Parameterization of a fuzzy classifier for the diagnosis of an industrial process

The aim of this paper is to present a classifier based on a fuzzy inference system. For this classifier, we propose a parameterization method which is not necessarily based on an iterative training. This approach can be seen as a pre-parameterization which allows the determination of the rules base and the parameters of the membership functions. We also present a continuous and derivable version of the previous classifier and suggest an iterative learning algorithm based on a gradient method. An example using the learning basis IRIS, which is a benchmark for classification problems, is presented showing the performances of this classifier. Finally this classifier is applied to the diagnosis of a D.C. motor showing the utility of this method. However in many cases the total knowledge necessary to the synthesis of the fuzzy diagnosis system is not, in general, directly available. It must be extracted from an often considerable mass of information. For this reason, a general methodology for the design of a fuzzy diagnosis system is presented and illustrated on a non-linear plant.