High-speed train state monitoring method with Broad Learning System

Abstract

In this paper, based on Broad Learning System (BLS) that is proposed recently, a real-time monitoring method is designed for the high-speed train (HST) braking system. Due to the high efficiency of the BLS, it is possible to update the state monitoring model in time to adapt to the changes during the HST operation. Moreover, the actual data is highly imbalanced, thus boosting ensemble learning framework is applied to optimize the model to obtain a higher generalization accuracy. In the monitoring system based on BLS integrated with boosting algorithm (noted as B-BLS), the data collected by sensors are added to the model training process in real-time, which makes the anomaly detection more suitable for the current state of the HST braking system. Compared with off-line training models, i.e. artificial neural networks and convolutional neural networks, experimental results demonstrate that the B-BLS has relatively higher adaptivity and efficiency, showing that the proposed monitoring method is feasible.