Comparison of supervised classification algorithms combined with feature extraction and selection: Application to a turbo-generator rotor fault detection

The goal of this paper consists in applying pattern recognition methods to turbo-generators. Previous works have shown that a monitor, thanks to pattern recognition, is practical on asynchronous machines. This procedure has rarely taken advantage of these methods for turbogenerator. The statistical model has been obtained from harmonics extracted from flux probes and from stator current and voltage. For this purpose, the main way is to build a learning matrix to predict the functional state of a new measurement. Finally, three classifiers have been compared: k Nearest Neighbors, Linear Discriminant Analysis and Support Vector Machines. The best classification result is obtained by Linear Discriminant Analysis combined with Factorial Discriminant Analysis achieving a score of 84.6%.

I. Nomenclature

Stator voltage phase and Stator current phase and Standard deviation of with or (also ) Deformation of the normalized characteristic With or Number of prototypes Number of test data Number of features , the learning matrix , a prototype (also x_i) , the test matrix a test data feature , a column of (x_i) The class of x_i Class Set of all classes card , mean of the class Random variable corresponding to the samples Random variable corresponding to the classes The membership degree of to the class nearest prototypes of The covariance matrix of the prototypes The within and between classes covariance matrix of the prototypes The within and between classes covariance matrix of the class