Abstract: By combining with the finite element simulation,digital modeling and experimental modal analysis,the frequency response function of the stator-end winding were predict to solve the end vibration of large turbogenerator and promote the development of intelligent turbine.The finite element model of the stator-end winding was established,and the modal parameters were obtained in good agreement with the measured results.The harmonic response analysis was further carried out to calculate the frequency response function of each point,which was used as the data source for subsequent research.Then,a single-step prediction model of frequency response function was constructed based on the Light Gradient Boosted Machine（Light GBM）algorithm,and the Light GBM model was compared with BP neural network,random forest and XGBoost models.On the basis of data-driven model,a mechanical-data hybrid model based on mode-ratio K was innovatively established according to the physical laws of the system.The high efficiency,low memory consumption,accuracy,tunability and parallelization support of Light GBM were more suitable for solving the prediction task on frequency response function in vibration problem of large turbine generator end.The frequency domain and space domain characteristics of displacement response under different modes were presented by the response surface.Finally,the hybrid data-driven model of frequency response function of integrated winding was established.By comparing with the frequency response function calculated by the finite element model and evaluating the prediction accuracy of non-sample spatial data,the reliability and generalization of the hybrid data-driven model were verified.The proposed model could reflect the vibration quickly and accurately,and help to take measures in time.

Key words: stator-end winding, finite element, frequency response, mechanical-data hybrid drive, light gradient boosted machine

摘要： 结合有限元仿真、数字化建模和试验模态分析等方法,预测定子端部绕组的频响函数,以解决大型汽轮发电机端部振动问题,从而推动智慧汽轮机的发展。首先,通过建立定子端部绕组三维精细有限元模型,获取与实测吻合较好的模态参数。进一步进行谐响应分析,计算得到各点频响函数,并作为后续研究的数据来源。其次,模拟试验模态锤击实验法,基于轻型梯度提升机（Light GBM）算法构建单阶频响函数预测模型,并将Light GBM模型与BP神经网络、随机森林和XGBoost模型进行比较分析;在数据驱动模型基础上,创新地基于系统物理规律建立了基于振型比例K的机理-数据混合驱动模型。Light GBM的高效性、低内存消耗、准确性、可调性和并行化支持,更适合用于解决大型汽轮发电机端部振动问题中的频响函数预测任务。通过响应曲面,展示了不同模态下位移响应的频域和空间域特性。最后,建立整体绕组频响函数混合数据驱动模型。通过与有限元模型计算的频响函数进行比较,并评估非样本空间数据的预测精度,验证了该混合数据驱动模型的可靠性和泛化能力。该模型能够快速、准确反映发电机端部振动情况,协助维护人员及时发现问题并采取必要措施。结合有限元仿真、数字化建模和试验模态分析等方法,预测定子端部绕组的频响函数,以解决大型汽轮发电机端部振动问题,从而推动智慧汽轮机的发展。首先,通过建立定子端部绕组三维精细有限元模型,获取与实测吻合较好的模态参数。进一步进行谐响应分析,计算得到各点频响函数,并作为后续研究的数据来源。其次,模拟试验模态锤击实验法,基于轻型梯度提升机（Light GBM）算法构建单阶频响函数预测模型,并将Light GBM模型与BP神经网络、随机森林和XGBoost模型进行比较分析;在数据驱动模型基础上,创新地基于系统物理规律建立了基于振型比例K的机理-数据混合驱动模型。Light GBM的高效性、低内存消耗、准确性、可调性和并行化支持,更适合用于解决大型汽轮发电机端部振动问题中的频响函数预测任务。通过响应曲面,展示了不同模态下位移响应的频域和空间域特性。最后,建立整体绕组频响函数混合数据驱动模型。通过与有限元模型计算的频响函数进行比较,并评估非样本空间数据的预测精度,验证了该混合数据驱动模型的可靠性和泛化能力。该模型能够快速、准确反映发电机端部振动情况,协助维护人员及时发现问题并采取必要措施。

关键词: 定子端部绕组, 有限元, 频响函数, 机理-数据混合驱动, 轻型梯度提升机