Research On Robust Seismic Reflection Pattern Analysis Method Based On Data Uncertainty Conditions

Seismic reflection pattern analysis is a method to obtain geological structure information by analyzing the differences between different seismic reflection signals.Seismic reflection pattern analysis is divided into two types: pre-stack seismic reflection pattern analysis and post-stack seismic reflection pattern analysis.Both types of reflection pattern analysis are affected by noise,which causes uncertainty in seismic data.In order to reduce the impact of this uncertainty on the seismic reflection pattern analysis.This thesis conducts a study on the robust seismic reflection mode analysis method under the uncertainty of seismic data from both pre-stack and post-stack directions.The noise in the pre-stack seismic signal has a severe impact on the seismic reflection pattern analysis is Erratic noise and Gaussian random noise.Since the traditional denoising methods cannot fundamentally suppress the Erratic noise,the objective of this thesis is to propose a new algorithm to suppress the Erratic noise and Gaussian random noise from the noisy pre-stack seismic data and then complete the robust pre-stack seismic reflection pattern analysis by a self-organizing mapping clustering algorithm.The main objective for post-stack seismic data is to propose a denoising method for footprint noise and Gaussian random noise and then complete the post-stack seismic reflection pattern analysis by clustering after the denoising is completed.Based on these two requirements,two methods of robust deep convolutional autoencoder and robust U-net network are targeted in this thesis.The specific innovative research is as follows.(1)A deep convolutional autoencoder is introduced for Erratic noise and Gaussian random noise,which provides a network architecture for noise removal.The Welsch norm is introduced as the loss function of the network for the first time in the field of seismic data denoising,and this norm can give smaller weights to significant errors in the iterative process,and this feature gives it the ability to suppress Erratic noise,The Total Variation regular term is also introduced to enhance the denoising ability of the deep convolutional autoencoder,and the Total Variation regular term can also enhance the local smoothing ability of the network for two-dimensional seismic data.After denoising the seismic data,the clustering of the seismic data is continued,and the robust pre-stack seismic reflection pattern analysis is finally completed.(2)The U-net network is introduced for footprint noise and Gaussian random noise to construct a post-stack seismic data processing network architecture.Due to the special network structure of the U-net network,it has a powerful information processing capability and requires only a small data set to complete the training of the network.This is certainly suitable for the post-stack seismic data,which is much smaller than the prestack seismic data.It is also the first time introducing a Unidirectional Total Variation regular term joint mean square error as the loss function in the U-net network to remove the noise more thoroughly and build a network model RU-net with noise attenuation capability of post-stack seismic data.Combining RU-net with the SOM algorithm can improve the robustness of post-stack seismic reflection pattern analysis.In this thesis,experiments are conducted in two-dimensional noisy seismic data,prestack noisy data,and post-stack noisy data.The effectiveness and superiority of the method proposed in this thesis are verified by comparing it with existing methods,which promotes the development of the field of robust seismic reflection pattern analysis methods under data uncertainty.