AN ENSEMBLE MODEL FOR STATISTICAL MONITORING OF PATTERNS IN BIVARIATE PROCESSES BASED ON MULTIPLE ARTIFICIAL NEURONAL NETWORKS

Abstract

Multivariate control graphics detect signals out of control of the process. These signals are special patterns of joint variation, but they do not allow to determine what types of variation patterns take place in the individual variables. The referred problem has been treated through models of pattern recognition (PR) by Artificial Neural Networks (ANN). There are important advances in solving the problem to univariate cases, but not so in multivariate cases. There is no research which affirms that a single ANN can identify a multivariate out-of-control signal and recognize the special variation types of the variables individually. This research presents a model of PR of special variation in bivariate processes, and is based on an organized assembly of different types of ANN which are activated sequentially. With this work, it is possible to obtain a diagnosis of the bivariate process control that simultaneously recognizes the type of variation of the variables involved. This novel model provides the basis of new knowledge about statistical control of bivariate processes by PR through ANN. The model had two stages of training: experimental and industrial. The first one worked with data generated by Montecarlo simulation and the second one with data from a process that performs manufactured operations on metal bars used in the speed system in automobiles. Keywords: Statistical quality control, Pattern recognition, Artificial neural networks, Perceptron, Backpropagation