旋转刀尖点频响函数的迁移学习预测技术

Transfer learning prediction technology for frequency response function prediction of rotary tool tip

摘要:

针对刀尖点频响函数受机床主轴位置、主轴转速和刀具参数的影响较大的难点,为快速准确地获取机床刀尖点频响函数,文中引入迁移学习提出了一种基于少量试验样本来获取不同刀具参数的旋转刀尖频响函数预测模型的方法。首先,生成机床主轴位置和转速的正交规划表,基于空运行自激励法和卷积神经网络（CNN）算法,建立与机床加工位置和主轴转速相关的刀尖频响函数预测模型。其次,考虑刀具伸长量、直径和种类等参数的影响,利用少量的相关数据样本,基于迁移学习训练出不同刀具工况的刀尖频响函数预测模型。最后,基于加工中心VMC80IV开展了锤击实验和空运行自激励实验,采用实验数据对预测模型进行训练,以各阶次模态参数为模型输出值,通过模态叠加法重构出刀尖点频响函数,并对比模型预测值和实际测量值。结果表明,对于不同刀具工况下的旋转刀尖频响函数预测模型,各阶次固有频率的预测误差不超过2%,阻尼比的预测误差不超过5%,验证了该预测模型的有效性和准确性。

Abstract:

Tool tip frequency response function (FRF) changes with the position of the machine tool spindle, spindle speed, and tool parameters. In order to quickly and accurately obtain the frequency response function of the machine tool tip, a method based on a small number of experimental samples to predict the rotating tool tip FRFs under different tool parameters is proposed by introducing transfer learning. Firstly, an orthogonal programming table for the position and speed of the machine tool spindle is generated, and the prediction model of the rotating tool tip FRFs related to the machining space and the spindle speed is established based on the combination of the self-excitation method and the convolutional neural network （CNN） algorithm; Then, considering the influence of parameters such as tool elongation, diameter and type, transfer learning is adopted to train the prediction model of tool tip FRFs under different tool conditions by a small amount of data. Finally, the prediction model is trained using experimental data from hammer impact experiments and self-excitation experiments at the machining center VMC80IV. The modal parameters of each order output by the model are reconstructed to obtain the tool tip FRFs using the modal superposition method, and the predicted values of the model are compared with the actual measured values. The experimental results show that for the prediction model of the FRFs of the rotating tool tip under different tool working conditions, the prediction error of each order of natural frequency did not exceed 2%, and the prediction error of damping ratio did not exceed 5%, which verifies the effectiveness and accuracy of the prediction model.