Seismic NN Training Using Reflectivity Images for Sparse Bin Recovery
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The quality of neural network (NN) models in seismic data processing is limited by the quality of training data, as simulated data often differs significantly from acquired data, leading to suboptimal processing efficiency and accuracy.
Innovation Solution
Employing real seismic data preprocessing to generate training output data, including steps like denoising, deblending, deghosting, and demultiple processing, followed by interpolation and regularization to create a reflectivity image, which is then used to train NNs to map the processed data into fully populated bins, enhancing the quality of seismic data images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If simulated data is used to train neural networks for seismic data processing, then training can be performed without real data, but the quality and accuracy of processing results deteriorates due to significant differences between simulated and acquired data
Solution Approach 1:
The patent performs preliminary processing on real seismic data to generate training output data before training the neural network. This includes denoising, deblending, deghosting, and demultiple processing to create high-quality reference data that accurately represents the target processing outcome, thereby resolving the contradiction between ease of training and processing accuracy
Solution Approach 2:
The patent creates a copy of real seismic data through preprocessing to generate training output data that mirrors the characteristics of actual processed seismic data. This copying approach ensures the neural network learns from data that closely resembles real-world scenarios, maintaining high processing accuracy while enabling systematic training
2Reliability
If traditional seismic processing methods are used to generate training output data, then processing steps are well-established, but the time required for training data generation increases
Solution Approach 1:
The patent applies preliminary processing steps (denoising, deblending, deghosting, demultiple) to real seismic data to generate training output data in advance. This preprocessed data serves as reliable training targets, reducing the need for repeated processing during model development and thereby reducing overall training data generation time
Solution Approach 2:
The patent applies multiple preprocessing steps beyond what might be minimally required, including comprehensive denoising, deblending, deghosting, and demultiple processing. This excessive action ensures high reliability of training data while establishing a standardized workflow that can be efficiently replicated
Data Source
AI summary
Methods of seismic data processing employ neural networks and use a reflectivity image based on the acquired seismic data to generate output training datasets. The neural networks thus trained are used for generating production datasets, without ghosts, source effects, multiples and/or populating a predetermined set of bins in inline-crossline plane for a set of offset classes.