On-board vibration-based robust and unsupervised degradation detection in railway suspensions under various travelling speeds via a Multiple Model framework

Abstract

A robust data-driven method for on-board vibration-based degradation detection in railway suspensions of running vehicles is introduced. The method employs two lateral vibration acceleration sensors per vehicle half, one on the bogie and one on the vehicle body. It is based on Transmittance Function data-driven models of the AutoRegressive with eXogenous excitation type within an unsupervised Multiple Model framework and aims at effective detection of early-stage component degradation while achieving robustness to varying Operating Conditions. The method is validated via thousands of Monte Carlo simulation experiments under three distinct travelling speeds. Through them, perfect detection performance is demonstrated for ‘small’ level degradation, characterised by $20\mathrm{\%}$ reduction in the properties of suspension components, while even ‘minor’ degradation, characterised by $10\mathrm{\%}$ reduction, is detectable but somewhat less effectively. The very good performance characteristics of the method are confirmed via field tests as well, while its superiority over alternative schemes is demonstrated via comparisons with a state-of-the-art entropy-based approach.

Keywords:

• Railway suspension
• condition monitoring
• vibration-based methods
• robust methods
• varying operating conditions
• early-stage degradation detection

Acknowledgments

Special thanks are due to I.A. Iliopoulos and N. Kaliorakis of the University of Patras for their help with the on-vehicle measurements, to A. Deloukas, G. Leoutsakos, C. Giannakis, E. Chronopoulos and I. Tountas of Attiko Metro S.A., Athens, for useful discussions, as well as to C. Mamaloukakis, K. Katsiana and K. Sarris of STASY S.A., Athens, who provided technical support for unit installation and the on-board measurements.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 Unsupervised in the sense that training is based on exclusively healthy (nominal) condition signals.

2 t designates normalised by the sampling period discrete time and N the signal length in samples.

3 In the sense that training is based on signal sets measured under the healthy condition only.

Additional information

Funding

The study has been partly funded by the European Union and Greek national funds through the Operational Program Competitiveness, Entrepreneurship and Innovation, under the call RESEARCH-CREATE-INNOVATION (The MAIANDROS project with code T1EDK-01440), and partly by the Hellenic Foundation for Research and Innovation (HFRI) under the 3rd Call for HFRI Ph.D. Fellowships (Fellowship Number 6714).