A paradigm shift towards condition-based and predictive maintenance (CBM and PM) is undergoing in the railway system. This will increase the maintenance efficiency and ultimately increase the reliability and availability of railway assets. As an essential part of CBM and PM, condition monitoring along with intelligent data processing algorithms determines the up-to-date asset conditions to support maintenance decision-making. Condition monitoring in railway applications usually requires real-time data processing for fault diagnosis. As train drivers or infrastructure operators should be immediately informed, once severe failures are detected. As many railway assets have not been electrified, the onsite infrastructure for power supply and data communication is absent. This results in further challenges for data processing with regard to power consumption and computational complexity. Furthermore, the operating conditions of railway assets vary in a large range. Condition variations reflect in the monitoring data and cause the distribution shift, which may induce the robustness problem of diagnostic models. To cope with these problems encountered in practice, we have conducted extensive research towards robust real-time condition monitoring and fault diagnosis for railway assets. We propose to use lightweight convolutional neural networks (LCNN) to realize real-time capability. During the model training, data augmentation is introduced for robustness enhancement. This general data processing procedure is demonstrated in two distinct railway applications. They have been described in three scientific publications, which constitute the present cumulative dissertation. The first application deals with wheel flat detection (WFD) for vibration monitoring on freight wagons, which supports wagon maintenance. As freight wagons are not electrified, the algorithm should be executed in real time on embedded systems powered by batteries, which have limited computation power. In the first paper, we propose to automatically search a one-dimensional (1D) LCNN for real-time WFD with the optimal tradeoff between computational complexity and detection accuracy. In the second paper, the robustness problems induced by the variation of vehicle speeds, monitored wagons and track conditions are investigated. A novel data augmentation framework, incorporating multibody dynamic simulation for physical modeling and fast weighted feature-space averaging to augment simulation data, is proposed for robustness enhancement. The second application concerns track geometry monitoring, supporting track maintenance. Our work published in the third paper is the first attempt to use a deep learning based computer vision solution for track geometry monitoring. Virtual point tracking for real-time target-less dynamic displacement measurement is proposed to track the lateral movement of the wheel on the rail, in order to calculate track alignment. It is mainly realized by a 2D LCNN for virtual point detection within each video frame, achieving frame rates of above 30 frames per second on edge devices. In addition, data augmentation based on image corruption is applied to enhance the robustness against different weather conditions and contaminations. Real-time requirements and robustness problems are the general topics of condition monitoring. The proposed methods for real-time data processing and robustness enhancement are not confined to the two exemplary applications. They can be adapted into similar scenarios. The potential for further improvement are discussed.
Towards robust real-time condition monitoring and fault diagnosis for railway assets
2023
Miscellaneous
Electronic Resource
English
DDC: | 625 |
Condition monitoring of railway pantographs to achieve fault detection and fault diagnosis
SAGE Publications | 2020
|Mapping of the electromagnetic environment on the railway: Condition monitoring of signalling assets
SAGE Publications | 2020
|Railway point mechanisms: Condition monitoring and fault detection
Tema Archive | 2010
|Railway point mechanisms: condition monitoring and fault detection
Online Contents | 2010
|Railway point mechanisms: Condition monitoring and fault detection
SAGE Publications | 2010
|