Research Of Seismic Data Denoising Method Based On Self-Supervised Learning

Seismic exploration technology is one of the important methods to obtain information about underground strata structure,and is an effective mean to search for underground mineral resources such as oil and natural gas.Over time,shallow and easily exploitable oil and gas reservoirs have been exhausted,and the focus of petroleum exploration has gradually shifted to developing deep and complex oil and gas reservoirs.However,seismic data from complex oil and gas reservoirs often contains extremely low signal-to-noise ratios;It has caused great difficulties in the subsequent interpretation and analysis of seismic data.However,traditional methods cannot achieve good noise reduction results when processing seismic data of complex oil and gas reservoirs.With the rapid development of artificial intelligence,methods based on deep learning have achieved great success in fields such as computer vision.Therefore,many geophysicists use deep learning methods for seismic data noise reduction.Due to the lack of clean seismic data in the field of seismic exploration,this thesis proposes two self-supervised learning methods for seismic data noise reduction.The specific work is as follows:(1)In the self-supervised learning seismic data denoising method,using odd traces of single section seismic data as input and even traces as pseudo tags to construct training data is a classic and effective method for training seismic data denoising models.However,this method does not take into account the differences between odd and even traces,making the model ignore some structural information of odd traces during training and focus on constructing the stratigraphic characteristics of even traces.This results in the method suppressing seismic noise while suppressing some peak information that reflects the underground structure,which leads to an over-smooth problem.To address this issue,this thesis proposes a regularization term that adds the difference information between odd and even channels,which can offset the impact of odd and even channel differences on model training under ideal noise reduction conditions,and solves the existing over-smooth problem.The proposed regularization term improves the signal to noise ratio in the experiment of suppressing synthetic noise;On real data,the proposed method preserves a higher signal amplitude while suppressing real noise.(2)Compared to single profile structural training data pairs to train depth learning models,multiple profiles can contain more stratigraphic structure information.Therefore,we propose a method to suppress noise using adjacent similar seismic profiles.Aiming at the problem that current seismic data noise reduction models cannot handle multiple profiles,this thesis introduces the Fast DVDnet network model,which can input multiple adjacent seismic profiles to reconstruct the middle seismic profile.However,Fast DVDnet has supervised training,and seismic data lacks a large amount of clean data to train Fast DVDnet.If multiple adjacent profiles are used as input and the middle profile is a pseudo label training model,there will be an identity mapping problem.Therefore,this article randomly occludes the middle section and reconstructs the randomly occluded portion through adjacent sections,avoiding the identity mapping problem.Compared with traditional methods,this method achieves higher signal-to-noise ratio in the experiment of eliminating synthetic noise;When suppressing real noise,cleaner noise reduction results and lower average local similarity values are obtained.