Abstract
In the last decade, deep learning (DL) has outperformed model-based andstatistical approaches in predicting the remaining useful life (RUL) ofmachinery in the context of condition-based maintenance. One of the majordrawbacks of DL is that it heavily depends on a large amount of labeled data,which are typically expensive and time-consuming to obtain, especially inindustrial applications. Scarce training data lead to uncertain estimates ofthe model's parameters, which in turn result in poor prognostic performance.Quantifying this parameter uncertainty is important in order to determine howreliable the prediction is. Traditional DL techniques such as neural networksare incapable of capturing the uncertainty in the training data, thus they areoverconfident about their estimates. On the contrary, Bayesian deep learninghas recently emerged as a promising solution to account for uncertainty in thetraining process, achieving state-of-the-art performance in many classificationand regression tasks. In this work Bayesian DL techniques such as Bayesiandense neural networks and Bayesian convolutional neural networks are applied toRUL estimation and compared to their frequentist counterparts from theliterature. The effectiveness of the proposed models is verified on the popularC-MAPSS dataset. Furthermore, parameter uncertainty is quantified and used togain additional insight into the data.