Research On The Key Technology Of Velocity Spectrum Picking In The Process Of Seismic Wave Normal Movement Correction

With the expansion of oil exploration and the improvement of technology,the seismic data in the exploration area are increasing year by year.In the seismic data processing process,the huge volume of data has put a great burden on the manual velocity spectrum picking,which makes the velocity spectrum picking process extremely time-consuming and requires a change in computer-aided data processing technology.To solve the time-consuming problem of velocity spectrum picking,the automation of velocity spectrum picking has become one of the hot research topics at this stage.The existing velocity spectrum automatic picking algorithms can be divided into semi-automatic velocity spectrum picking algorithms and automatic velocity spectrum picking algorithms.Semi-automated algorithms mainly include Monte Carlo method,nonlinear function optimization method,conjugate gradient method and simple deep learning algorithm.Semi-automatic algorithms are not suitable for complex geological scenarios because they require constraints and seed points,and their wrong constraints can lead to accumulation of errors in picked points,so semi-automatic algorithms are gradually replaced by automatic algorithms.Automatic algorithms mainly consist of deep learning algorithms.The automatic algorithm mainly consists of deep learning algorithm,which has the characteristics of self-learning habit and high efficiency,and can learn the effective information of energy heaps through different levels of texture.However,in practical engineering,the picked points obtained by the automatic algorithm do not sufficiently consider the sequence feature information of the stratigraphic structure,and the data such as multiple wave energy heaps and non-energy heaps centered picked points cannot be processed effectively,resulting in less accurate picked points and even geological interpretation bias.Therefore,this study aims to obtain accurate picked points and makes the following explorations for the problem of inaccurate picked points caused by multiple wave mispicking and non-energy heaps centered picked points.First,to address the problem of inaccurate picked points due to multiple wave energy heaps mis-picking in current neural network algorithms,this paper proposes a object detection network FLD that is robust to multiple wave energy heaps interference.Considering that the multiple wave energy heaps are extremely similar to the correct energy heaps morphology and texture,this study adds a spatial attention module for learning the global correlation features of the velocity spectrum energy heaps to ensure that the rough picked results of the FLD network can obtain valid texture details and avoid the influence of redundant information.Second,to address the problem that the current neural network cannot effectively pick up the picked points that deviate from the center of the energy heaps,this study proposes a context-sensitive PA module for fine-tuning the reasonable positions of the picked points to solve the problem of deviating from the center of the energy heaps.Considering that the acquisition of picked points that deviate from the center of the energy heaps requires reference to data such as superimposed segments and adjacent velocity spectrum,the PA module effectively fuses the contextual information of stratigraphic changes through the weighted summation of the image matrix to finely adjust the rough picked points of the FLD's output.Integrating the thinking of rough picking and refinement adjustment,the FLD-PA algorithm is proposed in this study.The algorithm feeds the rough picking result of FLD into the PA module for refinement,which fully integrates the contextual information of energy heap characteristics and stratigraphic changes.Model validation shows that the object detection algorithm FLD proposed in this study improves the accuracy index by19.53% compared to the YOLOv3 network.In addition,the FLD-PA algorithm proposed in this study improves the accuracy metric by 1.64% and reduces the root mean square error metric by 1.9 pixels compared to the YOLOv3-LSTM.Therefore,the FLD-PA algorithm proposed in this study can solve the problem of multiple wave energy heaps interference and large errors in the picked results,and effectively improve the velocity spectrum picking accuracy.Thus,it can provide accurate velocity parameters for the subsequent seismic exploration processing process.