Rotor extraction is an important work to be performed in the periodic overhaul of nuclear power units. However

excessive overhaul can lead to the damage risk of both stator and rotor and increase of operating cost. It is found by the research that the feasibility of stator and rotor defect detection in case of not extracting the rotor of the generator is of significance in reducing the maintenance costs. The tracked robot is used to detect the images collected by four cameras in real-time through the edge computing equipment. Based on the YOLOv7 original network

the effects of various optimization methods on the network’s generalization ability are investigated by using Mosaic-9 data augmentation and optimizing the Wise-IoU loss function. The results show that under the condition of constant recognition speed

the optimized YOLOv7 has superior performance metrics

faster detection speed

and fewer model parameters. Consequently

it is more suitable for deployment on the edge module. Furthermore

multiple onsite tests also show that the robot system is capable of detecting faults in the stator and rotor without the need for rotor extraction.