Dynamics-Based Estimation of Wheel-Rail Friction Coefficient using Deep CNN

This paper presents an estimation scheme for rail-wheel friction coefficients applying a multi-channel deep convolutional neural network on axlebox accelerations.Different to conventional approaches, the multi-channel deep convolutional neural network does not depend on any slip or creep measurements, nor knowledge of vehicle parameters.It is trained using axlebox longitudinal and lateral acceleration measurements and known rail friction coefficient measurements obtained from running a rail vehicle on friction-modified tracks with five different friction levels at four different speeds.The experimental test data includes both straight and curved track scenarios, and independent validation data shows that the friction coefficient can be very accurately estimated under normal running conditions in almost all of the validation data sets.