Transfer Learning Based Fault Detection Approach for Rail Components

Abstract

Railroad track fasteners are used to connect rail components together. Control of fasteners is great importance for travel safety. Missing, broken or deformed fasteners should be detected and repaired. In this study, a new method for fault detection is proposed by using a dataset consisting of railway images recorded using an autonomous drone. In deep learning, which has the potential of self-learning from the available data, the most important factor affecting model performance is data. In this study, obtaining the rail fastener images with an autonomous drone has provided an advantage compared to the existing studies in the literature. Deep learning training was conducted with Vgg16 and ResNet101V2, which are transfer learning models, in order to determine the faults caused by the lack of fasteners. The performances of the trained models in detecting faultless and missing/faulty fasteners were compared. In the results obtained, it was seen that the training made using the ResNet101V2 model with 99% accuracy produced results with higher accuracy.