ResNet-SE-Transformer: A Novel Leakage Detection Model for Fresh Air Pipelines

It is difficult to capture weak leakage signals due to inevitable noise in the real fresh pipeline leakage detection scenario. To solve the problem, this paper proposes a novel leakage detection model ResNet-SE-Transformer, which not only has the ability of feature extraction, but also has channel weighting and global modelling. Concretely, the proposed model first converts the audio signal into a spectrogram, uses ResNet to extract the leakage features, introduces the Squeeze-and-Excitation (SE) module to enhance the time-frequency of the leakage features, and then utilizes the Transformer's self-attention mechanism to model global temporal dependencies, thereby achieving an efficient separation of leakage signals and noise. To validate the effectiveness of the proposed model, we conducted three different types of experiments on the real-world leakage dataset collected from the engineering scenario of simulating Jinping underground laboratory. The experimental results show that the proposed model outperforms two existing baselines.