Multi-step Ahead Real-time Prediction of Tool Wear Based on YOLOv11-Seg and Transformer Model

doi:10.3969/j.issn.1004-132X.2025.12.017

Abstract

Abstract:

To address the problems of low prediction accuracy， poor generalization capability， and difficulty in achieving real-time prediction of tool wear states in traditional methods， a multi-step forward real-time prediction method for tool wear was proposed by integrating the YOLOv11-Seg model with a Transformer model. An on-machine monitoring experimental platform for tool wear of CNC machines was constructed， and PMC programming was employed to realize automatic cutting of tools and automatic photographing of wear regions. An improved YOLOv11-Seg model was adopted for tool wear measurement， where the CoordAtt coordinate attention mechanism and the Shape-IoU loss function were introduced to improve the accuracy of wear region segmentation. Based on tool wear time-series data and real-time wear measurement data， an improved Transformer multi-step forward real-time prediction model was established under the MFTWP mode， where the residual error correction mechanism was introduced to enhance the accuracy and stability of MFTWP mode prediction. The proposed model was tested on both of public datasets and experimental datasets， and was compared with traditional prediction models. The results show that the proposed multi-step forward real-time prediction model exhibits high accuracy and good generalization capability.

Key words: tool wear monitoring, multi-step forword tool wear prediction（MFTWP）, YOLOv11-Seg model, Transformer model

摘要：

针对传统刀具磨损预测精度低、泛化能力差、难以实现磨损状态实时预测的问题，提出一种融合YOLOv11-Seg模型与Transformer模型的刀具磨损多步向前实时预测方法。搭建数控机床刀具磨损在机监测实验平台，采用PMC编程实现刀具自动切削和磨损区域自动拍照。采用改进YOLOv11-Seg模型进行刀具磨损量检测，引入CoordAtt坐标注意力机制和Shape-IoU损失函数以提高磨损区域分割的精度。基于刀具磨损时序数据和实时磨损量数据构建改进Transformer多步向前刀具磨损预测（AFTWP）模型，并在改进Transformer模型中引入残差校正机制，提高了MFTWP模型预测的精度和稳定性。采用公开数据集和实验数据集测试改进模型，将结果与传统预测模型进行比较，验证了提出的多步向前实时预测模型的精确性和泛化能力。

关键词: 刀具磨损监测, 多步向前刀具磨损预测, YOLOv11-Seg模型, Transformer模型

CLC Number:

TG71

Yufeng XIAO, Chaoyong ZHANG, Saixiyalatu, Yifan MENG, Chuanjun ZHU. Multi-step Ahead Real-time Prediction of Tool Wear Based on YOLOv11-Seg and Transformer Model[J]. China Mechanical Engineering, 2025, 36(12): 2944-2951.

肖御风, 张超勇, 赛希亚拉图, 孟一帆, 朱传军. 基于YOLOv11-Seg与Transformer模型的刀具磨损多步向前实时预测方法[J]. 中国机械工程, 2025, 36(12): 2944-2951.

Figures/Tables 16

Fig.1 Tool wear experiment platform

Fig.2 Wear images at different stages

Fig.3 Image annotation

Fig.4 Real-time tool wear prediction model

Fig.5 CoordAtt attention mechanism

Tab.1 Prediction results of MFTWP model without RCS

参数		$σ R M S E$ / $μ m$			$σ M A P E$ / $%$			$σ M A E$ / $μ m$			$R 2 / %$
$T 1$	$T 2$	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$
10	5	13.25	110.12	8.87	8.36	85.26	6.22	10.08	96.35	8.14	90.33	$-$ 6.40	94.52
	10	5.99	12.24	7.84	4.91	11.06	4.36	4.56	10.89	6.06	98.15	88.20	95.87
	15	3.94	10.98	8.64	3.47	8.04	6.54	3.63	8.63	7.58	98.91	89.76	95.32
20	5	7.76	154.37	9.40	4.10	75.77	6.30	5.96	101.53	8.24	95.88	$-$ 17.31	94.20
	10	7.19	13.07	7.78	4.02	10.68	5.73	5.53	11.39	7.06	96.31	85.49	95.16
	15	4.01	100.66	4.13	2.20	58.29	2.86	3.08	72.07	3.52	98.77	$-$ 8.75	98.34
30	5	4.42	53.26	13.04	3.17	34.17	7.93	3.61	42.09	10.96	98.96	$-$ 1.22	87.03
	10	3.76	15.71	8.58	2.92	9.93	5.51	2.96	11.93	7.27	99.23	78.21	94.33
	15	12.99	5.68	18.42	10.29	4.50	8.53	12.64	4.72	13.52	88.39	97.24	74.09

Tab.1 Prediction results of MFTWP model without RCS

参数		$σ R M S E$ / $μ m$			$σ M A P E$ / $%$			$σ M A E$ / $μ m$			$R 2 / %$
$T 1$	$T 2$	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$
10	5	13.25	110.12	8.87	8.36	85.26	6.22	10.08	96.35	8.14	90.33	$-$ 6.40	94.52
	10	5.99	12.24	7.84	4.91	11.06	4.36	4.56	10.89	6.06	98.15	88.20	95.87
	15	3.94	10.98	8.64	3.47	8.04	6.54	3.63	8.63	7.58	98.91	89.76	95.32
20	5	7.76	154.37	9.40	4.10	75.77	6.30	5.96	101.53	8.24	95.88	$-$ 17.31	94.20
	10	7.19	13.07	7.78	4.02	10.68	5.73	5.53	11.39	7.06	96.31	85.49	95.16
	15	4.01	100.66	4.13	2.20	58.29	2.86	3.08	72.07	3.52	98.77	$-$ 8.75	98.34
30	5	4.42	53.26	13.04	3.17	34.17	7.93	3.61	42.09	10.96	98.96	$-$ 1.22	87.03
	10	3.76	15.71	8.58	2.92	9.93	5.51	2.96	11.93	7.27	99.23	78.21	94.33
	15	12.99	5.68	18.42	10.29	4.50	8.53	12.64	4.72	13.52	88.39	97.24	74.09

Tab.2 Prediction results of MFTWP model with RCS

参数		$σ R M S E$ / $μ m$			$σ M A P E$ / $%$			$σ M A E$ / $μ m$			$R 2 / %$
$T 1$	$T 2$	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$
10	5	0.85	2.59	2.02	0.69	1.95	1.35	0.63	1.80	1.27	99.94	99.49	99.75
	10	1.89	2.92	2.07	1.51	2.36	1.19	1.32	2.17	1.39	99.92	99.56	99.64
	15	1.68	3.44	3.21	1.33	3.05	2.01	1.22	2.79	2.35	99.30	98.84	99.29
20	5	1.48	3.39	2.59	1.07	3.02	1.85	1.21	2.80	2.24	99.31	99.22	99.54
	10	1.50	2.94	3.06	1.16	2.55	2.05	1.21	2.42	2.50	99.46	99.54	99.61
	15	1.09	5.58	2.67	0.77	4.42	1.81	0.84	4.23	2.23	98.67	97.40	99.42
30	5	2.24	2.71	3.03	1.61	2.50	2.16	1.71	2.45	2.66	99.31	99.27	98.96
	10	2.64	2.93	3.26	1.98	2.49	2.14	1.94	2.48	2.69	99.45	98.64	99.19
	15	3.16	3.39	3.09	1.97	2.79	1.91	1.78	2.79	2.51	99.24	99.38	99.81

Tab.2 Prediction results of MFTWP model with RCS

参数		$σ R M S E$ / $μ m$			$σ M A P E$ / $%$			$σ M A E$ / $μ m$			$R 2 / %$
$T 1$	$T 2$	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$
10	5	0.85	2.59	2.02	0.69	1.95	1.35	0.63	1.80	1.27	99.94	99.49	99.75
	10	1.89	2.92	2.07	1.51	2.36	1.19	1.32	2.17	1.39	99.92	99.56	99.64
	15	1.68	3.44	3.21	1.33	3.05	2.01	1.22	2.79	2.35	99.30	98.84	99.29
20	5	1.48	3.39	2.59	1.07	3.02	1.85	1.21	2.80	2.24	99.31	99.22	99.54
	10	1.50	2.94	3.06	1.16	2.55	2.05	1.21	2.42	2.50	99.46	99.54	99.61
	15	1.09	5.58	2.67	0.77	4.42	1.81	0.84	4.23	2.23	98.67	97.40	99.42
30	5	2.24	2.71	3.03	1.61	2.50	2.16	1.71	2.45	2.66	99.31	99.27	98.96
	10	2.64	2.93	3.26	1.98	2.49	2.14	1.94	2.48	2.69	99.45	98.64	99.19
	15	3.16	3.39	3.09	1.97	2.79	1.91	1.78	2.79	2.51	99.24	99.38	99.81

Tab.3 The impact of δ on model prediction accuracy （C4）

$δ$	$n$	$σ R M S E$ / $μ m$	$σ M A P E$ / $%$	$σ M A E$ / $μ m$	$R 2 / %$
100	1	56.17	49.58	46.99	$-$ 1.35
10	6	7.58	7.27	6.70	96.36
1	45	2.59	1.95	1.80	99.49

Tab.3 The impact of δ on model prediction accuracy （C4）

$δ$	$n$	$σ R M S E$ / $μ m$	$σ M A P E$ / $%$	$σ M A E$ / $μ m$	$R 2 / %$
100	1	56.17	49.58	46.99	$-$ 1.35
10	6	7.58	7.27	6.70	96.36
1	45	2.59	1.95	1.80	99.49

Tab.4 Performance metrics of tool wear prediction model

模型	$σ R M S E$ / $μ m$			$σ M A P E$ / $%$			$σ M A E$ / $μ m$			$R 2 / %$
模型	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$
CNN	1.37	1.62	1.37	1.51	1.60	1.26	0.74	1.06	0.95	99.87	99.40	99.58
TCN	2.01	2.53	1.95	2.21	2.33	2.18	1.15	1.30	1.28	98.75	98.47	97.53
BiLSTM	0.88	1.10	1.13	0.54	0.81	0.86	0.80	0.96	0.98	99.96	99.91	99.91
GRU	0.85	1.70	1.73	0.78	1.59	1.43	0.69	1.50	1.50	99.92	99.17	99.32
Transformer	0.48	1.01	1.03	0.32	0.98	0.85	0.33	0.84	0.86	99.98	99.94	99.94

Tab.4 Performance metrics of tool wear prediction model

模型	$σ R M S E$ / $μ m$			$σ M A P E$ / $%$			$σ M A E$ / $μ m$			$R 2 / %$
模型	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$	$C 1$	$C 4$	$C 6$
CNN	1.37	1.62	1.37	1.51	1.60	1.26	0.74	1.06	0.95	99.87	99.40	99.58
TCN	2.01	2.53	1.95	2.21	2.33	2.18	1.15	1.30	1.28	98.75	98.47	97.53
BiLSTM	0.88	1.10	1.13	0.54	0.81	0.86	0.80	0.96	0.98	99.96	99.91	99.91
GRU	0.85	1.70	1.73	0.78	1.59	1.43	0.69	1.50	1.50	99.92	99.17	99.32
Transformer	0.48	1.01	1.03	0.32	0.98	0.85	0.33	0.84	0.86	99.98	99.94	99.94

Tab.5 Ablation results

模型	CoordAtt	Shape-IoU	$P ¯ m A P$ /%	时间/ms
YOLOv11-Seg	×	×	86.90	65.85
	√	×	88.02	66.32
	×	√	88.31	66.69
	√	√	90.79	67.70

Tab.5 Ablation results

模型	CoordAtt	Shape-IoU	$P ¯ m A P$ /%	时间/ms
YOLOv11-Seg	×	×	86.90	65.85
	√	×	88.02	66.32
	×	√	88.31	66.69
	√	√	90.79	67.70

Tab.6 Performance comparison of the improved YOLOv11-Seg model with other models

模型	$P / %$	$R / %$	$P ¯ m A P$ /%
FCN	89.56	91.03	80.93
Lraspp	92.12	89.79	83.28
DeepLabV3	93.45	95.31	85.15
YOLOv5s-Seg	93.57	96.46	84.60
YOLOv11-Seg改进	98.54	99.72	91.26

Tab.6 Performance comparison of the improved YOLOv11-Seg model with other models

模型	$P / %$	$R / %$	$P ¯ m A P$ /%
FCN	89.56	91.03	80.93
Lraspp	92.12	89.79	83.28
DeepLabV3	93.45	95.31	85.15
YOLOv5s-Seg	93.57	96.46	84.60
YOLOv11-Seg改进	98.54	99.72	91.26

Fig.6 Wear area width measurement

Tab.7 Prediction results of MFTWP model with and without RCS

模型		$σ R M S E$ / $μ m$	$σ M A P E$ / $%$	$σ M A E$ / $μ m$	$R 2 / %$
MFTWP	$E 1$	12.72	8.27	11.07	89.76
	$E 2$	109.83	83.67	94.96	-7.99
	$E 3$	8.97	6.85	8.03	94.33
MFTWP-RCS	$E 1$	1.69	1.28	1.40	99.24
	$E 2$	2.51	2.20	2.00	99.67
	$E 3$	2.59	1.57	1.09	99.95

Tab.7 Prediction results of MFTWP model with and without RCS

模型		$σ R M S E$ / $μ m$	$σ M A P E$ / $%$	$σ M A E$ / $μ m$	$R 2 / %$
MFTWP	$E 1$	12.72	8.27	11.07	89.76
	$E 2$	109.83	83.67	94.96	-7.99
	$E 3$	8.97	6.85	8.03	94.33
MFTWP-RCS	$E 1$	1.69	1.28	1.40	99.24
	$E 2$	2.51	2.20	2.00	99.67
	$E 3$	2.59	1.57	1.09	99.95

Tab.8 Performance metrics of tool wear prediction model

模型	$σ R M S E$ / $μ m$			$σ M A P E$ / $%$			$σ M A E$ / $μ m$			$R 2 / %$
模型	$E 1$	$E 2$	$E 3$	$E 1$	$E 2$	$E 3$	$E 1$	$E 2$	$E 3$	$E 1$	$E 2$	$E 3$
CNN	9.3	11.6	16.6	1.12	1.35	2.18	4.4	6.1	8.7	99.86	99.70	99.51
TCN	26.7	27.2	48.9	54.36	35.73	27.95	21.6	24.0	36.4	98.75	98.47	97.53
BiLSTM	6.5	9.4	20.1	0.94	1.25	2.27	4.4	5.3	11.2	99.93	99.79	99.37
GRU	10.9	10.3	13.6	1.62	1.92	2.82	6.8	6.4	9.6	99.81	99.77	99.66
Transformer	4.4	5.8	6.6	0.85	0.94	1.19	2.9	3.3	4.3	99.97	99.93	99.92

Tab.8 Performance metrics of tool wear prediction model

模型	$σ R M S E$ / $μ m$			$σ M A P E$ / $%$			$σ M A E$ / $μ m$			$R 2 / %$
模型	$E 1$	$E 2$	$E 3$	$E 1$	$E 2$	$E 3$	$E 1$	$E 2$	$E 3$	$E 1$	$E 2$	$E 3$
CNN	9.3	11.6	16.6	1.12	1.35	2.18	4.4	6.1	8.7	99.86	99.70	99.51
TCN	26.7	27.2	48.9	54.36	35.73	27.95	21.6	24.0	36.4	98.75	98.47	97.53
BiLSTM	6.5	9.4	20.1	0.94	1.25	2.27	4.4	5.3	11.2	99.93	99.79	99.37
GRU	10.9	10.3	13.6	1.62	1.92	2.82	6.8	6.4	9.6	99.81	99.77	99.66
Transformer	4.4	5.8	6.6	0.85	0.94	1.19	2.9	3.3	4.3	99.97	99.93	99.92

Tab.9 Model prediction time

模型	数据集
	$E 1$			$E 2$			$E 3$
	$δ$ =100	$δ$ =10	$δ$ =1	$δ$ =100	$δ$ =10	$δ$ =1	$δ$ =100	$δ$ =10	$δ$ =1
	$n$ =0	$n$ =2	$n$ =20	$n$ =1	$n$ =6	$n$ =45	$n$ =0	$n$ =3	$n$ =36
Transformer	0.0237	0.0191	0.1452	0.0631	0.0822	0.1843	0.0129	0.0201	0.1465
CNN	0.6219			1.3425			0.4993
BiLSTM	0.6850			0.7012			0.6095

Tab.9 Model prediction time

模型	数据集
	$E 1$			$E 2$			$E 3$
	$δ$ =100	$δ$ =10	$δ$ =1	$δ$ =100	$δ$ =10	$δ$ =1	$δ$ =100	$δ$ =10	$δ$ =1
	$n$ =0	$n$ =2	$n$ =20	$n$ =1	$n$ =6	$n$ =45	$n$ =0	$n$ =3	$n$ =36
Transformer	0.0237	0.0191	0.1452	0.0631	0.0822	0.1843	0.0129	0.0201	0.1465
CNN	0.6219			1.3425			0.4993
BiLSTM	0.6850			0.7012			0.6095

Fig.7 Improved transformer model prediction curve

References 12

[1]	LINS R G， GUERREIRO B， MARQUES de ARAUJO P R， et al. In-process Tool Wear Measurement System Based on Image Analysis for CNC Drilling Machines［J］. IEEE Transactions on Instrumentation and Measurement， 2020， 69（8）： 5579-5588.
[2]	钱海章，张强，李帅. “十四五”规划下中国制造供给能力及发展路径思考［J］. 数量经济技术经济研究， 2022， 39（1）： 28-50.
	QIAN Haizhang， ZHANG Qiang， LI Shuai. Research on Development Path of China's Intelligent Manufacturing Supply Capacity under Background of 14th Five-Year Plan［J］. The Journal of Quantitative & Technical Economics， 2022， 39（1）： 28-50.
[3]	刘强. 数控机床发展历程及未来趋势［J］. 中国机械工程， 2021， 32（7）： 757-770.
	LIU Qiang. Development History and Future Trends of Numerical Control Machine Tools［J］. China Mechanical Engineering， 2021， 32（7）： 757-770.
[4]	李维刚，叶欣，赵云涛，等. 基于改进YOLOv3算法的带钢表面缺陷检测［J］. 电子学报， 2020， 48（7）： 1284-1292.
	LI Weigang， YE Xin， ZHAO Yuntao， et al. Strip Steel Surface Defect Detection Based on Improved YOLOv3 Algorithm［J］. Acta Electronica Sinica， 2020， 48（7）： 1284-1292.
[5]	曹大理，孙惠斌，张纪铎，等. 基于卷积神经网络的刀具磨损在线监测［J］. 计算机集成制造系统， 2020， 26（1）： 74-80.
	CAO Dali， SUN Huibin， ZHANG Jiduo， et al. In-process Tool Condition Monitoring Based on Convolution Neural Network［J］. Computer Integrated Manufacturing Systems， 2020， 26（1）： 74-80.
[6]	秦奥苹. 基于机器视觉的刀具磨损在机自动检测系统研究［D］. 成都：西南交通大学， 2021.
	QIN Aoping. Research on On-machine Automatic Detection System of Tool Wear Based on Machine Vision［D］. Chengdu：Southwest Jiaotong University， 2021.
[7]	张春霖，周婷婷，胡天亮，等. 刀具切削变工况数字孪生模型构建方法［J］. 计算机集成制造系统， 2023， 29（6）： 1852-1866.
	ZHANG Chunlin， ZHOU Tingting， HU Tianliang， et al. Construction Method of Digital Twin Model for Cutting Tools under Variable Working Conditions［J］. Computer Integrated Manufacturing Systems， 2023， 29（6）： 1852-1866.
[8]	MARANI M， ZEINALI M， SONGMENE V， et al. Tool Wear Prediction in High-speed Turning of a Steel Alloy Using Long Short-term Memory Modelling［J］. Measurement， 2021， 177： 109329.
[9]	何彦，凌俊杰，王禹林，等. 基于长短时记忆卷积神经网络的刀具磨损在线监测模型［J］. 中国机械工程， 2020， 31（16）： 1959-1967.
	HE Yan， LING Junjie， WANG Yulin， et al. In-process Tool Wear Monitoring Model Based on LSTM-CNN［J］. China Mechanical Engineering， 2020， 31（16）： 1959-1967.
[10]	朱锟鹏，黄称意，李俊. 刀具磨损状态的多步向前智能预测［J］. 计算机集成制造系统， 2024， 30（9）： 3038-3049.
	ZHU Kunpeng， HUANG Chengyi， LI Jun. Multi-step Forward Intelligent Prediction of Tool Wear Condition［J］. Computer Integrated Manufacturing Systems， 2024， 30（9）： 3038-3049.
[11]	张贺，邵琪越，薄翠梅. 基于注意力机制混合网络模型的刀具磨损监测［J/OL］. 控制工程， 1-8 ［2025-03-29］.
	ZHANG He， SHAO Qiyue， BO Cuimei. Tool Wear Monitoring Based on a Hybrid Network Model with Attention Mechanism［J/OL］. Control Engineering of China， 1-8 ［2025-03-29］.
[12]	张宇杰，蔡乐才，成奎，等. 改进YOLOv5金属表面缺陷检测方法［J］. 四川轻化工大学学报（自然科学版）， 2022， 35（4）： 32-41.
	ZHANG Yujie， CAI Lecai， CHENG Kui， et al. An Improved YOLOv5 Method for the Detection of Metal Surface Defects［J］. Journal of Sichuan University of Science & Engineering （Natural Science Edition）， 2022， 35（4）： 32-41.