Automated Defects and Failure Inspection in Remanufacturing and Maintenance

Abstract

The quality of the key components of manufacturing equipment, such as dies and moulds in the manufacturing industry, is highly relative to the quality of final products. Since dies/moulds work in harsh situations, subsurface defects happen and will gradually develop into fatal surface cracks. Therefore, regular maintenance and remanufacturing are necessary for dies/moulds. Inspection is a key process in implementing maintenance and remanufacturing that can detect and prevent the components from fatal damage. As an important non-destructive testing (NDT) inspection method, ultrasonic testing (UT) can detect subsurface cracks. Currently, most industrial UTs are still carried out manually. In this study, an automated robotic UT on dies/moulds is studied to improve the efficiency and effectiveness of inspection in remanufacturing/maintenance. In contact industrial UT, the probe must contact the object surface and be normal to the surface to guarantee the amplitude of the received ultrasonic waves, optimising the inspection result. The moving speed on locations, such as edges, should also be considered to maintain smooth scanning. Therefore, the results of UT heavily depend on the expertise level of operators. In this study, the objective is to implement UT with a UT probe attached to the end-effector of a robotic arm. The object is placed at the assigned location on a desk surface. The robotic arm will approach the assigned waypoint to scan the complex surface of the object without computer vision. A reinforcement learning (RL) model is introduced to control the orientation and moving speed of the UT probe during UT scanning. Considering the safety reason and precision of operation, a collaborative robotic arm, UR5e, is used to carry out UT. Only a 6 DOF force/torque sensor measures contact force/torque between the end-effector and the surface. The measured forces are used as interactions between the robot and the environment to give feedback to the RL model so that it can make action decisions to adjust the orientation and the moving speed of the UT probe. At each waypoint, the end effector's orientation adjustment and the moving speed will be planned in real-time by the “brain”, i.e., an A2C RL model. As the “muscle”, a model-based compliance controller will be used to maintain the contact force between the end-effector of the robotic arm and the object surface constant to keep the probe contacting the object surface. A control software platform based on the robot operation system (ROS) is established to implement the whole methodology in simulation and the real world. It can be shown that the proposed method has been implemented and adapted to different objects. The probability of detection for the proposed method can reach 80%, and the trajectory's traceability is more accurate than manual UT. The limitation of the results is that it only considered the object at the assigned location, the object localisation and path planning to the object can be studied in future research.