3D Seismic Fault Segmentation Using Synthetic Data and Balanced Loss
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for detecting faults in seismic images are inefficient due to sensitivity to noise and stratigraphic features, and require computationally expensive calculations, especially in 3D seismic data analysis.
Innovation Solution
An efficient machine learning system, specifically a convolutional neural network (CNN), is used for 3D fault segmentation, trained with synthetic seismic images and corresponding binary fault labeling, employing a class-balanced binary cross-entropy loss function to address the imbalance in fault and non-fault pixel distribution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional fault detection methods use seismic reflection continuity attributes (semblance, coherency), then faults can be detected, but the methods are sensitive to noise and stratigraphic features which also show reflection discontinuities
Solution Approach 1:
The patent segments the fault detection process into multiple independent attribute calculations (coherency, variance, gradient magnitude) and combines them through a voting mechanism. Each attribute detects different aspects of fault characteristics, and their combination through voting improves reliability by requiring multiple independent indicators to confirm a fault, thereby reducing false detections from noise or stratigraphic features.
Solution Approach 2:
The patent creates a composite fault detection approach by combining multiple seismic attributes (coherency, variance, gradient magnitude) into a unified fault likelihood volume. This composite method leverages the strengths of each individual attribute while compensating for their individual weaknesses, particularly improving robustness against noise and stratigraphic features that may mimic fault signatures in single-attribute analysis.
2Measurement precision
If fault-oriented semblance is calculated by smoothing over all possible combinations of fault strikes and dips, then fault detection accuracy improves, but computational cost increases significantly
Solution Approach 1:
The patent performs preliminary calculations of multiple seismic attributes (coherency, variance, gradient magnitude) at standard orientations before combining them. This preliminary computation approach allows the system to prepare fault likelihood indicators in advance using efficient algorithms, avoiding the need for computationally expensive iterative smoothing over all possible strike and dip combinations while still achieving improved fault detection accuracy.
Solution Approach 2:
The patent uses multiple copies of the seismic volume, each processed to calculate a different attribute (coherency volume, variance volume, gradient magnitude volume). These copied and independently processed volumes are then combined through voting to produce the final fault likelihood. This copying approach distributes computational work across multiple simpler operations rather than requiring one complex iterative computation.
3Measurement precision
If pixel-wise fault classification is performed using CNN with local window or cube at every image pixel, then fault detection capability improves, but computational expense becomes highly expensive especially in 3D fault detection
Solution Approach 1:
The patent segments the 3D seismic volume into multiple 2D slices (inline slices and crossline slices) and processes each slice independently through the CNN classifier. This segmentation reduces the computational burden compared to processing the entire 3D volume with cubic windows at every pixel, while still capturing local fault characteristics within each 2D slice. The results from multiple slices are then integrated to produce the final 3D fault detection output.
Solution Approach 2:
The patent transforms the 3D fault detection problem into a series of 2D classification problems by processing inline and crossline slices separately. This dimensionality reduction approach allows the use of efficient 2D CNN classifiers instead of computationally intensive 3D CNNs with cubic windows, significantly reducing computational expense while maintaining fault detection capability through the integration of results from multiple 2D perspectives.
Data Source
AI summary
A machine learning system efficiently detects faults from three-dimensional (“3D”) seismic images, in which the fault detection is considered as a binary segmentation problem. Because the distribution of fault and nonfault samples is heavily biased, embodiments of the present disclosure use a balanced loss function to optimize model parameters. Embodiments of the present disclosure train a machine learning system by using a selected number of pairs of 3D synthetic seismic and fault volumes, which may be automatically generated by randomly adding folding, faulting, and noise in the volumes. Although trained by using only synthetic data sets, the machine learning system can accurately detect faults from 3D field seismic volumes that are acquired at totally different surveys.