Unsupervised Domain Deep Transfer Learning Approach for Rolling Bearing Remaining Useful Life Estimation

Abstract

Abstract Accurate estimation of remaining useful life (RUL) becomes a crucial task when bearing operates under dynamic working conditions. The environmental noise, different operating conditions, and multiple fault modes result in the existence of considerable distribution and feature shifts between different domains. To address these issues, a novel framework TSBiLSTM is proposed that utilizes 1DCNN, SBiLSTM, and attention mechanism (AM) synergically to extract highly abstract feature representation, and domain adaptation is realized using the MK-MMD (multi-kernel maximum mean discrepancy) metric and domain confusion layer. One-dimensional CNN (1DCNN) and stacked bidirectional LSTM (SBiLSTM) are utilized to take advantage of spatiotemporal features with attention mechanism (AM) to selectively process the influential degradation information. MK-MMD provides effective kernel selection along with a domain confusion layer to effectively extract domain-invariant features. Both experimentation and comparison studies are conducted to verify the effectiveness and feasibility of the proposed TSBiLSTM model. The generalized performance is demonstrated using IEEE PHM data sets based on root mean squared error, mean absolute error, absolute percent mean error, and percentage mean error. The promising RUL prediction results validate the superiority and usability of the proposed TSBiLSTM model as a promising prognostic tool for dynamic operating conditions.