Abstract: Quality of constructed graph data affects directly performance of the graph data-driven bearing remaining useful life(RUL) prediction. At present, traditional RUL prediction methods usually use spatio-temporal graphs to represent multi-sensor network data at different times. However, how to construct graph data representing the degradation state of bearing performance in a single-sensor monitoring application scenario and ensure its quality is still an open problem. Therefore, a path graph attention network-based bearing RUL prediction method for single-sensor monitoring application scenarios is proposed. First, a path graph is constructed by using the time-domain statistical features of the vibration signals in the whole life cycle of the bearing, in which the edges in the path graph are used to connect the vibration signals at adjacent moments. On this basis, a graph attention long short-term memory network is designed to extract the time-series vibration signal features and time dependencies hidden in the graph features(nodes, edge connections) of the path graph, so as to deeply reflect the whole bearing life degradation process. Verification and comparison experiments are carried out on the public bearing lifetime dataset. Results show that the proposed path graph construction clarifies the physical meaning of edge connections and improves graph data representation performance. Meanwhile, the proposed prediction method effectively captures the graph features and time dependence, characterizing the bearing degradation state and providing a reference for solving the problem of bearing performance degradation prediction in single-sensor monitoring application scenarios.

Key words: rolling bearing, remaining useful life, path graph, graph attention network, single-sensor monitoring application