Defects Detection for Laser Welding Seams of Magnesium Alloy Based on Improved Reconstruction Network

doi:10.3969/j.issn.1004-132X.2026.07.011

China Mechanical Engineering ›› 2026, Vol. 37 ›› Issue (7): 1624-1634.DOI: 10.3969/j.issn.1004-132X.2026.07.011

Defects Detection for Laser Welding Seams of Magnesium Alloy Based on Improved Reconstruction Network

TANG Kun¹(), PENG Haibin¹, ZHU Yongjian²(), ZHANG Hang¹, ZENG Gang³, XIAO Hongchao⁴, ZHOU Xiaojie¹, TANG Weidong¹, ZHANG Mingjun¹, MAO Cong¹

^1.Hunan Provincial Key Laboratory of Intelligent Manufacturing Technology for High-performance Mechanical Equipment，Changsha University of Science and Technology，Changsha，410114
^2.College of Engineering Physics，Shenzhen Technology University，Shenzhen，Guangdong，518118
^3.Hunan Provincial Technology Innovation Center of Aerospace New Light Alloy Materials，Changsha，410205
^4.Hunan Provincial Engineering Research Center of Wrought Magnesium Alloys and Surface Protection，Changsha，410205

Received:2024-10-12 Revised:2026-02-03 Online:2026-07-25 Published:2026-08-18
Contact: ZHU Yongjian

基于改进重建网络的镁合金激光焊缝缺陷检测

唐昆¹(), 彭海滨¹, 朱勇建²(), 张航¹, 曾钢³, 肖宏超⁴, 周小杰¹, 唐伟东¹, 张明军¹, 毛聪¹

^1.长沙理工大学机械装备高性能智能制造关键技术湖南省重点实验室, 长沙, 410114
^2.深圳技术大学工程物理学院, 深圳, 518118
^3.湖南省航天航空新型轻合金材料技术创新中心, 长沙, 410205
^4.湖南省变形镁合金材料及表面防护工程技术研究中心, 长沙, 410205

通讯作者: 朱勇建
作者简介:唐昆，男，1980年生，副教授、硕士研究生导师。研究方向为精密/超精密加工技术、机器视觉与深度学习。发表论文20余篇。E-mail： tangkun@csust.edu.cn
朱勇建^*（通信作者），男，1979年生，教授、博士研究生导师。研究方向为2D/3D智能感知技术、基于人工智能的机器视觉检测、激光3D测量。发表论文30余篇。E-mail： zhuyongjian@sztu.edu.cn。
基金资助:
国家自然科学基金(51405034);国家自然科学基金(51375297);湖南省高新技术产业科技创新引领计划(2022GK4027);长沙市自然科学基金(kq2402020)

Abstract

Abstract:

To tackle the problems of complex morphology and inconspicuous features in surface defects of magnesium alloy laser welding seams， which often lead to high miss-detection and false-positive rates， this paper presents an improved reconstruction network-based defect recognition approach. The network incorporates ASPP， CBAM， SSPCAB， and a multiscale feature fusion module （MSFFM） to enhance feature extraction and anomalous target localization， thereby improving reconstruction accuracy at defect sites and feature information fusion. Experiments conducted on both self-built and public datasets validate the proposed method. Results demonstrate that the approach exhibits strong generalization and can effectively identify， accurately segment， and localize defects characterized by small sample sizes and complex surface morphologies.

Key words: laser welding, defect detection, unsupervised learning, abnormal localization, reconstruction

摘要：

针对镁合金激光焊缝表面缺陷形态复杂、特征不明显，以及漏检与误检率高等问题，提出了基于改进重建网络的无监督缺陷检测方法。通过ASPP、CBAM、SSPCAB模块的引入，以及多尺度特征融合模块MSFFM的添加，提高了网络特征提取与异常目标定位的能力，增强了缺陷位置的重建精度与相邻网络层间的特征信息融合，并在自建与公共数据集上进行了验证。所提方法可对镁合金焊缝表面的小样本复杂形态缺陷进行有效识别、精准分割与定位，且具备良好的通用性。

关键词: 激光焊接, 缺陷检测, 无监督学习, 异常定位, 重建

CLC Number:

TG441.7

TANG Kun, PENG Haibin, ZHU Yongjian, ZHANG Hang, ZENG Gang, XIAO Hongchao, ZHOU Xiaojie, TANG Weidong, ZHANG Mingjun, MAO Cong. Defects Detection for Laser Welding Seams of Magnesium Alloy Based on Improved Reconstruction Network[J]. China Mechanical Engineering, 2026, 37(7): 1624-1634.

唐昆, 彭海滨, 朱勇建, 张航, 曾钢, 肖宏超, 周小杰, 唐伟东, 张明军, 毛聪. 基于改进重建网络的镁合金激光焊缝缺陷检测[J]. 中国机械工程, 2026, 37(7): 1624-1634.

Figures/Tables 24

References 32

[1]	李翠，周丹，孟晓明，等．不锈钢板激光光束摆动叠焊工艺研究［J］．中国机械工程，2019，30（11）：1359-1366．
	LI Cui， ZHOU Dan， MENG Xiaoming，et al ．Research on Laser Beam Swing Stacking Welding Process of Stainless Steel Plate ［J］．China Mechanical Engineering，2019，30（11）：1359-1366．
[2]	李坤，马锐津，戴启雷，等．高性能镁合金激光制造技术研究现状及展望［J］．航空制造技术，2022，65（Z2）：30-47．
	LI Kun， MA Ruijin， DAI Qilei，et al ．Research Status and Prospects of High Performance Magnesium Alloy Laser Manufacturing Technology ［J］．Aerospace Manufacturing Technology，2022，65（Z2）：30-47．
[3]	洪景山，祝颖丹，宋康康，等．基于深度学习的工业缺陷检测研究进展［J］．计算机科学，2024，51（10）：261-275．
	HONG Jingshan， ZHU Yingdan， SONG Kangkang，et al ．Research progress on industrial defect detection based on deep learning ［J］．Computer Science，2024，51（10）：261-275．
[4]	ZHANG Hui， WU Zuxuan， WANG Zheng， et al. Prototypical Residual Networks for Anomaly Detection and Localization［C］∥2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Vancouver， 2023： 16281-16291.
[5]	梁明，张明路，吕晓玲．基于特征金字塔匹配和自监督的表面缺陷检测［J］．激光与光电子学进展，2023，60（4）：367-375．
	LIANG Ming， ZHANG Minglu， LV Xiaoling ．Surface Defect Detection Based on Feature Pyramid Matching and Self-supervision［J］．Advances in Laser and Optoelectronics，2023，60（4）：367-375．
[6]	刘永江，陈斌．基于多尺度记忆库的像素级无监督工业异常检测［J］．计算机应用，2024，44（11）：3587-3594．
	LIU Yongjiang， CHEN Bin ．Pixel Level Unsupervised Industrial Anomaly Detection Based on Multi-scale Memory Bank ［J］．Computer Applications，2024，44（11）：3587-3594．
[7]	冯志远，陈莹．基于判别式增强的蒸馏学习自监督缺陷检测［J］．系统仿真学报，2025，37（6）：1499-1511．
	FENG Zhiyuan， CHEN Ying ．Distillation Learning Self Supervised Defect Detection Based on Discriminant Enhancement ［J］．Journal of System Simulation，2025，37（6）：1499-1511．
[8]	ROTH K， PEMULA L， ZEPEDA J， et al. Towards Total Recall in Industrial Anomaly Detection［C］∥2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. New Orleans， 2022： 14298-14308.
[9]	YANG Jie， SHI Yong， QI Zhiquan. Learning Deep Feature Correspondence for Unsupervised Anomaly Detection and Segmentation［J］. Pattern Recognition， 2022， 132： 108874.
[10]	YANG Jie， SHI Yong， QI Zhiquan ．DFR：Deep Feature Reconstruction for Unsupervised Anomaly Segmentation［J］．Neurocomputing，2020，424：9-22．
[11]	蒋婉，杨凯，邱春蓉，等．基于图像修复的无砟轨道道床异常检测算法［J］．激光与光电子学进展，2024，61（12）：410-421．
	JIANG Wan， YANG Kai， QIU Chunrong，et al ．Anomaly Detection Algorithm for Ballastless Track Bed Based on Image Restoration ［J］．Advances in Laser and Optoelectronics，2024，61（12）：410-421．
[12]	王栋欢，于艾洋，肖洪．基于无监督学习的航空发动机铸造涡轮叶片缺陷检测方法［J］．航空动力学报，2025， 40（6）： 1000-8055（2025）06-20230800-12．
	WANG Donghuan， YU Aiyang， XIAO Hong ．Defect Detection Method for Casting Turbine Blades in Aircraft Engines Based on Unsupervised Learning ［J］．Journal of Aeronautical Power，2025， 40（6）： 1000-8055（2025）06-20230800-12.
[13]	乐华钢，王文武，朱磊，等．融合自注意力与缺陷凸显的缺陷检测方法［J］．电子测量与仪器学报，2023，37（9）：85-92．
	LE HuaGang， WANG Wenwu， ZHU Lei，et al ．Defect Detection Method Integrating Self Attention and Defect Highlighting ［J］．Journal of Electronic Measurement and Instrumentation，2023，37（9）：85-92．
[14]	ZAVRTANIK V， KRISTAN M， SKOČAJ D. DRÆM -a Discriminatively Trained Reconstruction Embedding for Surface Anomaly Detection［C］∥2021 IEEE/CVF International Conference on Computer Vision （ICCV）. Montreal， 2021： 8310-8319.
[15]	GEORGESCU M I. Masked Autoencoders for Unsupervised Anomaly Detection in Medical Images［J］. Procedia Computer Science， 2023， 225： 969-978.
[16]	LIU Tongkun， LI Bing， DU Xiao， et al. Component-aware Anomaly Detection Framework for Adjustable and Logical Industrial Visual Inspection［J］. Advanced Engineering Informatics， 2023， 58： 102161.
[17]	CHEN L C， PAPANDREOU G， KOKKINOS I， et al. DeepLab： Semantic Image Segmentation with Deep Convolutional Nets， Atrous Convolution， and Fully Connected CRFS［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2018， 40（4）： 834-848.
[18]	WOO S， PARK J， LEE J Y， et al. CBAM： Convolutional Block Attention Module［C］∥ European Conference on Computer Vision. Munich， 2018： 3-19.
[19]	RISTEA N C， MADAN N， IONESCU R T， et al. Self-supervised Predictive Convolutional Attentive Block for Anomaly Detection［C］∥2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. New Orleans， 2022： 13566-13576.
[20]	YANG Minghui， WU Peng， FENG Hui. MemSeg： a Semi-supervised Method for Image Surface Defect Detection Using Differences and Commonalities［J］. Engineering Applications of Artificial Intelligence， 2023， 119： 105835.
[21]	THOMINE S， SNOUSSI H. Distillation-based Fabric Anomaly Detection［J］. Textile Research Journal， 2024， 94（5/6）： 552-565.
[22]	ZHANG Hongwei， LIU Shuaibo， LU Shuai， et al. Knowledge Distillation for Unsupervised Defect Detection of Yarn-dyed Fabric Using the System DAERD： Dual Attention Embedded Reconstruction Distillation［J］. Coloration Technology， 2024， 140（1）： 125-143.
[23]	BERGMANN P， BATZNER K， FAUSER M， et al. The MVTec Anomaly Detection Dataset： a Comprehensive Real-world Dataset for Unsupervised Anomaly Detection［J］. International Journal of Computer Vision， 2021， 129（4）： 1038-1059.
[24]	LEE S， LEE S， SONG B C. CFA： Coupled-Hypersphere-based Feature Adaptation for Target-oriented Anomaly Localization［J］. IEEE Access， 2022， 10： 78446-78454.
[25]	LIU Zhikang， ZHOU Yiming， XU Yuansheng， et al. SimpleNet： a Simple Network for Image Anomaly Detection and Localization［C］∥2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Vancouver， 2023： 20402-20411.
[26]	ZHANG Xuan， LI Shiyu， LI Xi， et al. DeSTSeg： Segmentation Guided Denoising Student-teacher for Anomaly Detection［C］∥2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Vancouver， 2023： 3914-3923.
[27]	TAN Pengjie， WONG W K. Unsupervised Anomaly Detection and Localization with One Model for all Category［J］. Knowledge-Based Systems， 2024， 289： 111533.
[28]	BATZNER K， HECKLER L， KÖNIG R. EfficientAD： Accurate Visual Anomaly Detection at Millisecond-level Latencies［C］∥2024 IEEE/CVF Winter Conference on Applications of Computer Vision （WACV）. Waikoloa， 2024： 127-137.
[29]	DU Xiao， LI Bing， LIU Tongkun， et al. Efficient Visual Anomaly Detection Model with Adaptive Wavelet Transform［J］. Optics and Lasers in Engineering， 2024， 182： 108457.
[30]	YANG Lingxiao， ZHANG Ruyuan， LI Lida， et al. SimAM： a Simple， Parameter-free Attention Module for Convolutional Neural Networks［C］∥International Conference on Machine Learning. Vienna，2021：11863-11874.
[31]	OUYANG Daliang， HE Su， ZHANG Guozhong， et al. Efficient Multi-scale Attention Module with Cross-spatial Learning［C］∥2023 IEEE International Conference on Acoustics， Speech and Signal Processing （ICASSP）. Rhodes， 2023： 1-5.
[32]	SHI Pengfei， HE Qi， ZHU Sisi， et al. Multi-scale Fusion and Efficient Feature Extraction for Enhanced Sonar Image Object Detection［J］. Expert Systems with Applications， 2024， 256： 124958.

参数	数值
中心光束平均激光功率P_A/W	1000
中心光束调制频率/Hz	200，400，600，800，1000
中心光束调制振幅/W	100，300，500，700，900
环光束激光功率/W	1000
焊接速度/（mm·s^-1）	30
离焦量/mm	0
上表面保护气体流量/（l·min^-1）	20
背保护气体流量/（l·min^-1）	15

参数	数值
中心光束平均激光功率P_A/W	1000
中心光束调制频率/Hz	200，400，600，800，1000
中心光束调制振幅/W	100，300，500，700，900
环光束激光功率/W	1000
焊接速度/（mm·s^-1）	30
离焦量/mm	0
上表面保护气体流量/（l·min^-1）	20
背保护气体流量/（l·min^-1）	15

网络层	轻量化操作	输入图像尺寸
1	3→1	1×H×W
2	128→64	64×H/2×W/2
3	256→128	128×H/4×W/4
4	512→256	256×H/8×W/8
5	1024→512	512×H/16×W/16

网络层	轻量化操作	输入图像尺寸
1	3→1	1×H×W
2	128→64	64×H/2×W/2
3	256→128	128×H/4×W/4
4	512→256	256×H/8×W/8
5	1024→512	512×H/16×W/16

	镁合金数据集	Mvtec AD数据集
学习率	0.0001	0.001
批大小	2	8
学习率调整（衰减率）	0.2
迭代次数	800
优化器	Adam
损失函数	MSE+SSIM
图像大小（通道数×长×宽）	1×256×256
激活函数	ReLu

Defects Detection for Laser Welding Seams of Magnesium Alloy Based on Improved Reconstruction Network

基于改进重建网络的镁合金激光焊缝缺陷检测

RichHTML

PDF

Knowledge

Abstract

Cite this article

share this article

Figures/Tables 24

References 32

Related Articles 15

Recommended Articles

Metrics

真实标签	预测结果
真实标签	有缺陷	无缺陷
有缺陷	真阳性	假阴性
无缺陷	假阳性	真阴性

方法	模型	DET-AUROC	SEG-AUROC	R_PRO
基于特征方法	CFA	81.1	88.4	74.2
	Simple-Net	80.5	86.7	55.1
	DFC	67.6	89.1	51.6
	DeSTSeg	60.7	78.8	54.9
	Ms-Flow	75.3	89.9	65.4
基于重建方法	AE-MSE	41.0	53.0	9.0
	AE-SSIM	43.0	74.0	24.0
	DRAEM	62.7	83.1	40.9
	FAIR	86.0	79.6	40.2
	本文方法	94.9	84.5	58.1

	DET-AUROC	SEG-AUROC	R_PRO
飞溅	97.0	96.2	82.1
焊瘤	92.2	84.9	61.8
塌陷	98.7	93.0	77.6
咬边	91.7	85.5	58.3

类别		DET-AUROC				SEG-AUROC				R_PRO
类别		CFA	Simple-Net	FAIR	本文方法	CFA	Simple-Net	FAIR	本文方法	CFA	Simple-Net	FAIR	本文方法
纹理类	地毯	97.6	99.6	100	98.0	98.7	98.6	99.4	99.1	94.9	94.5	97.6	95.6
	网格	97.9	99.7	98.3	96.6	98.4	98.1	98.8	99.0	94.6	92.4	95.6	96.2
	皮革	98.9	100	99.7	99.3	98.7	99.0	99.6	99.6	95.6	96.2	98.8	98.8
	瓷砖	97.8	100	100	100	98.1	97.0	98.3	98.4	93.6	92.0	95.1	95.0
	木材	98.0	100	98.3	93.0	97.9	92.0	94.9	90.3	93.4	76.7	90.3	79.4
	均值	98.0	99.7	99.3	97.4	98.4	97.0	98.2	97.3	94.4	90.4	95.5	93.0
对象类	瓶口	100	100	99.2	99.7	98.5	98.2	98.1	98.3	95.2	93.2	93.6	94.3
	电缆	99.8	98.7	97.9	93.4	98.6	97.7	96.6	95.9	94.7	89.8	87.8	83.4
	胶囊	99.0	99.6	88.5	92.8	98.7	99.1	88.8	93.5	94.9	93.3	76.4	79.6
	榛子	98.6	100	99.6	99.4	98.5	98.5	99.1	99.4	95.0	88.7	96.1	95.4
	齿轮	99.1	100	91.1	92.9	98.6	98.7	93.2	96.7	95.2	89.4	74.2	87.7
	药片	98.9	97.8	96.1	97.6	98.6	98.5	98.4	98.8	95.5	92.9	94.9	95.6
	螺钉	97.5	93.1	82.4	94.9	98.5	98.7	98.3	98.5	94.9	93.2	92.1	89.0
	牙刷	97.8	96.1	83.9	92.5	98.2	96.9	95.8	98.9	93.5	84.6	80.4	93.6
	晶体管	97.9	100	99.3	97.8	98.2	96.3	93.5	91.1	93.6	88.1	97.3	83.5
	拉链	98.1	99.9	97.8	99.4	98.0	98.5	98.8	99.1	93.6	94.9	96.1	97.1
	均值	98.7	98.5	91.6	94.8	98.4	98.1	96.1	94.8	94.6	90.8	87.9	90.0

ASPP	SSPCAB	CBAM	MSFFM	DET-AUROC	SEG-AUROC	R_PRO
				100	51.8	39.5
√				100	53.5	43.3
√	√			98.6	60.0	42.2
√	√	√		95.4	78.8	48.6
√	√	√	√	94.9	84.5	58.1

SimAM	ECA	ELA	CBAM	DET-AUROC	SEG-AUROC	R_PRO
√				83.6	81.5	52.6
	√			69.2	70.5	45.4
		√		91.5	63.5	35.8
			√	94.9	84.5	58.1

	FAIR	DRAEM	本文方法	CFA	Simple-Net
损失	0.004	0.249	0.126	0.256	0.027
参数量/10⁶	69.0	98.2	23.3	11.2	68.9
浮点计算量/10⁹	159.9	198.2	69.3	1.8	11.4
训练时长/s	2895	4063	2684	646	1541
推理时长/s	7	9	6	646	1541

[1]	YANG Yongjun, YANG Renyou, LI Yufeng, QIN Hao. Underwater Target 3D Reconstruction Method Based on AUV Equipped with Monocular-Line Laser [J]. China Mechanical Engineering, 2026, 37(5): 1210-1217.
[2]	MEI Qizhen, SUN Xuan. A Method of Multi-type Job Identification Based on Improved Faster-RCNN [J]. China Mechanical Engineering, 2026, 37(5): 1226-1235.
[3]	HU Xinyu, ZHANG Junwei, AI Jia, LI Yunling, YAN Shuang. A Small-sample Defect Detection Method for Screen-printed Characters with Adaptive Target Region Extraction [J]. China Mechanical Engineering, 2025, 36(12): 2968-2977.
[4]	Hongchen WU, Xiaorong CHEN, Baiyang LI. Multi-threaded Semi-global Stereo Matching Method for Internal Inspection of Workpieces [J]. China Mechanical Engineering, 2025, 36(11): 2783-2791.
[5]	Siqian FENG, Jiaxu WANG, Xin ZHANG, Xinyue HUANG. Rolling Bearing Fault Diagnosis Using a New Resonance Demodulation Method [J]. China Mechanical Engineering, 2025, 36(09): 2022-2031.
[6]	Jun WANG, Guibing GAO. A Method for Detecting Surface Defects on Wind Turbine Blades Based on Improved YOLOv5s [J]. China Mechanical Engineering, 2025, 36(09): 2108-2116.
[7]	LUO Liang1, 2, LANG Xiao1, ZU Guoqing2, ZHANG Nong1, YANG Lin3, SHEN Xiongwei4. A Cylinder Liner Defect Detection Method Based on Improved YOLOv8n [J]. China Mechanical Engineering, 2025, 36(05): 1054-1064.
[8]	ZHU Chuanjun1, LIANG Zeqi1, FU Qiang1, ZHANG Chaoyong2. Label Assignment Algorithm for Metal Defect Samples Based on Distance-awareness [J]. China Mechanical Engineering, 2024, 35(09): 1634-1641.
[9]	LIU Mandong, PENG Zhenrui. Truss Damage Identification Based on Joint Multiple Reconstructions Autoencoders [J]. China Mechanical Engineering, 2024, 35(05): 840-850.
[10]	MEI Jie, QIN Jiarui, CHEN Dingfang, CHEN Kun, . SLAM-based Underwater Image Enhanced Visual 3D Reconstruction Method [J]. China Mechanical Engineering, 2024, 35(02): 268-279.
[11]	XU Peng, LIU Bailin, CHEN Yaxiong. Rail Crack Detection Method Based on Differential Flexible Eddy Current Probe [J]. China Mechanical Engineering, 2023, 34(20): 2419-2427.
[12]	DU Wenjie, ZHENG Huifeng, ZHANG Kaisheng, TANG Jiaxuan. Research on Far-focused Pixel-based Imaging Method of Defects inside Multilayer Media [J]. China Mechanical Engineering, 2023, 34(09): 1045-1051.
[13]	SHU Linsen, GONG Jiangtao, DONG Yue, SU Chengming, WANG Xin. Research on Laser Remanufacturing Method and Experiments of Valve Core Parts [J]. China Mechanical Engineering, 2023, 34(04): 446-453.
[14]	MENG Kuo, SUN Guangkai, ZHOU Kangpeng, HE Yanlin, ZHU Lianqing. Three-dimensional Deformed Multi-core Optical Fiber Reconstruction Technology for Skins of a Near-space AirshipLI Mohan [J]. China Mechanical Engineering, 2023, 34(03): 324-331.
[15]	LIU Jun, RUAN Xiaodong, YANG Pengliang, WU Di, ZHENG Minyi. Control of Foot Landing Flexibility in Metamorphic Robot Reconstruction and Steering [J]. China Mechanical Engineering, 2022, 33(24): 2917-2926.