基于多尺度窗口生成器网络的抽油机噪声压制OA北大核心CSTPCD

The noise of the pumping unit strongly interferes with the exploration and development of old oil fields and seriously reduces the signal-to-noise ratio of seismic data.Therefore,a pumping unit noise suppres-sion method based on a multi-scale window generator network is proposed.The constructed network is mainly composed of a double-layer encoder-decoder structure,and accurate denoising results can be obtained by com-bining characteristic information of different layers.The utilization of different-sized windows in different layers for feature extraction can effectively expand the sensing range of the neural network and extract more useful fea-tures from the pumping unit noise.To prevent the degradation of the network,residual connections are used re-spectively in each block of the encoder and decoder.The residual block of the encoder adopts the anti-bottle-neck design with a large amount of convolution kernels in the middle and small at both ends to extract more fea-tures of seismic data.The decoder uses one-fifth of the convolutional layers of the encoder,speeding up model training and seismic data reconstruction.The network constructed in this way can effectively suppress pumping unit noise in seismic data by using multi-scale semantic information.Both simulated data and real data experi-mental results show that compared with DnCNN,GAN,and MLGNet,the proposed method can obtain high-quality denoising results and retain valid data to the greatest extent.