Research On Noise Attenuation Method Of Seismic Data In Time-Frequency Domain Based On Deep Neural Network

Noise attenuation is one of the key steps in seismic data processing.Compared with traditional denoising methods,the deep learning-based noise attenuation method for seismic data is more efficient,and the denoising result has a higher signal-to-noise ratio.Existing denoising methods based on deep learning usually process seismic data in the time and space domain,but the feature difference between the effective signal and noise in the time-frequency domain is more obvious,which is beneficial to network training and noise attenuation.Combining deep learning technology with 2D discrete wavelet transform and Curvelet transform,we propose two random noise attenuation methods for seismic data.The main research contents are as follows:(1)Taking advantage of the sparsity and multi-scale of seismic data in the 2D wavelet domain,and combined with the 2D discrete wavelet transform and the U-Net network,this thesis propose a random noise attenuation method of seismic data in the2 D wavelet domain based on the U-Net network(Dwt-U-Net).we use the 2D wavelet coefficients of seismic data as the network input and output for the network training,and the denoising results are obtained by reconstructing the denoised wavelet coefficients.Tested on simulated and actual seismic data,and compared with different denoising methods,the results show that the denoising results of the Dwt-U-net method have higher signal-to-noise ratio and fidelity under different noise levels.In addition,compared with the traditional U-net network denoising method,Dwt-U-Net reduces the network training time by half while improving the signal-to-noise ratio.(2)Taking advantage of the sparsity,multi-scale and directionality of seismic data in the Curvelet domain,and combined with the Curvelet transform and the attention UNet network,this thesis propose a random noise attenuation method of seismic data in the Curvelet domain based on the attention U-Net network(Curvelet-AUnet).Firstly,CBAM is added before downsampling in the traditional U-Net network to enhance the spatial attention and channel attention of the network.Secondly,we use the Curvelet coefficients of seismic data as network input and output for the network training,and the denoising results are obtained by reconstructing the denoised Curvelet coefficients.Finally,Curvelet-AUnet is applied to simulated and actual seismic data random noise attenuation,and compared with Curvelet transform,traditional time-domain U-Net denoising and Dwt-U-Net method.The test results show that Curvelet-AUnet has the best denoising result among several noise attenuation methods.