A novel zero-shot bearing fault diagnosis method based on physics-informed digital twin model and domain-alignment strategy

Abstract This paper proposes a novel zero-shot bearing fault diagnosis method that tackles the critical challenge of data scarcity, a long-standing bottleneck in fault diagnosis, particularly for industrial applications.By integrating a physics-informed digital twin (PIDT) model with a domain-alignment strategy (DAS), the proposed method generates high-fidelity virtual vibration signals that comply with the inherent physical dynamics of bearing systems, and effectively aligns the noisy real-world signals with the virtual signals. The PIDT model fuses a 4-DOF bearing dynamic model with a neural network, incorporating two complementary loss functions: a physical residual loss, which enforces consistency with the governing physical equations, and a frequency prior loss that guarantees spectral fidelity. The DAS bridges the domain discrepancy between noise-free virtual data and real-world measured signals by aligning the real-world signals with the virtual domain. This integrated framework thereby enables reliable fault classification even without labeled fault data during model training. Experimental results validate the method's robustness and generalizability, achieving average zero-shot diagnostic accuracies of 95.53±1.46% on a self-built test rig and 95.15±1.89% on the CWRU dataset, respectively, significantly outperforming existing state-of-the-art methods in both accuracy and stability across various operational conditions. Furthermore, ablation studies rigorously validate the critical contributions of both the PIDT's loss components and the DAS in to the zero-shot performance. The proposed method provides an effective solution to the data scarcity dilemma in bearing fault diagnosis and offers significant potential for real-world industrial applications.