Outlier detection in rotating machinery under non-stationary operating conditions using dynamic features and one-class classifiers

Abstract

The main goal of condition-based maintenance is to describe the machine state under current operating regimes, which can be non-stationary depending of load/speed changes. Besides, damaged machine data are not always available in real-world applications. This paper proposes a methodology of outlier detection in time-varying mechanical systems based on dynamic features and data description classifiers. Dynamic features set is formed by spectral sub-band centroids and linear frequency cepstral coefficients extracted from time-frequency representations. One-class classification is carried out to validate performance of the dynamic features as descriptors of machine behavior. The methodology is tested with a data set coming from a test-rig including different machine states with variable speed conditions. The proposed approach is validated on real recordings acquired from a ship driveline. The results outperform other time-frequency features in terms of classification performance. The methodology is robust to minimal changes in the machine state and/or time-varying operational conditions.