Data Augmentation for Fault Classification of Railway Track Irregularities in Track-Vehicle Scale Model

Fault diagnosis of railway track irregularities (TI) via supervised learning algorithms is a difficult task due to the lack of suitable, labeled datasets. Class imbalance in the data poses an additional issue. While it may be possible to continuously monitor the condition of the railway track using sensors mounted on regular in-service vehicles, the presence and location of faults is unknown (unlabeled data). At the same time, most of the collected information will correspond to the nominal, non-faulty condition (imbalanced dataset), which impacts on the performance of Machine Learning classifiers. In this paper, a novel method to generate synthetic TI using advanced Generative Adversarial Networks (GAN) is presented, in conjunction with a numerical model of the track-vehicle interaction, to perform data augmentation and obtain a large, labeled and balanced dataset, suitable for supervised learning classification. Inertial measurements from a vehicle-track scale model are then used as test dataset to validate the data augmentation process.