Fly Aware: Lightweight Real-Time Anomaly Detection and Classification in UAV Telemetry

Unsupervised anomaly detection in UAV telemetry is essential for ensuring safe and reliable autonomous flight. This paper presents a sequence-to-sequence LSTM autoencoder with a bidirectional encoder and Bahdanau-style attention decoder to capture short-term temporal dependencies in multivariate telemetry data. The model is trained on fixed-length sequences extracted from MAVLink logs containing 17 key features related to attitude, velocity, and position. To increase data diversity and coverage, we generated additional training sequences through high-fidelity flight simulation. A custom QGroundControl plugin was developed to efficiently filter and log only the most relevant MAVLink packets, enabling streamlined and targeted data acquisition. By learning compact latent representations, the autoencoder distinguishes nominal behavior from anomalies using reconstruction error. For deployment, the model is compiled to ONNX and served via a FastAPI-based inference system with WebSocket support for real-time monitoring. Sensitivity analysis reveals which sensor modalities most influence reconstruction quality. Our system demonstrates competitive performance on both simulated and real-world datasets, offering a scalable, interpretable, and lightweight solution for UAV anomaly detection.