Neural network classifier of time series: A case study of symbolic representation preprocessing for Control Chart Patterns

Abstract

Detecting abnormalities in manufacturing process at an early stage is advantageous. This can be done by monitoring its Control Chart Patterns (CCPs). These patterns can be considered as time series data. Therefore, accurate classification of CCPs is vital in the process control. There have been several types of CCPs classifiers with various degrees of accuracy. This work is concerned with implementation of a CCPs classifier based on neural networks and symbolic representation as its preprocessing method. CCPs are generated by the commonly used Generalized Autoregressive Conditional Heteroskedasticity (GARH) Model. Several of both feed forward and recurrent networks are investigated for the classifier. The symbolic representation of time series known as Symbolic Aggregate Approximation (SAX) is selected as its application was satisfactory in numerous similar works. In feed forward network, the Multilayer Perceptron network yields the best performance while Time-lag network yields the best performance for recurrent network. The results of neural network classifier which utilizes SAX in preprocessing in this work are superior than previous works which used the same CCPs model.