Temporal multi-resolution hypergraph attention network for remaining useful life prediction of rolling bearings

Abstract

Accurate remaining useful life (RUL) prediction of rolling bearings plays a vital role in ensuring the safe operation of mechanical equipment. Graph-based models have become an emerging trend in RUL prediction by converting monitoring samples into graph structures to capture samples’ relationships effectively. However, graph-based models only use pairwise samples to model the relationships between samples and cannot capture the non-pairwise high-order relationships between multiple samples. Besides, graph-based models rely heavily on predefined graphs to aggregate relevant features. The bearing monitoring datasets have no explicit structure, and the predefined graph structures cannot characterize datasets. Aiming at these issues, a temporal multi-resolution hypergraph attention network (T-MHGAT) is proposed. Firstly, the bearings’ monitoring samples are established and fused into a multi-resolution hypergraph (MHG) to characterize the potential structure of bearings monitoring datasets. Then, a hypergraph attention network (HGAT) is designed to mine the high-order relationships between signal samples on hypergraph data. Meanwhile, multiple gated recurrent units (GRUs) are constructed to capture the signal samples’ temporal information. Finally, the linear layer is built after GRUs to output RUL prediction values. Many experiments on two rolling bearing datasets showed the effectiveness of T-MHGAT, which can lay the foundation for predictive equipment maintenance.