Desert Seismic Random Noise Reduction Based On Multi-level Wavelet Convolution Neural Network

Due to the development of science and technology and the progress of the times,the demand for oil and natural gas has become more and more large.At present,the proven and easy to exploit oil and gas reserves have been basically surveyed,so the exploration of the complex environment,difficult to exploit but relatively rich oil and gas reserves has become the focus of seismic exploration.Desert areas are often rich in oil and gas resources.In Tarim Region of Xinjiang,there are abundant oil and gas resources.Therefore,it is very important to explore the area,analyze the geological structure and understand the distribution of oil and gas inside.After surveying,there are a lot of random noises in the seismic records obtained in this area.However,high signal-to-noise ratio seismic data is the premise of analyzing stratum structure and attribute.Improving seismic signal-to-noise ratio is an important part of seismic signal processing.At present,there are some methods to suppress the random noise in the desert seismic signal.However,due to the complexity of the desert seismic noise,the processing effect of these methods is not stable,and some parameters need to be adjusted manually.Therefore,these methods often have shortcomings such as incomplete noise suppression and poor amplitude maintenance of effective signals.So these methods are difficult to meet the requirements of modern high precision degree of detection requirements.There is an urgent need for a better and widely applicable method to remove random noise.The increase of receptive field enables it to obtain more overall information of the events.According to the above problems,multi-level wavelet convolution neural network is used in this paper.The multi-level wavelet convolution neural network is a combination of two-dimensional discrete wavelet transform and convolution neural network.The two-dimensional wavelet transform is used to replace the pooling layer and the upper convolution of U-net.Because the pooling layer of the U-net will reduce the difficulty of training and increase the receptive field,but it will cause loss of information,which will affect the denoising effect.The two-dimensional wavelet discrete wavelet transform can not only achieve the same effect as the pooling layer,but also recover the data lossless due to its orthogonal characteristics,so as to realize the trade-off between the receptive field and the computational efficiency.At the same time,compared with other neural networks,the multi-level wavelet convolutional neural network increases receptive field without increasing the difficulty of training.By adjusting the training set and structure of multilevel wavelet convolutional neural network,it is suitable for the suppression of random noise in desert seismic signals.The advantages of multi-level wavelet convolution neural network have been proved by a large number of experiments.No matter from the overall denoising results,F-K spectrum and single channel effect,the multi-level wavelet convolution neural network has achieved good denoising results.At the same time,the advantage is also reflected in the versatility of denoising.Compared with the traditional band pass filter,f-x prediction filtering,wavelet transform and empirical mode decomposition,which have great differences in the processing effect of random noise with different intensity,Multi-level wavelet convolutional neural network shows suppression effect and stability to random noise of different intensities.And in the face of noise intensity is relatively large,the signal-to-noise ratio can be increased from-11.255 d B to 16.235 d B.At the same time,the trained multi-level wavelet convolution neural network also has the effect of removing surface wave.Therefore,the multi-level wavelet convolution neural network can also achieve good denoising effect for the seismic events submerged in the surface wave.