Paper Alert

In this work, led by Cyprien Hoelzl (SMM group), in collaboration with Giacomo Arcieri (SMM group), Lucian Ancu (SBB), Stanislaw Banaszak (SBB), Aurelia Kollros (SBB), Vasilis Dertimanis (SMM group), and Eleni Chatzi (SMM group), we fuse features derived from Axle Box Acceleration (ABA) measurements, collected from train vehicles, with expert feedback, in order to refine detection of faulty (defect) welds.

With support of the Swiss Federal Railways (SBB), we propose an automated approach for extraction of indicators on the condition of welds on the basis of on board monitoring (OBM) measurements. The OBM data are derived from Axle Box Accelerometers mounted on specialized, or in service, train vehicles. In this work, we use expert feedback as a complementary information source to further refine assessment. Three models are employed to this end; Binary Classification and Random Forest (RF) models, as well as a Bayesian Logistic Regression (BLR) scheme. We show that the RF and BLR models yield improved prediction accuracy compared against Binary Classification, while the BLR model further delivers a probabilistic assessment, quantifying the confidence we can attribute to the assigned labels. We explain that the classification task necessarily suffers high uncertainty, which is a result of faulty ground truth labels, and explain the value of continuously tracking the weld condition.

This research work was supported by the Swiss Federal Railways (SBB) as part of the ETH Mobility program. We would like to thank, in particular, our partners from the Metrology (MUD) and Strategic Asset Management of the Track departments (SAFB), as well as the track experts of the SBB, who supported the development of this Proof of Concept dataset of defect welds.

Read more in external page Sensors 2023, 23(5), 2672