Rolling bearing prognostic analysis for domain adaptation under different operating conditions

Abstract

In real-world applications, machinery operates under non-stationary conditions such as operating environment, failure modes, and noise where domain shift problems generally arise. Hence, deep learning methods are trained on one working condition cannot generalize effectively on different conditions. Also, the suitability of prognostic features significantly affects the prediction results. To address these issues, this paper proposes a transfer learning-based bi-directional Long Short-Term Memory (TBiLSTM) network for extracting prognostic sensitive features, and domain adaptation is realized using multi-kernel maximum mean discrepancy (MK-MMD). Therefore, the proposed TBiLSTM network can be utilized for RUL prediction of bearings under multiple working conditions. The superiority and effectiveness of the TBiLSTM method are validated by both experimentation and comparison with state-of-art methods. In addition to this, prediction results demonstrate its effectiveness on IEEE PHM challenge datasets. Hence, the results demonstrate that the prognostic features are more reliable and domain-invariant for RUL prediction.