Dynamic linear models-based time series decomposition and its application on bearing fault diagnosis

Abstract

Although various methods have been presented in the literature for rotating machine fault detection, it still remains a huge challenge to accurately extract features from non-stationary vibration signals with a high noise level, typically in the case of rolling element bearings faults diagnosis. Due to its random and non-stationary nature, fault related features are difficult to extract by common techniques and are usually overwhelmed by noise and macro-structural vibrations. In this paper, a new time series decomposition method based on a dynamic linear model is proposed, which provides time domain decomposition of a non-stationary signal into collections of latent components. The signal of a damaged bearing consists of exponentially decaying ringing that occurs periodically at the bearing characteristic defect frequency. Dynamic linear models with time-varying cyclical components provide a more generalized and adaptive description on rolling element bearings’ vibration signal with time-varying cyclical behavior. This allows the precise isolation of latent, quasi-cyclical bearing fault components via inferences on the time-varying parameters which characterize these components. An accelerated whole lifetime test of bearing has been performed to collect vibration data, which is utilized to validate the effectiveness of the proposed method.