Study On VSP Data Processing Method Based On Deep Learning

VSP data can provide accurate time-depth relationship and high-precision imaging results,the key processing steps of it include first-break picking,denoising and wave field separation.Affected by the assumptions,parameter settings,signal-to-noise ratio and other factors,the conventional methods still have some problems,such as large error,poor anti noise ability,effective signal damage and so on,and the processing accuracy needs to be further improved.Deep learning technology has been successfully applied in the field of surface seismic data processing and interpretation.How to introduce deep learning method into the key steps of VSP processing is of great significance to improve the accuracy of data processing.As for the conventional processing process,there are errors in picking up the first arrival time and poor anti-noise performance,it is easy to cause damage to the effective signal in the aspect of denoising,and there are problems in the aspect of wave-field separation,such as incomplete separation and artifacts contained in the separation result.In this paper,the underground medium model is constructed based on the well data.By setting different dominant frequencies of wavelet,the VSP full wave field,up-going wave field and down-going wave field data with different frequency components are generated by forward.Combined with the real data,the training sample data is constructed by manual picking and adding noise.The firstbreak picking processing of pixel level VSP is realized by using encoder-decoder nerwork.In order to solve the problem that one-dimensional convolution only focuses on local information in time direction,two-dimensional convolution is introduced to increase the attention of local information in space and reduce the number of misclassified pixels in the prediction result.The full convolution residual network is used to denoise VSP data.Aiming at the problem of error information in the denoising results of U-net network,the residual block structure is introduced to improve the network performance on the basis of ensuring that the original network does not degenerate.The asymmetric generative adversarial network is constructed to separate the upgoing and down-going waves of VSP.The discriminator network is used to modify the prediction result of the generator network.The asymmetric convolution kernel is introduced into the generator to reduce the prediction error of the generator.Combined with the balanced loss function designed based on the mean square error loss function,the separation error is further reduced.The model data and real data tests show that the average picking error of VSP first-break picking based on encoder-decoder network is less than that of conventional methods,which achieves higher accuracy and stronger anti noise ability.Compared with U-net network and conventional methods,the denoising results based on full convolutional residual network have better performance.The VSP wave field separation results based on asymmetric generative adversarial network have less effective signal loss and lower low-frequency information error.Compared with U-net and conventional methods,the VSP wave field separation method based on GAN achieves higher accuracy.The VSP data processing method based on deep learning research and design realizes higher precision and higher fidelity processing,and provides technical support for intelligent and high precision processing of well exploration data.