An Unsupervised Method based on Support Vector Machines and Higher-Order Statistics for Mechanical Faults Detection
Fernando Elias de Melo Borges, Andrey Willian Marques Pinto, Diogo Aranha Ribeiro, Tássio Spuri Barbosa, Daniel Pereira, Ricardo Rodrigues Magalhães, Bruno Henrique Groenner Barbosa, Danton Diego Ferreira
Abstract
In this paper an unsupervised method to detect mechanical faults using support vector machines and higher-order statistics is proposed. The method extracts compact vector features - based on higher-order statistics - from vibration signals and uses the one-class support vector machine to build a closed region around the data from the health structure. The method was evaluated considering two cases: fault detection in a cantilever beam and in a three-phase induction motor. In both cases, the vibrations were collected by a 3 axis accelerometer sensor. The acquisition system was controlled by an open-source electronic prototyping ARDUINO <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">®</sup> platform. After collecting the data, higher-order statistics-based features were extracted. These features were presented to the one-class support vector machine for fault detection. The proposed method was capable of identifying a closed region in a two-dimensional space so that events inside this region are signed as no faults and events outside this region are signed as faults. The method has two important characteristics: (i) it requires only healthy mechanical structures to be designed, and (ii) it operates in a low dimensional space (only two) constructed by the higher-order statistics features, which requires low computational cost in the operational phase.