Anomaly Detection in Railway Radio Signaling: Synthetic Data Generation and AI-Based Models

The increasing reliance on radio-based communication in modern railway systems, combined with stringent safety requirements, makes the design and validation of Wi-Fi networks a critical and complex task. In the context of the Torino Linea 1 metro line, even minor anomalies in radio signal performance can significantly impact operational safety and network reliability. This thesis presents the development of an automated Radio Network Anomaly Detection tool designed to assist network designers during the early stages of Wi-Fi planning. The tool leverages an ensemble of Artificial Intelligence techniques, including Autoencoder, Gaussian Mixture Models, and Isolation Forest, to identify anomalies in radio signal. A key contribution of this work is the creation of a synthetic data generation pipeline. Multiple models were extensively tested to determine the most effective method for producing high-quality, anomaly-free data, which is scarce in early design phases. The combination of a Tabular Variational Autoencoder (TVAE) with a rejection sampling strategy, guided by a Random Forest discriminator, emerged as the most effective pipeline, significantly enhancing the performance of anomaly detection models. The experiments show that the Autoencoder model achieves the highest performances. Overall, the developed tool provides network designers with a reliable solution for evaluating radio communication performance, reducing design time, minimizing costly rework, and ultimately contributing to the construction of safer railway infrastructures.