| Traditional seismic data acquisition methods usually use independent source and multiple receiver points to receive seismic data.In order to obtain high-quality seismic data,it always requires a lot of time and cost.The emergence of multi-sources blended acquisition method successfully overcomes this shortcoming and greatly improves the efficiency of seismic data acquisition.However,this approach also poses another problem : In seismic records,the wave fields of multi-sources are blended with each other,which can lead to process seismic data difficultly and degraded imaging quality.Therefore,it is necessary to separate the source wave fields.At the same time,in the process of seismic data acquisition,due to environmental and other factors,it is inevitable to acquire random noise signal,and the suppression of random noise has been a difficult problem for the seismic data processing stage.Therefore,this article aims to explore ways to solve these two difficult problems.Deep learning is a new research field in machine learning.In recent years,various neural networks have played a key role in image processing,speech recognition and other fields.Based on the relevant theoretical knowledge of deep learning,this paper proposes a deep convolutional auto-encoder neural network by improving the convolutional auto-encoder and adding a symmetric jump connection structure.The problems of random noise suppression and blended data separation is deeply explored through deep convolutional auto-encoder neural network.The characteristics of random noise signals always appear irregular,and they are mixed with useful signals in seismic data,seriously affecting the signal to noise ratio of the data.Therefore,this paper uses a deep convolutional auto-encoder neural network to learn the characteristics of random noise signals through a large amount of dataset training and achieve the purpose of suppressing random noise.In our research,we found that the use of skip connection structures accelerates and improves the training process of the network,improving the level of suppression of random noise.In the test dataset of processing effect of random noise,it is found that the network has excellent attenuation effect for different levels of noise,and can restore the structural characteristics of the original signal.In addition,we compared the random noise suppression capabilities of deep convolutional auto-encoder neural network with Dn CNN and traditional denoising methods,and finally found that the noise processing capability of deep convolutional auto-encoder neural network are more outstanding than Dn CNN and the processing effect is significantly better than traditional denoising methods.Therefore,deep convolutional auto-encoder neural network can effectively suppress random noise in seismic data,and reduce the dependence on parameters compared to traditional methods.It is a more as a suitable intelligent method for processing random noise.After processing such as pseudo-separation,the seismic source wave field to be separated can be used as a valid signal for blended acquisition data from multi-sources,while the wave field from other sources becomes a random pulse like blended noise signals.The separation of blended acquisition data can be achieved by removing the blended noise.On this basis,since deep convolutional auto-encoder neural network have demonstrated strong suppression capabilities of random noise,we use them in the study of separation issues of blended acquisition data.We independently made a dataset for training network,and used another dataset of blended collected data as a test dataset to study the effects of blended noise suppression and network generalization capability of deep auto-encoder neural network.Finally,we found that the network can effectively restore the wave field signals of the source to be separated and remove blended noise,achieving the separation of blended acquisition data.At the same time,it is also compared with the method based on the principle of noise suppression to separate blended data,and it is found that the denoising results are far inferior to deep convolutional auto-encoder neural network.In addition,we have also denoised for high intensity aliasing noise by the network,and the results show that deep convolutional auto-encoder neural network still have excellent noise suppression effects,which further proves the excellent generalization of the network extremely and it is an intelligent method and suitable for blended acquisition data separation.Based on deep convolutional auto-encoder neural network,this paper proves that deep convolutional auto-encoder neural network can effectively suppress random noise and blended noise through corresponding exploration for two different noise suppression problems,which achieve complete preservation of useful signals and separate blended acquisition data,thereby providing some assistance for subsequent processing and interpretation work. |
