International Journal for Quality and Productivity Management, vol. 11, n. 1, 2014 Superville 1 Outlier detection in autocorrelated manufacturing processes Claude Superville, PhD, CQE, FRSS 1 ABSTRACT In this simulation study, different schemes for monitoring production processes with of autocorrelated data are compared. A time series forecast is applied to the autocorrelated process and the resultant residuals are monitored by a control chart and two common tracking signals. Performance comparisons of different monitoring schemes have been typically based on the Average Run Length (ARL) criterion but we offer the the Cumulative Distribution Function (CDF) as an alternative and more informative performance evaluation criterion. This study compares the performance of the Individuals Control Chart (ICC), the Smoothed Error Tracking Signal (ETS), and the Cumulative Sum (CTS) Tracking Signal (CTS) in terms of their ability to detect the presence of additive outliers. Based on the CDF, we found that the Individuals Control Chart offers the greatest probability of early detection of an additive outlier in an autocorrelated process. Keywords: autocorrelation, control charts, tracking signals Introduction The occurrence of large unusual observations is not uncommon in time series data. These outliers may be due to recording errors or to one-time unique situations such as an unexpected change in demand for a product or a change in a production system. Fox (1972) defines two types of outliers that may occur in practice; additive outliers, corresponding to external disturbances that affect the value of a single observation; and, innovational outliers, or step shifts, refering to internal disturbances that change the value of an observation and all other successive observations. Typically in process control environments, the performance of monitoring schemes have been compared based on only in terms of their ability to detect step shifts or innovational outliers (Montgomery, 2013). However, which monitoring scheme detects the presence of an additive outlier most quickly is also of great interest in determining abnormal process behavior, a measurement error, a recording error or incorrect specifications based on distributional assumptions. (Walfish, 2006). Autocorrelation implies the existence of a relationship between the outcomes produced in different time periods by the same process. With advances in measurement technology along with more frequent sampling, today‟s manufacturing processes often yield observations that are autocorrelated. The inertia effects present in most manufacturing processes coupled with the advent of automated gages that sample processes more frequently, often render most process data autocorrelated (Montgomery and Mastrangelo 1991, Woodall and Faltin 1993). 1 Texas Southern University, USA