Research On Desert Seismic Random Noise Suppression Based On Convolutional Autoencoder Neural Network With Attention Module

Desert areas contain a large amount of oil and gas resources,which are mainly and effectively explored by seismic exploration methods.Random noise in seismic data collected from desert areas include environmental noise and random noise induced by the movement of human and vehicles.Affected by the loose sand layers and changeable sand dune shapes,seismic random noise appears as low-frequency colored noise with similar waveforms to the effective signal.It also has non-stationary and non-Gaussian characteristics.The complex noise characteristics make it difficult for traditional seismic denoising methods to effectively suppress desert noise,which will adversely affect subsequent seismic signal processing and imaging.Therefore,the noise reduction method suitable for desert seismic random noise is studied in this thesis to improve the quality of seismic data.An improved deep convolutional autoencoder(IDCAE)is proposed for suppressing desert random noise.Through residual learning,the encoding network of the IDCAE extracts noise features from seismic data to obtain the hidden representations,which is reconstructed through the decoding network to achieve the noise suppression.IDCAE adds a skip connection with threshold shrinkage into the structure of deep convolutional autoencoder(DCAE).The noise features obtained from the shallow layer of IDCAE are passed to the deep layer through threshold shrinkage,alleviating the problem of seismic events distortion caused by the deep network structure.Due to the lack of seismic signal in the field seismic data,the training set is simulated for desert random noise suppression.IDCAE uses a large number of data to train parameters of the model and adopts a transfer learning strategy designed for desert seismic noise suppression.In this thesis,the model trained by synthetic data is used as a pre-training model and its parameters are transferred to the field seismic denoising task.The actual desert noise data is used as the target to fine-tune the parameters of the new model,which can improve the ability of the IDCAE in preserving the field seismic signals.The denoising results of the synthetic and the actual desert seismic data show that the IDCAE obtains better denoising performance in the suppression of random noise and preservation of seismic signals than F-K mehod and DCAE.The signal-to-noise ratio and mean square error of the filtered data by the IDCAE have been significantly improved.The IDCAE also enhances the effective signals with the transfer learning.In order to better recover the complex structure of the seismic events from desert seismic random noise with weak similarity,an attention-guided deep convolutional autoencoder(AIDCAE)is proposed based on IDCAE.AIDCAE embeds the attention module(AM)in the IDCAE network.After contacting the input data through the long connection and the output of the decoding network,the attention weights are generated through the convolution and Sigmoid function in AM.This attention weights integrate the global features of the input and the local features of the decoding noise,and give different attention to the different features associated with the seismic events and random noise,the influence of the error feature on the output of the model is gradually weakened through the constraint of the weight coefficient in training.The AIDCAE is applied to the synthetic data and actual seismic data,and denoised results show that AIDCAE has achieved better signals recovery,even for the field desert seismic signals heavily destroyed by noise.This improvement is because the AM assigns different weights to the features.In addition,the attention weights can be used as a gradient filter in training,which can effectively suppress the gradient generated by the signal features,thereby improving the training efficiency of the model.The AIDCAE has10 iterations less than the IDCAE to achieve the optimal denoising results.Compared with traditional methods,the AIDCAE method has better denoising results.It requires fewer training iterations than the common deep learning methods and is more suitable for desert seismic noise suppression.