BiCoT-VTASA: A Transformer-Enhanced Hybrid Model for Predicting Remaining Useful Life of Aircraft Engines

Accurate prediction of the Remaining Useful Life (RUL) of aircraft engines is crucial to ensuring both aviation safety and effective maintenance management. Traditional RUL prediction models often face difficulties in modeling long-term dependencies within complex time-series data and in extracting latent features, which limits their predictive accuracy. To address this challenge, this paper proposes the BiCoT-VTASA hybrid model, which integrates two innovative modules, Bidirectional Contextual Transformer (BiCoT) and Variational Trend-Aware Self-Attention (VTASA), to improve the model’s ability to model complex time-series data. The BiCoT module improves the Transformer encoder by incorporating the Bidirectional Long Short-Term Memory (BiLSTM) network, innovatively strengthening the model’s ability to exploit long-range sequential relationships in data and effectively overcoming the traditional Transformer’s limitations in capturing global dependencies in time-series data. On the other hand, the VTASA module incorporates the Variational Autoencoder (VAE) into the Trend-Aware Self-Attention (TASA) mechanism, accurately extracting latent features from the data, further enhancing the model’s robustness in handling complex and noisy data. The synergistic interaction of these two modules enables BiCoT-VTASA to comprehensively and deeply mine the temporal information and latent structures in the data, significantly improving predictive accuracy. Ablation and prediction experiments conducted on the widely used C-MAPSS benchmark dataset validate the model’s rationality and effectiveness. Compared to other traditional and advanced models, BiCoT-VTASA outperforms the baseline models in terms of Root Mean Squared Error (RMSE) and Score metrics on the more complex FD002 and FD004 datasets, demonstrating the model’s strong performance in aircraft engine RUL prediction tasks. This study provides a more efficient and precise solution for addressing RUL prediction problems in complex time-series data.