Recognition performance of imputed control chart patterns using exponentially weighted moving average

Abstract

Performance of control chart pattern recogniser (CCPR) is dependent on the quality of data. Furthermore, when data is partially missing, false alarms and misclassification rate are high. This paper studied CCPR with incomplete data and investigated effectiveness of the exponential smoothing in restoring the patterns aiming to increase the recognition accuracy. The results demonstrated that average overall recognition accuracy degrades from 99.57 (without missingness) to 76.33 in severe missingness. Classification errors in the incomplete random and trend patterns increased up to 38 and 44 times, respectively. Exponential smoothing with a constant of 0.9 is found to be an effective imputation technique. In 50% missingness, recognition accuracy of imputed dataset improved by 99.2% and 19.4% in stable and unstable patterns respectively. Type I error in trend and type II error in random and cyclic patterns were reduced significantly with EWMA imputation. Sensitivity tests proved pattern recognition using proposed imputation technique resulted in superior robustness performance.