Discovering Riding Phases via Unsupervised Analysis of Motorsport Telemetry

In the competitive context of motorsport, it is important to take decisions quickly and effectively to optimize the performance of vehicles on the track. Modern racing motorcycles are equipped with several sensors that generate telemetry data for performance analysis. During a riding session, many actions and events occur, including riding maneuvers, vehicle movements and unexpected situations. A variety of behaviors must be identified and interpreted. Furthermore, telemetry channels are inherently time-dependent and each of them is correlated with the others. The result is a set of multiple signals with additional complexity that must be inspected together. These factors make the work of experts a time-consuming and demanding process. To provide practical support to race engineers, this thesis explores a solution aimed at capturing both temporal dependencies and cross-signal correlations within telemetry data. The core of our methodology is the application of multivariate unsupervised clustering to deal with the challenge of unlabeled telemetry data. We introduced this framework to automatically segment race laps into interpretable sub-sections, each corresponding to a distinct riding phase. Different algorithms were tested and compared, including both traditional clustering methods and approaches designed to capture temporal correlation. To effectively capture vehicle dynamics and maximize semantic consistency in the resulting clusters, feature engineering and different training strategies were involved in the pipeline. The obtained phases were validated through a real-world case study in collaboration with the 2WheelsPoliTo racing university team from Polytechnic of Turin. Experimental evaluation demonstrates that accurate channel selection plays a crucial role in segmentation quality and methods focused on time correlation produce outcomes closer to human expert analysis. Five distinct and recurrent riding phases have been identified: Straight, Braking, Corner Entry, Corner Exit and High Speed Cornering. The strength of this approach lies in its ability to exploit correlations among signals. This is what allows the framework to capture the underlying patterns of riding phases, enabling comparative analyses to support decision-making in racing environments.