Research Of Seismic Reservoir Image Interpretation Method Based On Deep Learning

In the current stage of industrialization,oil and gas has played an important role as a necessary strategic material.Reservoir characterization and seismic interpretation plays a vital role in the field of oil and gas exploration.Therefore,how to improve mining efficiency and obtain more resources at a lower cost has become an urgent problem to be solved.Faced with the dilemma of less and less shallow oil and gas resources,people began to try to develop deep layers.However,in the face of more complex geological structure and the steep increase of mining difficulty,high signal-to-noise data is more difficult to obtain.The algorithm is optimized to obtain the desired results with seismic reflection data with a low signal-to-noise ratio,and it has become a research hotspot.As two key research objects of seismic reservoir image interpretation methods,horizon tracking and wave impedance inversion can have a positive impact on future production and research.Horizon tracking is a basic problem in the interpretation of seismic data.For the above problems,the manual marking of the horizon is a huge workload,which seriously affects the efficiency of marking.Therefore,the proposal of automatic horizon tracking is very important for this field.Taking deep learning,data driving and feature extraction as the basic methodologies,it proposes a method that is different from the traditional automatic horizon tracking based on coherence or waveform similarity,while the datadriven approach facing complex horizons,it can still guarantee High accuracy.Wave impedance inversion is an important work in seismic data interpretation.Elastic wave impedance inversion based on well logging constraints using seismic reflection and well logging data to perform a variety of elastic parameters related to lithology,which is a hotspot in inversion research in recent years.Conversion of various data such as constraints brought by logging data can better distinguish lithology,and the results can have a more accurate correspondence with the reservoir.At present,deep learning methods are used to inverse elastic wave impedance is relatively scarce,and deep learning based on data can better extract effective information from seismic reflection and other information to construct elastic wave impedance.Therefore,this thesis did a study on elastic wave impedance inversion using deep learning as a methodology.1.This thesis proposes an automatic horizon tracking method for deep convolutional neural networks based on identity shortcut mapping.Aiming at the fact that traditional horizons are prone to cross to other horizons in complex multi-fault horizons,based on the characteristics of seismic amplitude reflection,for the characteristics of complex seismic waveform information,feature extraction is performed through deep convolutional neural network,and use identical shortcut mapping to guides gradient flow,solves the problem of gradient dispersion that is prone to occur in deep networks,and more effectively transmits fine-grained seismic amplitude reflection features to the end of the network.Compared with traditional convolutional neural networks,it greatly enhances the ability of identification of horizons.This algorithm can guarantee good accuracy with less training data,and has good generalization performance for horizons with spatial long-distance.2.This thesis presents the well logging constrained elastic wave impedance inversion based on the feedforward attention mechanism.In view of the sparse logging data itself,although the seismic reflection is known to be easier to obtain,and has lower signal noise and lower resolution compared to the logging data,in order to extract effective information,a bidirectional long-term and short-term memory network is used in filtering,extracting and integrating the fixed depth of layers,which effectively connects the correlation of the layers,and captures and influences the information of the layers that are far away from each other,which suppresses the situation that the unidirectional gradient is easily dispersed.Subsequently,a feedforward attention mechanism was used to further calculate the weight effect of each layer in a fixed depth on the final fitting result,which greatly improved the prediction accuracy of the elastic wave impedance.Through the effective generalization ability of the model,it is applied to the entire section to construct a full-section elastic wave impedance.Experiments verify that the model has good reconstruction performance.