Recognition of Control Chart Patterns Using Decision Tree of Multi-class SVM

Abstract

This paper aims to realize the automatic recognition of abnormal patterns of control charts in a statistical process control system. A novel multi-class SVM is proposed to recognize the control chart patterns, which include six basic patterns (i.e. normal, cyclic, up-trend, down-trend, up-shift, and down-shift pattern). Unlike the commonly used One-Against-All (OAA) implementation methods, the structure of proposed multi-class SVM is same as a special decision tree with each node as a binary SVM classifier, which is built via recursively dividing the training dataset of six classes into two subsets of classes. The proposed multi-class SVM can increase recognition accuracy and resolve the unclassifiable region problems caused by OAA methods. Based on this, Monte Carlo simulation is used to generate training and testing data samples. The results of simulated experiment show that the problem of false recognition has been addressed effectively, and the proposed decision tree of multi-class SVM is more effective in detecting unnatural patterns on control charts than the traditional OAA methods.