Prediction of Bearings Remaining Useful Life Across Working Conditions Based on Transfer Learning and Time Series Clustering

Recently, data-driven remaining useful life (RUL) prediction has become a promising tool in prognostics and health management for rolling bearings. In many actual applications, however, it is not easy to collect whole-life degradation data of bearings, while training with an insufficient amount of data would result in a biased RUL prediction model. If the monitoring data historically accumulated under different working conditions are introduced to facilitate model training, it probably tends to get inadequate prediction performance due to the violation of the precondition of independent and identical distribution (
${i.i.d.}$
). To solve this problem, a new bearing RUL prediction approach is proposed by utilizing the transfer learning strategy. First, a new time series clustering algorithm is proposed to exploit clusters of different degradation series. By integrating phase space warping and dynamic time warping, this algorithm is able to measure the similarity by using global degradation information and then get better clustering performance. Second, a new temporal domain adaptation method is proposed to obtain the pivot feature set of different bearings based on meta-degradation information which is represented by the principal curve of each cluster. Finally, support vector machine is run to establish the prediction model by means of these features, and RUL of the target bearing can be predicted via the same feature adaptation. A theoretical analysis is also provided to prove that the proposed approach has an upper bound of information loss in the transfer process. Experimental results on the IEEE PHM challenge 2012 dataset demonstrate the effectiveness of the proposed approach.