Effectiveness of Different Preprocessing Techniques on Classification of Various Lengths of Control Charts Patterns

Abstract

Abstract Control Chart Patterns (CCPs) can be considered as time series. They are used in monitoring the control process, therefore an ability to recognize these patterns are essential in manufacturing as abnormalities can then be detected at early stage. Feeding CCPs directly to classifiers has proven unsatisfactory, especially in existence of noises. Therefore, there have been different kinds of preprocessing of CCPs to aid the classification. This research has two main objectives, first, is to study how the lengths of CCPs affect the performance of the classification. Second, the study attempts to determine the most suitable preprocessing techniques that have been applied to CCPs. Three preprocessing techniques are selected, these are Kalman filter, statistical features and symbolic representation known as Symbolic Aggregate ApproXimation (SAX). The Minimum Descriptive Length (MDL) algorithm for selecting SAX parameters is also investigated. Neural network is chosen as the tool to implement classifiers. The study concludes that longer patterns are more preferable than shorter ones. Statistical features is found as the best preprocessing techniques.