A zero-fault sample diagnosis method of rolling bearings via dynamic model and physics-guided neural network

Accurate fault diagnosis of rolling bearings is essential for the safe and stable operation of rotating machinery. However, in critical equipment such as aero-engines and high-speed rail systems, real fault samples are often unavailable, which poses a major challenge to conventional diagnosis methods that rely on labeled fault data. To address this issue, this paper proposes a rolling bearing fault diagnosis framework for scenarios without real fault samples, by integrating dynamic modeling, feature extraction, and physics-guided deep learning. First, a four-degree-of-freedom dynamic model of the target bearing is established according to its structural and operating parameters, and simulated vibration signals under different fault conditions are generated to provide prior fault information for model training. Second, vibration signals are processed using Hilbert envelope analysis and order-spectrum-based harmonic feature extraction, and an order-domain feature vector is constructed to characterize outer-race, inner-race, and rolling-element faults. Furthermore, a Physics-Guided Enhanced Convolutional Neural Network (PGE-CNN) is developed. By introducing a physics-informed feature generation branch and a physical-prior fusion strategy, the proposed network improves diagnostic robustness and interpretability under the condition that no real fault samples are available for training. Experimental results indicate that the proposed method can achieve effective fault identification without real fault samples and shows promising generalization ability and interpretability.