An Unsupervised Method based on Support Vector Machines and Higher-Order Statistics for Mechanical Faults Detection

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 (R) 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.