基于改进的EfficientNetV2和UNetTSF的刀具磨损状态识别及预测方法

doi:10.3969/j.issn.1004-132X.2026.03.016

摘要/Abstract

摘要：

针对刀具在机磨损状态监测问题，为提高刀具磨损预测的准确性，提出了一种融合格拉姆角场（GAF）、改进的EfficientNetV2轻量级网络和UNetTSF时序预测模型的新型监测模型GAF-iEfficientNetV2-UNetTSF。该模型采用先分类后预测的策略，采集刀具加工过程中的力信号，并使用分段聚合技术实现特征降维，利用格拉姆角场分别对三向力信号进行编码，得到三组单通道图像，并将同一时序下的三组单通道图像堆叠成三通道图像。构建改进的EfficientNetV2训练网络进行特征的自动提取和分类以实现刀具磨损状态识别。针对最关键的刀具磨损状态，利用UNetTSF模型进行磨损值预测，以实现精确预判。通过对比实验，验证了该模型在刀具磨损状态识别任务中的高准确率以及磨损值预测方面的高精度，为刀具磨损状态监测领域提供了一种更精准的监测方法，对提高工业生产效率和降低维护成本具有重要意义。

关键词: 格拉姆角场, EfficientNetV2, UNetTSF, 刀具磨损状态, 刀具磨损监测

Abstract:

In order to improve the accuracy of tool wear prediction for the problems of tool in-machine wear condition monitoring， a new monitoring model named GAF-iEfficientNetV2-UNetTSF was proposed integrating GAF， the improved EfficientNetV2 lightweight network， and the UNetTSF time-series prediction model. The model adopted the strategy of first classification and then prediction. Firstly， the force signals were acquired during machining processes by tool， and the feature dimensionality reduction was realized by segmented aggregation technique. Then GAF was used to encode the three-directional force signals respectively， and three groups of single-channel images were obtained. The three groups of single-channel images under the same time sequence were stacked into three-channel images. Subsequently， an improved EfficientNetV2 training network was constructed to automatically extract and classify features to recognize the tool wear states. Finally， for the most critical tool wear states， the UNetTSF model was utilized for wear value prediction in order to achieve accurate prediction. Through comparative experiments， the high accuracy of the model in the task of tool wear state recognition and the high precision in wear value prediction were verified. The results provide an efficient and accurate monitoring method in the field of tool wear state monitoring， and is of great significance for improving industrial production efficiency and reducing maintenance costs.

Key words: Gramian angular field （GAF）, EfficientNetV2, UNetTSF, tool wear state, tool wear monitoring

中图分类号:

TH165

张曦, 朱红, 张龙佳, ABDELTAWAB Ahmed. 基于改进的EfficientNetV2和UNetTSF的刀具磨损状态识别及预测方法[J]. 中国机械工程, 2026, 37(3): 668-678.

ZHANG Xi, ZHU Hong, ZHANG Longjia, ABDELTAWAB Ahmed. Tool Wear State Identification and Prediction Method Based on Improved EfficientNetV2 and UNetTSF[J]. China Mechanical Engineering, 2026, 37(3): 668-678.

导出引用管理器 EndNote|Ris|BibTeX

链接本文: https://www.cmemo.org.cn/CN/10.3969/j.issn.1004-132X.2026.03.016

https://www.cmemo.org.cn/CN/Y2026/V37/I3/668

图/表 25

图1 某时序下x轴方向受力信号的GAF图

Fig.1 GAF plot of the force signal in the x-axis direction for a certain time series

图2 MBConv模块结构

Fig.2 MBConv module structure

图3 Fused MBConv模块结构

Fig.3 Fused MBConv module structure

图4 FPN方法

Fig.4 FPN model

图5 UNetTSF模型结构

Fig.5 UNetTSF model structure

图6 GAF-iEfficientNetV2结构

Fig.6 GAF-iEfficientNetV2 structure

图7 SCConv模块结构

Fig.7 SCConv module structure

图8 GAM模块结构

Fig.8 GAM module structure

图9 基于GAF-iEfficientNetV2-UNetTSF的刀具磨损状态监测模型结构

Fig.9 Structure of tool wear state monitoring model based on GAF-iEfficientNetV2-UNetTSF

表1 切削参数

Tab.1 Cutting parameters

参数	数值
主轴转速/（r·min^-1）	10 400
进给速率/（mm·min^-1）	1555
径向切深/mm	0.125
轴向切深/mm	0.2
采样频率/Hz	50 000

图10 Vb值与输入信号的斯皮尔曼相关性分析结果

Fig.10 Results of Spearman's correlation analysis between Vb value and input signal

图11 C1铣刀后刀面磨损量变化曲线

Fig.11 C1 milling cutter back face wear variation curve

图12 模型训练效果

Fig.12 Model training effects

图13 改进模型准确率验证曲线

Fig.13 Improved model accuracy validation curves

表2 改进前后模型性能对比 (%)

Tab.2 Comparison of model performance before and after improvement

模型	准确率	精确率	召回率	F₁
EfficientNetV2	95.30	95.31	95.30	95.31
SC-EfficientNetV2	96.21	96.24	96.21	96.23
G-EfficientNetV2	96.50	96.51	96.50	96.51
iEfficientNetV2	97.95	97.96	97.95	97.96

图14 EfficientNetV2在测试集上的混淆矩阵

Fig.14 Confusion matrix of EfficientNetV2 on the test set

图15 iEfficientNetV2在测试集上的混淆矩阵

Fig.15 Confusion matrix of iEfficientNetV2 on the test set

图16 稳定磨损阶段模型预测效果

Fig.16 Stable wear phase model prediction effects

图17 加速磨损阶段模型预测效果

Fig.17 Accelerated wear phase model prediction effects

图18 改进模型准确率验证曲线

Fig.18 Improved model accuracy validation curves

表3 各模型性能对比 (%)

Tab.3 Comparison of the performance of each model

模型	准确率	精确率	召回率	F₁
GAF-ResNet18	95.09	95.08	95.09	95.09
GAF-VGG16	95.77	95.76	95.77	95.77
GAF-ShuffleNetV2	94.41	94.46	94.41	94.44
GAF-MobileNetV2	95.60	95.59	95.60	95.60
GAF-iEfficientNetV2	97.95	97.96	97.95	97.96

图19 对比模型在稳定磨损阶段的预测结果

Fig.19 Predictive results of contrast models during stable wear stage

图20 对比模型在加速磨损阶段的预测结果

Fig.20 Predictive results of contrast models during accelerated wear stage

表4 稳定磨损阶段各模型性能对比

Tab.4 Comparison of the performance of each model in stable wear stage

模型	RMSE	MAE	MAPE/%	R²
BiLSTM	1.50	0.90	5.63	0.93
TCN	1.57	1.02	5.69	0.92
SVR	1.63	1.05	5.68	0.91
UNetTSF	1.26	0.54	5.58	0.95

表5 加速磨损阶段各模型性能对比

Tab.4 Comparison of the performance of each model in accelerated wear stage

模型	RMSE	MAE	MAPE/%	R²
BiLSTM	1.47	0.96	6.09	0.95
TCN	1.82	1.20	6.05	0.93
SVR	1.95	1.42	6.12	0.95
UNetTSF	1.25	0.54	6.10	0.95

参考文献 19

[1]	陈仁祥，吴志元，胡小林，等. 深度特征联合匹配的不同刀具间磨损状态识别［J］. 仪器仪表学报， 2020， 41（12）： 138-145.
	CHEN Renxiang， WU Zhiyuan， HU Xiaolin， et al. Wear State Recognition for Different Tools Based on the Joint Matching of Depth Characteristics［J］. Chinese Journal of Scientific Instrument， 2020， 41（12）： 138-145.
[2]	尹晨，周世超，何建樑，等. 基于多源同步信号与深度学习的刀具磨损在线识别方法［J］. 中国机械工程， 2021， 32（20）： 2482-2491.
	YIN Chen， ZHOU Shichao， HE Jianliang， et al. Tool Wear Online Recognition Method Based on Multi-source Synchronous Signals and Deep Learning［J］. China Mechanical Engineering， 2021， 32（20）： 2482-2491.
[3]	ZHOU Yang， LIU Changfu， YU Xinli， et al. Tool Wear Mechanism， Monitoring and Remaining Useful Life （RUL） Technology Based on Big Data： a Review［J］. SN Applied Sciences， 2022， 4（8）： 232.
[4]	MARTÍNEZ-ARELLANO G， TERRAZAS G， RATCHEV S. Tool Wear Classification Using Time Series Imaging and Deep Learning［J］. The International Journal of Advanced Manufacturing Technology， 2019， 104（9）： 3647-3662.
[5]	YANG Yinfei， GUO Yuelong， HUANG Zhiping， et al. Research on the Milling Tool Wear and Life Prediction by Establishing an Integrated Predictive Model［J］. Measurement， 2019， 145： 178-189.
[6]	LIU Min， YAO Xifan， ZHANG Jianming， et al. Multi-sensor Data Fusion for Remaining Useful Life Prediction of Machining Tools by IABC-BPNN in Dry Milling Operations［J］. Sensors， 2020， 20（17）： 4657.
[7]	BERGS T， HOLST C， GUPTA P， et al. Digital Image Processing with Deep Learning for Automated Cutting Tool Wear Detection［J］. Procedia Manufacturing， 2020， 48： 947-958.
[8]	林晨. 基于目标检测与语义分割的立铣刀磨损状态检测方法［D］. 杭州：杭州电子科技大学， 2024.
	LIN Chen. Method for End Mill Wear Status Detection Based on Object Detection and Semantic Segmentation［D］. Hangzhou： Hangzhou Dianzi University， 2024.
[9]	滕瑞，黄海松，杨凯，等. 基于图像编码技术和卷积神经网络的刀具磨损值在线监测方法［J］. 计算机集成制造系统， 2022， 28（4）： 1042-1051.
	TENG Rui， HUANG Haisong， YANG Kai， et al. On-line Monitoring Method for Tool Wear Based on Image Coding Technology and Convolutional Neural Network［J］. Computer Integrated Manufacturing Systems， 2022， 28（4）： 1042-1051.
[10]	梁军华. 镍基高温合金切削过程中刀具磨损及其状态监测技术研究［D］. 成都：西南交通大学， 2022.
	LIANG Junhua. Research on Tool Wear and Its Condition Monitoring Technology during the Cutting of Nickel-based Superalloys［D］. Chengdu： Southwest Jiaotong University， 2022.
[11]	LI Chu， XIAO Bingjia， YUAN Qingping. UnetTSF： a Better Performance Linear Complexity Time Series Prediction Model［EB/OL］. ［2024-09-19］. .
[12]	孙皓章，孔繁星，陈娜，等. 基于GAF-CNN的刀具磨损程度识别研究［J］. 机械工程师， 2023（8）： 7-10.
	SUN Haozhang， KONG Fanxing， CHEN Na， et al. Research on Recognition of Tool Wear Degree Based on GAF-CNN［J］. Mechanical Engineer， 2023（8）： 7-10.
[13]	PARK S， PARK Y H， HUH J， et al. Deep Learning Model for Differentiating Acute Myeloid and Lymphoblastic Leukemia in Peripheral Blood Cell Images via Myeloblast and Lymphoblast Classification［J］. Digit Health，2024，10：20552076241258079.
[14]	DENG Liwei， SUO Hongfei， LI Dongjie. Deepfake Video Detection Based on EfficientNet-V2 Network［J］. Computational Intelligence and Neuroscience， 2022， 2022（1）： 3441549.
[15]	LI Jiafeng， WEN Ying， HE Lianghua. SCConv： Spatial and Channel Reconstruction Convolution for Feature Redundancy［C］∥2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Vancouver， 2023： 6153-6162.
[16]	ZHAO Song， CAI Taiwei， PENG Bao， et al. GAM-YOLOv8n： Enhanced Feature Extraction and Difficult Example Learning for Site Distribution Box Door Status Detection［J］. Wireless Networks， 2024， 30（8）： 6939-6950.
[17]	KOU Rui， LIAN Shiwei， XIE Nan， et al. Image-based Tool Condition Monitoring Based on Convolution Neural Network in Turning Process［J］. The International Journal of Advanced Manufacturing Technology， 2022， 119（5）： 3279-3291.
[18]	龙佳宁，张昭，刘晓航，等. 利用改进EfficientNetV2和无人机图像检测小麦倒伏类型［J］. 智慧农业（中英文）， 2023， 5（3）： 62-74.
	LONG Jianing， ZHANG Zhao， LIU Xiaohang， et al. Wheat Lodging Types Detection Based on UAV Image Using Improved EfficientNetV2［J］. Smart Agriculture， 2023， 5（3）： 62-74.
[19]	YANG Rui， LU Xiangyu， HUANG Jing， et al. A Multi-source Data Fusion Decision-making Method for Disease and Pest Detection of Grape Foliage Based on ShuffleNet V2［J］. Remote Sensing， 2021， 13（24）： 5102.