Research On Dispersive Noise Suppression Method Based On Generating Adversarial Network

In the field of seismic exploration,seismic data is the direct carrier of underground structural information,and the simulation method for seismic data is the fundamental core link of exploration work.The improvement of efficiency and accuracy in seismic data simulation directly promotes the theoretical research progress and practical application effects of methods such as conventional seismic data processing,migration imaging,and reservoir inversion.The finite difference method is widely used in seismic wave field simulation due to its high computational efficiency,but the difference method has numerical dispersion noise due to its own limitations.The one-step extension method based on polynomials can significantly suppress dispersion noise but has lower computational efficiency.By introducing the multiple angle formula to assist in wave field continuation,computational efficiency can be improved to a certain extent.Overall,the issue of the trade-off between computational accuracy and efficiency in earthquake wave field simulation still needs to be addressed.Based on the above situation,this paper proposes a dispersion noise suppression method based on the generation of countermeasures network.This method uses the finite difference seismic wave field with serious dispersion noise and the one-step continuation seismic wave field without dispersion noise as the training data.The neural network is constrained by the idea of cyclic consistency training,so that the neural network can accurately identify the dispersion noise.The final form is a denoising model that can significantly suppress scattered noise and quickly output calculation results.In this paper,the finite difference method and one-step continuation method are used to generate batch of seismic wave field data for generating countermeasures network training to form a denoising network model.The denoising effects were tested on simple velocity models,complex terrains such as grabens,horsts,horizontal layers,and Marmousi models.The results show that the method proposed in this paper has a similar suppression effect on scattered noise compared to the one-step continuation method,and the computational efficiency is improved by 2-4 times compared to the one-step continuation method.