Hybrid Acoustic Fault Diagnosis in UAVs Using Wavelet Scattering Transform and Deep Learning

Unmanned aerial vehicles (UAVs) are increasingly employed in defense, agriculture, and logistics, where ensuring operational safety is critical. Propeller damages and rotor screw looseness represent two major fault types that can compromise flight reliability. Previous studies have typically addressed these faults separately, often relying on vibration analysis or conventional acoustic features. This paper introduces a hybrid model that combines Wavelet Scattering Transform (WST) and Long Short-Term Memory (LSTM) to simultaneously detect both fault types using microphone-recorded acoustic signals. Unlike traditional Fourier-based approaches or deep learning models utilizing Mel-Frequency Cepstral Coefficients (MFCCs) or spectrograms, the proposed framework leverages WST to extract deformation-stable, multi-resolution features, which are then modeled through an LSTM network. The model was trained and tested on a dataset comprising 750 one-second acoustic segments, approximately balanced between problem-free and faulty classes using stratified sampling and cross-validation. By integrating time–frequency-based multi-layered features into a time-sequential deep learning framework, the proposed model achieves reliable classification of both propeller faults and rotor screw looseness, reaching a high accuracy of 98.93%. These results highlight the potential of the WST-LSTM framework as a robust and innovative solution for UAV fault diagnosis, particularly in acoustic-based monitoring scenarios