Abstract: To improve the real-time prediction accuracy of contour errors in CNC machine tools,a mechanism-data hybrid-driven digital twin model was proposed.A mechanistic model integrating both the servo and mechanical systems was constructed to describe the dynamic response of the feed system under CNC interpolation commands.A time-series data-driven model integrated with multi-task learning was developed to compensate for the prediction residuals of the mechanistic model.In particular,to enhance the compensation capability of the data-driven model,a feature dataset integrating both simulation data and real-time data was constructed for model training.Finally,multiple sets of spatial motion trajectories were designed to conduct model validation experiments,during which the predictive performance of various models was systematically compared.The results indicated that the proposed hybrid twin model maintains high prediction accuracy across multiple time-step contour error prediction tasks.

Key words: contour error prediction, digital twin model, multi-task learning, multi-step prediction

摘要： 为了提升数控机床轮廓误差的实时预测精度,提出一种机理数据融合驱动的数字孪生模型。首先构建包含伺服系统和机械系统的机理模型,用于描述进给系统在数控系统插补指令下的动态响应;其次,建立结合多任务学习的时序数据驱动模型,构建融合机理模型孪生数据和实时采集数据的特征数据集进行模型训练,通过多时间步残差补偿修正机理模型的位置预测值,从而提高孪生模型的预测精度。最后,设计了多组空间运动轨迹进行了模型验证实验,并对不同模型的预测能力进行了对比。结果表明,所提融合孪生模型能够在轮廓误差多个时间步的预测任务中均保持较高预测精度。

关键词: 轮廓误差预测, 数字孪生模型, 多任务学习, 多时间步预测