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

To streamline fast-track processing of large data volumes, we have developed a deep learning approach to deblend seismic data in the shot domain based on a practical strategy for generating high-quality training data along with a list of data conditioning techniques to improve the performance of the data-driven model. We make use of unblended shot gathers acquired at the end of each sail line, to which the access requires no additional time or labor costs beyond the blended acquisition. By manually blending these data, we obtain training data with good control of the ground truth and fully adapted to the given survey. Furthermore, we train a deep neural network using multi-channel inputs that include adjacent blended shot gathers as additional channels. The prediction of the blending noise is added in as a related and auxiliary task with the main task of the network being the prediction of the primary-source events. Blending noise in the ground truth is scaled down during the training and validation process due to its excessively strong amplitudes. As part of the process, the to-be-deblended shot gathers are aligned by the blending noise. Implementation on field blended-by-acquisition data demonstrates that introducing the suggested data conditioning steps can considerably reduce the leakage of primary-source events in the deep part of the blended section. The complete proposed approach performs almost as well as a conventional algorithm in the shallow section and shows a great advantage in efficiency. It performs slightly worse for larger traveltimes, but still removes the blending noise efficiently.