Abstract

Bearings are one of the more widely used elements in rotating machinery, reason why they have focused the attention of many researches in the last decades. The aim is to obtain a methodology that allows a reliable diagnosis of this kind of elements without dismounting them from the machine, and detecting the failure in incipient stages before a critical failure occurs. This manuscript develops and improvement of a technique showed in [1] of automated diagnosis of bearings through vibration signals, using the coefficients of the Multirresolution Analysis (MRA) and Multilayer Perceptron (MLP) neural network (NN). Data were obtained from a quasi-real industrial machine, where bearings were supporting axial and radial loads while rotating at different speeds. This technique offered very good results when diagnosing healthy and faulty bearings, nevertheless the reliability decreased when distinguishing between different kinds of failures. The novel technique showed in the present work, increases the success rates obtained using the same data: not only allows detecting early faults but also their location with high accuracy. The methodology exposed in this work is based on the use of the relative energy of the Wavelet Packets Transform (WPT), and NN, concretely, the RBF.