Forecasting remaining useful life: Interpretable deep learning approach via variational Bayesian inferences