Research On Noise Reduction And Fault Identification Of Seismic Data Based On Deep Learning

Deep learning has developed rapidly in recent years as a step up from neural networks,and has created a third wave of artificial intelligence(AI)with its many benefits.Deep learning is underpinned primarily by theoretical mathematical knowledge and computer technology.And is close to a semi-supervised mode of learning target approximation,often relying on large amounts of training data.Traditional algorithms are predominant in research algorithms for seismic exploration problems,but the direction of research is also gradually drifting towards AI.This paper focuses on the problem of seismic data noise reduction and fault identification using the strong feature learning capability and transferability of deep learning techniques,as described below:The U-Net residual network is a combination of feedforward denoising and coding-decoding structures.The U-Net residual network is a combination of feedforward denoising network and coding-decoding structure with symmetric connections for better feature fusion,and the proposed improved algorithms change the network structure on this basis.The proposed improved algorithm changes the network structure on this basis.For the fault identification problem,a Rotated-Unet network with inverse sampling blocks is proposed.The U-Net model with encoding-decoding structure has better fault identification performance for seismic data,but the encoding process tends to lose the underlying features while expanding the sensory field.For this reason,the inverse Rotated-Unet model is proposed to be used for fault identification,which differs from the U-Net model in that it incorporates a shallow network with upscaling and then downscaling before the encoding-decoding structure,with the aim of improving the completeness of the features.Experiments have been conducted to verify the conjectures in both of the above problems.After comparison,it is concluded that of the improved U-Net residual network algorithms proposed in this paper based on the seismic data noise reduction problem.The asymmetrically connected U-Net residual network has better noise reduction performance,and the Rotated-Unet network proposed based on the fault identification problem presents stronger generalization in the actual work area tests.