Deep Learning‐Based Recognition Method for Powder Spreading State in Additive Manufacturing

Mingyi Zuo; Hongxuan Guo; Huaixue Li

doi:10.3788/CJL241164

Chinese Journal of Lasers, Volume. 52, Issue 8, 0802301(2025)

Deep Learning‐Based Recognition Method for Powder Spreading State in Additive Manufacturing

Mingyi Zuo^1,3, Hongxuan Guo², and Huaixue Li^3、*

Author Affiliations

¹College of Software Engineering, Southeast University, Nanjing 210018, Jiangsu , China

²School of Integrated Circuits, Southeast University, Nanjing 210018, Jiangsu , China

³Aeronautical Science and Technology Key Laboratory of Additive Manufacturing, AVIC Manufacturing Technology Institute, Beijing 100192, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(13)

Fig. 1. Overall task flowchart

Download full size

Fig. 2. Image examples under different states. (a) Normal; (b) scratch; (c) uneven powder coating; (d) insufficient powder coating; (e) excessive powder coating

Download full size

Fig. 3. AlexNet training and validation curves

Download full size

Fig. 4. Training and validation metrics. (a) VGG; (b) ResNet; (c) EfficientNet

Download full size

Fig. 5. Confusion matrixes of classification results. (a) VGG-16; (b) ResNet-101; (c) EfficientNetV2-XL

Download full size

Fig. 6. Confusion matrixes of validation set. (a) VGG-16; (b) ResNet-101; (c) EfficientNetV2-XL

Download full size

Fig. 7. Heatmaps of different models. (a) Original image; (b) VGG-16; (c) ResNet-101; (d) EfficientNetV2-XL

Download full size

Table 1. Dataset categories and quantities
View table
Table 1. Dataset categories and quantities
Category Quantity Category No.
Total 1327
Normal 108 1
Scratch 301 2
Uneven powder coating 302 3
Insufficient powder coating 295 4
Excessive powder coating 321 5

Table 2. Model loss functions, optimizers, and initial learning rates
View table
Table 2. Model loss functions, optimizers, and initial learning rates
Model Criterion Optimizer Learning rate
VGG-16 Cross entropy loss Adam 0.0001
ResNet-101 Cross entropy loss Adam 0.0001
EfficientNet-XL Cross entropy loss SAM 0.0001

Table 3. Comparison of classification results of validation set

View table

Table 3. Comparison of classification results of validation set

Category No.	VGG-16			ResNet-101			EfficientNetV2-XL
Category No.	Precision	Recall	F1-score	Precision	Recall	F1-score	Precision	Recall	F1-score
1	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000
2	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000
3	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000	0.9841	0.9919
4	1.0000	0.9831	0.9915	1.0000	0.9831	0.9915	1.0000	1.0000	1.0000
5	0.9851	1.0000	0.9925	0.9851	1.0000	0.9925	0.9846	1.0000	0.9923
Accuracy	0.9963			0.9963			0.9963
Macro average	0.9970	0.9966	0.9968	0.9970	0.9966	0.9968	0.9969	0.9968	0.9968
Weighted average	0.9964	0.9963	0.9963	0.9964	0.9963	0.9963	0.9964	0.9963	0.9963

Table 4. Comparison of classification results of test set

View table

Table 4. Comparison of classification results of test set

Category No.	VGG-16			ResNet-101			EfficientNetV2-XL			Quantity
Category No.	Precision	Recall	F1-score	Precision	Recall	F1-score	Precision	Recall	F1-score	Quantity
2	1.000	0.900	0.947	1.000	0.967	0.983	1.000	0.967	0.983	30
3	0.722	1.000	0.839	0.632	0.923	0.750	0.591	1.000	0.743	13
4	1.000	0.700	0.824	1.000	0.500	0.667	1.000	0.650	0.788	20
5	0.930	1.000	0.964	0.914	1.000	0.955	1.000	0.981	0.990	53
Accuracy	0.922			0.897			0.922
Macro average	0.913	0.900	0.893	0.886	0.847	0.839	0.898	0.899	0.876
Weighted average	0.937	0.922	0.921	0.919	0.897	0.890	0.954	0.922	0.926

Table 5. Testing results with datasets collected from other devices
View table
Table 5. Testing results with datasets collected from other devices
Model Precision Recall F1-score Number of Samples
VGG-16 0.902 0.500 0.632 66
ResNet-101 0.321 0.364 0.312 66
EfficientNetV2-XL 0.785 0.561 0.653 66

Table 6. Comparison of model performance
View table
Table 6. Comparison of model performance
Model Size /MB Test set accuracy /% Inference time of test set based on Tesla T4 /s Frame rate /（frame/s）
EfficientNetV2-XL 792.7 92.2 6.84 16.96
VGG-16 512.3 92.2 1.70 68.24
ResNet-101 162.8 89.7 2.54 45.67

Tools

Get Citation

Copy Citation Text

Mingyi Zuo, Hongxuan Guo, Huaixue Li. Deep Learning‐Based Recognition Method for Powder Spreading State in Additive Manufacturing[J]. Chinese Journal of Lasers, 2025, 52(8): 0802301

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites