To address the challenges of multi-channel feature fusion and the lack of physical constraints in deep feature extraction under data-scarce conditions, this paper proposes a Physics-Informed Meta-Learning Network (PIFMN) for rolling bearing fault diagnosis. Traditional few-shot cross-domain diagnosis methods mostly rely on pure data-driven paradigms, which lack macro-physical boundary constraints during feature aggregation and suffer from channel noise induced by transmission path attenuation during hard-splicing fusion. To solve these issues, Bayesian Model Averaging (BMA) is first introduced at the data frontend to achieve high-fidelity adaptive reconstruction of multi-source signals and suppress the adverse influence of noisy channels through confidence-based weighting. Subsequently, an operating boundary gated mechanism is designed to project macro-operating parameters into the latent feature space as explicit physical constraints, which adaptively modulate the spectral features via residual learning. Experimental results on the Case Western Reserve University (CWRU) dataset and a laboratory self-built dataset demonstrate that the proposed PIFMN achieves high classification accuracy under few-shot and cross-domain conditions, while exhibiting superior noise robustness and generalization capability

Rolling bearing; Fault diagnosis; Meta-learning; Physics-informed; Multi-source information fusion