detecting small-to-moderate shifts and has the most stable run-length property. Designing to identify the source of the shift, the NNI performs more stably than the Z chart under high autocorrelation. The NNI's source identification property can be further improved with the devised alternative decision heuristics. A pair-wise modular approach is also proposed to extend the NNI for multivariate processes.