Bayesian Deep Learning Framework with Variational Inference for Uncertainty Quantification in RUL Prediction

Uncertainty in Remaining Useful Life (RUL) prediction must be quantified to provide confidence in the results. Balancing the quantification of uncertainty with the need for accurate predictions is a challenge. To this end, this paper presents a novel Bayesian Deep Learning (BDL) method with Variational Inference (VI) for predicting the RUL of mechanical equipment and quantifying the associated uncertainty. The proposed framework introduces a new Evidence Lower Bound (ELBO) loss function, controlling the trade-off between the negative log-likelihood loss and the Kullback-Leibler (KL) divergence via a tunable weight parameter. The proposed approach integrates a Long Short-Term Memory (LSTM) network to effectively handle sequential sensory data, and enhances prediction accuracy by incorporating attention mechanism that emphasizes critical time steps. A case study utilizing a turbofan engine dataset demonstrates the superiority of this method, achieving lower Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) compared to other Bayesian approaches.