Research On Multi-frequency Convolutional Neural Network Inversion

With the requirements of high-accuracy seismic exploration,the reservoir inversion technique of thin beds and complex reservoirs with large lateral variations is becoming more and more important.However,the current inversion methods are mainly based on the convolutional model,so they have poor resolution.The fine characterization of the complex reservoirs,for example looking for the thin interbedded clastic and Heterogeneous carbonate reservoir,requires high-resolution seismic inversion results.Now the widely used inversion method in the industry is sparse pulse inversion method that makes the seismic data for the hard data.But its vertical resolution can not reach the requirements of fine characterization of the reservoirs.The stochastic inversion method which makes the logging data for the hard data has a high vertical resolution in the inversion profiles.However,the horizontal distribution and the changes of the stratigraphic occurrence have greater access with the seismic profile.As a result,this method didn't get the promotion.Therefore,how to improve the resolution of the inversion results becomes one of the urgent problems.As one of the algorithms to solve the problem of non-linear mapping,the neural network can fully exploit the lithology and physical properties which are contained in the seismic data.So we can establish the non-linear mapping relationship between the seismic data and the reservoir properties parameters.However,the mapping relationship between the input and output of the neural network which is currently used in the seismic inversion is the inner product operator,which limits the resolution of the inversion results.Based on this,this paper introduces the convolutional mapping operator to the neural network structure to replace the original inner product operator,which establishes the contact between the input parameters and the output parameters to a greater extent.Based on the convolutional neural network,this paper gives the optimization algorithm of the mapping operator and applies it to the seismic inversion.Because that this inversion method is totally driven by the seismic data and not limited by the convolutional model,the resolution of inversion result is greatly improved.For the convolutional neural network inversion method,if the original seismic profile itself has a higher resolution,the resolution of the corresponding inversion results will be improved.In this paper,with the help of the matching pursuit algorithm,we remove the strong reflection from the original seismic profile and extract the seismic profile which has the target enhancement to highlight the target layer and provide better raw material for the subsequent convolutional neural network inversion.Unfortunately,the enormous calculation of the current matching pursuit algorithm restricts its wide range of applications.On the basis of the complex fast matching pursuit algorithm,we perfect the calculation method of the instantaneous frequency and resolve the problem of the negative instantaneous frequency presence.Besides,in order to improve the operational efficiency of the matching pursuit algorithm,we use the exponential distribution function to fit the amplitude envelope to get the instantaneous frequency.Finally,we apply the convolutional neural network inversion method to the maximum likelihood AVO inversion.We can get the band-limited reservoir parameters profiles from the pre-stack seismic data and combine the matching pursuit algorithm to process the seismic data to highlight the target layer.So we can provide more perfect original data for the convolutional neural network inversion and present a complete maximum likelihood AVO inversion technique.