Abstract:

To reduce unexpected failures and maintenance costs of complex industrial equipment, this paper proposed a BiLSTM-based ensemble model for RUL prediction with uncertainty quantification. First, a convolutional VAE (CVAE) was designed to extract degradation features using residual connections and global feature aggregation, with a dynamic Kullback Leibler (KL)-divergence weighting strategy to balance reconstruction loss and latent space regularization. Next, a BiLSTM-Transformer temporal model was constructed, where BiLSTM captured long-range dependencies and multi-head attention focused on critical degradation stages. Finally, a multi-quantile prediction subnet was trained via QR to generate RUL interval forecasts, while kernel density estimation (KDE) estimated the probability density distribution. Experiments on National Aeronautics and Space Administration (NASA) C-MAPSS dataset show that, compared with state-of-the-art methods, the proposed approach achieves better results in both point prediction (4.28% lower in RMSE, 21.91% lower S-score) and interval prediction (14.12% higher in coverage, 14.09% narrower in average prediction interval), demonstrating its effectiveness and reliability in complex industrial scenarios.

Key words: bidirectional long short-term memory networks (BiLSTM), remaining useful life (RUL), variational autoencoder (VAE), ensemble model, Transformer, quantile regression (QR)