Seismic fault detection is one of the critical steps in seismic interpretation. Identifying faults is crucial for characterizing and finding the potential oil and gas reservoirs. Machine learning holds promise for eliminating some of the tedious and repetitive steps in fault interpretation. Seismic amplitude data serves as input for automatic fault detection and deep learning Convolutional Neural Networks (CNN) perform well on fault detection without any human interactive work.

This presentation shows an integrated CNN-based fault detection workflow that enhances the final fault detection volume by applying pre- and post-processing and an unsupervised seismic classification to ultimately isolate faults within a 3D volume. The pre and post-processing objectives were to suppress noise or stratigraphic anomalies subparallel to reflector dip and to sharpen fault and other discontinuities that cut reflectors. To suppress cross-cutting noise as well as sharpen fault edges, a principal component edge-preserving structure-oriented filter is first applied. The conditioned amplitude volume is then fed to a pre-trained 3D synthetic CNN model to compute fault probability.

Finally, a 3D Laplacian of Gaussian filter is applied to the CNN fault probability to enhance fault images. The resulting fault detection volumes (fault probability, fault dip magnitude and fault dip azimuth) compare favorably with traditional human interpretation and in complex structural settings, provide a more complete and unbiased image of faults. Finally, the fault volume is input into an unsupervised machine learning seismic classification (SOM) to generate a 3D volume in which the faults volumes are classified into discrete neurons with known values. This provides superior final results which can subsequently be used to generate geobodies of individual faults or used directly as input to other fault surface extraction tools.