Machinery multimodal heterogeneous-aware RUL prediction: A spatial-temporal graph network for cross-temporal information propagation and dynamic fusion

Graph neural networks (GNNs) have emerged as a promising technology in the remaining useful life (RUL) prediction. Despite the substantial progress achieved by GNN-based approaches, several limitations remain: 1) Existing methods fail to adequately account for the multi-attribute heterogeneous associations inherent in multi-sensor data, thereby constraining effective feature extraction from diverse sources. 2) Conventional methods generally construct graph structures based solely on single-timepoint sensor associations, thereby limiting the exploitation of rich temporal information. To address the above challenges, a heterogeneous graph convolutional network with cross-temporal information propagation, termed HGCN-CIP, is proposed to implement the RUL prediction for mechanical systems. First, a cross-temporal heterogeneous graph architecture is proposed to enable the joint modeling of sensor synchronization and interaction relationships across time steps. Two types of heterogeneous meta -paths are introduced to explicitly capture the diverse dependencies within the sensor network. Second, a multi-scale heterogeneous graph convolutional network (MSHGCN) incorporating the information propagation operation is proposed, which generates heterogeneous sub-features by temporally aggregating node features from temporal graphs under different meta -paths. A dynamic feature fusion layer is then employed to integrate the sub-features, enabling effective feature fusion and accurate RUL prediction. Experimental results demonstrate that the proposed method surpasses multiple state-of-the-art approaches in RUL prediction tasks.