Echoformer: An echo state-embedded transformer for robust reconstruction of railway trackside noise on urban metro lines

Abstract

Railway rolling noise on straight railway lines has become a crucial environmental impact of railway systems. The vibration of the rail tracks is the primary contributor to the formation of rolling noise. Developing a surrogate model to capture the reflectional relationship between track vibrations and trackside noise is desired in two perspectives. Firstly, it offers a solution for noise monitoring with data loss, or when field conditions are restrictive for sensors’ deployment. Secondly, such a model can facilitate the design and optimization of noise control devices in laboratory, where the actual trackside noise is intricate to simulate. However, it is a dauting task to reveal the underlying relationship between track vibration and trackside noise. This work introduces Echoformer, a novel framework that blends echo states with the transformer architecture, designed to perform time series mapping. Comprehensive testing shows that the Echoformer outperforms conventional RNN architectures in reconstructing both near-field and far-field trackside noise. Moreover, the Echoformer exhibits remarkable resilience against information loss and noisy signal scenario, ensuring a robust reconstruction for the task. This study underscores the Echoformer’s potential as a steadfast tool in the realm of railway noise analysis.