On-Board Wheel Profile Classification Based on Vehicle Dynamics - From Physical Effects to Machine Learning

Abstract

Composite brake blocks often cause hollow worn wheel profiles of freight wagons. The resulting shorter maintenance inspection intervals could be elongated if the wheel profile conditions can be estimated via on-board monitoring systems in service. On the one hand such on-board monitoring algorithms must be very accurate and on the other hand very robust against unknown influencing effects.

This paper shows the process from understanding the physical effects by studying the wheel rail contact conditions of new and worn profiles under several operating conditions up to the application of machine learning algorithms for a robust classification of the wheel profile states. In the first step, the physical effects are investigated by varying different operating conditions with a multi body dynamics model of a freight wagon. In the second step, this generated knowledge is used to find suitable features of measured vehicle response quantities to classify new and worn wheel profiles with machine learning algorithms. In the third step, the accuracy of classification results is analysed for different states of available track information like track irregularities or rail profile conditions.

These investigations show very promising results due to a high accuracy of the developed methodology based on machine learning algorithms. Based on the knowledge of the physical effects, these pre-trained algorithms will be verified with measurement data collected in the Shift2Rail project FR8RAIL II.