A Distributional Perspective on Remaining Useful Life Prediction with Deep Learning and Quantile Regression

• Published in: IEEE open journal of instrumentation and measurement Vol. 1; p. 1
• Main Authors: Zhang, Ming, Wang, Duo, Amaitik, Nasser, Xu, Yuchun
• Format: Journal Article
• Language: English
• Published: IEEE 2022
• Subjects: Data models; Deep learning; Degradation; distributional remaining useful lifetime (RUL) prediction; Distributional RUL prediction; Measurement; Predictive models; Quantile Regression; Rolling Bearing; Training; Uncertainty
• ISSN: 2768-7236; 2768-7236
• DOI: 10.1109/OJIM.2022.3205649

Abstract-With the rapid development of information and sensor technology, the data-driven remaining useful lifetime (RUL) prediction methods have been acquired a successful development. Nowadays, the data-driven RUL methods are focused on estimating the RUL value. However, it is more important to quantify uncertainty associated with the RUL value. This is because increasingly complex industrial systems would arise various sources of uncertainty. This paper proposes a novel distributional RUL prediction method, which aims at quantifying the RUL uncertainty by identifying the confidence interval with the cumulative distribution function (CDF). The proposed learning method has been built based on quantile regression and implemented from a distributional perspective under the deep neural network framework. The results of the run-to-failure degradation experiments of rolling bearing demonstrate the effectiveness and good performance of the proposed method compared to other state-of-the-art methods. The visualization results obtained by t-SNE technology have been investigated to further verify the effectiveness and generalization ability of the proposed method.