Exploring the Impact of GAN-Based Data Augmentation and FGSM-Refined Images on Wear Size Estimation of Railway Switches and Crossings

Abstract The switch and crossing (S&C) is a crucial component of the railway infrastructure network, significantly affecting traffic delays and maintenance costs. This study aimed to predict wear across the entire S&C using medium-range accelerometer sensors. Vibration data were collected, processed, and converted into spectrograms to develop accurate data-driven models. However, due to weather constraints such as ice and snow, our database remains limited. To ensure the proper generalization of deep learning models, it is essential to expand this dataset. Therefore, we plan to employ Generative Adversarial Networks (GANs) to address this objective. GANs are a powerful class of networks capable of generating realistic new images from a specified database. This study aims to explore the use of deep learning techniques, particularly GANs combined with Convolutional Neural Networks (CNNs), for the classification of wear levels in switch and crossing (S&C) components. To enhance the robustness of CNN models, we employ the fine-tuning technique in conjunction with the adversarial images using Fast Gradient Sign Method (FGSM).