Identification of Defective Railway Wheels from Highly Imbalanced Wheel Impact Load Detector Sensor Data

Abstract

The problem solving competition organized by the Railway Application Section of the Institute of Operations Research and Management Sciences (INFORMS) in 2017 was to predict the values of load exerted by wheels on the track, when a currently empty rail car would be loaded in the next trip. The organizers provided Wheel Impact Load Detector (WILD) data i.e. value of peak force along with other input variables such as train number, car number, axle side, wheel age, loaded or empty status etc.

In this work, the original prediction problem is converted into a classification problem on the basis of peak force values in order to detect defects in railroad wheels. Peak force values greater than or equal to threshold value (≥ 90 Kilo Pound Force (kips)) define one class, while its values less than threshold value (< 90 kips) define its complement. Given data set is highly imbalanced as about 99.23% of the peak force values fall below threshold values and the remaining 0.76% peak force values fall into its complement. The statistical methodologies that have been attempted to come up with a classification rule include (1) Zero-Inflated Binomial (ZIB) regression model, (2) ZIB regression with L ₁ norm regularization model, and (3) ZIB regression with L ₂ norm regularization model. Out of these three methods, ZIB regression with L ₂ norm model yielded satisfactory results with False Positive Rate reduced to 13.06% and False Negative Rate to 07.75% with accuracy of 87%.