Lightweight UAV Propeller Fault Detection Through Audio Signals Measurements

With the advance of aerial technology, Unmanned Aerial Vehicles (UAV) are applied in many industrial and civil fields. As UAVs demonstrate their effectiveness in many contexts, their safety needs to be improved to avoid accidents and danger-ous situations. To this aim, the detection and the classification of occurred damages play a key role: once a fault is revealed, a countermeasure must be applied on time. To accomplish this, in this work a methodology based on the classification of audio signals emitted by the UAV is proposed to detect two different types of fault on a drone propeller. The classification technique is developed by exploiting audio signals acquired in a certified semi-anechoic chamber to ensure the traceability of audio measures. The lightweight Machine Learning (ML) architecture used is designed to be deployed on low cost embedded units, for future adoption on board a drone, and shows very interesting results: a mild fault on a propeller is correctly classified with an accuracy higher than 81% by the quantised ML model, with just a 486.0 ms execution time on the target, low-cost, microcontroller.