Fault detection by local grouping based on MEWMA control charts

Abstract

In modern process industries fault detection is becoming more important. Many fault detection methods are based on Hotelling's T2 statistic. With an increasing number of variables, the Q statistic can detect only larger failures in the variables. A new method is introduced to improve the fault detection of T2 and MEWMA (multivariate exponential weighted moving average) control charts. The basic idea is to group variables locally. This method should not be confused with sub-grouping of data with individual observations. By local grouping, the dimension of the measured data will be reduced and the failures in the non-grouped variables can be detected more easily. The effectiveness of the new method is shown also by simulations of the average run length. The new method is illustrated by the fault detection of an industrial CO 2 sensor.