Research On Random Noise Reduction Algorithm For Seismic Exploration Based On Modular Cascading Convolutional Neural Network

Seismic exploration is one of the important means to judge the status of oil and gas reserves,and it is widely used in the exploitation of many oil and gas fields at home and abroad.With the gradual extension of oil and gas resources exploration to deeper and unconventional fields in recent years,the actual seismic exploration records currently obtained usually show the characteristics of weak effective signal,strong noise interference and complex noise wave field,which makes it difficult to achieve or fail to meet the requirements of high signal-to-noise ratio,high resolution and high fidelity in the oil and gas industry.Therefore,how to reduce the background noise in seismic data,restore effective signals,and then improve the overall quality of seismic data is extremely critical,which is of great significance to promote the exploration process of oil and gas resources.Convolutional Neural Network(CNN),as the most representative technology in deep learning,can adaptively extract the intrinsic features of data based on a large number of sample learning and multiple convolution operations,and it has achieved excellent results in a variety of advanced tasks.At present,a large number of studies have shown that this kind of algorithm has great development potential in the field of noise reduction of seismic data.Therefore,this paper studies the noise reduction algorithm of seismic data based on CNN for the surface desert seismic records collected by traditional moving coil geophones and the Vertical Seismic Profile(VSP)records collected by Distributed optical fiber Acoustic Sensor(DAS).The main research of this paper is as follows:(1)In the desert ground seismic data obtained in the Tarim area,the effective signal is often interfered by random noise and surface wave.The data has strong noise energy,low frequency,and a certain degree of frequency band aliasing with the signal,and the overall signal-to-noise ratio(SNR)is low.Aiming at the problem of noise reduction of desert ground seismic data,this paper proposes a Multiscale Enhanced Cascade Residual Network(MECRN),which introduces dilated convolution and multi-scale modules to extract features of different scales and levels,and fuses the feature information from the two paths,while making full use of the shallow feature information through cascade structure and global residual path.Complements the loss of detail as the network deepens,resulting in complete reduction of random noise and area waves,as well as complete recovery of valid signals.(2)Compared with traditional electronic detectors,DAS has the advantages of high spatial resolution,large-scale long-term deployment,and anti-electromagnetic interference.However,the actual obtained DAS seismic data generally have serious noise interference,usually manifested as weak signal,strong interference,large local SNR difference,and low overall resolution.Aiming at the problem of DAS-VSP recording noise reduction,this paper proposes a Residual Modular Cascaded Heterogeneous Network(RMCHN),which obtains feature information of different scales through heterogeneous convolutional residual fusion,introduces hop connections to make full use of the influence of shallow features on the network,and combines the network components of cascading different features to enhance feature extraction capabilities,thereby improving network denoising performance.Based on CNN,this paper constructs a corresponding denoising algorithm for the noise reduction problem of two complex seismic data.Both simulation and practical experiments show that the two deep learning-based seismic data noise reduction algorithms proposed in this paper can effectively reduce a variety of complex seismic background noise,and at the same time achieve complete restoration of effective signals under low SNR conditions,and have good advantages in SNR improvement and signal amplitude retention of seismic data.