Analysis of railway vibration signals using supervised machine learning for the development of exposure-response relationships

Abstract

The aim of this work is to investigate the applicability of the use of supervised machine learning methods to classify unknown railway vibration signals within a measurement database. The results of this research will be implemented in the development of exposure-response relationship for annoyance caused by freight and passenger railway vibration, so as to better understand the differences in human response to these two sources of environmental vibration. Data for this research comes from case studies comprising face-to-face interviews with respondents and measurements of their vibration exposure collected during the University of Salford study "Human Response to Vibration in Residential Environments". Vibration data from this study are then classified into freight and passenger categories using supervised machine learning methods. Finally, initial estimates of exposure-response relationships are determined using ordinal probit modelling. The results indicate that the annoyance response due to freight railway vibration may be significantly higher than that due to passenger railway vibration, even for equal levels of exposure. The implications of these findings for the potential expansion of freight traffic on rail are discussed. [Work funded by the Department for Environment, Food and Rural Affairs (Defra) UK, and EU FP7 through the CargoVibes project]