Abstract: Aiming at the insufficient accuracy problem of thermal error prediction model of direct drive feed shaft,a thermal error prediction method of direct drive feed shaft based on Chaos Improved Multi-Strategy Gorilla Troops Optimization algorithm（CIMGTO）optimized Deep Neural Network（DNN）was proposed.To reduce the multi-collinearity between temperature data,the Variance Inflation Factor（VIF）stepwise method was used to screen the temperature sensitive points.To improve the accuracy of thermal error prediction model,CIMGTO was used to optimize DNN and establish thermal error prediction model.CIMGTO introduced chaotic mutation factor,Cauchy Inverse Cumulative Distribution function（CICD）and Tangent Flight Operator（TFO）on the basis of GTO algorithm,which improved the initial population diversity,local search ability and balanced search ability of the algorithm.Through the thermal error experimental data of the linear motor gantry test bench,the VIF stepwise method was verified effective in screening temperature sensitive points,and the CIMGTO-DNN thermal error prediction model had high fitting accuracy.Its  value was 0.9961,which was 20.96 % higher than the BP model,10.65 % higher than the SVR model,2.15 % higher than the DNN model,1.26 % higher than the GTO-DNN model,and 0.82 % higher than the PSO-DNN model.The CIMGTO-DNN model value was 0.9905,which was verified by the thermal error experimental data of the linear motor gantry machine tool.The high-precision prediction of the thermal error of the direct drive feed shaft was realized,which provided data support for error compensation.

Key words: thermal error, direct drive feed shaft, deep neural network, artificial gorilla troops optimizer

摘要： 针对直驱进给轴热误差预测模型精度不足问题,提出一种基于混沌改进多策略大猩猩算法（CIMGTO）优化深度神经网络（DNN）的直驱进给轴热误差预测方法。为降低温度数据间的多重共线性,采用方差膨胀因子（VIF）逐步法筛选温度敏感点;为提高热误差预测模型精度,使用CIMGTO优化DNN,并建立热误差预测模型。CIMGTO在人工大猩猩部队优化算法（GTO）的基础上引入了混沌变异因子、柯西分布逆累积函数（CICD）和正切飞行算子（TFO）,提升了初始种群多样性、算法局部搜索能力和平衡搜索能力。经直线电机龙门测试台架热误差实验数据验证,VIF逐步法筛选温度敏感点有效,且CIMGTO-DNN热误差预测模型拟合精度高,其R2值为0.9961,较BP模型提升了20.96%,较SVR模型提升了10.65%,较DNN模型提升了2.15%,较GTO-DNN模型提升了1.26%,较PSO-DNN模型提升了0.82%。再经直线电机龙门机床热误差实验数据验证,CIMGTO-DNN模型R2值为0.9905,实现了直驱进给轴热误差的高精度预测,为误差补偿提供了数据支撑。

关键词: 热误差, 直驱进给轴, 深度神经网络, 大猩猩算法