Seismic Random Noise Attenuation Based On Deep Learning

The influence of instrument,environment and other factors makes it inevitable to introduce random noise in the seismic data acquisition process,and the existence of random noise affects the subsequent seismic data processing process.Therefore,random noise attenuation is a very important step in seismic signal processing.Although some traditional methods can suppress the noise in seismic data,there are problems such as effective signal loss and noise residue.Therefore,it is necessary to develop a new effective denoising method.In recent years,artificial intelligence technology represented by deep learning has developed rapidly and has achieved remarkable results in image processing,speech recognition and other fields.At present,the deep learning method based on label data is the mainstream method of seismic random noise suppression,but its denoising effect depends on the huge and complex label data.Therefore,the development of an unsupervised learning method that does not rely on label data is of great significance for seismic data noise suppression.To this end,this paper proposes a new denoising method based on unsupervised learning.This method does not require additional label training set,but only needs to use the training set composed of raw noise data to train the network.(1)Research on denoising based on denoising autoencoder.Aiming at the problem that supervised deep learning denoising requires a lot of time and energy to make a training sample library,a denoising method based on denoising autoencoder is proposed.This method first randomly damages the input data to a certain degree,and then transports the damaged data to the encoding and decoding framework.The encoding and decoding framework of the denoising autoencoder is composed of multiple fully connected layer networks.The encoding framework encodes the input data into a compressed feature expression,and the decoder reconstructs the compressed feature expression into noisefree seismic data.Since there is no given label data,this method uses the error between the reconstructed seismic data and the original seismic data as the cost of convergence for model training.The experimental results of synthetic data and actual data show that this method can effectively suppress random noise and improve the profile quality.(2)Research on denoising based on convolutional denoising autoencoder.Since the denoising autoencoder does not take the structure of the 2D image into consideration,a denoising method based on the convolutional denoising autoencoder is proposed.The coding framework of this method is composed of multiple convolutional layers and pooling layers,and the decoding framework is composed of multiple upsampling layers and convolutional layers.The encoding framework is responsible for capturing the waveform characteristics of seismic data and eliminating noise accordingly;the decoding framework can expand the feature map and restore the detailed information of the seismic data,thereby obtaining reconstructed seismic data.Considering the complexity and particularity of seismic data,a multi-scale convolution module is used in the encoding and decoding stages to extract seismic data features.The experimental results of synthetic data and actual data show that this method can significantly improve the signal-to-noise ratio,and the denoising effect is better than traditional methods.