Detecting singular track defects by time-frequency signal separation of axle-box acceleration data

Abstract

Singular railway track irregularities, such as squats and corrugation have a major impact on the ride stability, noise emission, comfort and safety of freight and passenger trains. Therefore, the detection and monitoring of such defects play an important role in railway track maintenance. Embedded low-cost sensors on in-service vehicles provide the opportunity of quasi-continuous condition monitoring of railway tracks and can thus enhance existing track maintenance strategies. In this paper we demonstrate a processing sequence to detect singular track defects from noisy axle-box acceleration (ABA) data. The data are acquired with a multi-sensor prototype measurement system on a shunter locomotive operating on the industrial railway network of the inland harbor of Braunschweig (Germany). A blind signal separation (BSS) algorithm based on non-negative matrix factorization is applied to the ABA data in the time-frequency domain. It is completely data-driven and hence does not rely on a priori knowledge or physical models. The algorithm makes use of different time-frequency characteristics of the signal components and is thus able to separate quasi-continuous band-limited signal components from transient broad-band components. The magnitude of the transient components reflects the strength of track singularities along the track and can hence be used to detect and quantify short track defects. Through georeferencing the identified defects can be localized, mapped on the track and be used to guide specific maintenance actions. Additionally, the BSS algorithm shows the potential to reduce the dimensionality of the data without significant loss of information.