Predicting failure times of railcar wheels and trucks by using wayside detector signals

Nowadays railway networks are instrumented with various wayside detectors. Given massive amount of data collected from electronic wayside detectors, railcar failure prediction has recently attracted great attention in order to reduce rolling stock inspection and maintenance costs and improve railway safety. In this work, we present a methodology to predict the failure times of railcar wheels and trucks, by fusing sensor signals from three types of wayside detectors, including Wheel Impact Load Detector (WILD), Machine Vision (MV) systems, and Optical Geometry Detectors (OGD). In data preprocessing, missing values are handled by missForest, a Random Forest based nonparametric missing value imputation algorithm, and a variety of new features are generated to capture the signal characteristics. Several state-of-the-art regression models are built and compared to predict the lifetime of railcar wheels and trucks in a US national railway network.