Deep Learning-Based Acoustic Fault Detection for Unmanned Aerial Vehicles

Abstract

Rapid expansion of the unmanned aerial vehicle (UAV) industry has brought operational reliability to the forefront as a critical bottleneck for widespread adoption. UAVs currently experience failure rates approximately 48 times higher than manned aircraft, with propulsion systems accounting for 27--38% of failures. Acoustic-based fault diagnosis offers unique advantages over traditional vibration-based methods, providing non-invasive, remote sensing capabilities ideal for weight-s UAV systems.This study introduces the application of the deep learning models in the field of UAV acoustic fault detection. Our classification approach employed an advanced LSTM-attention deep learning model to real-time identify three distinct drone conditions： normal, motor fault, and propeller fault using the K-Drone dataset comprising a total of 957,600 audio segments.The LSTM-attention model demonstrates superior performance, robust diagnostic capabilities, and real-time inference efficiency across all tasks. Our LSTM-attention approach with filter bank features achieves an F1-score of 98.93%, representing a 3.40% improvement over the state-of-the-art research outcomes. The standalone LSTM model with filter bank features achieves 98.56% F1-score, while attention-only approaches reach 98.13%. In addition, binary classification experiments show discriminative ability with 98.98% for motor fault detection and 99.96% for propeller fault detection. The novel model supports in-memory computing, achieving a latency of 1.59 milliseconds in real-time inference conditions.This research could be broadly applied across the drones, enhancing aviation safety, stability, and economic viability. The deep learning models may reduce operational risks, and financial losses. Beyond industrial benefits, it contributes to public safety in autonomous UAV systems.