Statistical multivariate monitoring of rolling bearings working under varying operational conditions

Abstract

Abstract In recent years, statistical data‐driven approaches have been developed for health monitoring of rolling bearings working under variable operational conditions. Due to the reliance on reference models, most of them encounter with the model mismatch problem. A novel statistical multivariate health monitoring framework is proposed in this article to determine the initial degradation of rolling bearings. On the one hand, multivariate statistical parameters are extracted from original vibration signals to indicate bearing's degradation trajectories from different perspectives, a dynamic autoencoder neural network is extended with a matrix augmentation trick for discovering both dynamic and nonlinear latent representation in the high‐dimensional features. On the other hand, an eccentricity index is introduced for measuring information variations in the latent representation, such that bearing health development can be revealed timely. Then, an adaptive threshold is exploited for online monitoring of bearing damage. The effectiveness of the proposed work is substantiated using real vibration signals acquired from two different bearing degradation test rigs. The monitoring performance of the proposed work is evaluated in terms of false alarm rate (FAR). Experimental results confirm that the proposed work can identify early bearing damage under variable working conditions, by circumventing the model mismatch problem. It is found to be more superior than existing techniques.

Keywords

• Other topics in statistics
• Signal processing and detection
• Other topics in statistics
• Digital signal processing
• Civil and mechanical engineering computing
• Statistics