Anomaly Detection in MotoGP Telemetry Data

Anomaly detection methods aim to identify deviations from expected data patterns, often in the presence of few labels, severe class imbalance, and non-stationarity. In competitive fields such as MotoGP, maximizing analysis efficiency between sessions and being able to anticipate potential motorcycle malfunctions is crucial for manufacturers’ racing departments to remain competitive; this is one reason why motorsport has increasingly been approaching artificial intelligence in recent years. In this thesis, realized in collaboration with Aprilia Racing S.r.l., anomaly detection on high-frequency multivariate telemetry data will be discussed. Starting from known examples of anomalies in the RPM domain, the first step was to define a coherent set of channels that are likely to be involved in broadly similar types of anomalies. Then, the possibility of identifying such anomalies using methods that go beyond manual rules is investigated, with the goal of reducing analysis time and enabling the discovery of previously unseen cases. To this end, several semisupervised anomaly-detection methods have been trained and tested, first using machine-learning algorithms and then deep-learning approaches, in particular forecasting-based and reconstruction-based models. The resulting errors were subsequently converted into per-feature alerts, with the aim of flagging not only the anomalous time interval but also the specific channels involved. Particular attention was devoted to data preparation and to defining a training set that is low in (ideally free of) anomalies—an essential requirement for a reliable model response. At the end of the study, the results show that the best configuration achieves satisfactory performance not only on anomalies labeled a priori, but also generalizes to unseen sessions and conditions, identifying new anomalies with good stability and interpretability.