Fault Prediction of On-board Train Control System: An Data Balance-enhanced AdaBoost Approach

Abstract

This paper concerns the fault prediction modeling for Condition-based Maintenance (CBM) of on-board railway train control systems. Based on the field data from the CTCS2-200H on-board equipment, the imbalance problem of the fault and fault-free samples is investigated, which makes it difficult to realize effective model training for a fault prediction purpose. A data-driven system fault prediction framework is established. Based on that, a data balance-enhanced AdaBoost (Adaptive Boosting) approach is proposed, which uses an integrated Borderline-SMOTE method to improve the class distribution of the training samples. With this method, effective modeling using the AdaBoost algorithm can be enabled with enhanced datasets. Field data collected from the practical on-board train control equipment is utilized to demonstrate the proposed approach. Comparisons with the different dataset conditions and training algorithms illustrate the superior performance of the proposal and show its potential in the fault prediction-enabled intelligent maintenance systems.