Research On Seismic Stacking Velocity Picking Method Based On Deep Learning

Since the 14th Five-Year Plan,the battle for technological innovation in oil and gas exploration technology in China has been in full swing,and complex geological areas such as deep water have become the focus of attention due to many problems such as large longitudinal and lateral variations and local anomalies in stratigraphic velocity.In order to get accurate information of this part of the region,it is often necessary to carry out large area and high-density 3D seismic exploration work,and how to obtain effective information efficiently and accurately from the huge seismic data has become a problem that needs to be solved in today’s oil and gas exploration industry.In seismic data processing,velocity analysis is the core process,which provides the optimal superposition pickup speed that directly affects the subsequent processes such as multiple wave suppression and offset imaging.The stack acceleration pickup in the traditional velocity analysis process is performed manually,but it is inefficient,time-consuming,and subjectively influenced by professionals in today’s world,and needs to be improved urgently.Therefore,this paper combines the seismic stacked acceleration pickup with neural networks,solves the problem of inconsistent output feature layers and missing information in the energy extremes of the velocity spectrum by improving the feature fusion structure of YOLOv5 s,and adds an attention mechanism to improve the allocation of feature learning weights in the tiny energy extremes of the velocity spectrum by YOLOv5 s,and obtains a set of YOLOv5s-AS for low signal-to-noise regions such as deep water.The YOLOv5s-ASFF-CBAM deep learning model is obtained for low SNR regions such as deep water.The model not only automatically detects the optimal stacked acceleration in the seismic velocity spectrum with regularity,but also improves the performance in terms of accuracy and recall compared with other mainstream stacked acceleration pickup models.The main research of this paper is as follows:(1)The seismic stack acceleration pickup is transformed into a target detection work in the energy extremum region,and YOLOv5 s is selected as the base model for this experiment.The center of the energy extremum region of the velocity spectrum corresponds to the optimal stack acceleration pickup point,and the optimal stack acceleration is picked up by using the target detection algorithm to extract the value corresponding to the center coordinates of the energy extremum region.(2)The YOLOv5 s overfitting problem is effectively solved using data augmentation and migration learning.In view of the small size,type,and relatively single label of the homemade seismic velocity spectrum dataset,this paper uses data augmentation as well as migration learning methods,and the model overfitting problem is found to be effectively solved by comparing the stacked acceleration pickup experiments.(3)Using adaptive feature fusion mechanism and hybrid domain attention mechanism to effectively enhance the detection and recognition effect of YOLOv5 s for small energy extrema regions.In this paper,the feature fusion mechanism of YOLOv5 s is improved and the attention mechanism is added to obtain the YOLOv5s-ASFF-CBAM model,which can accurately identify the small energy extremum region with low signal-to-noise ratio by stacked acceleration pickup experiments and has good overall performance.(4)A comprehensive demonstration of the performance of the YOLOv5s-ASFF-CBAM model.Through ablation experiments,YOLOv6 and YOLOv8 s stacked acceleration pickup experiments and comparison with manual pickup speed on reflection interface correction effect,it is found that the overall performance of YOLOv5s-ASFF-CBAM model is good and can meet the manual pickup standard,and it is extremely efficient to meet the needs of stacked acceleration pickup in deep water and other complex geological regions.