Application Of Cnn Based On Low-Rank Decomposition In Time-frequency Domain In Desert Seismic Signal Processing

Seismic prospecting is an important means to explore underground energy and resources.Through this way,we can obtain prospecting data that can accurately reflect the stratigraphic structure.With the development of our country,the oil and gas exploitation in conventional areas has entered the later stage,and the exploitation is gradually developing to the areas with complex geological conditions,which are very difficult to exploit.The Tarim Basin,located in Northwest China,is rich in oil and gas.The first task of developing these oil and gas resources is to accurately analyze its geological structure and explore its oil and gas reserves.However,desert seismic data are often characterized by low signal-to-noise ratio(SNR)due to the fickle surface conditions and desert random noise with non-stationarity,nonlinearity,low-frequency and non-Gaussian characteristics.This low SNR is likely to affect the following inversion and interpretation.Therefore,robust noise reduction is crucial and has farreaching significance for the exploration and development of oil and gas resources.Convolutional neural network(CNN)can obtain the denoising model with strong noise suppression ability through model learning.In addition,the model also has strong generalization ability,and does not need to adjust the parameters when processing different noise level data.Low-rank decomposition is an important method of matrix decomposition,which can decompose the matrix according to the properties of its internal elements.We can use low-rank decomposition to extract the signal features.In this paper,we combine CNN and low-rank decomposition to suppress the desert seismic noise.We regard the noisy desert seismic data as a two-dimensional matrix,in which the noise component is low-rank and the signal component is sparse.According to this feature,this paper uses low-rank decomposition to extract prior information,and then uses CNN to further learn the signal features.We propose a noval denoising method: Alternating direction method of multipliers based denoising convolutional neural network(ADMM-CNN).Because the ADMM-CNN is a supervised neural network,we use the synthetic seismic data that resembles the characteristics of field desert seismic data and desert noise to construct a training set.In the model learning process,we use ADMM to decompose the data into three layers(low-rank,sparse,and perturbation)which are used as the inputs of three channels neural network.After the training,we can obtain a denoising model with strong ability to suppress desert noise,and then we can use this model to suppress desert noise.In order to verify the effectiveness,superiority and generalization ability of the proposed method,we apply it to synthetic and real desert seismic data denoising,and evaluate its denoising performance by using quantitative indexes such as SNR and mean square error,as well as qualitative analysis such as time domain analysis and frequency domain analysis.Both synthetic field data tests demonstrate that the robuster performance of ADMM-CNN compared to traditional methods,also that the ADMMCNN method suppresses the desert random noise,surface wave,and desert ringing more thoroughly and recovers the effective reflections better.In the synthetic record test,the SNR improvement of the proposed method for noisy records with different noise levels can reach about 20 d B,and the amplitude preservation of effective reflections is more than 95%,which shows that the proposed method has strong suppression ability and generalization ability for desert noise.