Suppressing Well-Pump Noise From Seismic Data Based on Multilayer Generator Network

During the secondary seismic exploration in the maturing oil field, due to the complex ground conditions, well-pump noise strongly reduced the signal-to-noise ratio (SNR) and affected the subsequent application of seismic data. At present, deep learning is mainly used in suppressing random noise and ground-roll noise. Due to the obvious difference between the features of well-pump noise with random noise and ground roll, the existing deep learning seismic denoising method, which only extracts the features on one scale, is not suitable for the suppression of well-pump noise. Therefore, we adopted the coarse-to-fine denoise strategy and presented a new method using multilayer generator network (MLGnet) to suppress well-pump coherent noise. In the proposed method, the network mainly consists of multilayer encoder–decoder architecture, which could combine different layer feature information to obtain accurate denoised results. Meanwhile, we split the noisy seismic data into multiple patches in each layer, which could effectively expand the receiving range of denoising network and extract more useful features from well-pump noise. In this way, the proposed network can effectively utilize the multiscale semantic information to suppress well-pump noise from seismic data. Experimental results on synthetic data and field data illustrated that our approach could obtain high-quality denoised results and retain the valid data to the greatest extent, compared with denoising CNN (DnCNN) and generative adversarial network (GAN).