基于注意力机制的时空卷积数控机床热误差模型研究

Spatiotemporal Convolution Thermal Error Model of CNC Machine Tools Based on Attention Mechanism

为了提高数控机床热误差模型的精度与泛化性,提出了基于注意力机制的长短时记忆卷积神经网络(Long short term memory convolutional neural network based on attention mechanism,AM-CNN-LSTM)热误差模型。利用卷积神经网络提取高维数据空间状态特征的能力和长短时记忆网络提取长时间序列状态特征的能力,构建具有2个支路的热误差模型,分别提取特征后输入到注意力机制中进行特征重要性重构,建立原始数据与热误差的特征映射,最后通过全连接层进行热误差预测。采用G460L型数控机床进行实验数据采集,将不同季节采集到的温度数据和热误差作为模型输入,采用循环学习率与正则化优化方法对模型进行训练。与LSTM、ConvLSTM和CNN-LSTM热误差模型对比,结果表明,AM-CNN-LSTM模型对特征还原能力最强,残差波动范围最小,其残差范围较最大值下降62.09%,模型预测精度在2.4μm以内。

In order to improve the accuracy and generalization of the thermal error model of CNC machine tools,a thermal error model of and long short term memory convolutional neural network based on attention mechanism(AM CNN LSTM)was proposed.A thermal error model with two branches was established by using the ability of convolutional neural networks to extract the space state features of high-dimensional data and the ability of long short term memory networks to extract long-term sequence state features,and the extracted features were input into the attention mechanism to reconstruct according to the importance,and then a feature map of original data and thermal error value was established.Finally,the thermal error prediction value was performed through the full connect layer.The G460L CNC lathe was used to collect experimental data,the temperature and thermal error collected in different seasons were used as the model input,and the model was trained using the cyclic learning rate and regularization optimization method.Compared with the thermal error model of LSTM,ConvLSTM and CNN LSTM,the results showed that AM CNN LSTM model had the strongest ability to restore features and the smallest residual error range.It was decreased by 62.09%,and the prediction accuracy of the model was within 2.4μm.