Acceleration-based friction coefficient estimation of a rail vehicle using feedforward NN: validation with track measurements

Abstract

Low friction can lead to poor adhesion conditions between the rail and wheel, which is detrimental to rail vehicle operation and safety. Up to date knowledge of the rail-wheel friction level is currently not available across rail networks, meaning planning mitigation strategies is difficult. This paper presents a real-time friction coefficient estimation algorithm based on a feed-forward neural network (FNN). Unlike conventional methods, the FNN does not depend on slip/adhesion curves or creep force models, and only requires wheelset longitudinal acceleration and speed. The wheelset acceleration and friction measurements are obtained by running a two-car rail vehicle on a friction-modified track with five different levels of friction conditions at four different vehicle speeds. Four different FNNs are trained for four speed conditions, and their estimation performance were validated by training multiple FNNs and testing them in each speed case using new sets of data. Validation results show that the average mean absolute errors from the four FNNs remains below 0.0083.