Common distribution discrepancy knowledge distilling: A new out-of-distribution generalization framework for machinery RUL prediction
Quan Qian, Xuezhong Chen, Jingdong Chen, Yi Qin
Abstract
Plenty of remaining useful life (RUL) transfer prediction methods have been developed to tackle with distribution shift and knowledge migration. However, most of them completely rely on the presupposition that the testing target-domain data sample can be accessible in advance during the training process, which cannot satisfy the real-time requirement in some engineering scenarios. To this end, a new out-of-distribution generalization framework named common distribution discrepancy knowledge distilling (CDDKD) is proposed, which can implicitly achieve the distribution alignment between source and unseen target domains via mining the common discrepancy knowledges embedded in the heterogenous mechanical equipment. In the CDDKD framework, an information decoupling network (IDA) model is constructed to eliminate the negative influence caused by the interfering noise, thereby extracting the inherent degradation trend features. Furthermore, a new vibration characteristic-based high-order moment distance (HMD) is proposed to boost the discrepancy representation ability in both original sample feature space and high-dimensional Hilbert space. An asynchronous hierarchical update strategy is designed to ensure the effectiveness of IDA model. Ultimately, the experimental results on gear and XJTU-bearing life cycle datasets validate the advantages of proposed CDDKD compared with other typical and state-of-the-art RUL transfer prediction methods.