基于注意力机制的数控机床热误差深度学习预测方法

为提高热误差预测精度和鲁棒性,提出一种基于注意力机制和深度学习网络的数控机床热误差预测模型。采用数据转化策略,将数控机床原始温度数据转化为温度图像,直接作为深度学习网络的输入；提出一种基于注意力机制的温度敏感点识别网络,根据温度测点与热误差关联程度,赋予各温度测点不同的权值,避免了温度测点的人为选取弊端；建立12层深度CNN学习预测网络,利用其强大的图像特征学习能力,挖掘温度图像与热误差的非线性映射关系,无需对温度关键点进行预选择,保留了更多的热误差与机床温度特征关系,显著提高了模型预测精度。为了提高热误差模型的精度与泛化能力,引入Dropout正则化方法和Adam优化算法,对深度卷积神经网络的结构与参数进行了优化。该方法在针对G460L型数控车床的热误差验证中表现出较高的预测精度。通过与BP神经网络和多元回归等传统热误差模型进行对比,深度卷积神经网络框架下的热误差模型在泛化性指标上表现更优。

Deep Learning Prediction for Thermal Error of CNC Machine Tools Based on Attention Mechanism

Thermal error prediction and compensation of CNC machine tools is an important technology to improve the machining accuracy and reliability of CNC machine tools. The thermal error of machine tool is time-varying and nonlinear. To improve the accuracy and robustness of thermal error prediction, a numerical control machine tool thermal error prediction model based on attention mechanism and deep learning network was proposed. Using the data conversion strategy, the original temperature data of CNC machine tool was transformed into temperature image, which can be directly used as the input of deep learning network. The complete information of the temperature field of the machine tool was retained by converting the temperature field data into the temperature image points. At the same time, the nonlinear and coupling problems between the temperature measuring points were avoided by using the deep learning modeling method. A recognition network of temperature sensitive points based on attention mechanism was proposed. According to the correlation degree between temperature measuring points and thermal error, different weights were given to each temperature measuring point to avoid the disadvantages of artificial selection of temperature measuring points. A 12-layer deep CNN learning prediction network was established to mine the nonlinear mapping relationship between temperature image and thermal error by using its powerful image feature learning ability. This method does not need to pre select the key temperature points, retained more relationship between thermal error and machine temperature characteristics, and can significantly improve the prediction accuracy of the model. In order to improve the accuracy and generalization ability of thermal error model, dropout regularization method and Adam optimization algorithm were introduced to optimize the structure and parameters of deep convolution neural network. The method has showed high prediction accuracy in the thermal error verification of G460L CNC lathe. Compared with traditional thermal error models such as BP neural network and multiple regression, the thermal error model based on deep convolution neural network performs better in generalization index.