Research On The Construction Method Of Intelligent Facies-controlled Impedance Model

In oil and gas data processing,model-based inversion is highly dependent on the initial model.Establishing an initial model that is more consistent with the geological situation is conducive to improving the inversion accuracy and reducing the multiple solutions.Conventional interpolation-based modeling algorithms only use well logging data and horizon information,but the rich geological information contained in seismic data has not been effectively utilized,resulting in low model accuracy,slow convergence speed,obvious bull-eye phenomenon,and multiplicity of solutions,etc.Therefore,the idea of joint modeling of well-seismic data has research value,but due to the fact that the geological structure of the actual work area is often complex and presents heterogeneous characteristics,the well-seismic relationship corresponding to different seismic facies is not the same.If the overall well-seismic relationship is directly extracted in the unit of work area,the model accuracy will be reduced due to the average effect.Aiming at the above problems,this thesis studies the initial impedance model construction method under the control of seismic facies.First,the seismic facies identification is carried out based on the seismic waveform characteristics,and then the well-seismic relationship is extracted by facies and the initial impedance model is constructed.(1)Aiming at the problem of seismic facies identification,a seismic facies identification method based on the spatial structure characteristics of seismic waveforms is studied.Due to the high dimensionality of seismic data,direct seismic facies identification will face dimensional disaster,and traditional data dimensionality reduction methods cannot effectively extract spatial structure information in seismic waveforms.Therefore,this thesis proposes a clustering algorithm based on spatial structure features.In the first training stage,the multi-scale convolutional autoencoder module can effectively extract the deep features of seismic waveforms and realize data dimensionality reduction;in the second training stage,by adopting the training method of multi task learning,the auxiliary distribution is constructed as the training target,and the autoencoder reconstruction error is introduced as the constraint term,so that the dynamic clustering layer module can take into account the effective representation of waveform characteristics when continuously optimizing the clustering center.Finally,based on this algorithm,the seismic facies identification results with high coincidence with geological data and obvious regional characteristics are achieved.(2)Aiming at the extraction of well-seismic relationship and the construction of initial impedance model,a method for extracting well-seismic relationship based on the self-attention mechanism of seismic data is studied and the initial impedance model is established.Specifically,the modeling idea under the control of seismic facies is adopted to avoid the average effect between different facies,so that the corresponding relationship between well-seismic characteristics in a single facies is closer.In the extraction of well-seismic relationship,in view of the problem of information forgetting caused by too long seismic data,this thesis adopts a long short-term memory network with a bidirectional structure to realize the bidirectional extraction of features and alleviate information forgetting.Aiming at the non-local similarity features contained in seismic data,this thesis introduces a self-attention mechanism module to achieve variable distribution of network attention weights and to mine the intrinsic correlation of underground geological features.Finally,the algorithm constructs an initial impedance model with high coincidence with the seismic facies zone,obvious feature correspondence,clear geological significance and high resolution in the actual work area.Finally,through the comparative analysis with a variety of related algorithms,the effectiveness of the algorithm proposed in this thesis is proved,and new ideas and technical solutions are provided for the construction of the initial impedance model.