Toward Few-Shot Leakage Detection in Fresh Air Ducts: A CB-UNet-Enabled Data Augmentation Approach

Fresh air duct leakage detection is critical for environmental safety. However, existing work in both industry and academia mainly focuses on improving the performance of leakage detection models, with little consideration given to the challenge of obtaining leakage data in real-world engineering applications. It has significantly hindered the practical implementation of air duct leakage detection methods. To overcome this bottleneck, this paper proposes a data augmentation method based on a denoising diffusion probabilistic model. First, we use a short-time Fourier transform (STFT) to convert leakage acoustic signals into time-frequency spectrograms, capturing the abrupt spectral features at the leakage onset. Second, a conditional branch embedding CB-UNet model is introduced to generate multi-condition samples, incorporating residual connections and channel attention modules to ensure the detailed fidelity of the generated samples. Third, a lightweight EfficientDet network is employed for leakage detection on the augmented dataset to ensure inference speed. Finally, the method is validated on a real fresh air duct dataset for few-shot leakage detection tasks. Comparative results show that the proposed method outperforms other advanced data augmentation methods, achieving a leakage detection accuracy of 91.5% on the augmented dataset.