Research And Application Of Seismic Signal Fine Processing Method Based On Machine Learning

Oil and gas resources are important energy related to the national economy and the people’s livelihood.With the increasing complexity of exploration targets,the difficulty of seismic exploration is also increasing.Moreover,seismic signal processing is an essential part of seismic exploration.There is an urgent need for a set of scientific method system to process the complex seismic signals that can reflect the characteristics of oil and gas reservoirs,so as to realize the high-efficiency,highprecision and high fidelity processing of seismic data.In practical seismic exploration,the fine processing and interpretation of seismic signals usually face the following three problems.First,the seismic signals usually contain a variety of noise interference,and the traditional denoising technology based on mathematical transformation can not adaptively deal with a variety of complex noise;Second,the traditional fault identification technology which mainly depends on manual interpretation can not meet the requirements of intelligent identification of multi-scale complex fault system;Third,the reservoir prediction method in less well area needs to be further integrated into the intelligent prediction mode to improve the recognition accuracy.In order to solve the above problems,it is particularly important to study a set of seismic signal fine processing method system with high efficiency,strong adaptability and high intelligence.Based on the artificial intelligence machine learning method,this paper studies the trace set optimization processing technology based on seismic signal denoising,multi-scale fault fine identification and reservoir dessert attribute comprehensive identification,so as to form a set of scientific intelligent processing and interpretation scheme,which provides technical support for adaptive processing,fine interpretation and high-precision prediction in the target work area.The main achievements and conclusions of this paper are as follows:1.Complex noise suppression of pre-stack seismic signal.Based on morphological component analysis(MCA)and dictionary learning technology,this paper proposes a three-step denoising process:(1)The seismic signal is decomposed into local singular signal and smooth texture signal by MCA algorithm.(2)For the two extracted components,KSVD dictionary learning is used to remove noise and obtain effective signal and noise.(3)The effective signal is extracted again from the removed noise and fused with the above effective signal.The theoretical model and real data processing results show that: the method proposed in this paper can effectively suppress multiple types of noise interference in pre-stack signals at the same time and retain effective signals to the greatest extent.It is especially suitable for seismic trace set optimization of pre-stack complex seismic signals with low signal-to-noise ratio.2.Random noise suppression of post-stack seismic signal.Based on the theory of deep learning,this paper proposes a denoising method process based on self-supervised learning:(1)In this paper,the autoencoder model is used to extract the training data directly from the target work area,which solves the problem of preparing label data.(2)The method of tied-weights noise suppression is proposed,which effectively reduces the risk of network over fitting,improves the ability of model learning and rapid convergence,and improves the signal-to-noise ratio of seismic signal while effectively suppressing random noise.This method provides an effective solution for seismic poststack random noise processing.3.High precision prediction of multi-scale seismic faults.Based on the excellent model UNet in the field of natural language processing,this paper proposes a set of multi-scale fault recognition methods:(1)The seismic data is separated by sepctral decomposition algorithm,and the low,medium and high data that can best reflect the morphology of different scales of the stratum are selected to explain the faults of different scales.(2)Based on the Res UNet Plus neural network model,the residual module(Res Net)is proposed to strengthen the learning ability of the network to the original signal and improve the learning efficiency.Self-Attention mechanism is introduced to enhance the recognition ability of small-scale faults and improve the recognition ability of fault continuity and morphology.The loss function is improved to adapt to the imbalance of fault data and improve the accuracy of fault recognition.(3)Through multi-scale fault fusion,the fault recognition resolution is effectively improved,and the high-precision recognition of complex faults is realized.4.Accurately detect the oil and gas sweet spot area of the reservoir in the small sample work area.In this paper,a set of reservoir gas detection method is proposed by using deep learning algorithm and nonlinear attribute model:(1)Through seismic well logs attribute analysis,attribute parameters sensitive to natural gas production are found.(2)Based on Res DNN neural network model,the learning degradation problem of multi-layer network is solved by residual module.The class weight parameters are used in the training parameters to effectively solve the imbalance of multi category label data of sweet spot detection.(3)In the work area,the method in this paper is used to detect the sweet spot production,and the prediction accuracy is up to about 95%,which provides a reliable reference for the efficient exploration and drilling in the target work area.