GAN-based data augmentation of time series for fault diagnosis in railway track
Héctor A. Fernández-Bobadilla, Yahya Bouchikhi, Ullrich Martin
Abstract
Abstract Supervised learning classification has arisen as a powerful tool to perform data-driven fault diagnosis in dynamical systems, achieving astonishing results. This approach assumes the availability of extensive, diverse and labeled data corpora for training. However, in some applications it may be difficult or not feasible to obtain a large and balanced dataset including enough representative instances of the fault behaviors of interest. This fact leads to the issues of data scarcity and class imbalance, greatly affecting the performance of supervised learning classifiers. Datasets from railway systems are usually both, scarce and imbalanced, turning supervised learning-based fault diagnosis into a highly challenging task. This article addresses time-series data augmentation for fault diagnosis purposes and presents two application cases in the context of railway track. The case studies employ generative adversarial networks (GAN) schemes to produce realistic synthetic samples of geometrical and structural track defects. The goal is to generate samples that enhance fault diagnosis performance; therefore, major attention was paid not only in the generation process, but also in the synthesis quality assessment, to guarantee the suitability of the samples for training of supervised learning classification models. In the first application, a convolutional classifier achieved a test accuracy of 87.5% for the train on synthetic, test on real (TSTR) scenario, while, in the second application, a fully-connected classifier achieved 96.18% in test accuracy for TSTR. The results indicate that the proposed augmentation approach produces samples having equivalent statistical characteristics and leading to a similar classification behavior as real data.