Autoencoder-based Real-time Thrust Anomaly Detection and HILS Verification for Multicopter UAVs

This study proposes a real-time thrust anomaly detection algorithm for multicopter-type unmanned aerial vehicles (UAVs) using an autoencoder-based semi-supervised learning approach. The performance of this system is verified through a hardware-in-the-loop simulation (HILS) environment. The proposed method detects anomalies without additional sensors by using only the onboard signals of the flight controller including the attitude, angular rate, and motor pulse-width modulation data. The autoencoder was trained solely on normal flight data and determines faults by monitoring reconstruction errors. The developed algorithm accurately detected an induced motor failure within 0.06 s while showing no false alarms during normal operation. The HILS system successfully replicated the dynamic behavior of the UAV and validated the real-time performance and reliability of the proposed model. This approach enables the efficient and safe pre-flight validation of fault detection systems. Future work will involve real flight tests to evaluate the robustness of the algorithm under various fault scenarios and environmental conditions.