A Novel Transfer Ensemble Learning Framework for Remaining Useful Life Prediction Under Multiple Working Conditions
Jilun Tian, Yuchen Jiang, Jiusi Zhang, Shimeng Wu, Hao Luo
Abstract
Data-driven remaining useful life (RUL) prediction is critical for industrial devices. There is an important assumption for classic machine learning methods that the training and test sets need to follow independent and identical distribution (IID), which does not hold under multiple working conditions. To relax the IID assumption, transfer learning is a key technique, which is also limited by the knowledge of the target domain data distribution. This paper proposes a novel transfer ensemble learning (TEL) framework, which can effectively utilize the information of source domain and improve the generalization ability of the model to unknown target domain. The framework mainly relies on the knowledge of metric learning, and adopts Kullback-Leibler (KL) divergence to measure the differences in data distributions. A domain dissimilarity metric is proposed to ensure that sub-models of similar datasets have a greater impact on the results. To verify the performance of this framework, a real filtering system from the PHM 2020 competition is used. Meanwhile, the information of time series data can be fully utilized by using the bidirectional long short-term memory (Bi-LSTM) model. Experimental results show that the proposed TEL-Bi-LSTM method outperforms existing machine learning methods.