Temporal attention mechanism in latent space for RUL prediction with bidirectional trend consistency

Abstract Smart ship regulations call for shore-based intelligent operation and maintenance of critical ship components. However, due to communication delays and related issues, existing remaining useful life (RUL) prediction methods often struggle with estimating RUL from disordered health monitoring time series. To address this, this paper proposes a latent-space temporal cycle-consistent attention method designed to extract reliable degradation features from incoherent time series. This method constructs a shared latent space that captures both forward and reverse sequence features. A cycle-consistent temporal attention mechanism is developed within this space to ensure that the relationships between forward and reverse sequences, as well as their chronological order, are maintained. The proposed method proves effective in generating consistent degradation representations, even under highly disordered conditions. Experimental results demonstrate that, under severely disordered conditions, the proposed method achieves a coefficient of determination ( R 2 ) of 0.8363 for degradation feature extraction and 0.8512 for RUL prediction, showing strong robustness and high accuracy in non-sequential data scenarios.