In railway management, it is essential to ensure that monitoring systems are both effective and efficient, promoting greater safety with minimized maintenance and operational costs. In this context, adopting an indirect approach such as Vehicle to Infrastructure (V2I) to identify track damage is more advantageous than an Infrastructure to Infrastructure (I2I) approach. This is due to the capability of a single vehicle’s ability to monitor an entire railway line. The purpose of this work is to propose a methodology based on machine learning techniques to detect varying levels of track irregularities using acceleration records measured at the axlebox, bogie, and carbody of a freight vehicle. First, numerical data are obtained from a multibody vehicle model developed in Simpack commercial software, considering a virtual on-board monitoring system. Then methodology for damage detection developed in this work comprises six steps (i) data acquisition, (ii) data segmentation, (iii) feature extraction, (iv) feature normalization (v) data fusion and (vi) feature discrimination. The measurements are transformed into alternative information by CWT, and, for the automatic geometrical track damage detection, an outlier analysis is performed. Furthermore, a sensitivity analysis is performed to analyse the influence of the number of sensors, allowing them to define an optimal number of sensors to obtain good results.