16 December 2025 in IEEE Transactions on Instrumentation and Measurement
The graph neural networks (GNNs) have drawn much attention for predicting the remaining useful life (RUL) due to their excellent performance in processing correlation relationship among data. Spatio-temporal evolution information and spatial correlation are key factors in the bearing degradation process. Multi–graph structures can comprehensively leverage the spatial correlation and temporal evolution characteristics inherent in different graph topologies to extract multi–scale spatio–temporal information. In this paper, a new graph isomorphism network (GIN-GIG) based on graph framework in graph is proposed to predict bearing RUL. Firstly, the feature spatial graph is constructed according to the correlation coefficients between the optimal features. Its output is used as the node feature of the new temporal path graph. The graph-in-graph (GIG) framework is constructed by embedding and fusing the feature spatial graphs and the temporal path graphs. Then, the graph isomorphism network (GIN) is established to capture the spatial information in the first-hierarchy spatial graph and the local temporal information in the second-hierarchy temporal path graph. Meanwhile, the self-attention mechanism is introduced to achieve adaptive hierarchical fusion of GIN output, so as to enrich the feature information. Finally, the time convolution network (TCN) is established to further extract the global temporal information for performing the RUL prediction. The effectiveness and superiority of the GIN-GIG method is validated using the PHM2012 and IMS bearing datasets, and some comparative experiments have been performed to show its advantage for the RUL prediction.