Real-time Fault Detection on Small Fixed-Wing UAVs using Machine Learning

Murat Bronz (1) , Elgiz Baskaya (1, 2) , Daniel Delahaye (1) , Stéphane Puechmorel (1)

1 ENAC - Ecole Nationale de l'Aviation Civile
2 ENGIE Ineo - Safran RPAS Chair

In this study, we have highlighted the main challenges of real-time fault diagnosis on small scale fixed-wing UAVs. The feasibility of real-time fault prediction has been shown in real flight conditions experiencing noisy measurements, communication limitations, and wrapped wing structure that breaks the geometric symmetry. A total of eleven flight logs have been recorded and shared publicly for future potential use by other researchers on fault and anomaly detection. Our proposed method uses a data driven algorithm, SVM, in order to classify the behavior of the vehicle in nominal flight phase and faulty phase. Feasibility of a basic binary classification is shown, despite the well-known over-fitting problem caused by limited data. We have shown that geometrical imperfections that are common in small UAVs can cause particular effects on the prediction performance, and we used it in our advantage to improve the detection on multi-class classification. The SVM algorithm with proposed feature trajectories was capable to detect variation of loss of control effectiveness faults up to an accuracy of 95% in real flights. The data-set and all related programs can be downloaded from: (https://github.com/mrtbrnz/fault_detection).

Mots clés: Real-time Fault-Diagnosis; SVM; Paparazzi; UAV; Drones; Real-time Fault-Detection; Machine Learning

Domaines: Machine Learning [stat.ML]

Murat Bronz, Elgiz Baskaya, Daniel Delahaye, Stéphane Puechmorel. Real-time Fault Detection on Small Fixed-Wing UAVs using Machine Learning. DASC 2020 AIAA/IEEE 39th Digital Avionics Systems Conference, Oct 2020, San Antonio, United States. pp.ISBN:978-1-7281-8088-5, ⟨10.1109/DASC50938.2020.9256800⟩. ⟨hal-03018053⟩