How to Choose Proper Prognostic Knowledge and Transfer It ? A New Deep Sub-domain Adversarial Adaptation Method for Few-shot Remaining Useful Life Prediction Based on Wavelet Scattering Network

With huge success in remaining useful life (RUL) prediction of different rotatory machines, deep transfer learning techniques generally require sufficient source domain data to extract prognostic knowledge and to transfer them as well. In practical applications, however, it is not easy to accumulate enough run-to-failure degradation data from source domain for model training. To guarantee the transfer effect, two concerns intuitively rise up: what prognostic knowledge can be transferred, and, how to transfer it? To address these concerns, a deep sub-domain adversarial adaptation method is proposed for RUL prediction in few-shot environment. With the merit of wavelet scattering network (WSN) in expanding feature space, a complete feature set is first constructed for the few-shot whole-life degradation data by applying WSN with different frequency bands (decomposition scales) and rotation orientations. An orientation-first selection strategy is further built to determine the optimal orientation features in all scales, with sub-domains divided by different scales in source domain and target domain. Second, a new frequency contribution indicator (FCI) is designed in terms of geometric similarity to quantify the contribution of each frequency band to the transfer process. Finally, for the sub-source domain and sub-target domain at the same scale, a deep sub-domain adversarial network is built orderly with the domain discriminator weighted by FCI. The selective transfer of prognostic knowledge is then achieved. Experiments are conducted on a widely-used bearing dataset IEEE PHM Challenge 2012. The results demonstrate that the selective transfer from different frequency bands is beneficial to improve RUL prediction performance in few-shot environment. The results also prove that the orientation information of wavelet feature performs more critical than the scale information in the transfer process.