Modeling train timetables as images: A cost-sensitive deep learning framework for delay propagation pattern recognition

Abstract

As a vital component of train operational control, train delay propagation pattern discovery is critically important for both railway controllers and passengers. In this study, we present a carefully designed deep learning model, called FCF-Net, that comprises fully connected neural networks (FCNN) and convolutional neural networks (CNN) for train delay propagation pattern recognition in railway systems. FCF-Net first uses a CNN component that handles train timetables as images to capture interactions of train events and an FCNN component to capture the influence of non-operational features separately; then it uses another FCNN component to combinedly learn the dependencies between operational and non-operational features. In addition, considering the imbalance of train delay data, a cost-sensitive technique that assigns different misclassification costs for different class was used to better deal with the imbalanced data. The main goal of the FCF-Net is to realize efficient and accurate train delay propagation pattern recognition by mining potential knowledge from train operation data. The predictive and computational performance of the model was tested and evaluated on data from two high-speed railway lines with different operational features in China. The results show that FCF-Net, once trained with sufficient data, outperforms conventional deep learning with common loss and other state-of-the-art deep learning models for train delay propagation pattern recognition, indicating its capability in knowledge discovery from train operation data. In addition, the computational results show that FCF-Net exhibits more efficient training process than existing state-of-the-art deep learning models.