Setting Parameters of Control Chart Pattern Classifier for Effective Patterns Recognition

Abstract

This research study examines the dynamic field of control chart pattern recognition, which has seen tremendous growth in research over the last few decades. The field has seen advancements in efforts to improve the accuracy of artificial neural network (ANN) based classifier for control chart patterns recognition, including the use of statistical features, wavelet-based denoising techniques for input representation, and the development of both integrated and modular recognizer designs. This study applies design of experiments (DOE) methodology to select the setting parameters for the ANN-based classifier. This approach is meant to aid in the improvement and optimization of pattern recognition systems within the context of statistical process control. The main goal of this study is to assess the effectiveness of an ANN classifier in identifying different kinds of abnormal patterns within statistical process control. The simulated control chart pattern (CCP) samples were generated using programming software which is MATLAB. The multilayer perceptron neural network model was utilized as the CCP classifier. The setting parameters were analysed and optimized using design of experiment technique. The result from this study shows that with low window size, it is enough to achieve 100% for normal pattern and cyclic pattern. From the New Set 3 that has parameter low recognition window size (20), high amount of training dataset for normal pattern (1500), low amount of training dataset for shifts pattern (50), high quality data abnormal pattern for shifts patterns (0.1) and low quality data abnormal pattern for trends pattern (0.002), it can conclude that the factor is superior for recognising shifts patterns also known as balance recognition.