Research And Application On Land Seismic Noise Reduction Algorithm Based On Deep Learning

Seismic exploration is the mainstream means of oil and gas exploration and is widely used at home and abroad.Structural information at different geological depths can be observed and analyzed in the process of seismic exploration,thereby indirectly inferring the location of oil and gas reserves.In recent years,the trend of seismic exploration in our country is unconventional deep oil and gas reservoirs,which means that the geology of oil and gas reservoirs will be more complex,deeper,and the inhomogeneity of medium will be stronger.Therefore,the quality of the seismic data collected in the field is not high.A large amount of regular noise and irregular noise are mixed in the seismic records,which makes it difficult to identify effective signals,and the resolution of seismic records is very low,which brings difficulties to the subsequent interpretation of seismic data.In order to solve this problem,it is of great significance to explore a feasible noise reduction scheme and obtain reliable and clear seismic information that meets the requirements of "three highs" for oil and gas exploration.At present,many seismic data noise reduction algorithms are applied in the processing of field seismic data,such as wavelet transform,multi-trace fitting,etc.Although they have certain effects in applications,their limitations cannot be ignored.In recent years,the era of artificial intelligence has come,and deep learning has shown its superior potential with its powerful learning ability.It is a data processing technology that starts from the data itself and uses a large amount of training data as the basical data.It can process high-dimensional and complex data by mining the inherent nonlinear representation of the input.Convolutional Neural Network(CNN)and Generative Adversarial Network(GAN)are two important models that are widely used in text recognition,image super-resolution,data augmentation and other fields but few applications in seismic data processing.Therefore,in this paper,for the desert seismic records collected in surface by traditional moving coil geophones and seismic records collected in wells by distributed optical fiber acoustic sensor(DAS),deep learning methods based on CNN and GAN are applied in two types of seismic data.In the surface desert seismic records obtained by traditional moving-coil geophones,the effective signals are often disturbed by irregular random noise and regular surface waves,and the noise in the records has strong energy,low frequency,and overlaps with the signal spectrum.We propose an unsupervised generative adversarial network based on relative attributes to address this issue.This method introduces the idea of attribute editing,encodes the data of different attributes in the seismic record,uses the relative attributes obtained by the difference between the target attributes and the original attributes as a guide,and three loss functions as constraints to control the generator to generate the data corresponding to the target attributes.By setting appropriate target attributes,signal and noise can be separated.Distributed optical fiber acoustic sensor(DAS)is an advanced geophone in modern exploration,which is deployed in wells.Compared with traditional moving coil detectors,DAS has attracted extensive attention of researchers due to its low cost,strong anti-interference ability,and high temperature and high pressure resistance.However,the DAS records obtained in the field show weak signals and strong interference,and the local amplitudes vary greatly.Compared with the data obtained by the traditional moving coil sensor,the resolution is lower and the noise is more complex.In addition,there are not only random noises,but also new types of noises such as optical system noise and fading noise.Therefore,DAS is more difficult to processing.In response to this problem,we designed a convolutional neural network based on operation-wise.The network uses different computing methods to obtain the feature maps of different sizes of receptive fields and combines them with the weights generated by the channel attention,which increases the context information,selectively enhances the most important feature information and suppresses useless feature,so that the network can locate the signal precisely.In this paper,corresponding strategies are respectively constructed for the noise suppression of two types of complex seismic data.In experimental part,we measure and quantify various aspects of the denoising results,differences,f-k spectra,etc.Finally,it is concluded that the two deep learning-based algorithms proposed in this paper not only effectively reduce the noise,but also completely restore the seismic signals and ensure their continuity.