GAN-based Data Augmentation of Railway Track Irregularities for Fault Diagnosis

Nowadays it is possible to perform automated monitoring on many engineering systems, leading to a huge data gathering. To analyze this information, Machine Learning (ML) is commonly applied. ML algorithms are data-driven, meaning that their capabilities depend, to a large extent, on the quantity and quality of the available data. These methods normally require large training datasets, diverse and rich enough for the algorithms to be able to generalize on previously unseen data. However, there are still applications where it is difficult, not affordable or even impossible to collect data in the form and volume suitable for ML, being railway systems a typical example. In such cases, the available information could be so sparse, that ML techniques yield unsatisfactory results. This article introduces and compares several schemes based on Generative Adversarial Networks (GAN) for Data Augmentation (DA) of Track Irregularities (TI), treated as time-series. Major attention was put not only on the generation process, but also on the synthesis quality assessment, to assure the suitability of generated samples for training of ML models for fault diagnosis. The results indicate that this approach could contribute to alleviate the issue of data scarcity and imbalanced datasets for supervised learning-based fault classification on the railway track.