Development of a log analyzer framework for real-time and post-flight anomaly detection in PX4-based UAVs

The increasing use of Unmanned Aerial Vehicles (UAVs) in domains such as delivery, surveillance, and search-and-rescue highlights the need for reliable mechanisms to ensure flight safety. Failures in sensors, actuators, or navigation systems can compromise both mission success and vehicle integrity, underscoring the need for anomaly detection as a central component of the control architecture. The majority of existing approaches are based on supervised methods, which are limited by scarce anomaly datasets, resulting in poor reproducibility and weak generalization across platforms and missions. Unsupervised methods, while able to flag outliers, lack semantic grounding and therefore offer limited interpretability. Moreover, many solutions are computationally demanding for resource-constrained embedded hardware. As a result, post-flight analysis and real-time monitoring are often treated separately. To address these challenges, this thesis presents Log Analyzer, a hybrid anomaly detection framework for UAV platforms operating on the PX4 autopilot. The system integrates a rule-based layer, offering interpretable and context-aware detection of known failures, with an unsupervised learning layer able to detect statistical anomalies. Together, these layers provide both reliability and adaptability in diverse scenarios and vehicle configurations. The framework supports dual execution modes, unifying post-flight analysis with real-time monitoring, while a command-line interface ensures structured user interaction and anomaly reporting. The system has been validated through simulation, laboratory, and real flight tests, demonstrating the feasibility of transitioning anomaly detection from research prototypes to reliable components of autonomous flight operations.