A Wasserstein generative adversarial network-based approach for real-time track irregularity estimation using vehicle dynamic responses

Accurate and timely estimation of track irregularities is the foundation for predictive maintenance and high-fidelity dynamics simulation of the railway system. Therefore, it's of great interest to devise a real-time track irregularity estimation method based on dynamic responses of the in-service train. In this paper, a Wasserstein generative adversarial network (WGAN)-based framework is developed to estimate the track irregularities using the vehicle's axle box acceleration (ABA) signal. The proposed WGAN is composed of a generator architected by an encoder-decoder structure and a spectral normalised (SN) critic network. The generator is supposed to capture the correlation between ABA signal and track irregularities, and then estimate the irregularities with the measured ABA signal as input; while the critic is supposed to instruct the generator's training by optimising the calculated Wasserstein distance. We combine supervised learning and adversarial learning in the network training process, where the estimation loss and adversarial loss are jointly optimised. Optimising the estimation loss is anticipated to estimate the long-wave track irregularities while optimising the adversarial loss accounts for the short-wave track irregularities. Two numerical cases, namely vertical and spatial vehicle-track coupled dynamics simulation, are implemented to validate the accuracy and reliability of the proposed method.

Automated summary

This paper presents a framework based on Wasserstein generative adversarial network for estimating track irregularities using axle box acceleration signal, combining supervised and adversarial learning for optimizing the network training process. This method has been successfully applied to dynamic simulations in railway systems.