An Unsupervised Multi-grade Domain-Adversarial Regression Adaptation Network for Online Remaining Useful Life Prediction of Rolling Bearing Across Machines

Recently, deep transfer learning technology has been introduced to solve remaining useful life (RUL) of rolling bearing across working conditions. These methods enhance the RUL prediction performance by utilizing the historically-accumulated degradation data from offline working conditions. However, for the problem of online RUL prediction across different machines, due to the changeable operating machines, it is hard to effectively solve the data with a large distribution difference by adopting domain adaptation methods across working conditions. To solve this problem, this paper proposes an unsupervised multi-grade domain-adversarial regression adaptation network for online remaining useful life prediction of rolling bearing across machines. First, by employing offline whole-life degradation data of different machines, a deep domain-adversarial multi-regression network is constructed as advanced adversarial model. Second, by updating the advanced adversarial model from pseudo-target bearing selected for each machine, the supervised information of the pseudo-target bearing is dynamically generated, and then intermediate model is constructed. Finally, the RUL of online target bearing under an unknown machine can be evaluated through feeding into the intermediate model. Experiments results on the IEEE PHM Challenge 2012 and XJTU-SY bearing datasets demonstrate that the proposed method obtains accurate prediction performance across different machines and provides a practical solution for online RUL prediction under an unknown machine.