High-dimensional Seismic Signal Concealed Feature Extraction And Pattern Recognition

seismic waves are artificially excited on the ground and during the propagation of seismic waves in the deep formation,various geological interfaces will produce reflected seismic waves.The seismic waveform(also known as seismic signals)are received by a large number of sensors deployed on the ground and be processed by complex technological procedure.Finally,high-dimensional seismic signals that reflecting underground geological conditions are obtained.The geological structure,lithology,oil and gas property not only affect the amplitude,frequency,phase of high-dimensional seismic signals and other characteristics,but also affect the characteristics of seismic signals changing with the incident angle and azimuth angle.These seismic signal characteristics form different seismic signal patterns.These seismic signal characteristics form different seismic signal patterns.Therefore,the study of high-dimensional seismic signal patterns can potentially invert various underground geological characteristics.Due to the particularity and interdisciplinary of high-dimensional seismic signal pattern analysis in comparison to other signal pattern analysis,the feature extraction and pattern analysis of high-dimensional seismic signals have become the frontier and challenging issues in the field of information processing and oil and gas exploration.With increased effort on the in-depth petroleum prospecting in China,relatively simple oil and gas reservoirs have been explored completely and existing seismic signal feature extraction and pattern analysis methods are difficult to extract concealed features,which is weak and subtle feature but related to reservoir characteristics.The existing methods meet challenges to effectively find underground complex oil and gas reservoirs in areas with complex geological structures or‘hidden’reservoirs.Under this background,petroleum exploration needs information processing field based on traditional information processing methods and develop a new generation concealed feature extraction and pattern analysis methods of high-dimensional seismic signals for the complex oil and gas reservoirs.In this dissertation,based on the existing signal sparse representation,deep neural network and various common signal analysis methods,we propose a number of concealed feature extraction methods of high-dimensional seismic signals to identify seismic signal patterns precisely and improve the recognition ability of complex oil and gas reservoirs.The main research contents and innovations of this dissertation are as follows:(1)Prior knowledge guided deep feature extraction and pattern analysis: Traditional seismic signal attribute analysis is able to reflect the underground geological conditions from different aspects.Yet,it is not comprehensive enough to resolve the hidden reservoirs.On the other hand,the evolving deep feature is advantageous in fully representing the information of seismic signals.However,the extracted features might not necessarily have difficulty to highlight the concealed features associated with the reservoir,and lacks pertinence for the underground lithology and oil and gas.In this dissertation,a shareautoencoder(S-AE)deep feature extraction method is proposed,which can not only fuse multiple seismic attributes,but also obtain relatively complete and comprehensive concealed features of seismic signals.The S-AE network structure adds multiple independent layers to ensure the reconstruction of multiple attributes and waveform data,and then completes the deep feature extraction guided by seismic attribute knowledge.The research results demonstrate that the proposed method can effectively obtain concealed deep features under the guidance of multiple seismic attributes,and has superior performance in seismic signal pattern analysis in comparison to the traditional seismic attributes analysis.(2)Seismic signal waveform structure feature extraction and pattern analysis based on sparse representation: Most of the oil and gas reservoirs are sedimentary formations.The lateral continuity of the formations brings the similarity of the seismic traces in the lateral direction,giving rises to the similarity,sparse characteristics of the seismic signal waveform structure and the limited seismic signal patterns.The sparse representation of signals is particularly suitable for such sparse and patterns limited seismic signals.To extract the concealed waveform structure feature caused by low SNR and narrow frequency band of seismic signals,a robust dictionary learning method based on sparse representation is proposed.The robust dictionary learning method first obtains the waveform structure feature dictionary of seismic signals in a data-driven fashion.Then,it selects the waveform structure features in the dictionary for weighted summation to achieve sparse representation of seismic signals and obtaion the waveform structure features.Simultaneous source error separation is used to improve the dictionary learning algorithm.In an iterative dictionary updating procession,the noise is gradually separated from the training data.The research results demonstrate that the proposed method improves the robustness of the dictionary learning algorithm and effectively extracts the concealed waveform structure features of the thin bed.(3)Inter-trace variation feature extraction and pattern analysis based on two-dimensional sparse representation: The lithology variation of underground rock strata results in the signal amplitude changing with incident angles,which easily affected by seismic waveform structure and amplitude.The traditional seismic signals pattern recognization method cannot give the concealed inter-trace variation feature.This dissertation proposes a feature extraction method for seismic signals with different incident angles based on twodimensional sparse representation.This method sparsely represents the pre-stack gather directly.The two-dimensional sparse representation realizes the vertical and horizontal feature decomposition of the pre-stack gather.Meanwhile,the horizontal feature characterizes the inter-trace variation feature of the pre-stack gather.The research results demonstrate that the proposed method effectively extracts the concealed inter-trace variation feature of the pre-stack gather,and the pattern analysis results of extracted inter-trace variation feature can better predict the reservoir lithology distribution in three-dimensional space.(4)Three-dimensional(3-D)seismic signal feature extraction and pattern analysis based on 3-D tensor analysis : The presence of oil and gas in the underground rock formations result in changes of seismic signal frequency.Moreover,the frequency variation with incident angles is not completely consistent.Therefore,the oil and gas reservoirs property analysis need to consider the inter-trace variation with different incident angles in time domain and frequency,which makes the seismic signal pattern analysis a pattern analysis problem of 3-D signals.To solve the characteristics of pre-stack signal hidden in3-D space,this dissertation proposes to use continuous wavelet transform to extract the 3-D tensor features of pre-stack gather,and then apply them to seismic signal pattern analysis of reservoir lithology and hydrocarbon prediction.The research results demonstrate that the extracted tensor features include three dimensions of time,amplitude and frequency,which can effectively highlight the weak features of the high-dimensional seismic signal related to the hydrocarbon fluids and lithology variation.And finally,this study intends to support a more accurately identification and interpretation of the seismic signal patterns.