关键词: 速度谱, 智能拾取, YOLOv8神经网络, 深度学习, 地震数据处理

Abstract:

Velocity spectrum picking is a crucial step in seismic data processing. Traditional velocity spectrum picking methods usually require manual intervention, which is time-consuming, labor-intensive, and prone to error. Therefore, an intelligent velocity spectrum picking method based on the YOLOv8 (You Only Look Once v8) neural network was proposed. This method transforms velocity spectrum data analysis into an image recognition task, therefore achieving automated and intelligent velocity spectrum picking. The core of this method is to convert velocity spectrum data into images, which are then input into the constructed YOLOv8 neural network model. The feature extraction network in the model learns the spatial information of energy clusters in the velocity spectrum images, and the feature fusion network fuses the extracted multi-scale features of energy clusters from shallow, intermediate, and deep layers to capture the energy cluster features in the images more comprehensively. The detection head of the model allows for refined predictions of energy cluster targets, obtaining pixel points corresponding to different picking positions in the velocity spectrum images. Then, the pixel points are converted to finally obtain the time-velocity pairs. For the exploration area GY of the Sinopec Jiangsu oilfield with developed igneous rocks and strong multiple interference, a dataset containing 1 200 velocity spectrum images was constructed. By optimizing training parameters, both the model accuracy and recall reached about 90%. The intelligent velocity spectrum picking technology based on the YOLOv8 neural network showed over 94% consistency with manually picked velocity curves in high-coverage areas, more than 90% consistency in areas above 3 500 ms, and about 92% consistency in areas with igneous rocks and fault development. Compared with traditional convolutional neural network (CNN) methods, the intelligent velocity spectrum picking technology based on the YOLOv8 neural network obtains more picking points with higher positional accuracy, and the processing time of a single velocity spectrum is only 10 ms, showing significant efficiency improvement. This technology provides an efficient and accurate intelligent solution for seismic data processing, demonstrating promising application and promotion value.

Key words: velocity spectrum, intelligent picking, YOLOv8 neural network, deep learning, seismic data processing