Author Affiliations
1The Institute of Technological Sciences, Wuhan University, Wuhan 430072, Hubei , China2Shanghai Spaceflight Precision Machinery Institute, Shanghai 201600, China3Hubei Institute of Measurement and Testing Technology, Wuhan 430223, Hubei , Chinashow less
Fig. 1. Overall workflow chart of defect detection
Fig. 2. Experiment system for laser welding defect detection
Fig. 3. Complete sample images of typical welds. (a) Sheet butt weld;(b) thick plate butt weld; (c) bead-on-plate weld
Fig. 4. Different forms of data during data preprocessing. (a) Preprocessing of point cloud HDM data; (b) RGB images of surface defects; (c) high-density point cloud data; (d) depth images including 3D profile information of defects
Fig. 5. Structural diagram of Faster R-CNN
Fig. 6. Detection results using Faster R-CNNs based on ResNet18, ResNet50, and ResNet101
Fig. 7. Statistical results of three models. (a) Loss evolution of different models; (b) point cloud detection precisions and recall rates with different models; (c) detection precisions and recall rates of defects for point clouds and RGB images with different models; (d) detection mAPs of defects for point clouds and RGB images with different models; (e) testing time of different models
Fig. 8. Typical false negative test results of Faster R-CNN model based on ResNet50
Fig. 9. Measurement process of defect sizes. (a) RGB images; (b) point clouds; (c) depth gray images; (d) threshold segmentation; (e) locating defect areas; (f) defect feature size measurement
Fig. 10. Relative errors of defect measurement results
Parameter | Value |
---|
Size of sensor /(pixel×pixel) | (2×103)×(2×103) | Repeatability in Z direction /μm | 3.3 | Resolution of image on XY plane /mm | 0.06‒0.09 | Field of view /(mm×mm) | 71×90 ‒100×154 | Clearance distance /mm | 164 | Measuring range /mm | 110 |
|
Table 1. Specific parameters of binocular structured light sensor
Parameter | Range | Value in online test |
---|
Sagging | Undercut |
---|
Laser power /kW | 1‒6 | 2 | 4 | Welding speed /(m/min) | 0.5‒10.0 | 2 | 5 | Defocus /mm | -5‒5 | -2 | -2 | Shielding gas flow rate /(L/min) | 1‒20 | 15 | 15 |
|
Table 2. Welding parameters
Image | Dataset | Undercut | Sagging | Defect-free | Total number | Data augmentation |
---|
Point cloud image | Total | 170 | 150 | 130 | 450 | Noise addition Mirroring | Training | 119 | 105 | 91 | 315 | Testing | 51 | 45 | 39 | 135 | RBG image | Total | 84 | 80 | 36 | 200 | Noise addition Mirroring | Training | 58 | 56 | 26 | 140 | Testing | 26 | 24 | 10 | 60 |
|
Table 3. Datasets after data augmentation
Model | Overall precision /% | Overall recall rate /% | mAP /% | Run time /s |
---|
ResNet18 | 79 | 77.0 | 72.4 | 0.191 | ResNet50 | 93 | 89.5 | 91.9 | 0.194 | ResNet101 | 73 | 66.5 | 62.4 | 0.253 |
|
Table 4. Performance parameters of three models based on point cloud analysis
No. | Defect type | Maximum depth /mm | Area /mm2 | Width /mm | Length /mm | Actual width /mm | Actual length /mm |
---|
1 | Sagging | 1.514 | 7.53 | 3.00 | 3.55 | 2.96 | 3.52 | 2 | 2.030 | 8.67 | 2.20 | 5.85 | 2.17 | 5.83 | 3 | 1.810 | 9.68 | 3.35 | 4.00 | 3.39 | 3.94 | 4 | Undercut | 0.438 | 18.97 | 1.45 | 30.55 | 1.47 | 30.64 | 5 | 0.565 | 33.66 | 1.20 | 56.25 | 1.18 | 56.34 | 6 | 0.473 | 23.86 | 0.95 | 39.25 | 0.92 | 39.32 |
|
Table 5. Measurement results of defect sizes