融合残差块与Swin-Transformer机制的刀具磨损监测方法  

Tool wear monitoring methods incorporating residual block and Swin-Transformer mechanisms

作　　者：李泽稷 周学良 孙培禄[2] LI Zeji;ZHOU Xueliang;SUN Peilu(School of Mechanical Engineering,Hubei Institute of Automotive Technology,Shiyan 442002,China;School of Mechanical Engineering,Yuncheng University,Yuncheng 044000,China)

机构地区：[1]湖北汽车工业学院机械工程学院,十堰442002 [2]运城学院机械工程学院,运城044000

出　　处：《现代制造工程》2024年第8期126-135,共10页Modern Manufacturing Engineering

基　　金：国家自然科学基金资助项目(52075107);湖北省高等学校优秀中青年科技创新团队计划项目(T2020018)。

摘　　要：为进一步提高切削加工过程刀具磨损值监测的精度,提出一种融合残差块与Swin-Transformer模型的刀具磨损监测模型。首先,采用分组卷积残差块提取信号的特征;然后,利用Swin-Transformer模型中的分块滑动窗口自注意力机制对提取的特征进行平移融合;最后,通过回归层实现刀具磨损值监测。试验结果表明,融合一层残差块与一层stage机制的Swin-Transformer模型可以有效融合刀具磨损状态监测信号的全局信息,相比其他Swin-Transformer模型,不仅模型结构简单,而且具有更高的监测精度,利用PHM2010数据集验证的MSE、MAE和R2分别达到4.471 9、1.467 5和0.995 8。To further improve the accuracy of tool wear value monitoring in the cutting machining process,a tool wear monitoring model that integrated the residual block and Swin-Transformer model was proposed.Firstly,the grouped convolutional residual block was used to extract the features of the signal.Then,the chunked sliding window self-attention mechanism in the Swin-Transformer model was used to translate the extracted features.Finally,the tool wear value prediction was realized through the regression layer.The experimental results show that the Swin-Transformer model fusing a layer of residual blocks with a layer of stage mechanism can effectively fuse the global information of tool wear state monitoring signals,which not only has a simple model structure but also has a higher monitoring accuracy compared with other Swin-Transformer models,and the MSE,MAE,and R 2 verified by utilizing the PHM2010 dataset reached 4.4719,1.4675,and 0.9958,respectively.

关 键 词：刀具 磨损监测 残差卷积神经网络 Swin-Transformer模型 

分 类 号：TG713[金属学及工艺—刀具与模具] TH162[机械工程—机械制造及自动化]