Intelligent Classification of Defective Rails Through GAN and Ensemble Predictive Model

Abstract

Detection of defects on the rail is an important part of rail inspection. Therefore, in recent years, automatic inspection systems are needed to inspect railway tracks quickly and accurately. Methods such as deep learning are used in automatic control systems. Deep learning methods need a lot of data sets for training. A traditional convolutional neural network cannot learn features with a small data set. Although it is easy to create a dataset of solid rail components for the training process, it is quite difficult to create a dataset of defective rail components. To create this type of dataset, it is necessary to collect images from hundreds of kilometers of railways. Therefore, in this study, images of artificial defective rail components were created using GAN, and the data set was reproduced. The quality and variety of images produced by the GAN architecture directly affect the performance of the proposed model. For this reason, the performance of the GAN model was evaluated with evaluation metrics. Then, the imperfections of the rail components with the neural network based on ensemble prediction were examined. In the study, there are three rail classes as healthy, cracked, and hole. The performances of the new images produced in each class were calculated. The basic idea of the study is to increase individual model performances by making ensemble predictions with models created as a result of training with different deep-learning methods. In the experimental results, an accuracy rate of 99.44% was obtained for the proposed method and its advantage over traditional methods was demonstrated. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.