Chapter 5. Deep-Learning-Based Defects Detection Model in Keyhole-mode Laser Welding of
5.3 Results and Discussion
In Figure 5.4, the OM images of cross-sectional weld bead cutting in the direction shown in Figure 5.2 (b) were presented to identify the continuous variations of penetration depth and these images were magnified to 1.5 times in height direction for emphasizing the variation of penetration depth. As shown, the penetration depth was fluctuated dramatically during the welding process and it revealed that the keyhole-mode welding in thin foil was extremely unstable process. For this reason, it was essential to measure the continuous variations in penetration depth in instability analysis of keyhole-mode thin foil welding. The continuous cross-section of the weld bead was the most accurate method for analyzing unstable welding process, but there was a limit to measure the entire welding processes due to the thin bead width. To overcome this limitation, Figure 5.5 shows the cross-sectional images cut by the direction in Figure 5.2 (a) at 10 mm intervals, presenting the variations in penetration depth welded by a laser power of 450 W, a scanning speed of 9 m/min, and a shielding gas flow rate of 10 l/min. The penetration depths according to the bead positions, which measured from the images in Figure 5.5, were plotted in Figure 5.6, and the average and standard deviation of penetration depth were 360.6 μm and 79.9 μm, respectively. Through the severe fluctuation of penetration depth shown in Figure 5.5 and Figure 5.6, it indicated the instability of keyhole-mode foil welding. To find optimized process condition, it is important to study the parameters (flow rates of shielding gas and experimental conditions) that affect the stability of keyhole-mode thin foil welding.
Figure 5.4. Continuous cross-sectional images at the bead positions of (a) 10 ~ 20 mm and (b) 65 ~ 75 mm under laser power of 450 W, scanning speed of 9 m/min, and shielding gas flow rate of 10 l/min. (The distance between each red and blue line is 100 μm and black line is the boundary of fused zone and base metal.)
Figure 5.5. Cross-sectional weld images at the bead positions of 10 mm intervals under laser power of 450 W, scanning speed of 9 m/min, shielding gas flow rate of 10 l/min.
Figure 5.6. The fluctuations of penetration depth according to the bead positions. The penetration depths were measured from the images shown in Figure 5.5 and the average penetration depth was marked with red dashed line.
To study the effect of shielding gas, the variations in the penetration depth corresponding to the five different flow rates of shieling gas were measured under three experimental conditions (laser power:
350, 400, 450 W and scanning speed: 13, 11, 9 m/min, respectively). Under the maximum experimental condition (laser power: 450 W and scanning speed: 9 m/min) used in this study, Figure 5.7 presents the average, min-max, and standard deviation of penetration depths according to the flow rates of shielding gas (0, 2.5, 5, 7.5 and 10 l/min). At the above conditions, joining was successful regardless of the flow rate of shielding gas. The averages of penetration depth were quite similar for all flow rate conditions, however, the min-max and standard deviation of penetration depth tended to increase when the flow rate of shieling gas was increased. This means that the welding process became more unstable as the flow rate of shielding gas increased. In fact, the purpose of shielding gas is for preventing oxidation on the bead during the welding process, but the excessive shielding gas causes greater fluctuations in the keyholes and melt pools. As shown in Figure 5.8, the dynamics of unstable melt pool, which were welded by a laser power of 450 W, a scanning speed of 9 m/min, and a shielding gas flow rate of 7.5 l/min, were observed using a high-speed imaging, and the recorded images were extracted at 35 frame intervals. Figure 5.8 (a) presents the keyhole and the melt pool in stable keyhole-mode welding and the melt pool was blown out by the shielding gas shown in Figure 5.8 (b). After eliminating the melt pool, the mode of welding process was changed into conduction mode welding in Figure 5.8 (c). The transition regimes could be seen in Figure 5.8 (d) and (e), and then the stable keyhole mode welding was performed again illustrated in Figure 5.8 (f). As the flow rate of shielding gas increased, these patterns occurred more frequently and it leaded to increase the variations of penetration depth during the welding process.
Figure 5.7. Average, min-max, and standard deviation of penetration depths vs flow rates of shielding gas (laser power: 450 W, scanning speed: 9 m/min, beam diameter: 100 μm)
Figure 5.8. Fluctuations of melt pool captured at 35 frames interval (Laser power: 450 W, Scanning speed: 9 m/min, Flow rate of shielding gas: 7.5 l/min)
As shown in Figure 5.9, the averages and standard deviations of penetration depth according to the five different flow rates of shielding gas (0, 2.5, 5, 7.5 and 10 l/min) were plotted for three experimental conditions (laser power: 350, 400, 450 W and scanning speed: 13, 11, 9 m/min, respectively). Note that the joining was completely failed under the lowest energy input (laser power:
350 W and scanning speed: 13 m/min) and the shielding gas flow rates of 7.5 and 10 l/min. Furthermore, the circles presented in Figure 5.9 mean the partially bonded conditions. In the highest energy input (laser power: 450 W and scanning speed: 9 m/min), the average penetration depth was slightly decreased under the shielding gas flow rate of 2.5 l/min due to slight oxidation prevention. It leaded to decrease the laser absorptance slightly. When the flow rate of shielding gas increased, the laser-induced plasma blew away by the shielding gas and the laser beam was absorbed without any disturbance. That was why the average penetration depth was increased under the high flow rates. Moreover, the standard deviation of penetration was dramatically increased after shielding gas flow rate of 5/min. The excessive shielding gas highly effected on the fluctuations of melt pool and penetration depth. On the other hand, welding with lower energy input (400 W, 11 m/min and 350 W, 13 m/min) had different trend compared with higher energy input. As the flow rate of shielding gas was increased, the average penetration depth was tended to decrease. Notably, welding process became extremely unstable under the condition, which was over the shielding gas flow rate of 5 l/min, resulting in either fail to weld or partial joining.
Under the lower laser energy input, it was obvious that the melt pool became smaller and was difficult to sustain. For this reason, the melt pool was easily blown out by the high flow rate of shielding gas, which presents in Figure 5.8 (b). Without and with small amount of shielding gas flow rate, the standard deviations of penetration depth were comparatively small, and the complete joining was succeeded under all three experimental conditions.
Figure 5.9. (a) Averages and (b) standard deviations of penetration depth according to the flow rates of shielding gas and the three experimental conditions shown in box at the top left of figure. (Partial joining conditions were marked by circles.)
To measure the oxidation on the welded bead, the EDS analysis was conducted, which can measure the chemical compositions on the surface. In Figure 5.10, the examples of EDS analysis result, which measured the chemical compositions on the bead welded by a laser power of 400 W and a scanning speed of 11 m/min without shielding gas (shown in Figure 5.10 (a)) and with a shielding gas flow rate of 5 l/min (presented in Figure 5.10 (b)). Figure 5.11 shows the variations of oxygen component ratio on the bead according to the flow rate of shielding gas. The stainless foil itself had an oxygen component of 1.12% and was expressed in black dotted line shown in Figure 5.11. Note that the EDS analysis was conducted only at the shielding gas flow rate of 5 l/min in two experimental conditions (laser power:
350, 450 W and scanning speed: 13, 9 m/min, respectively). For three experimental conditions, the oxygen ratio at the flow rate of 5 l/min was similar as a bare foil surface, which means that the prevention of oxidation on the bead was success. When the flow rate of shielding gas was exceeded 5 l/min, excessive shielding gas leaded to increase the instability of welding. However, the welding processes were already unstable at the shielding gas flow rate of 5 l/min, as shown in Figure 5.9.
Furthermore, at the shielding gas flow rate of 2.5 l/min, the degree of oxidation on the bead was reduced, but not completely shielded. As a result, it was determined that the optimized condition was not to use shielding gas with high average penetration depth and low standard deviation penetration depth. Even if oxidation on the bead could not be prevented, this condition was the most stable process in keyhole mode welding.
Figure 5.10. EDS analysis on the beads welded by a laser power of 400 W and a scanning speed of 11 m/min (a) without shielding gas and (b) with a shielding gas flow rate of 5 l/min.
Figure 5.11. Ratio of oxygen on the beads vs flow rates of shielding gas.
To find the optimized experimental condition in keyhole-mode foil welding, nine experimental conditions were conducted without shielding gas. For nine experimental conditions, Figure 5.12 presents the images of cross-sectional weld bead in the center of entire bead. The measured penetration depth shown in Figure 5.12 versus Ioti0.5 and ultimate load per width versus penetration depth were plotted in Figure 5.13 (a) and (b), respectively. As shown, the penetration depth beame deeper as the Ioti0.5 (which is propotional to surface temperature after laser irrdiation ) increased. Except for minimum penetration depth, the values of ultimate load per width were practically the same. As the penetration depth deepened, this value slightly increased and decreased after the maximum value. In keyhole mode welding, the maximum ultimate load per width was 82.6 N/mm under the optimized condition (laser power: 400 W, scanning speed: 13 m/min) without employing shielding gas and the penetration depth at this moment was 223.1 μm.
Figure 5.12. OM images of cross-sectional weld bead welded by nine experimental conditions and without shielding gas.
Figure 5.13. (a) Penetration depth vs Ioti0.5 and (b) ultimate load per width vs penetration depth. (The penetration depths were measured from the cross-sectional weld bead images in Figure 5.12)
Figure 5.14 (a) shows the training loss (blue) and the validation mAP (red) were plotted according to the epochs. Note that the mAP for validation was calculated per 20 epochs. The maximum mAP for validation set was 91.42% at 1340 epochs. In Figure 5.14 (b), the recall and precision at 1340 epochs were plotted according to the class probability threshold. Among the detected boxes, if the probabilities of detected objects exceed the class probability threshold, the detected objects are determined to be correct. As the probability threshold increased, the recall decreased while the precision increased as shown in Figure 5.14 (b). The optimized class probability threshold was selected at the maximum sum of precison and recall, and was 30%. Thus, the performance of trained defect detection model was evaluated by the test sets at probability threshold of 30%.
Figure 5.14. (a) Training loss (blue) and validation mAP (red) curves for defects detection model according to the epochs. (b) Recall, precision, and precision + recall curves according to the class probability threshold.
Figure 5.15 shows three cases of test results and presents success, partial success, and failure, respectively. Red bounding box indicates successful detection of defects, and blue box means failure of detecting defects. In the case of partial success, there were three defects, but only two defects were detected as shown. However, even in this case, it can be determined to be a success because it detected the defects. Hence, the accuracy of detection was calculated as follow:
100 (1 # )
of failures Accuracy
of test sets
= × − (5.1)
Figure 5.16 presents test results for each defect and stable bead case, with 30 failures for a total of 2000 test sets. Except for case 5 and 6, where the number of data is small, the accuracy was over 96% in all cases. For case 5 and 6, accuracy is expected to increase as the number of data increases, however the data belonging to case 5 and 6 were difficult to obtain. In conclusion, the trained defect detection model had a accuracy of 98.5% and the detection speed of 73 fps. The developed model was possible to detect
defects in real time with high accuracy. Training by the higher quality and quantity of data will enable to implement the defect detection model with higher accuracy. The accuracy of deep learning models highly depends on the training sets.
Figure 5.15. Examples of test results: (a) success, (b) partial success, and (c) failure. (Red box:
success to detect and blue box: fail to detect)
Figure 5.16. Test results for each defect and stable bead case. The number of success (blue), partial success (orange), and failure (red) cases are presented, and the accuracy of detection for each case is marked as black dots. The total accuracy of this model is expressed as red dash lines. Defect case 1-6 indicate the defects in Figure 5.3 (a)-(f).
In this chapter, a 100 μm- thick stainless-steel foil overlapped on a 0.5 mm-thick stainless-steel sheet was welded by a 2 kW multi-mode continuous fiber laser with a focused beam diameter of 100 μm. In thin foil welding, keyhole-mode welding was forcibly demanded when the penetration depth was greater than beam size as show in this study. However, the variation in penetration depth during the welding process was extremely fluctuated. Furthermore, as the flow rates of shielding increased, the change in penetration depth according to the bead positions became more severe and the joining failed under lower laser energy input conditions. Due to the tiny melt pool and keyhole in foil welding, the melt pool was frequently blown out during the welding process because of excessive flow rate of shielding gas. It leaded to increase the instability of keyhole mode welding. Through the EDS analysis results, the oxidation on the bead was completely prevented after the shielding gas flow rate of 5 l/min.
However, the joining was still partially, and the penetration depth fluctuated considerably. As a result, keyhole-mode welding without shielding gas was the most stable process in thin foil welding. At the optimized experimental condition (laser power: 400 W and scanning speed: 13 m/min), the maximum ultimate load per width was 82.6 N/mm, with a penetration depth of 223.1 μm. In addition, a YOLOv4- based defects detection model with a test accuracy of 98.5% and a detection speed of 73 fps was developed.