In the experiments reported in Scientific Reports, the defect detection model reached an mAP of 88.2% and the average inspection cycle per blade, including detection, manipulation and inspection, was approximately 4 seconds. The off-the-shelf baselines trained under identical conditions scored lower: SSD 68.5%, YOLOv5 71.4%, YOLOv6 75.2% and YOLOv7 78.7%. Per-class average precision was 99.5% for dents, 82.8% for scratches, 88.8% for nicks and 79.5% for corrosion.
Aeroengine blades are critical components that need consistent monitoring in maintenance, repair and overhaul (MRO) work, and traditional manual or borescope-based inspection is time consuming, labor intensive and susceptible to human error. The Khalifa University team built a system in which a container filled with blades is placed in front of the robot, which autonomously localizes each blade, picks it up, places it in an image acquisition box, detects surface defects and returns the blade to its original position.
The setup uses an ABB IRB 1600-10 robotic manipulator whose end-effector carries a Schunk gripper, an S-beam load cell, an Intel RealSense D405 depth camera and a fiducial marker. The image acquisition box has a diffused light source for uniform illumination and a high-resolution camera. Two datasets were collected, one for blade localization using a stereo depth camera from different angles and lighting, and one for surface defects such as corrosion, dents, scratches and nicks, captured under controlled illumination with a fixed camera setup. The datasets include approximately 73 unique blades, and the defect dataset has 282 samples. A separate model was trained on each dataset. The detector keeps a backbone similar to YOLOv5 but replaces the C3 module with the C2f module, and training used 640 x 640 images for 150 epochs with the Adam optimizer.
Localization accuracy was within ±1 mm under controlled conditions, and grasping and pick-and-place achieved a near 100% success rate under controlled blade orientations, measured over about 50 operations. Failures occurred mainly with improper initial blade placement, partial occlusions and extreme orientations. The paper notes that the system operates under controlled illumination and is trained on a predefined set of defect classes. Future work will focus on improving defect detection precision, expanding the dataset and optimizing the robotic system for real-world deployment in the aerospace industry. The research was funded by the Advanced Research and Innovation Center (KU-ARIC), jointly funded by Aerospace Holding Company LLC, a wholly owned subsidiary of Mubadala Investment Company PJSC, and Khalifa University.