A fault detection and diagnosis methodology for railway turnout systems based on parameter estimation

Kumpao, Tanapon (2024). A fault detection and diagnosis methodology for railway turnout systems based on parameter estimation. University of Birmingham. Ph.D.

Kumpao2024PhD.pdf
Text - Accepted Version
Available under License All rights reserved.
Download (52MB)

Abstract

Railway turnout systems are one of the most important signalling elements in railway operations, such that their failure during service hours can adversely impact service reliability and safety. Thus, effective maintenance must be applied to ensure that their failure is prevented. For such purposes, data-driven fault detection and diagnosis (FDD) methodologies have been commonly proposed as viable solutions. However, regardless of their exceptional accuracy, they usually
employ advanced data analysis algorithms which are difficult for maintenance staff to comprehend. Moreover, their overfitting nature requires them to undergo a sophisticated re-training process before they can be effectively transferable to diagnose other railway turnout systems.

Due to these difficulties in their practical applications, this thesis proposes an FDD methodology developed from a parameter estimation technique instead, where underlying system knowledge is used to simplify diagnostic reasoning. Once the proposed FDD methodology is applied to a real turnout, it can estimate physically interpretable turnout parameters. These parameters exhibit changes in patterns when a particular fault mode is present in the turnout system. Due to the interpretability of estimation results and specific changes in the parameters, an interpretable and comprehensible classifier can consequently be developed.

The classifier achieves more than 80% overall accuracy when it is applied to the training turnout system. For every diagnostic decision, the skilled personnel can also re-assess it based on the interpretable parameter estimates used in the classification. However, when the proposed methodology is immediately applied to diagnose a different turnout system, it only achieves approximately 60% overall accuracy. Due to the interpretable and traceable diagnostic reasoning used in the
classifier, a simplified cross-training and tuning procedure can be implemented. This improves the accuracy to nearly 90% in diagnosing the test turnout system.

In summary, this thesis proposes and develops a parameter estimation-based FDD methodology for railway turnout system applications, which realises interpretability and traceability as well. Achieving transferability is also possible, but the FDD methodology must undergo some degree of cross-training first for its application to be effective on different turnout systems.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):