Detection of damaged parts via transfer learning method in jet aircrafts

One of the most common methods used for defect detection of jet aircraft parts is X-ray imaging, which is a non-destructive inspection method. In this method, defects are identified in the analyses of X-ray radiographs made by specialized personnel. Defect detection is of great importance for flight safety and therefore damage inspection must be done very carefully. Transition of these defect detection methods into automated systems will bring technical innovations and convenience to the institutions, in terms of both increasing the sensitivity of analysis and reducing the dependence on human-sourced methods. For these reasons, in order to detect the defect using image processing algorithms and artificial intelligence elements, crack defects in military aircraft parts were analyzed on X-ray images within the scope of this thesis. In these analyses, algorithms trained by using deep learning methods from artificial intelligence elements were used. Within the available 100 radiography images (50 images with cracks and 50 images without cracks), 90 were used for the training dataset and 10 for the test dataset. With the help of convolutional neural networks, transfer learning was performed from pre-trained ResNet50, AlexNet, VGG19 and DenseNet169 networks, and crack detection was performed in the available data (images). The F-scores in these network models are 78,85%, 66,67%, 73,68% and 77,48%, respectively. Heat maps are also shown, in which ResNet50 convolutional neural network model, which analyzes the defects need to be detected in the images with the highest F-score, concentrates in its decision process. As a result, it is observed that the methodology can give results with a high rate of accuracy in the defect detection of radiographs. With the automated defect analysis application carried out in this thesis, a model was developed that helps to determine the defects by using transfer learning with a small data set in a case where data may be limited.