Process Mean Shift Detection Using Prediction Error Analysis

Abstract

Autocorrelation in 100 percent measurement data results in false alarms when the traditional control charts, such as X and R charts, are applied in process monitoring. A popular approach proposed in the literature is based on prediction error analysis (PEA), i.e., using time series models to remove the autocorrelation, and then applying the control charts to the residuals, or prediction errors. This paper uses a step function type mean shift as an example to investigate the effect of prediction error analysis on the speed of mean shift detection. The use of PEA results in two changes in the 100 percent measurement data: (1) change in the variance, and (2) change in the magnitude of the mean shift. Both changes affect the speed of mean shift detection. These effects are model parameter dependent and are obtained quantitatively for AR(1) and ARMA(2,1) models. Simulations and examples from automobile body assembly processes are used to demonstrate these effects. It is shown that depending on the parameters of the AMRA models, the speed of detection could be increased or decreased significantly.