Surrogate model construction and improved detection of rail corrugation in high-speed railways based on a hybrid neural network

Abstract

Rail corrugation in high-speed railways severely affects the safety and comfort of train operations. Existing intelligent detection technologies are predominantly confined to the classification of single corrugation features and fail to achieve quantitative detection of complex corrugations. To this end, this study proposes a high-precision quantitative detection method for rail corrugation based on a hybrid neural network surrogate model. Firstly, the spatial and frequency domain characteristics of the axle box acceleration (ABA) under the excitation of rail corrugation are obtained based on the vehicle-track coupled dynamics model. Then, a hybrid neural network surrogate model is proposed to construct the nonlinear mapping relationship between the multi-features of corrugation and the multi-features of the ABA signal in both spatial and frequency domains, achieving high-precision fitting of the response surface. Finally, an improved Newton-Raphson-based optimizer (NRBO) algorithm is applied to further enhance the detection efficiency of rail corrugation. Comprehensive comparisons with classical models across multiple detection performance indicators demonstrate the superiority of the proposed model. Specially, the detection rates of the amplitude and wavelength of single-feature rail corrugation, with an error rate of less than 1 %, are respectively 95 % and 99 %, and good adaptability to rail corrugation with composite features can also be identified. This study provides a novel approach for the real-time and quantitative detection of rail corrugation, holding beneficial engineering significance for intelligent railway maintenance.