A small-sample RUL prediction method based on physics-informed neural network with integrated domain-generalized features

Abstract Bearings are critical components in rotating machinery, and accurate prediction of their remaining useful life (RUL) is essential for prognostics and health management. However, complex and varying operating conditions in real industrial environments, together with the scarcity of full life-cycle failure data under target conditions, severely limit the small-sample prediction accuracy, cross-condition generalization capability, and practical applicability of existing RUL prediction models. To address these challenges, this paper proposes a hybrid RUL prediction framework that integrates unsupervised deep feature extraction with physics-driven modeling: the deep Boltzmann machine-physics-informed neural network (DBM-PINN). First, an unsupervised DBM is employed to automatically extract key degradation-related latent features from healthy data, effectively tracking the degradation trend while suppressing interference caused by varying operational conditions. Subsequently, an implicit PINN is constructed to embed degradation knowledge by modeling the intrinsic evolution of the degradation trajectory through nonlinear partial differential equations. Using DBM-extracted features together with time as inputs, the proposed framework overcomes the limitation of insufficient failure data and achieves high-precision, cross-condition RUL prediction from a single dataset. The effectiveness of the proposed DBM-PINN in terms of small-sample prediction accuracy and cross-condition generalization is validated using the XJTU-SY bearing dataset. Experimental results demonstrate that the proposed method enables the model to focus on the intrinsic degradation evolution, effectively capture essential degradation features, and reduce the influence of specific operating conditions, thereby achieving robust RUL prediction under complex and varying environments.