Li, Ruizhe (Beijingjiaotong University), Bu, Bing (State Key Laboratory of Rail Traffic Control and Safety, Beijing), Wu, Daohua (Beijing Jiaotong University)

A Diagnosis Method of Broadcast Storm Based on WGAN-GP-CNN for Data Communication System of CBTC

Scheduled for presentation during the Poster Session "Railway systems and applications" (ThBT13), Thursday, September 26, 2024, 14:30−16:30, Foyer

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on December 26, 2024

Abstract

In a communication based train control (CBTC) system, the data communication system (DCS) is deployed to transmit the status and control commands of trains. Typically, the backbone of a DCS is a layer 2 network with a ring topology. The network is susceptible to broadcast storms, which will block the communication among the connected equipment, bring catastrophic impacts on the operation of trains. In recent years, broadcast storms have occurred from time to time in urban rail transit. Although methods to effectively diagnose broadcast storms are desperately needed, the related researches are seriously hindered due to a severe scarcity of fault data. In this paper, we propose a wasserstein generative network with gradient penalty and a convolutional neural network (WGAN-GP-CNN) method to identify three status of a DCS, which are healthy, sub-healthy and faulty. A WGAN-CP model is proposed to solve the small size and imbalance problem of the data set, to enhance the performance of the CNN to discriminate three different status of a DCS. A ring network is setup. Different faults which may introduce broadcast storms are simulated. The status of switches and the performance of data transmission are collected to form the data set, which are used for the training, validating and testing of the proposed model. On the original data set, the CNN of the proposed model can detect the three different status of the network with 85.1% accuracy. The performance can be improved to 97.3% on the synthetic data set which is processed by the WGAN-GP model.