基于深度学习的井间地震CT初至自动拾取方法

Automatic Picking Method for First Arrival of Cross-well Seismic CT Based on Deep Learning

井间地震CT方法能够有效探测岩溶等不良地质,而拾取地震波的初至对其反演结果至关重要.以往井间地震CT初至拾取通过手动标定,效率低,无法满足模型训练要求.本文基于人工源地震初至的时距特征,提出了一种基于深度学习的地震初至训练样本库快速标定方法.利用人工合成数据训练初始网络模型,采用少量人工标定实测数据进一步训练,基于该网络模型对未标定的井间地震实测数据剖面进行初至拾取,采用长短时窗比法修正曲线拟合结果,对仍不理想的拾取结果进行人工判别和改正,并将准确标定数据加入训练集,形成新的网络模型;通过迭代实现数据集再生成并获得后续阶段的网络模型.基于上述方法完成了80000道数据的初至拾取,在充分训练后神经网络具备较高的拾取精度,采用3840组数据对该模型进行测试,发现预测结果与人工拾取结果对比的绝对误差值小于0.2ms的数据占91.8%,说明该方法是一种切实有效且高效的井间地震CT初至拾取方法.

The cross-well seismic computed tomography(CT)method can effectively detect bad geology,such as karst.The first arrival of the picked up seismic wave is very important to its inversion result.In the past,manual calibration in the first arrival of cross-well seismic CT was inefficient and could not meet actual engineering requirements.Therefore,based on the timedistance characteristics of the first arrival of an artificial source,this paper proposes a fast calibration method for the training sample database of the first arrival of an earthquake based on deep learning.First,use artificial synthetic data to train the initial network model,and use a small amount of manually calibrated measured data for further training.Based on this network model,the uncalibrated cross-well seismic data section is first picked.The long-short time window ratio method is used to modify the curve fitting results to manually distinguish and correct the still unsatisfactory picking results.Then add accurate calibration data to the training set to form a new network model.Through iteration,the data set is regenerated,and the network model of the subsequent stage is obtained.Based on the above method,the first arrival of 80000 channels of data has been picked up.After full training,the neural network has a high picking accuracy,and 3840 sets of data are used to test the model.Comparing the prediction results obtained by deep learning with the manual picking results,the data with an absolute error value of less than 0.2 milliseconds accounted for 91.8%.The result shows that the method proposed is a practical and efficient method for picking up cross-well seismic CT first arrivals.