The Study Of Full Waveform Inversion In Time-domain Based On Local-scale Matching Of Wavefield

Seismic exploration is the only geophysical method that has both large exploration depth and high resolution.It is recognized as the preferred exploration method in the field of deep resource exploration.Recently,the economy and society of our country are developing rapidly,and the corresponding demand for resources such as petroleum and minerals is also increasing year by year.With the location and reserves of shallow resources are basically proven,the focus of resource exploration has gradually shifted to deep complex geological structures.Under this circumstance,our country has put higher requirements for the depth and accuracy of seismic exploration.The main function of seismic exploration is to realize the proof of subsurface geological structure and lithology information after a series of data processing based on the seismic signals recorded by receivers deployed on the surface.Therefore,it is particularly important to obtain the distribution of physical parameters of subsurface media through seismic data.Among many physical parameters,the propagation velocity information of seismic waves in different media is a key factor that determines the quality of migration imaging.At present,the widely used subsurface medium velocity modeling methods are velocity analysis,tomography and full waveform inversion.Among them,velocity analysis and tomography only use the travel time information of seismic wave and ignore many other important information.Therefore,the results of velocity modeling based on velocity analysis and tomography cannot meet the needs of high-precision seismic exploration.Full Waveform Inversion(FWI)technique uses all the information of the pre-stack seismic wave field(including kinematics and dynamics information such as amplitude,phase,travel time,etc.)to achieve high-precision velocity modeling of underground media.In addition to modeling velocity parameters,full waveform inversion can also perform high-precision modeling of parameters such as density,quality factor,and anisotropic parameters.However,full waveform inversion is a highly nonlinear problem,which can be converted to a local optimization problem under the conditions of Born approximation,and this causes full waveform inversion tend to converge to a local extreme value.The cycle skipping phenomenon in the FWI process is the direct cause of FWI falling into the local extreme value.The main reasons for the cycle skipping problem are the lack of low-frequency components in the seismic data and the large differences between the initial velocity model and the subsurface true velocity model.In addition,the noise and wavelet inaccuracy in the field data also have a great influence on the inverted results.By studying the matching of wavefields in local-scale to deal with the complexity of seismic data changes over time,this paper solves the cycle skipping problem in full waveform inversion from the perspective of improving the degree of waveform matching and adaptive data identification.Firstly,this paper proposes a waveform local travel time correction method,which can improve the noise immunity of the algorithm while mitigating the cycle skipping problem.Then,this paper proposes new methods based on adaptive data recognition to solve the cycle skipping problem.These methods solve the problem without sacrificing computational efficiency,and can be combined with the encoded multi-source strategy to further reduce the calculation time.The methods proposed in this paper and the main achievements obtained can be summarized as follows:From the perspective of the traveltime difference between the observed and calculated data,and the basic idea of the digital image correlation method is used for reference,the waveform local travel time correction method is proposed to solve the cycle skipping problem.By studying the traveltime difference between the observed and synthetic data at different time samples,it is found that the traveltime difference increases with the increase of time,and the traveltime difference has polarity.A method to correct the traveltime of the synthetic data with the original observed data as the standard is proposed.The purpose is to reduce the traveltime difference between the observed and synthetic data in each time period to half a cycle to solve the cycle skipping problem.A rectangular time window is used to intercept observed and synthetic data.Calculating the traveltime difference of the waveform in the window and time shift the synthetic data in the window,and then slide the time window to correct the traveltime of the waveform in the next window until the time window covers all time samples.Discussed and concluded the optimal time window length and step length selection strategy,so that this method has greater general applicability.Due to the use of the cross-correlation algorithm,this method has strong noise immunity.This method is a kind of data preprocessing method.Compared with the traditional FWI,the local traveltime correction process barely increases the amount of calculation.In addition,the encoded multi-source strategy can further improve the computational efficiency.From the perspective of data adaptive recognition and the theory of incremental sign correlation in the image matching field,a full waveform inversion method based on amplitude increment coding is proposed,which greatly alleviates the cycle skipping problem.The waveform variation trend and the polarity of the waveform amplitude will be different between the cycle skipping events in different time periods.According to the relative amplitude value of different time samples,the amplitude increment coding is performed on each time sample and the amplitude polarity is introduced as a constraint item can adaptively identify the part where the observed and synthetic data match well and the part where the cycle skipping occurs.The coding matrix is constructed by performing amplitude increment coding on the observed and synthetic data,and then the data identification criterion is set according to the coding matrix to construct a binary zeroing matrix.Multiplying the zeroing matrix by the synthetic data can remove the cycle skipping events to reduce its interference on the gradient.The optimal number and the position of the time samples used for coding are given and proved.Numerical tests prove the effectiveness of this method for solving the cycle skipping problem.Explained the solution of this method when the observed data contain noise.From the perspective of data adaptive recognition,learned from the application of deep learning in the field of image matching that the convolutional layer in the convolutional neural network can extract image features,a full waveform inversion method based on convolutional coding is proposed,which effectively alleviates the cycle skipping problem in FWI and has strong noise resistance.According to the advantages that different convolution kernels in the convolutional neural network can extract different features of the image,Gaussian one-dimensional convolution kernels of different lengths are used to extract features from the observed and the synthetic data trace by trace.The extracted features represent the comprehensive response of wavefield amplitude,phase,and traveltime within the time range of a convolution kernel length,and the convolution value calculated for each movement of the convolution kernel is taken as the feature value of the time sample when it is aligned with the center point of the convolution kernel.Encoding each time sample according to the polarity of its feature value.By comparing the similarities and differences of the convolutional coding values of the same time sample between the observed and synthetic data,the parts of the synthetic data that do not match the observed data can be identified.By constructing an attenuation matrix,the identified data can be attenuated to reduce the interference of the data with cycle skipping on the gradient.The optimal number and length of convolution kernels are given and proved.Numerical tests have proved the effectiveness of this method for solving the cycle skipping problem and its strong anti-noise ability.