Manual measurements with hand-held measuring equipment have become ineffective due to their time consumption and disruption of the regular operation schedule. In recent years, new measuring methods have been established, using specialized or in-service vehicles to collect significant amounts of data on condition of railway or tramway track infrastructure. However, accessing large datasets requires an extensive amount of time to process and evaluate the data before providing valuable information on track infrastructure condition to the operator. Effective large dataset analysis method could simplify the maintenance and intervention plan for the tramway infrastructure and improve the quality of track monitoring system. Several researchers and authors have investigated the possibility of implementing various machine learning approaches to speed up and automate the evaluation of track condition data. Based on historic and real-time data from in-service vehicles, using machine learning it is possible to detect irregularities and update digital twin model of the track for predictive maintenance. As part of the project URITMIS - Urban Railway Infrastructure Maintenance System, machine learning techniques and digital twin models of the tramway track will be investigated to improve maintenance efficiency and track reliability and resilience of Zagreb's tramway network. This paper presents current state of art as well as case-study of weld detection and evaluation from in-service vehicle bogie acceleration signals using machine learning techniques.

track maintenance, machine learning, urban railway systems, vibroacoustic measurements