Research On Desert Low-frequency Noise Suppression Based On Generative Adversarial Network

Oil and gas is an important non-renewable strategic resource,which is closely related to the country's economic development.Seismic exploration is the main method to search for oil and gas.It is based on studying the propagation of elastic waves in the crust.Seismic records are formed by recording the seismic waves transmitted back to the surface,which serve as the basis for imaging and interpretation of stratigraphic structures.However,when collecting seismic data,noise is also acquired.Therefore,suppressing noise in seismic records is an important part of seismic data processing.At present,conventional oil and gas resources that are easy to exploit in shallow formations are gradually on the verge of exhaustion.Therefore,unconventional oil and gas resources that are relatively difficult to exploit have become the current focus of exploration.With the continuous improvement of exploration level and the continuous expansion of exploration demand,the target of seismic exploration has gradually shifted from shallow to deep,and from flat land to mountainous and desert areas.There are many deserts in northwest China,and these areas often contain large amounts of unexplored and undeveloped oil and gas resources.However,the geological conditions and survey conditions in the desert area are very bad,resulting in intricate noise in the acquired seismic records.Meanwhile,the spectrum overlapping between effective signals and noise is serious in the desert region.The complex noise seriously affects subsequent inversion,imaging and interpretation of the seismic record.Therefore,we urgently need a practical and effective denoising strategy for desert seismic data.Most conventional noise reduction methods distinguish between signal and noise based on the difference in energy,correlation or other physical quantities in the transform domain.However,when processing desert seismic data containing complex noise,most conventional algorithms have problems such as incomplete suppression of noise and difficulty in accurately recovering signals.In recent years,methods based on convolutional neural networks(CNN)can effectively suppress the complex noise of seismic data.Relying on a complex system and huge parameters,CNN can learn a complex mapping from noisy signals to pure signals,and the denoising effect is better than traditional methods.However,the denoising algorithm based on CNN also shows great limitations.In the algorithm,the goal is mainly to train the network to minimize the mean square reconstruction error of the pure signal and the signal after CNN denoising.Although the estimate has a high signal-to-noise ratio,the result is biased towards local perception.The global structure of the signal is difficult to be accurately restored,and the high-frequency details of the signal cannot be preserved.In addition,the high dependence on the dataset also makes it difficult to train in the absence of noisy data.Models based on deep learning are data-driven.In network training,paired noisy records and clean signals are needed to construct a one-to-one corresponding dataset.However,the matched data in seismic exploration is difficult to obtain,which also seriously affects the application of CNN in the denoising of real seismic records.Based on the above two problems,we respectively propose two denoising strategies based on generative adversarial network(GAN).First of all,drawing on the idea of GAN,we increase the discriminator network to guide the generator network.The whole system contains two parts: generator network and discriminator network.The generator network is used to learn the mapping from the noisy signal to the denoising signal.After the noisy data propagates forward in the network,the network will generate a pure signal.The discriminator network is trained to distinguish the denoising signal from the clean signal,and the generator network is forced to improve the denoising quality to deceive the discriminator network.The final denoising result of the network will have structural integrity.In addition,to solve the problem of the network's excessive dependence on the data set,we also designed a cycle denoising generative adversarial network to learn the domain mapping from the noisy data domain to the clean signal data domain under the constraint of cycle consistency.The cycle consistency ensures that the learned domain mapping can retain the event information in the seismic record to a great extent.The unsupervised training system no longer requires pairs of noisy records and clean signals.The seismic records collected in the field can be entered into the dataset to participate in network training,making the model more suitable for the processing of actual records.In real desert seismic data processing,we have verified the feasibility and effectiveness of the method proposed in this paper through comparative tests.The models trained under the two strategies are data-driven automatic denoisers that do not require manual adjustment of parameters,which are superior to conventional denoising methods.More importantly,compared with the current popular research on CNN algorithms,the methods in this article improved in response to the limitations of CNN mentioned.