Recognition of control chart patterns in auto-correlated process based on random forest

摘要

This paper proposed to use random forest as classifier for control chart pattern recognition in auto-correlated process. In order to construct the random forest, the bootstrap sample method is used to generate a great many different training datasets from the original dataset first. Based on each training data set, a CART (Classification and Regression Tree) is created to be base classifier, which gets each internal node split from an attribute subset randomly selected from whole attributes during growth to raise diversity. Then all base classifiers perform together, and all results are integrated through majority voting method to achieve the final result. An experiment on AR(1) process dataset created by Mont-Carlo simulation show that the proposed random forest approach perform well than BP neural networks both on recognition rate and noise resisting ability.