Abstract: The condition of the tool directly affects the quality of the workpiece,but the tool in the machine monitoring distance is too far to detect hardly with the small image of the cutter tip.To address this issue and enhance the precision of tool tip wear detection,a methodology for milling tool tip wear monitoring using deep learning was proposed,which involved the combination of YOLOv5 with the improved U-Net.The acquired milling cutter images were initially processed using YOLOv5 to identify and extract the tip portion of the cutter,and the tip wear region from the images was separated using the improved U-Net network.In the enhanced U-Net,the residual blocks was incorporated into the Encoder structure of the original network and an ECA attention mechanism was incorporated into the Decoder structure.The experimental results demonstrated that the enhanced U-Net yielded an mIOU improvement of 8.39%,7.1%,and 17.66% relative to the original U-Net,Deeplabv3+,and PSPNet respectively.Additionally,the detection time was reduced by 85.52% and 86.73% compared to Deeplabv3+ and PSPNet respectively.In conclusion,the proposed method addressed the challenge of detecting milling cutter tip wear due to the necessity of long-distance shooting for on-machine tool monitoring,while also markedly enhancing detection efficiency.

Key words: YOLOv5, improved U-Net, deep learning, tip wear, attention mechanism

摘要： 刀具的状态直接影响着工件质量,但刀具在机监测距离较远导致图像中铣刀刀尖过小难以检测。为解决该问题,进一步提高刀尖磨损的检测精度,提出利用深度学习进行铣刀刀尖磨损监测的方法。将YOLOv5与改进后的U-Net相结合,首先使用YOLOv5对采集到的铣刀图像进行处理,识别并提取刀尖部分,然后利用改进的U-Net网络对提取的图像进行刀尖磨损区域的分割。改进后的U-Net在原网络的Encoder结构中加入Residual Blocks残差块,在Decoder结构中加入ECA注意力机制。实验结果显示,改进的U-Net相比于原U-Net、Deeplabv3+和PSPNet,mIOU分别提高了8.39%,7.1%和17.66%。同时,其检测时间相比Deeplabv3+和PSPNet分别缩短了85.52%和86.73%。综上所述,该方法解决了由于刀具在机监测需远距离拍摄而导致铣刀刀尖磨损难以检测的问题,显著提高了检测效率。

关键词: YOLOv5, 改进的U-Net, 深度学习, 刀尖磨损, 注意力机制