Q-value Estimation And Seismic Attenuation Compensation Based On Deep Learning

Due to the anelasticity of the subsurface medium makes seismic waves undergo absorption attenuation during propagation,resulting in amplitude attenuation and phase distortion,which greatly affects the resolution and fidelity of seismic data.The quality factor Q is the most commonly used physical quantity to quantify the anelasticity of subsurface media,and is an important indicator for oil and gas prediction,which can characterize the strength of attenuation,and many attenuation compensation methods need to estimate accurate and reliable Q factors in advance.Attenuation compensation is an important method for high-resolution processing of seismic data,providing a reliable guarantee for subsequent inversion and interpretation.Both Q-value estimation and seismic attenuation compensation are ill-posed inverse problems that are difficult to solve and the robustness of the results is poor when noise is present.In response to the problems of poor stability and computational complexity of traditional methods,this thesis investigates the Q-value estimation and seismic attenuation compensation methods based on deep learning.The traditional Q estimation method requires high quality of seismic data,the calculation accuracy of the method is limited by the width of the time window and the selection of the calculation frequency band,it is sensitive to noise,and the estimation results for layer Q are not stable enough.To address the problems in the traditional methods,based on the feature engineering principle of deep learning and drawing on the experience of previous scholars who used complex seismic traces to analyze attenuation,this thesis combines raw attenuated seismic data and seismic envelopes as two-channel inputs,and uses a deep neural network with multiple inputs and single outputs for stable and efficient Q-value estimation.The simulated data test verifies the advantage of faster convergence of the multi-input network compared with the single-input network.The test comparison with the traditional frequency domain Q estimation method on simulated data verifies that the proposed method has higher accuracy and better noise immunity,and the successful application on field data by using transfer learning also verifies that the proposed method has practical application value.The essence of attenuation compensation is to establish a nonlinear mapping relationship between attenuated and non-attenuated seismic traces to achieve the recovery of amplitude energy and correction of phase distortion of attenuated traces.This thesis proposes an end-to-end deep neural network for attenuation compensation in a supervised learning framework.This method does not require pre-estimated Q values,and directly establishes a mapping relationship from attenuated seismic data to non-attenuated seismic data to achieve accurate and reliable simultaneous amplitude compensation and phase correction,which is a simpler computational process than traditional compensation methods.The simulated data test verifies that the supervised learning attenuation compensation method proposed in this thesis is more accurate and more robust to noise than the inverse Q filtering method,and the combination of transfer learning also successfully achieves the effective compensation of the field data.Considering that supervised learning requires a large amount of training data,and the training data synthesized by the attenuation model differs from the real stratigraphic model and the attenuation propagation law,this thesis proposes an attenuation compensation method under semi-supervised learning by constructing a dual generative adversarial network,one network for attenuation compensation and one network for attenuation forward modeling thus breaking the limitation of the attenuation model,and only a small amount of labeled training data is needed to achieve stable and accurate seismic attenuation compensation.This method does not require either pre-estimation of Q values or artificial synthesis of large amounts of training data,and both numerical simulation experiments and field data applications verify the effectiveness and practicality of the method proposed in this thesis.