Indirect Identification of Varying Conicity Levels of Wheel Tread Using Convolutional Neural Networks

  • Khurram Shaikh, Sanaullah Mehran Ujjan, Imtiaz Hussain Kalwar, Bhawani Shankar Chowdhry

Abstract

Maintenance of Railway wheelset conicity under a certain range is necessary to ensure safe and
comfortable operation of a railway vehicle. During a hectic schedule, wheelset could be worn out reducing
its conicity, which affects the ability of wheelset to align itself back on center position while facing track
disturbances. For designing a predictive maintenance strategy it is necessary to predict variations in
conicity levels to gather useful information about the wheel wear. In this research, an indirect conicity
identification technique based on Convolutional neural networks is developed. This technique exploits the
variations in dynamic response of wheelset to detect and identify the corresponding changes in conicity of
the wheel tread. A simulation model of a railway wheelset is developed to generate the required data at
different conicity levels. The variations in dynamic response are then recorded, split at fixed time intervals
to establish a dataset. A comparative analysis of supervised data driven algorithms on this dataset is
conducted using Python scientific computation libraries. Algorithms are trained using Keras and
Tensorflow and validated on test set. Final model, a 1D- Convolutional Neural network, is able to achieve
about 97% accuracy on conicity identification task.

Published
2020-05-20
How to Cite
Khurram Shaikh, Sanaullah Mehran Ujjan, Imtiaz Hussain Kalwar, Bhawani Shankar Chowdhry. (2020). Indirect Identification of Varying Conicity Levels of Wheel Tread Using Convolutional Neural Networks. International Journal of Advanced Science and Technology, 29(7), 2537-2547. Retrieved from http://sersc.org/journals/index.php/IJAST/article/view/18026
Section
Articles