Abstract
Unmanned aerial vehicles (UAVs) are cyber-physical systems equipped with multiple sensors for navigation and data collection, which have become increasingly prevalent in both military and civilian applications. Meanwhile, concerns over the security and reliability of UAVs have also grown due to the potential for internal system errors and external cyber-attacks. To address these challenges, recent research has focused on detecting anomalies in UAV flight data, typically using machine learning techniques to predict flight data based on historical data. However, these methods often ignore the prediction delay during status-changing periods, which can result in false alarms and impact the performance of UAVs. In this thesis, we aim to enhance the accuracy of anomaly detection and mitigate the impact of prediction delay. We explored approaches from two directions. The first approach is a hybrid detection method involving real-time data prediction and the second one combines convolutional neural networks with symbolic reasoning method for abnormal data pattern detection. We evaluate the proposed approaches on flight data collected from multiple UAV flight paths. Our evaluation results validate the effectiveness of our designs, which achieves both high anomaly detection accuracy and reliable recovery.