Estimation of railway track longitudinal irregularity using vehicle response with information compression and Bayesian deep learning

Abstract

AbstractIn railway transportation, track geometry irregularity is one of the main factors in controlling train safety. At present, railway practitioners typically use the track geometry car (TGC) based on the inertial navigation system to inspect track irregularities. However, TGCs are quite expensive, and their inspection interval is relatively long. Among a variety of emerging methods, using vehicle responses to estimate track irregularities seems very promising as it enables a cheaper and more efficient solution. In this work, an extended auto‐encoder (EAE) is proposed to estimate the track longitudinal irregularity through car body acceleration. The mean absolute percentage errors of the estimated results on the simulated and the real‐world dataset are 2.67% and 3.75%, respectively, which is 50%–55% lower than the traditional neural network. In the frequency domain, the characteristic wavelengths of 5.4 and 32 m can be effectively identified. Besides, the Bayesian deep learning (BDL) method is introduced to improve the EAE and estimate the confidence interval of the track longitudinal amplitude. A metric (coverage width and error) for evaluating and optimizing the performance of the estimated interval is proposed. The interval estimation result in the time domain has a 98% correct coverage rate of the ground truth and 93% in the frequency domain. Within the error range of plus/minus one standard deviation, the EAE model has an estimation accuracy of 94.2% for the standard deviation of track longitudinal irregularity, and the BDL‐EAE can even reach 100%. Compared with the existing methods, our proposed model only requires car body acceleration and has the potential to use ordinary in‐service trains for onboard track inspection.