A Mechanism-Data Hybrid-Driven Modeling Method for Vibration Response Prediction of the Magnetically Suspended Rotor System

In recent years, modern industry has seen growing demands for high-performance rotating machinery, imposing higher requirements on the dynamic performance and reliability of rotor systems. Developing an efficient and accurate rotor dynamic model can not only provide outputs that closely match real rotor dynamic characteristics and vibration responses, but also simulate potential vibration sources and corresponding control strategies. However, when vibration sources involve the interaction of multiple complex factors or arise from processes not yet fully understood, it becomes challenging to accurately describe and quantify them through a single mechanism-based model. To address current modeling requirements for high-precision control and long-term autonomous health management, this paper proposes a mechanism-data hybrid-driven modeling method for vibration response prediction of the magnetically suspended rotor system. It fully utilizes the mechanism model, measured data, and artificial intelligence algorithm to achieve high-precision simulation of the modeling object. First, a mechanism model incorporating mechanical rotor system, magnetic bearing control system, and internal vibration sources is established. The measured vibration data are used to analyze the unbalanced parameters to be identified in the mechanism model, achieving the first stage of mechanism model calibration. Second, an attention-based data mapping network is designed to extract harmonic disturbance information caused by external vibration sources from measured vibration data, which is then supplemented to the calibrated mechanism model, achieving the second stage of mechanism model correction. Finally, the vibration response prediction performance of the constructed hybrid mechanism-data model is compared with that of the physical prototype platform, verifying the effectiveness and superiority of the proposed mechanism-data hybrid-driven modeling method.