Spectral Kurtosis against SVM for best frequency selection inbearing diagnostics

Abstract

Rolling bearing is probably the most widely used component in rotating mechanicalequipments and its condition monitoring and fault diagnosis to prevent the occurrence ofbreakdown is growing in interest since many years. Vibration signal based methods are themost popular and have been adopted in many kinds of condition monitoring systems. Startingin the early 60, an immense range of different methods has been proposed on this basis, toperform diagnosis, fault identification and classification of bearing faults. Among theothers, one typical approach consists in deep analysis of the most informative frequencyrange output of the system under test; the identification of this band is notstraightforward because the fundamental task consists in finding out the band which is themost informative in contents which, in turn, might not be corresponding to that one of themaximum response, as claimed by some authors. In this paper, Spectral Kurtosis and SupportVector Machine are analysed and compared and it is shown that they typically reach similarresults, in spite of their totally different approach. A brief description of both methodsis given and laboratory data are analysed from a lab rig which uses spare parts of a fullsize power transmission gearbox, designed by AVIO. By taking advantage of thesecomparisons, the analyses are conducted using classical indicators applied to the specificbands suggested by previous analysis such as the RMS and other statistical quantities.Multi dimensional graphs are reported to show the reliability of the obtained results.