Denoising,Reconstruction And Compression Of Seismic Signals Based On Autoencoder In Deep Learning

Seismic signal processing includes important aspects such as denoising,reconstruction and compression and so on.Due to the poor exploration environment and the increase of exploration cost,the collected seismic signals are often accompanied by various noise and information missing phenomena,which will affect the subsequent seismic signal processing.In addition,as the exploration technology is improved,the amount of collected seismic signal data increases,which increases the cost of seismic signal transmission and storage.In recent years,deep learning has made great progress in the field of machine learning,especially the flexible end-to-end encoding-decoding features of Auto Encoder in deep learning,which has aroused the interest of researchers.How to use Auto Encoder in deep learning for seismic signal processing,especially denoising,reconstruction and compression,is a new research topic.In this paper,under the framework of Auto Encoder in deep learning,the researches on seismic signal denoising,reconstruction and compression are carried out.The main research contents are as follows:(1)Denoising algorithm for seismic signal with additive noise based on Auto Encoder network.In order to attenuate the additive noise in seismic signals and improve the signal-to-noise ratio,a denoising algorithm for seismic signal with additive noise based on Auto Encoder is proposed.The core of the algorithm is to build an end-to-end deep convolutional Auto Encoder network with skip connections(AE-SC).In the AE-SC input layer,the seismic signal with additive noise is used as the input of the network,the input layer is first convoluted,and then the Leaky Re LU activation function is processed;the AE-SC hidden layer contains 9 layers,the convolutional or deconvolutional operation is performed primarily in the first 8 hidden layers,and then the batch normalization operation is required before the Leaky Re LU activation function processing.The ninth hidden layer only performs the deconvolutional operation;the AE-SC output layer outputs the additive noise.For the three kinds of additive noises of Gaussian white noise,Rayleigh noise and Exponential noise,the denoising experiment of marine seismic signals is carried out,and the convolutional neural network(CNN)denoising algorithm,the conventional wave atom transform and the curvelet transform denoising algorithm are compared.The experimental results show that AE-SC can effectively remove additive noise,which can shorten the training time and improve the denoising accuracy compared with CNN.Compared with the conventional denoising algorithm,AE-SC has stronger denoising ability,which verifies the feasibility of the algorithm.(2)Seismic signal reconstruction algorithm combined with deep convolutional Auto Encoder and generative adversarial networks.In order to accurately reconstruct the seismic signal,a seismic signal reconstruction algorithm combined with deep convolutional Auto Encoder and generative adversarial networks is proposed.The algorithm network structure consists of two parts: a generative adversarial networks' generator composed of a deep convolutional Auto Encoder(self-encoding generator)and a generative adversarial networks' discriminator.The self-encoding generator reconstructs the input missing track seismic signal and outputs a generated seismic signal.Skip connections are introduced in the self-encoding generator to speed up training and improve reconstruction accuracy.In order to generate conditionally for generative adversarial networks,the pseudo-seismic signal is superimposed by the missing track seismic signal and the generated seismic signal via channel,and the true seismic signal is superimposed by the same missing track seismic signal and the original seismic signal as the generating condition via channel,and the discriminator is used to distinguish the pseudo-seismic signal and the true seismic signal.Through training,the generated seismic signal learns continuously the data distribution law of the original seismic signal,and finally the generated seismic signal is consistent with the original seismic signal.In the test,only the self-encoding generator is used to generate the reconstructed seismic signal.Taking the marine seismic signal as a sample,the experimental results are compared with the wave atom transform,curvelet transform and wavelet transform reconstruction algorithm.The results show that the algorithm has higher reconstruction accuracy.Therefore,the algorithm is feasible for seismic signal reconstruction.(3)Seismic signal compression algorithm based on deep convolutional Auto EncoderIn order to shorten the seismic signal transmission time and reduce its storage space,it is necessary to effectively compress the seismic signal.a seismic signal compression algorithm based on deep convolutional Auto Encoder(CAE)is proposed.The network structure of CAE consists of four parts: encoder,coding,quantization and decoder.Firstly,the seismic signal to be compressed is used as the input of the encoder.The encoder is constructed by a series of convolutional layers.Three residual blocks are introduced in these convolutional layers to reduce the clutter.After many convolutional layers,the quantization layer is used.The quantization layer is followed by an output layer for output coding.The input of the decoder is coding.The decoder consists of a series of convolutional layers and deconvolutional layers.In these convolutional layers and deconvolutional layers,three residual blocks are also introduced.Finally,the output of the decoder is compressed reconstruction seismic signal.Training networks' parameters use marine seismic signals as training data set.To test the effectiveness of CAE compression,compression experiments are performed on the marine seismic signal test data set and compared to the JPEG2000 compression algorithm.Experiments show that under the same compression ratio,the mean square error and peak signal-to-noise indexes of CAE are better than JPEG2000,and the reconstruction effect after CAE compression is good;and when the peak signal-to-noise ratio is the same,CAE has higher compression ratio.