Current students

Railway tracks are prone to damages such as rail wear, cracks and misalignment, which are typically caused by a set of different stresses they can be subject to as: bending, shear, contact, thermal and residual stresses. ; Current railway maintenance strategies rely heavily on periodic geometry/wear measurements taken by Track Recording Cars (TRCs), which make a full pass over the railway network with a set of laser/camera sensors, and then alert the infrastructure manager if some readings report values outside of current safety standards’ thresholds. ; However, because this type of maintenance is reactive, the track inspection needs to be conducted over the entire network, which implies both high costs and high occupation time. For this reason, TRCs are run with low frequency, typically bi-monthly, which increases the risk of missing track geometry deviations that exceed the intervention thresholds. ; To address both the inefficient maintenance strategy and overall data scarcity, we aim to shift the current paradigm by exploiting accelerometer data measured with each commercial train run using various techniques from Deep Learning, focusing specifically on the adaptation of Large Language Models to the analysis of vibration signals. With this, we aim to enable continuous monitoring of the railway network to address geometry defects before they become critical, improving both comfort and safety.