基于迁移学习与残差网络的刀具磨损状态监测

Tool Wear Monitoring Based on Transfer Learning and Deep Residual Network

针对现有基于深度学习的刀具磨损状态监测方法训练样本少、识别精度低的问题,建立基于迁移学习(TL)与深度残差网络(ResNet)的铣刀磨损状态监测模型。将刀具加工过程中的振动监测信号通过连续小波变换转换成能量时频图,作为网络模型的输入;将在ImageNet数据集上训练的ResNet50模型作为预训练模型,通过迁移学习的方法,应用到刀具磨损状态监测领域当中。实例验证表明:TL-ResNet模型的平均识别准确率达到98.52%,实现了刀具不同磨损状态下的智能识别,有效提高了刀具磨损状态监测的准确性和稳定性。

To solve the problem that the existing tool wear monitoring methods based on deep learning require too many samples and have low identification accuracy,a tool wear monitoring model based on transfer learning(TL)and deep residual network(ResNet)was established.The vibration monitoring signals during tool machining were converted into energy time-frequency diagram by continuous wavelet transform,which were used as the input of network model.The ResNet50 model trained on ImageNet dataset was used as the pre-training model,and was applied to the field of tool wear state monitoring by transfer learning method.The example verification shows that the recognition accuracy of TL-ResNet reaches 98.52%,realizing intelligent recognition under different tool wear states.So the accuracy and stability of tool wear state monitoring are improved effectively.