LPS-IRM: A Novel Denoising Method for Leakage Detection in Fresh Air Ducts Under Low-SNR Conditions

To capture invisible leakage signals hidden in high water flow noise in the leakage detection of fresh air ducts, this paper proposes a novel noise reduction method LPS-IRM (Log Power Spectrum-Ideal Ratio Mask) which can extract weak leakage signals under low signal-to-noise ratio (SNR) conditions. The proposed method first uses Short-Time Fourier Transform (STFT) to convert the time-domain leakage signals into a time-frequency domain representation to preserve the phase information of leakage signals. Then, a deep neural network is trained using the LPS of the time-frequency domain signals to obtain IRM that can preserve the energy of the leaked signals. Finally, the obtained IRM is applied to the amplitude spectrum with noise that is further combined with the original phase information using Inverse Short-Time Fourier Transform (ISTFT) to reconstruct, resulting in denoised leakage signals. The proposed method was validated on a real-world fresh air duct leakage dataset. The experimental results show that the logarithmic spectral distortion of LPS-IRM is significantly improved under different leakage conditions. More specifically, LPS-IRM reaches a 54.59% increase at a low SNR -10dB. Moreover, compared to traditional spectral mapping methods, there is an improvement of 26.86%, whereby it indicates that LPS-IRM can be applied to noise reduction in practical engineering applications of new air duct leakage detection.