Degradation assessment of bearings with trend-reconstruct-based features selection and gated recurrent unit network

Abstract

Degradation assessment (DA) is one of the most important technologies to implement the health management and predictive maintenance of rotating machinery. As the main task of DA, anomaly/degrade point detection and remaining useful life (RUL) prediction method of rolling element bearings is investigated in this paper. To detect the abnormal point more accurately, root mean square value is considered as the health monitoring indicator and 3σ rule is used to adaptively monitor the abnormal point. To predict precisely the RUL, a three-stage strategy is proposed. Firstly, twenty-four basic characteristics are extracted from vibration signal, which are reconstructed by using basic characteristics based complete ensemble empirical mode decomposition with adaptive noise (BC-CEEMDAN), and then the trend curves are extracted to reduce the fluctuation. Next, the most sensitive features are selected by employing a linear combination of monotonicity and correlation criteria. Finally, by input the selected features into the gated recurrent unit (GRU) neural network, we achieve the efficient health indicator with BC-CEEMDAN-GRU. To verify the effectiveness of the proposed approach, experiments on two bearing datasets are carried out, and the advantage is emphasized by comparison with the five existing methods.