| Abstrakt: |
Pre-stack depth migration has been widely used in petroleum exploration and it is of great significance to use depth-domain seismic data for reservoir prediction. Depth-domain seismic data is different from the timedomain, and the traditional time-domain inversion process will lead to the loss of high-frequency information in the data during the process of domain conversion, which is not conducive to the characterization of fine reservoirs. In order to directly invert acoustic impedance in depth-domain seismic data, we proposed a datadriven inversion method based on deep learning. Firstly, the two-dimensional convolutional neural network (2DCNN) is used as the basic framework of the inversion module to improve the horizontal continuity of inversion. The depth domain seismic data, initial model and logging data are input into the inversion module of the network model. Then, the output acoustic impedance data is used to synthesize the depth domain seismic records using the proposed method and the input seismic data are compared. Strong anti - noise activation function is used to improve the anti - noise of the network model. For the seismic data of few Wells, a transfer learning method is introduced to transfer the features from the training data set to the target data set through pre-training, which further improves the inversion accuracy. The application results of model test data and actual seismic data show that the inversion method has a better effect than one-dimensional neural network and timedepth conversion method, and the inversion results are in good agreement with the actual logging data. Therefore, deep learning can be well applied to seismic data in deep domain, which has certain feasibility. [ABSTRACT FROM AUTHOR] |
