Estimation and generation of training patterns for Control Chart Pattern Recognition

Most applications of machine learning (ML) algorithms to control chart pattern recognition (CCPR) have focused on pattern detection and identification, rather than obtaining more detailed information about the pattern, which is important for effective assignable cause analysis. If real control chart data is not available for training purposes, synthetic patterns must be generated. Furthermore, pattern recognition accuracies achieved by different CCPR systems are usually not comparable since these were developed with different training data. How to create a diverse range of patterns for designing CCPR systems that can be compared and that are able to recognise a greater variety of patterns is an issue that needs studying.
This paper presents a scheme to generate training patterns that addresses this issue of diversity and comparability. The scheme also comprises change point detection and mean change categorisation methods that implement nonlinear models (NLMs) for estimating abnormal pattern parameters. The effect of this new pattern generation scheme on the accuracy of pattern recognition has been studied using two ML algorithms: Support Vector Machine (SVM) and Probabilistic Neural Network (PNN).
With the proposed pattern generation scheme, the mean pattern recognition accuracy was increased by 6.90% and 8.42% when SVM and PNN were used, respectively.