A Multiscale Pooling Attention-Based Graph Attention Network for Remaining Useful Life Prediction

Owing to the intricate spatial and temporal relationships inherent in data collected from multiple sensors, achieving precise predictions of remaining useful life (RUL) becomes a challenging task. Recently, deep learning (DL)-based approaches have made substantial advancements in RUL prediction. However, the traditional neural network-based methods have encountered some trouble in extracting spatial features. Graph neural network (GNN) has demonstrated the ability to effectively capture the spatial dependencies between multisensor data, but current GNN-based approaches cannot achieve much in terms of the spatial-temporal dependencies at various scales. Motivated by this, a multiscale pooling attention-based graph attention network (MSPA-GAT) is proposed. First, a multi-GATv2 network is designed for the spatial dependencies modeling, and a bidirectional long short-term memory (BiLSTM) network is used for modeling the temporal dependencies. Second, a multiscale pooling attention (MSPA) mechanism is constructed to highlight the local details of different scales and capture multilevel information. Finally, the effectiveness of the proposed MSPA-GAT to consider spatial and temporal dependencies is validated using two datasets. Moreover, the experimental results have shown that MSPA-GAT outperforms current state-of-the-art methods in RUL prediction.