Research on inversion of wheel-rail force based on neural network framework

Abstract

Wheel-rail force is a critical parameter for vehicle-track-bridge (VTB) structure system, playing a pivotal role in ensuring the safe operation of trains and service performance of track and bridge structures. This paper proposes a methodological framework for efficient and accurate prediction of wheel-rail force, that bidirectional gated recurrent unit (Bi-GRU) neural network is developed and the hyperparameters of networks are optimized through grey wolf optimization (GWO). Based on the developed VTB system model, numerical experiments are carried out to produce datasets which are verified through field test. In order to make the data more universal and comprehensive, Latin hypercube sampling (LHS) is introduced to randomize the parameters of VTB model. The datasets generated from VTB system, with rail seat force as input and wheel-rail force as output, serve as the training data for framework. Additionally, the inversion effects under various working conditions are examined and discussed. The results indicate that the model can precisely inverse continuous wheel-rail forces from the moment the wheels enter the first test rail seat until leave the last test point of rail seat. When the neural network learns from the features of at least three or more rail seats on each track slab (with intervals not exceeding 2 rail seats), it can accurately inverse the time history curve of wheel-rail forces. The framework presents a viable alternative for the wheel-rail force analysis, and promising potential for the realization of online monitoring of wheel-rail force.