Condition Monitoring of Railway Tracks Using Instrumented Wheelsets: Development of Physics-Based Models and Time-Frequency/Data-Driven Methods

This thesis investigates whether IWS can serve as an effective and standalone diagnostic technology for track condition assessment, clarifying their capabilities, limitations, and the type of information they can provide beyond conventional sensors. A combination of finite-element modelling, state-space simulations, advanced signal processing, and machine learning is adopted to characterise the sensitivity of wheel–rail contact forces to long-, short-, and very-shortwavelength rail irregularities. The work demonstrates that contact-force measurements can detect and localise defects such as dipped welds, squats, corrugation, and ballast voids, and can provide meaningful insights into vertical alignment. A validated numerical model confirms an effective bandwidth above 200 Hz for single-bridge configurations, enabling the detection of fast-evolving dynamic phenomena. Time–frequency and wavelet analyses show that IWS outperform optical systems in capturing high-frequency irregularities, while a simplified 3-DOF model successfully reconstructs mean vertical geometry with good agreement to Track Recording Vehicle (TRV) reference data. Furthermore, a machine-learning framework based on Isolation Forest proves effective for railhead fault detection when using both space- and space–frequency-domain features. These results indicate that IWS can complement, and in some scenarios partially substitute, traditional inertial and optical solutions, offering actionable diagnostic information for infrastructure managers and certification bodies.