Prediction of wheel and rail wear using artificial neural networks

The prediction of wheel wear is a significant issue in railway vehicles. It is correlated with safety against derailment, economy, ride comfort, and planning of maintenance interventions, and it can result in delay, and costs if it is not predicted and controlled in an effective way. However, the prediction of wheel and rail wear is still a great challenge for railway systems. Therefore, the main aim of this thesis is to develop a method for predicting wheel wear using artificial neural networks.
Initial tests were carried out using a pin-on-disc machine and this data was used to establish how wear can be measured using an Alicona profilometer.
A new method has been developed for detailed wheel wear and rail wear measurements using ‘Replica’ material which was applied to the wheel and rail surfaces of the test rig to make a copy of both surfaces. The replica samples were scanned using an optical profilometer and the results were processed to establish wheel wear and rail wear. The effect of load, and yaw angle on wheel wear and rail wear were examined. The effect of dry, wet, lubricated, and sanded conditions on wheel wear and rail wear were also investigated.
A Nonlinear Autoregressive model with eXogenous input neural network (NARXNN) was developed to predict the wheel and rail wear for the twin disc rig experiments. The NARXNN was used to predict wheel wear and rail wear under deferent surface conditions such as dry, wet, lubricated, and sanded conditions.
The neural network model was developed to predict wheel wear in case of changing parameters such as speed and suspension parameters. VAMPIRE vehicle dynamic software was used to produce the vehicle performance data to train, validate, and test the neural network. Three types of neural network were developed to predict the wheel wear: NARXNN, backpropagation neural network (BPNN), and radial basis function neural network (RBFNN). The wheel wear was calculated using an energy dissipation approach and contact position on straight track.
The work is focused on wheel wear and the neural network prediction of rail wear was only carried out in connection with the twin disk wear tests. This thesis examines the effect of neural network parameters such as spread, goal, maximum number of neurons, and number of neurons to add between displays on wheel wear prediction.
The neural network simulation results were implemented using the Matlab program. The percentage error for wheel and rail wear prediction was calculated. Also, the accuracy of wheel and rail wear prediction using the neural network was investigated and assessed in terms of mean absolute percentage error (MAPE). The results reveal that the neural network can be used efficiently to predict wheel and rail wear.
Further work could include rail wear and prediction on a curved track.