The Research On Time-frequency Feature Optimization Of Seismic Data And Denoising Methods Based On Learning Models

Petroleum and natural gas resources serve as a fundamental pillar of China’s energy resource portfolio,proving essential to both industrial production processes and individual lifestyles across multiple sectors.These commodities exert a profound influence that transcends their substantial contribution towards shaping the nation’s overall economic growth trajectory;indeed,they hold crucial importance in safeguarding national energy security.Thus,the strategic significance and inherent worth of oil and gas resources demand rigorous assessment and substantial regard from all stakeholders’ point of views.With escalating oil and gas extraction initiatives comes more challenging exploration areas and environments.Field-collected seismic data is susceptible to varying factors,impacting data quality and hindering post-processing and interpretation.Consequently,to optimally harness these costly seismic data,advanced techniques are required to counteract these effects,enhancing data precision,assuring its veracity,and supplying secure data sustenance in oil and gas exploration and development.The noise within seismic records can hinder efficient seismic signal extraction,impede precise phase analysis,and cause waveform distortion.Therefore,effectively suppressing noise in seismic data,improving the signal-to-noise ratio and quality of seismic data are extremely necessary in both theory and practice.The dissertation is primarily focused on the research and processing of two types of complex seismic data: vertical seismic profile(VSP)data collected using distributed acoustic sensing(DAS),and desert seismic data collected in the Tarim Basin.Given the varied characteristics of real seismic noise,single-type method often fails to provide optimal filtering.Hence,we thoroughly analyze and utilize the specific characteristics of effective signals and different kinds of noise in the pending seismic data,and amalgamate methods from varying types for establishing robust noise abatement strategies of seismic data.It takes into account the characteristics of both one-dimensional and two-dimensional data in seismic records,leveraging the interconnections and patterns among features across dimensions.This study employs time-frequency analysis theory to represent and optimize one-dimensional seismic signals,and utilizes machine learning models to complete feature extraction and separation for two-dimensional seismic data.The ultimate objective of denoising is to enhance weak seismic signals,and to increase the signal-to-noise ratio of seismic data.The research initially focuses on the transformation and enhancement of signal and noise characteristics for seismic data in high-resolution time-frequency domain.By developing a high-quality feature tailored to the low-rank sparse model,we reduce the coherent noise within the same frequency band as signals.Subsequently,we analyze the inherent oscillatory characteristics of seismic signals and perform modal decomposition to reduce the constraint limitations of the low-rank model and improve its processing performance.This approach aims to effectively reduce low-frequency,weakly correlated noise.Finally,a deep neural network model is selected to learn the nonlinear characteristics of seismic data,addressing the complexities of parameter adjustment and the average generalization performance associated with low-rank models.To enhance the network’s attention and improve the model’s prediction accuracy,time-frequency prior analysis is incorporated into the network.This integration enables the high-resolution reconstruction of seismic data.The main body and achievements of the proposed methods can be summarized as follows:To address the coupling noise,which demonstrates strong coherence with the effective signal,along with the intricate background noise in VSP data gathered by DAS,the dissertation introduces a refined high-resolution time-frequency domain low-rank sparse matrix decomposition technique.Its objective is to efficiently extract seismic signals while suppressing noise.From the analysis of DAS-VSP data,it is observed that there is significant coherence between the effective seismic signals and coupling noise,with overlapping frequency bands,which makes it difficult to separate them directly in the time-space domain.However,differences can be observed in the time-frequency domain.General time-frequency representation methods have low resolution and weak energy concentration,making it impossible to clearly distinguish signal and noise characteristics.But when using high-resolution time-frequency analysis method to characterize signal and noise,it is observed that effective signals exhibit certain sparse characteristics,while coupling noise demonstrates pronounced low-rank structural features.In light of this,the adoption of a low-rank sparse matrix decomposition model,coupled with statistical analysis of the time-frequency coefficients of the low-rank and sparse matrices,can achieve the separation of effective signals from coupling noise and enhance computational accuracy.The use of SVD-free optimization techniques in low-rank decomposition models boosts computational speed and facilitates the handling of extensive DAS datasets.By leveraging low-rank constraints,we can effectively suppress complex and unevenly distributed background noise.The proposed denoising method achieves feature optimization in the transform domain of signal-to-noise characteristics,enhances the performance of the feature separation model,and improves the efficiency of data processing.To tackle the challenges associated with the noise of strong energy,prominent oscillatory patterns,weak correlation characteristics,and spectral overlap with effective signals in desert seismic data,a noise suppression method based on windowed low-rank constraints for multi-oscillatory component seismic signals is proposed.Due to interference from factors such as the strong energy of low-frequency noise,low-rank models are unable to effectively suppress noise by exploiting the spatial structural correlations in seismic signals.Exploiting differences in oscillatory properties between low-frequency random noise and effective reflection signals,modal decomposition of seismic signals with distinct oscillation patterns can be realized by combining geometric morphological analysis with tunable quality factor wavelet transform.This approach enhances the low-rank structural characteristics of each modal component.By meticulously adjusting the quality factor of the wavelet transform,a wavelet possessing a high quality factor and exceptional frequency resolution is employed for the analysis of noise components exhibiting pronounced oscillatory attributes.Conversely,a wavelet characterized by a low quality factor but offering superior time resolution is utilized to dissect effective signal components that possess more subdued oscillatory signatures.Leveraging low-rank constraints on the coefficients of distinct scale windows,which are associated with each modal component,enables signal-noise separation to be achieved within the same frequency band.The proposed denoising method effectively suppresses low-frequency random noise,enhances weak signals,and successfully reduce ground roll waves in actual seismic data.Considering the limited computational efficiency and the inability to guarantee optimal denoising results of sparse low-rank models,which rely on empirical and manual parameter settings,we propose a seismic noise reduction system for desert seismic data based on the integration of time-frequency priors and multi-scale long-term memory convolutional neural networks.The powerful nonlinear processing capability of convolutional neural networks can be utilized to achieve complex mappings from noisy seismic data to denoised seismic records.In view of the absence of mathematical interpretability in traditional neural network frameworks,optimal control theory in dynamic systems can be employed to articulate the fundamental principles of deep neural networks.Furthermore,the discretization of fractional differential equations serves as a guiding principle in the construction of long-term memory neural networks.We enable the multi-scale feature interaction mode to balance the overall structural distribution and local detailed features of seismic data,thereby improving the resolution capability of the data.To address the problem of network models’ limited attention to low-frequency,weak structural components in seismic data,we introduce seismic data time-frequency prior analysis and decomposition.By extracting features in the high-resolution time-frequency domain,we enhance the network’s focus on structural information of events across various frequency bands,ultimately achieving high-resolution reconstruction for seismic data.The constructed network model has effectively suppressed noise in seismic data with varying noise intensities,significantly improving the recovery of weak effective signals and enhancing seismic events resolution.The dissertation constructs a series of seismic noise suppression processes from three aspects: seismic data analysis,feature representation and optimization,as well as model selection and improvement.It is anticipated that by effectively combining these processes,the majority of seismic data denoising challenges can be addressed,thereby significantly improving the quality of seismic records,and providing accurate and reliable data support for the subsequent analysis and utilization of seismic data.