Transfer learning for remaining useful life prediction of multi-conditions bearings based on bidirectional-GRU network

Remaining useful life (RUL) prediction, has been a hotspot topic in the engineering field, which can ensure the security, availability, and continuous efficiency of the system. Different degradation trajectories of bearings under various working conditions may lead to the problem of inconsistent feature distribution and difficult acquisition of corresponding training labels, which affects the validity and accuracy of the prediction model. In this paper, a new transfer learning method based on bidirectional Gated Recurrent Unit (TBiGRU) is proposed to accurately predict the RUL of bearings under different working conditions. Firstly, based on dynamic time wraping (DTW) and Wasserstein distance to construct a comprehensive evaluation index of feature, the selection of transferable feature is carried out. Then a new index of energy entropy moving average cross-correlation based on maximal overlap discrete wavelet transform (MODWT) is proposed to realize adaptive recognition of bearings running states and the acquisition of corresponding training labels, which can also get rid of the constraint of setting threshold. Finally, transfer learning is carried out on the BiGRU model to solve the problem of distribution discrepancy, and timing information is also taken into account. The method is applied to the analysis of experimental data, and the results show that the framework can adaptively recognize different running states of bearings and obtain corresponding training labels, and at the same time realize better RUL prediction performance under different working conditions.

Automated summary

The research introduces a new transfer learning method to accurately predict the RUL of bearings under different working conditions, by constructing a comprehensive feature evaluation index and proposing a new energy entropy moving average cross-correlation index to achieve adaptive recognition of bearing running states and acquisition of corresponding training labels, and ultimately solving the problem of distribution discrepancy on the BiGRU model.