Classification of extended control chart patterns: a neural networks approach
Bunthit Watanapa, Jonathan H. Chan · 2006
by classifying six common control chart patterns into eight classes of time series data patterns. Incorporating two more patterns of bottom-out and peak-off can yield better insights for not only the traditional real time control environment but also the behavioral study in other time-domain systems such as money and security markets. This work reports the results of empirical study on the incorporation of new extracted features, especially those with a lesser extent of outliers ’ effect, for example median, robust regression and RMS value of the time series. The feedforward backprogation ANN is deployed and experimented using two different training schemes, namely the Levenberg-Marquardt method and the Bayesian regularization. The best performance generated by the ANN is 98 % classification accuracy. Technical insights into the model settings are also provided.