A rotor dynamic balancing method without trial weights based on multi-sensor feature fusion with an attention mechanism

Accurate identification of unbalance parameters lies at the core of rotor dynamic balancing technology. In this paper, a deep learning approach is integrated with rotor dynamic balancing theory, and a multi-sensor feature fusion dynamic balancing method without trial weights based on the attention mechanism is proposed. First, considering the unique nature of the dynamic balancing process, a rotor model with varying unbalance parameters was established, and the dataset for model training was augmented by adding bandpass noise. Subsequently, a neural network model for unbalance identification based on the multi-head attention mechanism was designed, and ablation experiments were conducted to verify its superiority over single-sensor models and the feature fusion model without attention. Finally, both a finite element simulation model and an experimental model of a triple-disk rotor system were established for dynamic balancing validation. The simulation and experimental results demonstrate that the proposed dynamic balancing method can substantially reduce the vibration amplitudes of the rotor across all critical speed regions. Moreover, compared with the influence coefficient method (ICM), the proposed method requires only a single run-up process to complete the dynamic balancing procedure, significantly improving the balancing efficiency and achieving superior balancing performance across all critical speed regions.