Surface Defect Recognition for Laser Welding of Magnesium Alloys Based on Active Semi-Supervised Learning

Kun Tang; Zhao Huang; Yongjian Zhu; Hang Zhang; Gang Zeng; Hongchao Xiao; Xiaojie Zhou; Weidong Tang; Mingjun Zhang; Cong Mao

doi:10.3788/CJL241344

Chinese Journal of Lasers, Volume. 52, Issue 8, 0802103(2025)

Surface Defect Recognition for Laser Welding of Magnesium Alloys Based on Active Semi-Supervised Learning

Kun Tang¹, Zhao Huang¹, Yongjian Zhu^2、*, Hang Zhang¹, Gang Zeng³, Hongchao Xiao⁴, Xiaojie Zhou¹, Weidong Tang¹, Mingjun Zhang¹, and Cong Mao¹

Author Affiliations

¹Hunan Provincial Key Laboratory of Intelligent Manufacturing Technology for High-Performance Mechanical Equipment, Changsha University of Science and Technology, Changsha 410114, Hunan , China

²College of Engineering Physics, Shenzhen Technology University, Shenzhen 518118, Guangdong , China

³Hunan Provincial Technology Innovation Center of Aerospace New Light Alloy Materials, Changsha 410205, Hunan , China

⁴Hunan Provincial Engineering Research Center of Wrought Magnesium Alloys and Surface Protection, Changsha 410205, Hunan , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(24)

Fig. 1. The framework of active semi-supervised learning

Download full size

Fig. 2. Schematic of adaptive dynamic threshold adjustment module (ADTAM)

Download full size

Fig. 3. Schematic of sample screening mechanism (SSM)

Download full size

Fig. 4. Experimental setup of laser welding. (a) Laser welding system; (b) welding principle

Download full size

Fig. 5. Image acquisition device. (a) Physical diagram; (b) schematic

Download full size

Fig. 6. Partial defect samples in the magnesium alloy weld dataset

Download full size

Fig. 7. Partial defect samples in the NEU-DET dataset

Download full size

Fig. 8. Network architecture of Faster R-CNN

Download full size

Fig. 9. Loss function curves. (a) Magnesium alloy weld dataset; (b) NEU-DET dataset

Download full size

Fig. 10. Effects of hyperparameter setting (magnesium alloy weld dataset). (a) Threshold ε; (b) adjustment coefficient η

Download full size

Fig. 11. Effects of hyperparameter setting (NEU-DET dataset). (a) Threshold ε; (b) adjustment coefficient η

Download full size

Fig. 12. Defect visualization (magnesium alloy weld dataset)

Download full size

Fig. 13. Defect visualization (NEU-DET dataset)

Download full size

Table 1. Chemical composition of AZ31B magnesium alloy
View table
Table 1. Chemical composition of AZ31B magnesium alloy
Element Mass fraction /%
Mg 96.040
Al 2.960
Zn 0.520
Mn 0.310
Si 0.160
Cu 0.006
Fe 0.003
Ni 0.001

Table 2. Welding parameters

View table

Table 2. Welding parameters

Parameter	Content
Average laser power of the center beam /W	1000
Modulation frequency of the center beam /Hz	200, 400, 600, 800, 1000
Modulation amplitude of the center beam /W	100, 300, 500, 700, 900
Laser power of ring beam /W	1000
Welding speed /(mm/s)	30
Defocus /mm	0
Upper surface protection gas flow /(L/min)	20
Back protection gas flow /(L/min)	15
Protection gas	Ar

Table 3. Platform parameters and hardware models

View table

Table 3. Platform parameters and hardware models

Parameter	Content
Range of motion	X-axis: 600 mm, screw Y-axis: 600 mm, belt Z-axis: 250 mm, screw
Movement speed /(mm/s)	50
Acquisition time /s	35
Collection accuracy /μm	8
Platform size /(mm×mm×mm)	910×775×1800
Industrial camera	MV-CH250-10GC
Industrial lens	MVL-KL2528M-12MP
Light source	MV-LCDS-18-18-W

Table 4. Experimental environment and hardware configuration
View table
Table 4. Experimental environment and hardware configuration
Name Content
GPU NVIDIA Geforce RTX3090
CPU AMD R9 5950X
Operating system Ubuntu 20.04
Video memory /GB 24
Compiler PyCharm
Deep learning framework PyTorch1.13.1
CUDA version 11.1

Table 5. Model parameter settings

View table

Table 5. Model parameter settings

Parameter	Magnesium alloy weld	NEU-DET
Activation function	ReLU	ReLU
Learning rate	0.01	0.01
EMA decay rate	0.9996	0.9996
Unsupervised loss weight	2	2
Iterations	30000	20000
Learning rate adjustment	Fixed	Fixed
Warm-up steps	2000	2000
Batch size	8 （4 labeled and 4 unlabeled images）	16 （8 labeled and 8 unlabeled images）
Loss function	Cross-entropy loss	Cross-entropy loss

Table 6. Overall recognition accuracy (magnesium alloy weld dataset)

View table

Table 6. Overall recognition accuracy (magnesium alloy weld dataset)

Method	AP₅₀ /%				mAP_50：95 /%
Method	Labeling ratio of 2%	Labeling ratio of 5%	Labeling ratio of 10%	Labeling ratio of 20%	Labeling ratio of 2%	Labeling ratio of 5%	Labeling ratio of 10%	Labeling ratio of 20%
Faster R-CNN	50.27	60.28	68.71	75.13	19.07	24.68	28.92	32.95
STAC	51.61	62.53	69.93	75.58	19.94	25.89	29.77	33.56
Unbiased Teacher	55.28	65.17	71.06	78.31	23.45	28.33	33.86	37.01
Soft Teacher	52.49	63.86	70.52	76.39	20.56	26.64	30.41	34.72
Active Teacher	57.39	67.89	73.81	80.74	24.41	29.27	34.59	37.78
Ours	62.43	73.79	80.65	86.32	25.52	30.84	36.68	39.34

Table 7. Overall recognition accuracy (NEU-DET dataset)

View table

Table 7. Overall recognition accuracy (NEU-DET dataset)

Method	AP₅₀ /%				mAP_50：95 /%
Method	Labeling ratio of 2%	Labeling ratio of 5%	Labeling ratio of 10%	Labeling ratio of 20%	Labeling ratio of 2%	Labeling ratio of 5%	Labeling ratio of 10%	Labeling ratio of 20%
Faster R-CNN	47.85	54.22	63.46	68.04	18.71	23.31	28.45	32.05
STAC	49.35	56.51	63.79	69.08	19.62	24.47	29.57	32.93
Unbiased Teacher	53.74	60.15	66.85	71.67	22.93	27.77	33.05	36.29
Soft Teacher	51.49	58.96	64.55	69.52	21.56	25.41	30.44	33.76
Active Teacher	56.69	62.49	68.34	73.36	23.85	28.53	33.92	37.07
Ours	60.48	66.76	73.54	78.11	24.57	29.44	35.21	37.93

Table 8. Single-class recognition accuracy (magnesium alloy weld dataset)

View table

Table 8. Single-class recognition accuracy (magnesium alloy weld dataset)

Defect class	mAP_50：95 /%
Defect class	Faster R-CNN	Unbiased Teacher	Active Teacher	Ours
Undercut	23.38	25.16	25.74	27.53
Spatter	35.06	37.58	38.27	39.81
Weld bead	34.56	39.45	40.41	41.86
Collapse	42.63	47.67	48.25	49.94
Others	29.12	35.19	36.23	37.56

Table 9. Single-class recognition accuracy (NEU-DET dataset)

View table

Table 9. Single-class recognition accuracy (NEU-DET dataset)

Defect class	mAP_50：95 /%
Defect class	Faster R-CNN	Unbiased Teacher	Active Teacher	Ours
Crazing	7.01	8.36	8.86	9.83
Inclusion	32.85	39.25	40.18	40.85
Patches	47.62	57.53	58.49	59.23
Pitted-surface	33.79	38.57	39.05	40.19
Rolled-in-scale	22.68	21.57	22.71	23.16
Scratches	48.32	52.48	53.12	54.31

Table 10. Ablation test results of key components (magnesium alloy weld dataset)
View table
Table 10. Ablation test results of key components (magnesium alloy weld dataset)
Method ADTAM SSM Sampling strategy uncertainty AP₅₀ /% mAP_50：95 /%
Baseline × × × 78.31 37.01
Ours √ × × 81.32 37.79
√ √ × 84.24 39.03
× √ √ 82.28 38.25
√ √ √ 86.32 39.34

Table 11. Ablation test results of key components (NEU-DET dataset)
View table
Table 11. Ablation test results of key components (NEU-DET dataset)
Method ADTAM SSM Sampling strategy uncertainty AP₅₀ /% mAP_50：95 /%
Baseline × × × 71.67 36.29
Ours √ × × 72.92 36.57
√ √ × 76.38 37.48
× √ √ 74.63 37.04
√ √ √ 78.11 37.93

Tools

Get Citation

Copy Citation Text

Kun Tang, Zhao Huang, Yongjian Zhu, Hang Zhang, Gang Zeng, Hongchao Xiao, Xiaojie Zhou, Weidong Tang, Mingjun Zhang, Cong Mao. Surface Defect Recognition for Laser Welding of Magnesium Alloys Based on Active Semi-Supervised Learning[J]. Chinese Journal of Lasers, 2025, 52(8): 0802103

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites