Challenges of Automatic Optical Inspection of Used Turbine Blades with Convolutional Neural Networks

This paper presents an automatic optical inspection task for used turbine blades. The defects arising are very small and occur very rarely. The paper analyzes to what extent state of the art deep learning methods of image processing help to solve the inspection tasks. A total of 34 different turbine blades were acquired image-wise for this work. For the localization and classification of the defects, detection methods such as YOLOv7 were used on the one hand, and segmentation methods such as Mask R CNN and QueryInst on the other. Despite a very small amount of data, the methods can be trained to learn the defects and recognize unseen defects. A maximum mAP 0.5 of 60.9% was achieved. Even though the inspection task was challenging in terms of defect characteristics and the number of training data was low, reliable models could be created. The accuracy is not sufficient for full automation, but it can initially generate useful suggestions for the workers and focus attention on critical areas.