基于CEEMDAN和改进轻量化时空网络的刀具状态监测  

Tool Condition Monitoring Based on CEEMDAN and BiLSTM-SN-ECA

作　　者：周鹏博 刘德平[1] ZHOU Pengbo;LIU Deping(School of Mechanical and Power Engineering,Zhengzhou University,Zhengzhou 450001,China)

机构地区：[1]郑州大学机械与动力工程学院,郑州450001

出　　处：《组合机床与自动化加工技术》2024年第3期177-181,186,共6页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：河南省重大科技专项项目(171100210300-01)。

摘　　要：针对刀具退化特征提取困难和传统时空网络模型参数多等问题,提出了基于自适应噪声完备经验模态分解(CEEMDAN)和改进轻量化时空网络(BiLSTM-SN-ECA)的刀具磨损监测模型。首先,将刀具振动信号经CEEMADAN分解得到若干模态分量,将模态分量与振动信号结合,构造特征矩阵;其次,利用ECA改进ShuffleNetv2基本单元,并优化ShuffleNetv2整体结构,构造BiLSTM-SN-ECA网络模型;最后,将特征矩阵输入模型进行特征学习与磨损预测。所提方法预测值的平均绝对误差和均方根误差分别为1.246μm和2.065μm,结果表明该方法在减少传统时空网络模型参数量与训练时间的同时,提高了预测准确度。Aiming at the difficulties in extracting tool degradation features and the many parameters of traditional spatiotemporal network models,a tool wear monitoring model based on adaptive noise-complete empirical mode decomposition(CEEMDAN)and improved lightweight spatiotemporal network(BiLSTM-SN-ECA)is proposed.Firstly,the tool vibration signal is decomposed by CEEMADAN,and the modal component is combined with the vibration signal to construct a feature matrix.Secondly,ECA is used to improve the basic unit of ShuffleNetv2,and the overall structure of ShuffleNetv2 is optimized to construct the BiLSTM-SN-ECA network model.Finally,the feature matrix is input into the model for feature learning and wear prediction.The MAE and RMSE of the predicted values of the proposed method are 1.246μm and 2.065μm,and the results show that the proposed method can reduce the number of parameters and training time of the traditional spatiotemporal network model,and improve the prediction accuracy.

关 键 词：刀具磨损监测 自适应噪声完备经验模态分解 轻量化时空网络 注意力机制 

分 类 号：TH16[机械工程—机械制造及自动化] TG71[金属学及工艺—刀具与模具]