748 M A T E R I A L S E V A L U A T I O N • J U L Y 2 0 2 1 in Figure 10. Surface profile analysis indicated clear periodic band structure was achieved in the FSW repairs without any obvious defects observed. After determining the prototype repair condition, three FSW repairs were performed on premachined cracks with widths from 0.2 mm up to 0.8 mm on the A108 steel plate. A W-Re tool with a pin length of 2.3 mm and a shoulder diameter of 10 mm was employed. Based on the weld appear- ance as shown in Figure 11, optimal repair parameters were determined: 500 rpm rotational speed, 80 mm/min travel speed, 20 mm/min plunge speed, 3 s dwell time, and 45 mm travel distance. It is noteworthy that the plunge depth was adjusted to accommodate the void volume of the cracks: the repair on the 0.2 mm wide crack demonstrates the detrimental effect of insufficient plunge depth the other two cases demonstrate satisfactory repairs with proper plunge depths. In terms of the loads such as vertical force required in the FSW process, the vertical force was found to be up to ~7.5 kN when using the pinless tool, and ~12 kN when using the tool with a pin length of 2.3 mm. The peak vertical force appeared during the plunge stage in all instances. In order to ensure the flexibility and mobility of the robot platform, it is necessary to reduce the demand for the vertical force to reduce the total weight. Therefore, an induction heating unit was added to the FSW system as a preheating tool, as shown in Figure 12. It could assist in reducing the vertical force by softening the materials during FSW and thus enabling a compact, live-repair robot. Meanwhile, the plunge speed was varied from 20, 10, 5, 3, 1, to 0.5 mm/min to study its effect on reducing the peak vertical force during the plunge stage. Figure 13 summarizes and compares the peak vertical force during the plunge stage as a function of plunge speed under 400 rpm and 500 rpm. It is obvious that the peak vertical force decreases as the plunge speed decreases under both conditions. Accordingly, the lowest value of peak vertical force both occurs at the plunge speed of 0.5 mm/min, which ME TECHNICAL PAPER w ai-enabled robotic nde for structural damage 500 μm 10 mm 500 μm Figure 10. Surface morphology of the friction stir welding (FSW) specimen. Crack length: 60 mm Crack length: 60 mm Crack length: 60 mm Crack width: 0.8 mm 10 mm Crack width: 0.4 mm Crack width: 0.2 mm 15 mm 10 mm 10 mm Figure 11. Surface appearance of friction stir welding (FSW) repairs for the cracks with various width. Magnetic composite Coil 48 mm 17 mm 20 mm FSW tool Alumina Figure 12. The setup for the induction-heating assisted FSW system: (a) bottom view (b) side view. (a) (b)

J U L Y 2 0 2 1 • M A T E R I A L S E V A L U A T I O N 749 is about 1.5 kN and 2 kN lower than the case of 20 mm/min, with and without induction preheating, respectively. Further- more, in Figure 13, comparing the peak vertical force at a certain plunge speed, it is observed that the reduction rate of the peak vertical force slows down with the decrease of plunge speed. This might be due to a more significant temper- ature drop caused by a longer plunge time. But the lowest peak vertical force is still obtained at the plunge speed of 0.5 mm/min. Maximum reduction in the peak vertical force occurs at the plunge speed of 20 mm/min in all cases. Take the 500 rpm case for example: the additional preheating assists to reduce the plunge force by 0.64 kN with a reduction rate of about 12% at 0.5 mm/min, and by 1.46 kN and 20% at 20 mm/min. It can be concluded that the peak vertical force can be efficiently reduced by optimizing the plunge speed and induction heating parameters, and a smaller peak vertical force could be obtained if continuing to optimize the induc- tion parameters, such as increasing the induction heating power or induction heating time. Based on the results mentioned previously, the induction- heating assisted FSW technique would potentially be the best candidate for the repair of cracks with improved mobility and longevity of repairs. Discussion and Conclusion Although this novel approach for autonomous robotic NDE and repair is proposed and demonstrated using the prelimi- nary results, there are still many technical gaps existing with further development desired to make this CPS system feasible in the field. The first improvement requires the need for a large-scale data set of representative damages for power plant boiler inspection, which requires an extremely large number of NDE data sets for training. Relative to what is possible for this report, obtaining this large number of data sets is not feasible, nor is the number of data sets to achieve a properly trained system known. Another drawback of this current technique is the computational complexity of the R-CNN method. This method requires significant overhead during training due to the need to run an individual CNN over various regions of the same image. There has been recent progress with methods such as Fast R-CNN and Faster R-CNN. These methods only require a single CNN per image, which would hasten the training of the model. Another consideration is to automatically determine if certain damages are more critical to system integrity. This would allow for the system to prioritize the repair of certain components and therefore optimize maintenance strategies. Methods for false positive and false negative errors for damage analysis are also considerations. Having the robot query for online feedback from a human expert should help in addressing repair prioritization and false determinations of damages however, the need for human feedback should be minimized as the model should learn and build upon itself to be fully automated. Another major consid- eration is that many power plant boiler house walls consist of an array of water tubes in addition to simple uniform steel surfaces. These surfaces present a challenge given the nonuni- form distance of the NDE coil array, which results in distorted sensor readings that cannot be accurately evaluated by the AI model. The last concern is the possibility of a scan along a crack that occurs parallel to the robot’s direction of travel. In this case, the sensor would detect a minimal change in unifor- mity, which may lead to damage being left unclassified. This problem may be addressed by performing multiple scans along the same surface but in perpendicular directions. The FSW process preliminarily repaired the cracks on the flat steel plate surface, and the weld without any surface defects was successfully obtained by using proper FSW parameters. This work has a certain guiding significance, which shows the feasibility of FSW for crack repair. However, in order to apply the robotic repair technology to the crack repair on the real boiler surface, some problems still need to 4.0 0 2 4 6 8 10 Plunge speed (mm/min) 12 14 16 18 20 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 500 rpm, without induction heating 500 rpm, with induction heating 4.0 0 2 4 6 8 10 Plunge speed (mm/min) 12 14 16 18 20 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 400 rpm, without induction heating 400 rpm, with induction heating 20% force reduction Figure 13. Peak vertical force during plunge stage as a function of plunge speed under the conditions of: (a) 400 rpm and (b) 500 rpm. (a) (b) Maximum vertical force (kN) Maximum vertical force (kN)