Real-World On-Board Uav Audio Data Set For Propeller Anomalies

Abstract

Detecting propeller damage in Unmanned Aerial Vehicles (UAV) is a crucial step in ensuring their operational resilience and safety. In this work, we present a novel real-world audio data set of propeller anomalies, and use several deep learning models to classify the damage. This data set consists of more than 5 hours of audio recordings, covering all configurations of intact and broken propellers in a UAV quadcopter. A microphone array was mounted onto a UAV, and numerous autonomous indoor missions were flown. Our on-board setup has provided clean audio recordings containing little background noise. We have developed classification models for this data set, using different deep learning architectures: Deep Neural Networks (DNNs), Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM), and Transformer Encoder (TrEnc). We conclude that the TrEnc outperforms other architectures, having 11k parameters, .57M Flops, 98.30% accuracy, .98 precision, and .98 recall. Finally, we make our data set publicly available here <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">⊙</sup> .