A low-cost wireless accelerometer approach to UAV propeller fault detection

Abstract Propeller failure in flight is one of the most safety-critical failure modes for small multirotor unmanned aerial vehicles (UAVs). It is sudden, generally undetectable on conventional telemetry and often catastrophic. This study addresses the problem through a simple onboard sensing scheme. A wireless triaxial accelerometer (ADXL345) mounted on the airframe streams vibration data to a ground station via an ESP32 microcontroller over Wi-Fi. Seven propeller conditions were investigated: one healthy reference and six deliberately introduced defects of varying severity. Each flight followed the same protocol consisting of hovering, controlled manoeuvring and harsh manoeuvring, with a nominal duration of 60 s per phase. After interquartile-range (IQR) based outlier removal, 27 statistical features were extracted from 10 s sliding windows (180 windowed samples in total, with two minority classes of fewer than 15 samples each) and supplied to four classical classifiers: random forest (RF), support vector machine (SVM), k k $k$ -nearest neighbors (kNN) and decision tree (DT). RF achieved the best performance. For the binary healthy-versus-faulty task, RF reached 94.44% test accuracy and 96.67% mean accuracy under five-fold cross-validation. For the seven-class task, RF reached 79.63% test accuracy (91.67% CV). The binary result indicates the potential to support an in-flight return-to-launch decision rule after validation under broader flight conditions, whereas the seven-class task remains bottlenecked by dataset size rather than by classifier choice. The implications of this gap and the path towards onboard deployment on a Pixhawk-class flight controller are discussed.