Zainal Arifin, Lilik Anifah, I.G.P. Asto Buditjahjanto
Autonomous fixed-wing UAVs are now a cornerstone of precision agriculture, especially for large-scale mapping. But even small aerodynamic disturbances or slight shifts in the center of gravity (CG) can introduce subtle instabilities in flight. These small changes may not seem dramatic, yet they can reduce mapping accuracy and potentially affect operational safety. In this study, we present an unsupervised deep learning framework for detecting flight anomalies using multivariate telemetry data (specifically, roll, pitch, and altitude signals). Three autoencoder architectures were tested: LSTM, GRU, and a 1D-CNN. The evaluation was based on telemetry collected from 13 real autonomous agricultural flights, where CG variations were intentionally controlled to create realistic instability conditions. Anomalies were identified using window-level reconstruction errors, and performance was evaluated through threshold-independent ROC-AUC analysis. All three models perfectly classified severe anomalies, achieving accuracies and F1 Scores of 1.00. When it came to milder CG deviations, differences became more apparent. The GRU autoencoder showed the strongest sensitivity, reaching an ROC-AUC of 1.00 and detecting over 98 of mild anomaly cases. The LSTM achieved near-perfect separability (AUC = 0.99), while the 1D-CNN delivered solid performance (AUC = 0.97) with the added advantage of lower computational complexity. Overall, the results confirm that recurrent autoencoder architectures are highly effective for real-time UAV anomaly monitoring. Among them, the GRU offers the best balance between detection performance and model efficiency, making it particularly well-suited for embedded agricultural UAV applications. This is an open access article under the CC BY-NC-SA 4.0 license. CC BY NC SA
Department of Electrical Engineering, Faculty of Engineering, State University of Surabaya, Surabaya, 60231, Indonesia
