A mechanism-based sample generation method utilizing morphological analysis and dynamic modeling for online monitoring of rotor systems

Abstract

Real-time monitoring is essential for mechanical rotor systems, but obtaining sufficient labeled samples to train intelligent models poses significant challenges in real-world scenarios. Moreover, existing data-driven sample generation techniques often fail to capture the physics of fault development, resulting in biased or inaccurate samples. To address this, a sample generation method based on fault mechanisms is proposed, leveraging morphological analysis and dynamic modeling to tackle the few-shot issue in condition monitoring. Specifically, a spatial lumped parameter model is constructed to generate fault mechanism samples. This model accounts for time-varying displacements and transient excitation force coupling, accurately reflecting defective bearing vibration responses. Subsequently, an adaptive optimization approach based on the Pearson correlation coefficient is employed to minimize the gap between real and generated data, ensuring the generated signals are more realistic. Finally, parameter transfer techniques are employed to train models with limited samples on both shallow and deep networks. The diagnostic results confirm that the proposed method effectively overcomes the few-shot challenge.