A Full-position Welding Pool Identification and Deviation Measurement Method Based on DeepLab-EMCAD

doi:10.3969/j.issn.1004-132X.2025.12.023

Abstract

Abstract:

A method for full-position welding pool identification and deviation measurement was proposed based on DeepLab-EMCAD. A lightweight MobileNetV3 network was adopted as the backbone of the model encoder， and the atrous spatial pyramid pooling（ASPP） module was optimized to reduce the model parameters and improve the segmentation efficiency. The EMCAD multi-attention mechanism was integrated into the decoder to enhance the segmentation accuracy of the welding pools. A deviation calculation method was proposed to quantitatively describe the deviation based on the segmentation results of the welding pools. Experimental results show that compared with the baseline model， the proposed model improves the average intersection over union and average pixel accuracy in welding pool segmentations by 5.72% and 5.5% respectively， and the inference time is reduced by 29.69 ms， with the number of parameters decreasing by 4.854×10⁷. Compared with classic segmentation networks， the proposed model has the best performance in handling the edges of the welding pools. The deviation detection errors are controlled within 0.1 mm.

Key words: pipe full position welding, efficient multi-scale convolutional attention decoder（EMCAD）, welding pool, semantic segmentation, deviation measurement

摘要：

提出了一种基于DeepLab-EMCAD的全位置焊接熔池识别与偏差测量方法。采用轻量级MobileNetV3网络作为模型编码器主干网络，并优化空洞空间金字塔池化模块，以减少模型参数量、提高分割效率。在解码器中集成EMCAD多注意力机制，提高熔池的分割精度。基于熔池分割结果，提出一种偏差计算方法来定量描述偏差。实验结果显示，相较于基线模型，所提出的模型在熔池分割中的平均交并比、平均像素精确度分别提高了5.72%和5.5%，推理时间缩短了29.69 ms，参数量减少了4.854×10⁷个；相比于经典分割网络，所提出的模型在熔池边缘处理上效果更优，焊接偏差检测误差控制在0.1 mm之内。

关键词: 管道全位置焊接, 高效多尺度卷积注意力解码器（EMCAD）, 熔池, 语义分割, 偏差测定

CLC Number:

TG409

Yecheng XIONG, Haisheng LIU, Zhongren WANG, Tielin SHI, Hai XIA, Hongbo YANG. A Full-position Welding Pool Identification and Deviation Measurement Method Based on DeepLab-EMCAD[J]. China Mechanical Engineering, 2025, 36(12): 2993-3001.

熊烨成, 刘海生, 王中任, 史铁林, 夏海, 杨洪波. 基于DeepLab-高效多尺度卷积注意力解码器的全位置焊接熔池识别与偏差测量方法[J]. 中国机械工程, 2025, 36(12): 2993-3001.

Figures/Tables 18

Fig.1 Schematic diagram of the all-position pipeline welding experimental platform

Tab.1 Welding parameter

名称	参数
焊接实心焊丝	ER50-6
焊接保护气体	80%Ar+20%CO₂
气体输送流量/（L·min $- 1$ ）	18
焊丝送丝速度/（cm·s $- 1$ ）	3.2
焊接机器人内外延时/ms	300
焊枪头摆动速度/（mm·s $- 1$ ）	30.0
焊接机器人移速/（mm·s $- 1$ ）	8.0
焊机设置电流I/A	140
焊机设置电压U/V	17.5

Tab.1 Welding parameter

名称	参数
焊接实心焊丝	ER50-6
焊接保护气体	80%Ar+20%CO₂
气体输送流量/（L·min $- 1$ ）	18
焊丝送丝速度/（cm·s $- 1$ ）	3.2
焊接机器人内外延时/ms	300
焊枪头摆动速度/（mm·s $- 1$ ）	30.0
焊接机器人移速/（mm·s $- 1$ ）	8.0
焊机设置电流I/A	140
焊机设置电压U/V	17.5

Fig.2 Principle of deviation measurement for a single welding cycle

Fig.3 Force diagram of the welding pool at all positions of the pipeline

Fig.4 Structure of the Deeplab-EMCAD segmentation model

Tab.2 Modified MobileNetV3-Small structure

输入层	操作符	扩展维度	输出通道	SE 模块	NL模块
224²×3	2d卷积，3×3	2d卷积，3×3	16		HS模块
112²×16	Bneck模块，3×3	Bneck模块，3×3	16	√	RE模块
56²×16	Bneck模块，3×3	Bneck模块，3×3	24		RE模块
28²×24	Bneck模块，3×3	Bneck模块，3×3	24		RE模块
28²×24	Bneck模块，5×5	Bneck模块，5×5	40	√	HS模块
14²×40	Bneck模块，5×5	Bneck模块，5×5	40	√	HS模块
14²×40	Bneck模块，5×5	Bneck模块，5×5	40	√	HS模块
14²×40	Bneck模块，5×5	Bneck模块，5×5	48	√	HS模块
14²×48	Bneck模块，5×5	Bneck模块，5×5	48	√	HS模块
14²×48	Bneck模块，5×5	Bneck模块，5×5	96	√	HS模块
7²×96	Bneck模块，5×5	Bneck模块，5×5	96	√	HS模块
7²×96	Bneck模块，5×5	Bneck模块，5×5	96	√	HS模块

Fig.5 Structure of different modules in EMCAD

Tab.3 The hyperparameter comparison experiments

编号	学习率	批大小	训练轮次	准确率/%	内存消耗/GB	时间/h
1	0.0005	8	200	98.5	14	4.0
2	0.0005	16	200	97.0	22	6.0
3	0.0010	8	200	96.8	18	4.5
4	0.0005	4	200	94.0	14	3.5
5	0.0005	8	100	93.5	16	3.0
6	0.0005	8	300	98.8	16.5	5.0

Fig.6 Experimental comparison of different backbone network models on different datasets

Tab.4 The compares experimental results

网络模型	平均交并比/%	平均像素精度/%	推理时间/ms	参数量/10⁶
SegNet模型	87.61	87.85	25.27	46.25
PSPNet模型	86.43	89.89	27.92	48.96
UNet模型	93.82	95.92	19.20	42.98
Yolov8模型	92.75	94.22	25.61	40.15
DeepLabV3+模型	92.14	93.21	37.92	55.10
本文模型	97.86	98.71	8.23	6.56

Fig.7 shows the segmentation effects of different models on the molten pool

Fig.8 The cutting angle θ1、 θ2 during normal welding

Fig.9 The cutting angle θ1、 θ2 when offset 1 mm to the right

Fig.10 The cutting angle θ1、 θ2 when offset 1.5 mm to the right

Fig.11 The cutting angle θ1、 θ2 when offset 1 mm to the left

Fig.12 The cutting angle θ1、 θ2 when offset 1.5 mm to the left

Fig.13 Scatter plot of the angle cut and welding deviation by the molten pool method

Tab.5 Comparison of predicted welding deviations and actual deviations

序号	实际值/ mm	\| d \|/mm	θ₁/（ °）	θ₂/（ °）	误差/mm	偏差方向
1	1.2	1.163	8.3	42.0	0.037	右
2	0.7	0.765	43.5	13.2	0.065	左
3	1.0	0.911	10.2	37.5	0.089	右
4	1.2	1.247	52.7	6.1	0.047	左
5	0.4	0.489	18.3	31.3	0.089	右
6	0.3	0.318	24.5	29.6	0.018	右
7	0.7	0.800	44.2	12.7	0.100	左
8	1.3	1.289	53.1	5.4	0.011	左
9	1.0	0.940	9.6	37.9	0.060	右
10	1.4	1.302	7.3	44.5	0.098	右

References 17

[1]	ROUT A， DEEPAK B B V L， BISWAL B B. Advances in Weld Seam Tracking Techniques for Robotic Welding： a Review［J］. Robotics and Computer-Integrated Manufacturing， 2019， 56： 12-37.
[2]	WANG Tianqi， MENG Kaiquan， CUI Liyang， et al. Weld Tracking Technology for All-position Welding of Pipes Based on Laser Vision［J］. Optics and Lasers in Engineering， 2025， 188： 108912.
[3]	HE Yinshui， MA Guohong， CHEN Shanben. Autonomous Decision-making of Welding Position during Multipass GMAW with T-Joints： a Bayesian Network Approach［J］. IEEE Transactions on Industrial Electronics， 2022， 69（4）： 3909-3917.
[4]	WANG Baicun， HU S J， SUN Lei， et al. Intelligent Welding System Technologies： State-of-the-art Review and Perspectives［J］. Journal of Manufacturing Systems， 2020， 56： 373-391.
[5]	唐昆，彭海滨，朱勇建，等. 基于改进重建网络的无监督镁合金激光焊接焊缝表面缺陷视觉检测［J/OL］. 中国机械工程， 2025： 1-12. ［2025-02-14］. .
	TANG Kun， PENG Haibin， ZHU Yongjian， et al. Unsupervised Visual Inspection of Surface Defects in Laser Welding Seams of Magnesium Alloy Based on Improved Reconstruction Network［J/OL］. China Mechanical Engineering， 2025： 1-12. ［2025-02-14］. .
[6]	HONG Yuxiang， JIANG Yuxuan， YANG Mingxuan， et al. Intelligent Seam Tracking in Foils Joining Based on Spatial-temporal Deep Learning from Molten Pool Serial Images［J］. Robotics and Computer-integrated Manufacturing， 2025， 91： 102840.
[7]	王林，潘骏，贺青川，等. 后倾式离心风机叶轮机器人焊接工艺优化［J］. 中国机械工程， 2020， 31（19）： 2379-2387.
	WANG Lin， PAN Jun， HE Qingchuan， et al. Welding Process Optimization on Robot Welding of the Backward Centrifugal Fan Impeller［J］. China Mechanical Engineering， 2020， 31（19）： 2379-2387.
[8]	XIA Lei， ZHOU Jianping， XUE Ruilei， et al. Real-time Seam Tracking during Narrow Gap GMAW Process Based on the Wide Dynamic Vision Sensing Method［J］. Journal of Manufacturing Processes， 2023， 101： 820-834.
[9]	王小刚，王中任，刘海生. 熔滴-焊缝同步视觉焊接偏差测定方法［J］. 计算机集成制造系统， 2022， 28（5）： 1352-1360.
	WANG Xiaogang， WANG Zhongren， LIU Haisheng. Measurement Method for Simultaneous Visual Welding Deviation of Droplet-weld Seam［J］. Computer Integrated Manufacturing Systems， 2022， 28（5）： 1352-1360.
[10]	HOU Huaishu， WANG Tong， LI Jinhao， et al. Vision-based Algorithm for Online TIG Welding Deviation Detection of Stainless Steel Welded Pipes［J］. Journal of the Optical Society of America A， 2024， 41（6）： 1175-1184.
[11]	MO Yujian， WU Yan， YANG Xinneng， et al. Review the State-of-the-art Technologies of Semantic Segmentation Based on Deep Learning［J］. Neurocomputing， 2022， 493： 626-646.
[12]	YU Rongwei， TAN Xiaxin， HE Shen， et al. Monitoring of Robot Trajectory Deviation Based on Multimodal Fusion Perception in WAAM Process［J］. Measurement， 2024， 224： 113933.
[13]	蒋宇轩，林凯，王瑶祺，等. 基于改进ViT的熔池识别与焊接偏差在线检测方法［J］. 现代制造工程， 2024（10）： 130-137.
	JIANG Yuxuan， LIN Kai， WANG Yaoqi， et al. Improved ViT-based Method for Molten Pool Recognition and Online Detection of Welding Deviation［J］. Modern Manufacturing Engineering， 2024（10）： 130-137.
[14]	CHEN L C， ZHU Yukun， PAPANDREOU G， et al. Encoder-decoder with Atrous Separable Convolution for Semantic Image Segmentation［C］∥Computer Vision – ECCV 2018. Cham： Springer， 2018： 833-851.
[15]	RAHMAN M M， MUNIR M， MARCULESCU R. EMCAD： Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation［C］∥2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. IEEE， 2024： 11769-11779.
[16]	HOWARD A， SANDLER M， CHEN Bo， et al. Searching for MobileNetV3［C］∥2019 IEEE/CVF International Conference on Computer Vision （ICCV）. IEEE， 2019： 1314-1324.
[17]	DAI Wei， LI Dayong， TANG Ding， et al. Deep Learning Assisted Vision Inspection of Resistance Spot Welds［J］. Journal of Manufacturing Processes， 2021， 62： 262-274.