Improving signal fidelity for deep learning?based seismic interference noise attenuation

Deep learning has shown a considerable potential to significantly improve processing efficiency but not yet been widely deployed production projects of seismic signal separation such as interference attenuation. The main reasons are: First, the industry high standards for fidelity, which are critical success subsequent imaging, and deep neural network methods have matched required level; second, network's interpretability issue affected many geophysicists sponsors’ trust in technique. To develop towards end benefiting real-world production, we first attempt better understand their performance, especially how they make use local global features data. A novel quantitative research overall model behaviour on synthetic data is conducted. We simulate three types coherent components shot domain, blend them together then train separate them. In this process, random noise, component having only learnable features, selectively injected into training pairs. Three models sharing same architecture trained individually, show distinctive behaviours when applied test Step-by-step analysis each reveals that with additional noise both input output channel desired can lead decent prediction based good and, meantime, preserve almost all information from being lost. propose key lesson learnt new method fidelity shot-domain attenuation, essentially task. Its effectiveness demonstrated field Africa comparison conventional physics-based attenuation used production.