Unsupervised Remaining Useful Life Prediction for Bearings with Virtual Health Index

Abstract

A particular interest in achieving Prognostics Health Management (PHM) for bearings has been developed in the scientific community and the industry, as they are critical components in generators and turbines. Majority of state-of-the-art methods used in prediction of Remaining Useful Life (RUL) require large amounts of run-to-failure data for training. While these methods offer accurate prediction, the usage barrier is particularly high to small-scale, downstream sector companies due to the significant amount of data needed. The goal of this paper is to demonstrate a novel unsupervised method to address this problem. The algorithm takes advantage of Convolution Neural Network (CNN) encoder-decoder to infer Virtual Health Indices (VHI) which are representative of the degradation pattern. Additionally, thresholds for these indices are determined with only End-of-Life (EOL) data, removing the need for run-to-failure experiments. The RUL is then obtained through the inference of the VHI. The suggested method is tested on various benchmark datasets, particularly the XJTU-SY bearing dataset, offering promising prediction results to reduce the barrier of usage for RUL algorithms.