Research On Multi-scale Feature Progressive Fusion Algorithm And Its Application In DAS Noise Suppression

Energy is an important support for the prosperity of national economy and the stability of national life.To ensure the security of national energy and meanwhile boost energy reserve,China has issued a number of policies on energy development to promote the stable output of old oil and gas fields,and increase the productivity in new oil and gas areas.In recent years,the target of seismic exploration has gradually changed toward deeper or more complex oil and gas reservoirs,thus the borehole seismic technology has been intensively studied and applied.Compared with surface seismic exploration,the receiving equipment of the borehole seismic technique is closer to the target oil and gas formation,so the acquired seismic waves would more clearly reflect the longitudinal changes of the stratigraphic profile,which is very beneficial for fine imaging of small-scale geological structures.Under the new situation and task of seismic exploration,the future development prospect of borehole seismic technology is promising.Vertical seismic profiling(VSP),as an borehole seismic observation method,has unique advantages in formation location calibration,velocity parameter extraction,and multiple wave identification,etc.However,affected by the high cost as well as the difficulty of construction,conventional borehole geophones are subject to many constraints in practical application.With the continuous innovation of fiber optic equipment and technology,the development of distributed acoustic sensing(DAS)is becoming more and more mature.As a low-cost,full-well coverage,electromagnetic interference-resistant and high-accuracy acoustic field detection technology,DAS is capable of replacing conventional geophones in VSP borehole and playing an important role in the borehole seismic technology.Despite the so many advantages of DAS technology,the strong energy of noise interference and the huge difference in amplitude between signal waves in DAS data can not be neglected,which bring a lot of obstacles to the subsequent seismic data interpretation and inversion work.In order to promote the application and development of DAS in the field of seismic exploration,effective methods are needed to suppress the complex noise in DAS data and reconstruct seismic signals that contain rich stratigraphic information.So far,many classical data processing algorithms,such as band-pass filtering,predictive filtering,algorithms based on transformation,decomposition and rank reduction,have been proven to perform well in seismic data processing.However,due to the limitations of the above traditional methods,they are not satisfactory when applied to process DAS data.In recent years,convolutional neural network(CNN)-based methods have shown great advantages and development potential in the field of seismic data processing.Therefore,it is of great significance to establish adaptive DAS feature learning model and explore intelligent noise suppression scheme.Taking the high accuracy processing requirements of DAS data into consideration,this paper describes a multi-scale feature progressive fusion algorithm for DAS data processing.With the help of deep learning network architecture,this algorithm explores the collaborative representation of complementary information between DAS noise and its multi-scale versions,and achieves accurate noise suppression by modeling DAS noise comprehensively.In the overall algorithmic architecture,at each scale,the recurrent calculation is taken to explore the potential correlation of complex DAS noise and learn its global texture;in addition to this,the feature rescaling strategy is applied to reassign the weights of feature maps on each channel to focus on the detailed structure of noise features.In each feature fusion stage,the information flow converges continuously from the bottom to the top layer of the network,fusing feature information at each scale to accomplish accurate estimation of DAS noise.Finally,the DAS noise suppression and signal reconstruction can be realized by removing the modeled noise from the noisy DAS data.Experiments based on synthetic DAS data and field DAS data show that the described algorithm performs better than some representative methods,not only the noise is more significantly suppressed,but also the reflected signal is better reconstructed.