Estimation of wheel-rail friction coefficient using deep CNN on axlebox accelerations

Low friction can lead to poor adhesion conditions between the rail and wheel, which is detrimental to rail vehicle operation and safety. This paper presents a friction coefficient estimation algorithm using a multi-channel deep convolutional neural network (MC-DCNN) on short-time Fourier transform of axlebox/wheelset accelerations during normal running condition. The estimator is developed by large-scale experimentally obtained axlebox accelerations and friction measurements data from running a rail vehicle on a friction-modified track with five different contact contamination conditions offering different friction levels at four different constant speeds, including both straight and curved tracks. A total of 50 experimental test runs providing data from both straight and curved tracks are employed. Four different input combinations to the MC-DCNN have been investigated: longitudinal, lateral, longitudinal and lateral, and longitudinal slip accelerations of the wheelsets. Multiple training and cross-validation studies were carried out using validation datasets from different experiments to show the MC-DCNN can accurately estimate the friction coefficients during normal operating conditions. The results take a significant step in realizing up-to-date monitoring of friction levels across the network for planning and implementing mitigation strategies against adverse effects of low adhesion conditions.