Intelligent UAV health monitoring: Detecting propeller and structural faults with MEMS-based vibration

In recent years, with the increasing prevalence of Unmanned Aerial Vehicles (UAVs), numerous studies have been published on mechanical fault detection and safety. However, most of these studies fail to adequately consider real environmental conditions and the behaviour of UAVs at different speed levels both in the air and on the ground. Furthermore, most of these studies are limited solely to propeller damage. Early detection of mechanical faults in UAV systems, including propeller damage, loosening of rotor screws, and UAV carrier arm screws, is crucial for system safety and operational efficiency. In this study, two cost-effective Microelectromechanical System (MEMS)-based models are proposed to effectively detect propeller damage, as well as rotor and carrier arm screw looseness, under real-world environmental conditions and at different rotor rotation speeds. These models are designed to perform fault detection both before and during flight based on the vibration data of a multirotor UAV. To achieve this, vibration signals were analysed using Fast Fourier Transform (FFT) to identify the operational frequency regions under both fault-free and faulty conditions. Additionally, eight different amplitude-frequency features within the significant frequency bands of the signals were examined and compared across different rotor speeds. For fault detection, the performance of Support Vector Machine (SVM), K-Nearest Neighbour (KNN), Decision Tree (DT), and Neural Network (NN) algorithms was evaluated. The model achieved a maximum accuracy of 99.40 % in detecting the severity of propeller damage. Furthermore, the study also investigated loosening conditions in the propeller rotor and UAV carrier arm screws, demonstrating that, in combination with propeller faults, other mechanical loosening problems can be detected with a maximum accuracy of 95.86 %, highlighting the superior performance of the proposed approach.

Abstract licence: CC BY-NC-ND

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.