Acoustic Recognition of Unmanned Aerial Vehicles in a Gas Discharge Noise Environment Based on Adaptive Time–Frequency Denoising and Multichannel Feature Fusion

Background: The acoustic recognition accuracy of Unmanned Aerial Vehicles (UAVs) is severely degraded by strong gas discharge noise in power substations and overhead transmission lines. Methods: This paper proposes a UAV acoustic recognition method based on adaptive time–frequency denoising and multichannel feature fusion. The method integrates adaptive Wiener filtering with time–frequency masking techniques, employs a six-channel microphone array for collaborative noise reduction, and constructs a CNN-LSTM hybrid deep learning model for UAV type recognition. Results: Experimental results demonstrate that the two-stage denoising achieves a signal-to-noise ratio (SNR) improvement of 19.08 dB and a spectral fidelity of 0.89. Under extreme noise conditions at −10 dB SNR, the proposed method achieves 87.4% recognition accuracy, representing a 23.7% improvement over the traditional MFCC-SVM method and a 12.3% improvement over single-stage denoising strategies. In unknown environments, the recognition accuracy remains at 87–91%, exhibiting strong robustness and generalization ability. Conclusions: The proposed method effectively mitigates strong discharge noise interference and provides a reliable acoustic recognition solution for UAV intrusion detection in power facilities.