An Interpretable Deep Transfer Learning-Based Remaining Useful Life Prediction Approach for Bearings With Selective Degradation Knowledge Fusion

Abstract

This article tries to answer the two questions of bearings’ remaining useful life (RUL) prediction with deep transfer learning: what bearing data in the source domain contribute more to transfer learning and how to quantify such contribution? From the perspective of sample-based interpretability, this article proposes a new deep transfer learning approach of RUL prediction. First, a new time series clustering algorithm based on multiscale degradation similarity is proposed. Comprehensively considering the geometry and tendency characteristics of degradation sequences, this algorithm can divide the source domain into multiple subsource domains. Second, a significance metric, named transfer domain validity index (T-DVI), is built to quantify the contribution of each subsource domain to transfer learning in terms of degradation similarity. Furthermore, a new stacked long short-term memory model with selective transfer learning is constructed. Running with the obtained T-DVIs, this model builds two new transfer strategies of weighted initialization and adaptive freezing to improve the transfer effect of degradation knowledge for RUL prediction. Experimental results on the two bearing datasets prove the effectiveness of the proposed approach. More importantly, the proposed approach makes the transfer learning process interpretable via identifying the significance of the bearings data, which helps transfer useful degradation knowledge and improve the RUL prediction performance as well.