A Fault Diagnosis Method for a Quadrotor UAV Based on Audio Signals

In complex flight environments, unmanned aerial vehicles (UAVs) are prone to various uncertain failures and damages during actual flight. Specifically, in the event of a propeller malfunction in a quadrotor UAV, it is crucial to quickly identify both the type and extent of the failure to prevent potential losses caused by loss of control. In this paper, the audio data is captured from the UAV's flight. To visualize the audio signal, short-time fourier transform (STFT) spectrograms and Mel spectrograms are used to extract the feature of the audio signals, respectively. The features from these two types of spectrograms are then combined together and the MobileNetV2 network is used to train the fused feature maps. Additionally, we introduce a spatial attention mechanism into the MobileNetV2 network to enhance its ability to focus on the most relevant features in the spectrograms, thereby improving the accuracy and effectiveness of fault diagnosis.