Research On Noise Reduction Algorithm For Seismic Exploration Data

As a vital material to human beings,energy and mineral resources are bases of survival and development.They play an important role in national economy,people's livelihood and national security.It is strategically significant to promote the revolution of energy geology exploration and exploitation for the sustainable development of economy and society.With the development of oil and gas exploration in recent years,the focus of global oil and gas exploration has shifted to the exploitation of complex oil and gas reservoirs and unconventional oil and gas.This has resulted in noise pollution in the data obtained from seismic exploration.Noise pollution will reduce the signal-to-noise ratio of seismic data,increase the difficulty of seismic data processing technology and affect subsequent seismic data interpretation,which will ultimately affect oil and gas judgment.As a result,denoise processing of raw seismic data is urgently required in relevant fields to ensure the quality of seismic data.Based on time-frequency analysis technology and machine learning theory,the corresponding denoising strategy is constructed for the research of noise suppression in desert seismic signal and micro-seismic data.Oil and gas resources play an important role in the development of national economy.As the largest oil and gas reserve base in China,Tarim Basin coved by desert has been the most important seismic exploration areas.However,the noise in the seismic prospecting data has the characteristics of non-stationary,Non-gaussian,nonlinear and low frequency.Furthermore,the band of low frequency noise overlaps with the desired signals,leading to more difficulties in noise suppression for desert areas.As a result,it is challenging for the existing signal extraction techniques to process the complex desert seismic data with low-quality.In order to effectively extract weak seismic signals from low quality seismic data,it is urgent to seek new techniques and methods to extract seismic signals more accurately and improve the signal-to-noise ratio.Empirical curvelet transform(ECT)is a new time-frequency analysis algorithm that combines the self-adaptability of empirical mode decomposition and the theoretical framework of Curvelet.A desert low frequency noise reduction method based on ECT was proposed in this thesis.The method first separated the Fourier spectrum of desert seismic signals and then adaptively constructs filter banks to obtain the empirical curved-wave coefficients in several scales and directions.The energy spectrum distribution of the direction coefficients at various scales of ECT was analysed and the coefficients were classified into signal concentrated sub-bands and noise concentrated sub-bands.The signal concentrated sub-bands were processed by the coherence-enhancing diffusion filter(CEDF)while noise concentrated sub-bands were processed by threshold processing.Ultimately,the signal and the noise were separated through coefficients classification.In this thesis,a new denoising method is proposed for noise in desert seismic exploration signal,while improving the filtering performance,it is necessary to consider the shape of the effective in-phase axis and the spatial distribution position of the in-phase axis in advance.The signal extraction technology based on the combination of in-phase axis location and filtering will become a new vision in seismic signal processing.Support vector machine(SVM)algorithm maps low-dimensional input space data to high-dimensional characteristic space using kernel function so that the linearly inseparable problems in low-dimensional space become linearly separable in high-dimensional space.In this thesis,a localization denoising system based on improved particle swarm optimization support vector machine is proposed.This method first detected the spatial position of the reflection in-phase axis through improved particle swarm optimization support vector machine according to the difference of the correlation between the desired signal and noise.Second-order total generalized variation(TGV)is then used to denoise the desired signals so that the non-damage localization of the spatial position and shape of the in-phase axis in the desired signals from desert area can be achieved together with noise reduction processing of the in-phase axis.It can be seen form the simulated and real seismic data that this algorithm has the advantage over traditional denoising techniques.In the above algorithm,we select feature vectors by a large number of experiments.In order to solve the problem of low rate of artificial feature extraction,this thesis further studies the feature extraction of desert seismic data and constructs a desert seismic signal denoising model based on unsupervised feature extraction and time-frequency analysis technology.This model first decomposed the desert seismic signal through the variational mode decomposition(VMD)method.Then each mode was sent to the unsupervised feature extractor for feature extraction.Finally,the desert seismic data is divided into desired signal and noise by the classifier,and the desired signal is selected for signal reconstruction.The results shown that this denoising model can effectively remove the desert low frequency noise meanwhile retain the desired signal.Next,this also research investigated the noise suppression algorithm of micro-seismic data.Unconventional oil and gas have become a new hotspot in the exploration and development of global oil and gas resources.They are abundant in China with great exploration prospect.Micro-seismic monitoring technology has become an important technique of unconventional oil and gas exploitation with its unique advantages.This technology is a geophysical technology that monitors the effect of production activities and underground state through observing and analysing the micro-seismic produced in production activities.The energy of micro-seismic is usually between-3 and +1 on the Richter scale,whose signals are easily affected or masked by the surrounding noise from fracturing construction.Consequently,the collected micro-seismic data is characterized with weak energy and strong noise,which is difficult to process using conventional wave filtering techniques.Shearlet transform is a new multi-scale geometric analysis theory and a nearly optimal sparse representation method of multi-dimensional functions.It has the characteristics of multi-resolution,multi-scale,multi-directionality and time-frequency locality,thus it can detect,locate,represent and process two-dimensional and higher dimensional data.Based on the shearlet transform theory,this research takes full advantage of the spatial and temporal characteristics of micro-seismic reflected signals and the difference between them and random noise in the shearlet domain to construct the optimal shearlet transform domain coefficient reconstruction algorithm to recover desired micro-seismic signals from strong noise.It has been studied that the micro-seismic signals and random noise have different singularities,which is the maximum modulus of shearlet system of desired signal in the same direction increases and the maximum modulus of random noise decreases when the scale increases.It can be seen from the results that this algorithm can recover desired micro-seismic signals from low SNR micro-seismic data effectively.In this thesis,the three time-frequency analysis algorithms were investigated including shearlet transform,VMD and ECT.We also discussed machine learning algorithms,including supervised SVM algorithm and unsupervised feature extraction algorithm.Optimized filtering schemes are applied to improve the SNR of data,which provides reliable and powerful technical support for micro-seismic exploration,desert seismic exploration application in petroleum exploration.