End-to-End Gas Leakage Acoustic Signal Noise Reduction Based on Improved Sandglasset Model

For addressing background noise interference during gas leakage detection, an end-to-end gas leakage acoustic signal noise reduction method based on the improved Sandglasset model is proposed. This model, tailored for enhanced modeling and sound source separation, incorporates the Temporal Convolutional Network (TCN) module and Coordinate Attention (CA) module, replacing Bi-directional Long Short-Term Memory (Bi-LSTM) and Multi-headed Attention (MA). TCN excels in capturing local information and boosting time series data modeling, while the CA module specializes in inter-segment sequences modeling and feature extraction from various time sequences, yielding superior global modeling capability. The process involves encoding the collected gas leakage acoustic signal with a one-dimensional convolutional encoder to generate the feature sequence, which is then process by the improved Sandglasset model and subsequently decoded to reconstruct the noise-reduced signal. The experimental results demonstrate that the method effectively isolates the background noise signal. Post-noise reduction, the SI-SNR and SDR reach 5.93dB and 8.17dB, respectively, thereby minimizing interference with further signal analysis and identification.