Research On Desert Seismic Random Noise Modeling Based On Deep Learning

In the process of seismic exploration,due to the complex surface conditions and acquisition environment,the acquired seismic records are often accompanied by a large amount of random noise.These noises reduce the signal-to-noise ratio of seismic data and is a key factor affecting the quality of seismic data,which brings challenges to the interpretation of geological structures.Studying the generation mechanism of random noise in desert seismic exploration is helpful to deeply understand and characterize the complex nature of random noise,and provide theoretical support for further improving seismic data processing technology by utilizing the characteristics of random noise.The random noise in desert seismic data is mainly generated by natural external forces or human activities during the exploration process,and the dynamic process of its propagation in the near surface can be modeled as two-dimensional fluctuations excited by various noise source functions.Using the Green's function and other methods to solve the wave equation can simulate seismic random noise in desert areas.The random noise model derived from this empirical observation and theoretical analysis usually assumes that the surface medium is homogeneous,isotropic,and completely elastic.However,the actual seismic exploration medium conditions are complex,and the ideal random noise wave equation cannot fully reflect the actual surface medium properties.Aiming at this problem,this paper introduces the partial differential equation learning method to learn the dynamic characteristics of random noise from the random noise data of desert seismic exploration.The paper proposes a deep random noise wave equation neural network and solves random noise modeling and solving problems based on a data-driven approach.Firstly,a general form of nonlinear wave equation under complex medium conditions is established based on the two-dimensional wave equation of random noise,and a deep random noise wave equation neural network(RNWENet)is constructed based on it.RNWENet consists of a differential convolution layer,a symbolic regression neural network and a connection module with a jumper structure.The differential convolution layer,which can learn parameters,can approximate the differential operators of different orders and different directions of the wave equation,and obtain various order differentials of random noise.The symbolic regression neural network can learn the nonlinear response of each differential term of the wave equation,which enables the network to describe the dynamic process of the random noise wave equation evolving with time.In order to make RNWENet learn the long-term dynamic change of random noise data,multiple random noise wave equation network units are connected in series and each network unit is connected by a jumper connection module.RNWENet training uses a cost function with regular term constraints,uses the BFGS algorithm to solve the optimization problem,so as to learn the exact form of the random noise wave equation for desert seismic exploration.The synthetic random noise data of different media verifies the effectiveness of RNWENet in learning the random noise model.Then the dynamic characteristics of the field random noise data are learned through the network model.Then the random noise wave equation model obtained by training is used to simulate different kinds of noise in this dissertation.A random noise source function model is established according to the actual seismic exploration and acquisition environment,and then the RNWENet model is stimulated by wind blowing,near-field and far-field human noise sources to generate different types of random noise in desert seismic exploration.The influence of noise source parameters on noise properties is further analyzed to obtain a noise source function that conforms to seismic exploration characteristics.In the end,the three kinds of noise are superimposed to generate simulated desert seismic exploration random noise.The simulated noise spectral characteristics,statistical characteristics and spatial similarity are basically consistent with the actual random noise characteristics.The experimental results verify that the RNWENet model can be used to obtain a large number of simulated noises that conform to the random noise characteristics of actual desert seismic exploration.