Fault detection of seismic data is a key step in seismic data interpretation. Many techniques use deep learning for seismic fault detection and have got good results. Traditional supervised learning assumes that the training data and the prediction data have a similar distribution.  However,  there are differences between the areas far apart from each other in the same work area, such as fault density and fault spatial distribution.  So it is difficult for deep learning model to get ideal result when the prediction data is far away from the training data. We propose progressive transfer learning to solve this problem: Considering the ability of transfer learning to reduce the impact of differences in data distribution, we transfer from training data to intermediate data between training data and target prediction data. Then we predict intermediate data and use the prediction result for automatic fault interpretation to get pseudo-labels, the training data set is updated with the intermediate data and the pseudo-labels. And the next intermediate data is processed progressively. With the distance between the updated training data and the target prediction data decreasing, the data distribution of them is more similar and the problem is solved. The experimental results demonstrate the effectiveness of our method.

Keywords --- Fault detection, Transfer learning, Deep learning