Exploring the impact of foot-by-foot track geometry on the occurrence of rail defects

Predicting rail defects is of great importance for safe railway transportation. Using foot-by-foot track geometry and tonnage data, this paper develops a new machine learning based approach to identify the track geometry parameters that contribute most to the prediction of rail defects occurrences. Taking more than 60 types of track geometry measurements into account, this study develops a Recursive Feature Elimination (RFE) algorithm for feature selection and compares its results with Singular Value Decomposition (SVD). In addition, to capture more knowledge from the geometry data, some additional features, including Track Quality Index (TQI), energy spectral density, and time-trend are extracted. This, in turn, facilitates the learning and predicting process. Moreover, since there exists a very limited number of rail defects, the Adaptive Synthetic Sampling Approach (ADASYN) is applied to overcome the issue of imbalance in the dataset. In terms of machine learning algorithms, the proposed approach employs an extreme gradient boosting (XGBoost) algorithm in which the hyper-parameters are optimized using a Bayesian optimization method. Furthermore, the proposed approach investigates the impact of each track geometry parameter as well as a subset of them on rail defects occurrences with Partial Dependence Analysis (PDA). Finally, our approach is implemented on a six-year dataset with over 60 million track geometry records collected from a 100-mile section of a U.S. Class I railroad to demonstrate its applicability and efficiency.