A neural network approach for the analysis of control chart patterns

Control chart pattern recognition is an important aspect of statistical process control (SPC). The presence of unnatural patterns indicates that a process is affected by assignable causes, and corrective actions should be taken. This paper describes two types of pattern recognizers based on different neural network architectures: a multilayer perceptron trained by back-propagation and a modular neural network. The pattern recognizers were developed to take the advantage of the fact that a particular unnatural pattern is often associated with a set of assignable causes. Identification of unnatural patterns can greatly narrow the set of possible causes that must be investigated, and thus the diagnostic search could be reduced in length. The performances of the proposed pattern recognizers were evaluated through Monte Carlo simulations on the basis of appropriate performance measures. An extensive evaluation indicates that the proposed pattern recognizers could recognize multiple unnatural patterns for which they were trained. The modular neural network provides a better recognition accuracy than back-propagation when there will be strong interference effects. The notable features of the proposed pattern recognizers include the compactness of input/output representation, the ability to detect unnatural patterns when the signal-to-noise ratio is low, the outstanding capability to detect unnatural patterns starting anywhere in the sequence of data and the directional invariance property in the sense that patterns of different orientations could be detected equally well.