From Detection to Diagnosis: Elevating Track Fault Identification with Transfer Learning

This paper presents a new application of transfer learning (TL) in improving the accuracy of track fault detection. Due to the diverse and complex nature of track faults it is hard to classify. There are various types of track faults such as Cracks, Flakings, Grooves, Joints, Shellings, Spallings, Squats. Classification of track fault could not be easily achieved using any simple machine learning (ML) algorithm. In this regard, this study uses pretrained models that are tuned to fit the peculiar nature of track fault. After acquiring track fault data from a widely cited repository of Mendeley from a research work already represented in ‘Data in Brief’. This application of TL achieves an unbeaten accuracy of 92%, which surpasses most previously generalized models with some explicit data augmentation strategies.