Real-time UAV motor fault detection with dual MFC piezoelectric sensors using feature-based and raw-data deep models
Artykuł w czasopiśmie
MNiSW
200
Lista 2024
| Status: | |
| Autorzy: | Stęgierski Rafał, Ambrożkiewicz Bartłomiej, Syta Arkadiusz, Ołdziej Daniel, Koszewnik Andrzej, Ambroziak Leszek, Dzienis Paweł, Karczmarek Paweł, Kiersztyn Adam, Dolecki Michał, Smoliński Konrad, Żmudzińska Alicja |
| Dyscypliny: | |
| Aby zobaczyć szczegóły należy się zalogować. | |
| Rok wydania: | 2026 |
| Wersja dokumentu: | Drukowana | Elektroniczna |
| Język: | angielski |
| Wolumen/Tom: | 267 |
| Numer artykułu: | 120531 |
| Strony: | 1 - 15 |
| Impact Factor: | 5,6 |
| Web of Science® Times Cited: | 0 |
| Bazy: | Web of Science |
| Efekt badań statutowych | NIE |
| Finansowanie: | The research was financed from the internal Fund of the Scientific Discipline (Grants no. FD-20/IT-3/004, FD-20/IT-3/002, FD-20/IT3/049, FD-20/IT-3/053). This work is supported by the University Work no WZ/WM-IIM/4/ 2023 of the Faculty of Mechanical Engineering, Bialystok University of Technology (A. Koszewnik, P. Dzienis, D. Ołdziej, L. Ambroziak). |
| Materiał konferencyjny: | NIE |
| Publikacja OA: | NIE |
| Abstrakty: | angielski |
| Real-time health monitoring is critical for Unmanned Aerial Vehicles (UAVs) operating in safety-critical and rapidly changing environments, where early detection of incipient faults can prevent mission loss or airframe damage. We instrument an octocopter with two piezoelectric patches mounted on the airframe and show that these high-bandwidth sensors provide rich, localized vibration signatures that enhance diagnostic sensitivity beyond traditional telemetry. MFCs capture structural response directly at the source, improving signal-to-noise for subtle faults while adding negligible mass and power. To safely and repeatably study failure modes in flight, we simulate motor degradation by perturbing the PWM. This protocol emulates partial torque loss, desynchronization, and wiring/intermittency, producing realistic transients without risking hardware damage. We compare two learning pipelines: (i) a feature based model that classifies windows of handcrafted time/frequency descriptors, and (ii) a raw data deep model that learns features end-to-end from synchronized sensor streams. For deployment, we stabilize decisions with a rolling median filter, chosen over mean filtering for its robustness to impulsive noise and outliers and its ability to preserve event onsets with minimal latency. Post-processing yields consistent gains across metrics; for example, accuracy improves from 0.9973 to 0.9995 for the feature based model and from 0.9910 to 0.9989 for the raw data model, alongside concurrent increases in precision, recall, and F1. These results demonstrate that combining airframe-mounted piezoelectric sensing with lightweight AI and robust temporal smoothing enables reliable, real-time UAV fault detection. |
