Approaches for diagnosis and prognosis of asset condition: application to railway switch systems

Chen, Qianyu ORCID: 0000-0002-1273-7988 (2021). Approaches for diagnosis and prognosis of asset condition: application to railway switch systems. University of Birmingham. Ph.D.

Preview

Chen2021PhD.pdf
Text - Accepted Version
Available under License All rights reserved.
Download (3MB) | Preview

Abstract

This thesis presents a novel fault diagnosis and prognosis methodology which is applied to railway switches. To improve on existing fault diagnosis, energy-based thresholding wavelets (EBTW) are introduced. EBTW are used to decompose sensor measurement signals, and then to reconstruct them within a lower dimensional feature vector. The extracted features replace the original signals and are fed into a neural network classifier for fault diagnosis. Compared to existing wavelet-based feature extraction methods, the new EBTW method has the advantage of an intrinsic energy conservation property during the wavelet transform process. The EBTW method localises and redistributes the signal energy to realise an efficient feature extraction and dimension reduction.

The presented diagnosis approach is validated using real-world switch data collected from the Guangzhou Metro in China. The results show that the proposed diagnosis approach can achieve 100% accuracy in identifying a railway switch overdriving fault with various severities, improving upon existing methods of conventional discrete wavelet transform (C-DWT) and soft-thresholding discrete wavelet transform (ST-DWT) by 8.33% and 16.67%, respectively.

The presented prognosis approach is constructed based on traditional data-driven prognosis modelling. The concept of a remaining maintenance-free operating period (RMFOP) is introduced, which transforms the usefulness of sensor measurement data that is readily available from operations prior to failure. Useful features are then extracted from the original measurement data, and modelled using linear and exponential regression curve fitting models. By extracting key features, the original measurement data can be transformed into degradation signals that directly reflect the variations in each movement of a switch machine. The features are then fed into regression models to derive the probability distribution of switch residual life. To update the probability distribution from one operation to the next, Bayesian theory is incorporated into the models.

The proposed RMFOP-based approach is validated using real-world electrical current sensor measurement data that were collected between January 2018 and February 2019 from multiple operational railway switches across Great Britain. The results show that the linear model and the exponential model can both provide residual life distributions with a satisfactory prediction accuracy. The exponential model demonstrates better predictions, the accuracy of which exceeds 95% when 90% life percentage has elapsed. By applying the RMFOP-based prognosis approach to operational data, the railway switch health condition that is affected by incipient overdriving failure is predicted.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):