A Hybrid Prognostic Model based on Bayesian Deep Learning Network and Uncertainty Quantification for Roller Bearings

Abstract

Good prognostics and health management (PHM) technology has been proven to provide safe and reliable operation for mechanical or electrical devices or equipment-critical components in industry. Using the collected vibration signals, many data-driven prediction approaches show good potential for the remaining useful life (RUL) prediction for roller bearings, but an open problem is still to develop a more accurate and practical prognosis method and further extent to the whole machinery. However, for the same bearing, the prediction results can differ due to many uncertainties, such as operation condition, model inaccuracy, noisy signal. Inspired by the Bayesian theorem and deep learning, a hybrid prediction model based on the Bayesian neural network and bidirectional long short-term memory (LSTM) is presented for bearing prognostics in this paper. The Bayesian machine learning and deep learning work together to provide accurate predictions with uncertainty quantification. In the experiment, the vibration datasets from the platform PRONOSTIA (PHM2012) are applied to the hybrid model. The experimental results and comparison with the LSTM indicate that the hybrid network provides more accurate prediction results and 95% confidence interval for roller bearings.