Application of Machine Learning in the Development of a Predictive Track-Monitoring System for Urban Railway Infrastructure

Urban tram track infrastructure, particularly switches and crossings (S&C), is subject to rapid, unevenly distributed wear, yet condition assessment still primarily relies on manual inspection, which can be slow, costly, and unsafe. Vibration-based monitoring methods are a viable alternative but have mostly been developed for mainline railway lines rather than the specific geometries of urban tram networks and typically depend on costly labelled wear data. This paper presents the URITMIS methodology, which uses vibroacoustic data recorded by an inertial measurement system (IMS) mounted on an in-service tram, with the goal of assessing S&C wear without requiring wear labels during training. A self-supervised contrastive 1D Convolutional Neural Network (CNN) encoder, conditioned with Feature-wise Linear Modulation (FiLM) and a gradient-reversal adversary to suppress confounds, was trained separately on data from 133 switches and 75 crossings, collected over one year of in-service monitoring. Wear risk scores were computed based on embedding anomalies and validated against manual wear measurements from 28 switches and 14 crossings. The final model achieved a moderate wear-rank correlation for crossings (ρ = 0.631) and a weak yet statistically significant correlation for switches (ρ = 0.311). The results demonstrate that self-supervised vibration embeddings can support reliable, scalable, and label-free prioritisation of S&C maintenance.