Progressive hybrid hypergraph attention network with channel information fusion for remaining useful life prediction of rolling bearings

Nowadays, remaining useful life (RUL) prediction of rolling bearings based on graph neural network has attracted extensive attention. However, most existing methods can only model pairwise correlations, while they have not yet concerned the complex higher-order relationships and lack the ability to learn global-local information. This paper proposes a progressive hybrid hypergraph attention network for RUL prediction of rolling bearings with multi-channel signals, where the pair-wise graphs and hypergraphs are incorporated to simultaneously capture high-order and low-order dependencies. Initially, the network constructs a dynamic graph structure to extract global to local temporal information with the progressive elimination of neighbor nodes. Meanwhile, the channel information is fused through the connectivity of hybrid hypergraph construction. Afterwards, node-level and hyperedge-level attention are emphasized to enhance the contribution of significant nodes. Eventually, the high-order and low-order features are blended via an adaptive fusion module. The experimental study on two benchmark datasets and a time-varying condition dataset of rolling bearings indicates the effectiveness, generalization, and comparable efficiency of the proposed approach. Besides, the impact of channel information variability on prediction results is explored with the considerable effectiveness of the fusion strategy.