Hu, Xiaoxi (State Key Laboratory of Rail Traffic Control and Safety, Beijing), Zhang, Xiaohan (Zhejiang University), CHEN, FEI (Wuhan University), Liu, Zongyang (Huazhong University of Science and Technology), Liu, Jin (University of Leeds), Tan, Lei (Beijing Municipal Engineering Research Institute), TANG, Tao (Beijing Jiaotong University)

Simultaneous Fault Diagnosis for Sensor and Railway Point Machine for Autonomous Rail System

Scheduled for presentation during the Invited Session "Control, Communication and Emerging Technologies in Smart Rail Systems II" (WeBT7), Wednesday, September 25, 2024, 15:50−16:10, Salon 15

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 October 7, 2024

Abstract

Since the safe and reliable behavior of a Railway Point Machine (RPM) is definitely pivotal for rail transportation, scholars and engineers have studied a considerable number of algorithms for RPM fault diagnosis via many kinds of sensor. However, few scholars have considered the possibility that sensors can also experience faults, which cannot be ignored in practical railway applications. To fill the gap, we propose an end-to-end deep learning network named Simultaneous Fault Diagnosis Network (SFDNet) for the simultaneous fault diagnosis for sensor and RPM. Our approach includes Information Aggregate and Update Module (IAUM) and Fault Diagnosis Module (FDM). The IAUM module leverages the concept of Spatial Attention Mechanism (SAM) and Channel Attention Mechanism (CAM) to effectively filter and aggregate information from diverse sensors. It enhances the salience of normal sensor observations while attenuating the influence of anomalous signals, thereby enabling the resultant features to more accurately capture the operational state of the RPMs. The features yielded by SAM are directly employed for sensor fault diagnosis. The FDM integrates Bidirectional Gated Recurrent Units (BiGRUs) and Time-series Feature Pyramid Network (TFPN) to respectively extract and integrate long-term dependencies and multi-scale features, enhancing fault diagnosis in RPM by capturing comprehensive temporal patterns and correlations. Lastly, ablation experiments and comparison studies shows the superiority of our SFDNet.