Multilayer perceptions for detecting cyclic data on control charts

SUMMARY

Cyclic data occur relatively frequently in manufacturing processes. Traditional approaches to detecting cyclic behaviour are mostly statistics-based, such as spectral analysis and time series analysis. In this paper, a special-purpose cyclic pattern recognition system applying neural networks is proposed. The system consists of multiple multilayer perceptrons with each perceptron dealing with cycles of a certain period. Thus, it incapable of identifying cycles of various periods. Multiple perceptrons may work concurrently, but a final decision is made through a unified decision rule. This type of special-purpose system is recommended when certain behaviour is known to exhibit more frequently in a given manufacturing process. Under the circumstances, an automatic assignable-cause interpretation system, which contains a special-purpose pattern recognition system as a core component, may be tuned to be more sensitive to this particular type of behaviour. Simulation indicates that a Special-purpose cyclic pattern recognizer performs comparably to a general-purpose pattern recognizer in detecting less noise-contaminated cycles, but performs superiorly in detecting cycles of higher noise and cycles of higher amplitudes.