Hierarchical attention temporal model with degradation mode-aware health indicator similarity matching for aeroengine remaining useful life prediction

Health indicator similarity matching(HISM)-based prognostic methods have achieved promising results in aeroengine remaining useful life (RUL) prediction, but they still face several limitations: 1) difficulty in constructing accurate HIs to effectively characterize aeroengine degradation because of insufficient consideration of long-term temporal dependencies in multidimensional time-series (MTS) data, particularly for aeroengines working under nonstationary conditions; and 2) neglect of the differences among degradation modes (DMs) within HISM, leading to mismatched HIs and inaccurate RUL predictions across diverse degradation scenarios. Accordingly, a novel hierarchical attention temporal model with DM-aware HISM framework (HATFormer) is proposed, which integrates spatiotemporal representation learning with DM-aware HISM (DMAHISM) strategy, to improve prognostic performance under complex DMs. First, a hierarchical attention-based spatiotemporal aggregation autoencoder is designed, which embeds multi-head self-attention into a specially designed time-series encoder-decoder to achieve spatial aggregation and temporal memory of MTS in an unsupervised manner, thereby enabling accurate HI extraction. Second, a novel DMAHISM strategy is designed, which decouples HISM into DM recognition and DM-specific HI matching, and performs RUL prediction by probability-weighted fusion of DM-specific predictions, significantly improving prediction accuracy and robustness. Finally, the effectiveness of HATFormer is validated by extensive comparative experiments on four aeroengine degradation datasets with diverse and complex DMs.