Random Noise Attenuation of the High-Speed Train Induced Seismic Data Based on an Improved U-Net Network

Summary It is a promising means to use the high-speed train (HST) seismic data to prospect the subsurface structures along the railway. The existence of noise significantly degrades the imaging accuracy of the HST seismic data. We have developed a random noise attenuation method for the HST seismic data based on an improved U-Net network, which combines the self-attention mechanism and the residual connection of ResNet with the classic U-Net deep neural network. The proposed method constructs the dataset by numerical modelling of the HST sources, uses the noisy data as the input and the noise-free data as the labels, yields the denoised results through network training, and then realizes the separation of HST signals and noise based on the constructed U-Net network. The processing results on synthetic and real HST data suggest that the proposed method can effectively suppress random noise without damaging seismic signals.