A method based on an attention-guided multi-branch residual network for coupled noise suppression in distributed acoustic sensing data

Distributed acoustic sensing vertical seismic profile (DAS-VSP) is an emerging exploration technology. However, the strong energy-coupled noise generated by the vibration of optical cables inside the well casing interferes with the signals collected, reducing the signal-to-noise ratio of seismic data. To address this issue, we have developed a new method based on a U-Net for coupled noise suppression called attention-guided multi-branch residual network (AGMRN). In this method, to capture and enhance multi-scale features of the data, diverse branch blocks (DBB) and convolutional block attention modules (CBAM) are introduced into the U-Net architecture. Additionally, pre-activation residual units (Pre_ARU) are incorporated into the network to reduce the number of parameters in each convolutional layer and improve the efficiency of network training. Furthermore, to ensure the accuracy of the predicted noise, a hybrid loss function combining the L1 norm and Multi-scale structural similarity index measure (MS-SSIM) is employed. The proposed method is based on a large training dataset that includes pairs of data with and without coupled noise. To diversify the noise in the training data, we use the proposed method of synthesizing coupled noise to generate various types of coupled noise. After training the neural network with the training data, the trained network can identify and suppress the coupled noise in DAS-VSP data.