Research On Desert Seismic Random Noise Suppression Based On Attention Neural Network

Seismic exploration is an important means of searching for oil and gas resources.In actual seismic exploration,due to the increase in depth of seismic exploration and the complexity of the acquisition environment,the obtained seismic data are inevitably affected by random noise,which greatly reduces the signal to noise ratio of seismic data,resulting in the impossibility of subsequent geological interpretation.The seismic exploration data collected in desert areas contain a large amount of random noise,with noise energy concentrated in low frequencies and characterized by non-Gaussian,non-stationary,and nonlinear characteristics,as well as weak similarity to effective signals.Therefore,effectively suppressing desert random noise and recovering clean seismic signals are of great significance for improving the quality of desert seismic data.In order to effectively suppress random noise in desert seismic data while recovering weak reflection signals,a self-attention based adaptive feature enhancement neural network(SAFE)is proposed,which includes a feature extraction module and a denoising module.The feature extraction module uses a branch structure to increase the width of the network,and the upper branch uses expanded convolution to generate a mask of complex feature information;the lower branch uses convolution to extract context features from seismic data.The two branches adaptively enhance the effective characteristics of seismic signals in a weighted manner,guiding the subsequent denoising tasks of the denoising module.The denoising module uses self-attention blocks to globally search for effective features,learning the dependency between any two positional elements,in order to aggregate non-local features and focus on features related to seismic signals.Finally,the denoising module uses convolution layers to obtain clean seismic signals and suppress similar noise.The denoising results of synthetic seismic data and actual seismic data show that compared to the denoising results of f-x deconvolution,Dn CNN,and REDNet,SAFE has the best effect in SNR and MSE,which can restore seismic events and effectively suppress desert random noise that is weakly similar to the signal.In order to further effectively recover the event with complex structure under low signal-to-noise ratio conditions,a desert noise suppression network based on multiscale attention interaction(MAINet)by combining multiscale strategy and attention mechanism is proposed.MAINet includes a multiscale feature extraction module,a multiscale feature fusion module,and a reconstruction module.The multiscale feature extraction module fully obtains multi-scale features of signals from seismic training data sets using dual branch cross blocks at different depths of the network to describe the characteristic differences between similar random noise and seismic signals.The features of adjacent scales are merged.For areas with strong noise,multi-channel features are still mixed with noise interference.The multiscale feature fusion module first emphasizes the importance of each feature channel and different regions by using shuffle attention to suppress noise features within the multi-channel and enhance the effective seismic signal characteristics.Then,the coordinate attention is used to extract features from the horizontal and vertical directions respectively to generate two independent direction-aware feature maps.These two feature maps embedded with specific direction information are applied to the input feature map complementarily,so as to locate the position of the seismic events more accurately and promote the accurate recovery of complex seismic signals by MAINet.Synthetic and actual desert seismic data processing results show that MAINet can suppress random noise in desert seismic data.Compared to traditional methods and two denoising networks,MAINet better recovers seismic events with complex structural features.