Neural ODE powered model for bearing remaining useful life predictions with intra- and inter-domain shifts

Abstract

In bearing remaining useful life (RUL) predictions, current domain adaptation (DA) and domain generalization (DG) methods are typically concerned with mitigating inter-domain shifts (DSs)—a type of DSs existing across the bearing degradation data sequences. Yet, intra-DSs along the bearing degradation data sequences, which are another type of DSs governing inter-DSs, have not attracted sufficient attention, thus hindering the applicability of existing methods. Moreover, many existing DG methods are developed based on multi-source domains, while bearing RUL predictions in reality often expect models of single-source DG capability. This study investigates the potential of the neural ordinary differential equation (ODE) for filling the aforementioned research gaps, leading to a novel neural ODE powered modeling (NOMI) scheme. First, the ODE characteristic of time invariance is utilized to address intra-DSs for learning time-invariant latent features from a single source bearing degradation data domain. Then, the gained time consistency could reduce heterogeneous intra-DS patterns, thereby decreasing inter-DSs and promoting model generalizability. The designed ODE module can be conveniently employed under DA and DG scenarios. Additionally, with a further gradient manipulation technique, the proposed model can be trained efficiently. Theoretical analyses demonstrate the benefits of intra-domain minimization for solving the data distribution problem. The experimental results based on multiple bearing datasets also verify the superiority of our proposed method compared with state-of-the-art approaches.