基于循环神经网络的切削力无传感器监测  

Sensorless Cutting Force Monitoring Based on Recurrent Neural Network

作　　者：程英豪 刘长青[1] 庄其扬 李光旭 郝小忠[1] CHENG Yinghao;LIU Changqing;ZHUANG Qiyang;LI Guangxu;HAO Xiaozhong(College of Mechanical and Electrical Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 210016)

机构地区：[1]南京航空航天大学机电学院,南京210016

出　　处：《机械工程学报》2025年第6期14-23,共10页Journal of Mechanical Engineering

摘　　要：切削力对切削状态的变化具有高度敏感性和快速响应能力,是用于加工状态监测和自适应加工最有价值的物理量。由于无须引入额外的传感元件,基于数控系统内部伺服监测信号的切削力在线预测方案,有潜力实现切削力的长期低成本准确监测。然而,伺服监测信号和切削力之间的关系十分复杂,为此提出基于循环神经网络的切削力在线预测方法。首先将基于机床伺服监测信号的进给轴切削力预测问题定义为自适应时滞的非线性动态系统建模问题,进而引入长短期记忆(Long short-term memory,LSTM)神经网络和门控循环单元(Gated recurrent unit,GRU)神经网络两类循环神经网络直接从端到端观测数据中学习动态预测模型。通过一组变转速变进给孔铣削试验构造时变工况下的切削激励数据集进行对比验证,结果表明,对于动态特性更为复杂的X轴,LSTM神经网络的预测性能更好,相对方均根误差为17.62%;对于动态特性相对简单的Y轴,GRU神经网络的预测性能更好,相对方均根误差为11.74%。Cutting force is highly sensitive and capable of rapid response to changes in cutting state,which is considered as the most valuable physical quantity for machining state monitoring and adaptive machining.Since there is no need to introduce additional sensing components,an online prediction cutting force solution based on inherent servo monitoring signals in CNC systems has the potential to achieve long-term,low-cost,and accurate monitoring of cutting force.However,the relationship between servo monitoring signals and cutting force is very complex.Therefore,a cutting force online prediction method based on recurrent neural networks is proposed.Firstly,the problem of feed-axis cutting force prediction based on machine tool servo signals is defined as a nonlinear dynamic system modeling problem with adaptive time delay.Then,two types of recurrent neural networks,long short-term memory neural network(LSTM NN)and gated recurrent unit neural network(GRU NN),are introduced to directly learn the dynamic prediction model from end-to-end observation data.A set of variable speed hole milling experiments are carried out to construct a cutting excitation dataset under time-varying working conditions for comparative verifications.For the X-axis with more complex dynamic characteristics,LSTM NN has better prediction performance,with a relative root mean square error of 17.62%.For the Y-axis with relatively more simple dynamic characteristics,GRU NN has better prediction performance,with a relative root mean square error of 11.74%.

关 键 词：切削力 加工状态监测 自适应加工 循环神经网络 伺服监测信号 

分 类 号：TG54[金属学及工艺—金属切削加工及机床]