Unsupervised Recurrent Network DAS Data Processing Based On Attention Mechanism

Oil is a familiar non-renewable resource,and the ratio of the net import weight of crude oil to the national oil consumption of a country indicates the dependence of a country’s oil consumption on foreign oil.China’s oil dependence has increased in recent years,so seismic exploration shows a very important position in the development of oil and gas resources.In the past two decades,seismic exploration has become one of the important tools for oil and gas exploration,mainly through analog geophones or digital geophones to obtain vertical seismic profile(VSP)data.However,there are some limitations when using conventional geophones for VSP data acquisition.On the one hand,borehole seismic exploration requires a large number of geophones,which consumes a lot of labor and cost.On the other hand,a high degree of coupling between geophones and ground machines is required,otherwise the reliability of the acquired seismic data will be directly affected,which greatly reduces the exploration efficiency.In recent years,a new sensing technology,Distributed Acoustic Sensing(DAS),has received a lot of attention in the field of seismic exploration.The shortcomings of the geophone.Although DAS has been successfully applied to borehole seismic data acquisition,the seismic data acquired by DAS usually contains complex types of noise and relatively high noise levels due to the spatially distributed deployment of DAS systems and complex exploration environments.For example,random noise,from current perturbations;fading noise due to interference cancellation of backscattered light;low frequency noise,influenced by temperature variations;and horizontal noise caused by potential leakage of electronic equipment.To solve the multi-type DAS noise suppression problem,this paper designs a deep learning framework with unsupervised cyclic training.Also,this paper introduces an attention module to enhance the extraction ability as well as feature recognition of signals,and achieves the recovery of high amplitude conservation of effective signals(low-frequency,high-frequency and faint signals)to improve the continuity of seismic events.In addition,the performance of supervised deep learning-based methods often relies on a large number of high-quality datasets with labels.To improve the dataset dependence problem,this paper constructs an augmented noise dataset that uses not only noise collected from some field data but also synthetic noise(e.g.,synthetic horizontal noise)to enrich the noise set.In the constructed network architecture,however,for seismic surveys,data labeling requires a lot of expertise and a lot of time,so only a small amount of labeled raw seismic data is available.Therefore,the insufficiency of labeled datasets becomes one of the main bottlenecks affecting most deep learning-based denoising methods.To avoid this situation,this study introduces adversarial loss and periodic consistent loss instead of the commonly used L1 loss or L2 loss to train the network.In addition,a hybrid training set containing synthetic seismic data and field seismic data is constructed in this study for the pre-training and fine-tuning process.In order to verify the feasibility and effectiveness of the proposed method,both synthetic data experiments and field data experiments are conducted in this study,and the results show that the research method effectively suppresses various types of noise on the experimental data and achieves signal recovery to ensure the continuity as well as the amplitude preservation of the signal.