Geophysical Data Interpretation Using Partial Labelling and Dynamic Distance Loss
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current seismic interpretation methods face challenges with accurate and efficient identification of geological features in seismic data, particularly due to the need for large amounts of labeled data and the difficulty in calibrating Deep Learning networks to nuanced seismic data, leading to inaccurate results and high computational costs.
Innovation Solution
The implementation of a partial labelling process and Dynamic Distance Loss method within a Fully Convolutional Network (FCN) architecture, allowing for the use of small labeled areas to train ANN networks and reducing the requirement for extensive labeled data, while also improving the accuracy of fault detection by dynamically adjusting error penalties based on distance and precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional seismic interpretation methods are used, then geological features can be identified, but large amounts of labeled data and high computational costs are required
Solution Approach 1:
The patent applies partial labeling by providing ground truth labels for only a subset of training patches rather than requiring complete labeling of all patches. This allows the neural network to be trained effectively with significantly reduced labeled data, directly resolving the contradiction between detection accuracy and data quantity requirements
Solution Approach 2:
The training data is divided into multiple patches, and the system processes these patches individually or in batches. This segmentation allows selective labeling of representative patches while maintaining the ability to train on the full dataset structure, reducing the overall labeling burden while preserving detection accuracy
2Measurement precision
If Deep Learning networks are calibrated to nuanced seismic data, then interpretation accuracy improves, but computational time and costs increase
Solution Approach 1:
The system performs partial calibration by processing only a subset of patches for ground truth generation and network training. This selective approach maintains interpretation accuracy by focusing on representative examples while dramatically reducing the computational time and resources required compared to processing the entire dataset
Solution Approach 2:
The system performs preliminary processing and filtering to identify and select representative patches for labeling and training before the main training process. This preliminary action reduces the overall computational burden by pre-identifying the most informative training samples, thereby reducing total computational time while maintaining accuracy
3Reliability
If manual labeling of all seismic data is performed, then training data quality is high, but the process is time-consuming and labor-intensive
Solution Approach 1:
The patent implements partial labeling where only a subset of patches requires manual ground truth labeling. The system selects representative patches for manual labeling while using automated methods or representative sampling for the remainder, maintaining training data quality for critical examples while dramatically reducing the time and labor required compared to complete manual labeling
Solution Approach 2:
The system incorporates automated ground truth generation and patch selection mechanisms that reduce reliance on manual labeling. By using algorithms to identify representative patches and generate preliminary labels, the system serves itself for portions of the data preparation process, reducing human intervention time while maintaining quality through selective manual verification
Data Source
Figure 1~2
Figure 3(a)~3(b)
Figure 4~5
AI summary
A computer-implemented method is provided for interpreting geophysical data utilising an Artificial Neural Network (ANN), performed by electronic operations executed by a computing device, comprising: performing a training processing step on at least one training-data set, comprising the steps of: (a) generating a first label-data by segmenting said at least one training-data set into at least a first region, representing a known first region having at least one identified geological feature, and/or a second region, representing a known second region having at least one unidentified geological feature, and a third region, representing an unknown region; (b) generating a first ANN model output for a dynamically adaptable Region of Interest (ROI) of said first label-data, said dynamically adaptable ROI including said first and/or second region; (c) generating an updated label-data by selecting at least a first portion of any one of said first, second and third region, and labelingly append at least said first portion to any one of said first, second and third region; (d) generating an updated ANN model output for an updated dynamically adaptable ROI of said updated label-data; (e) repeating steps (c) and (d) until a predetermined condition is met, providing a final ANN model output; and then applying said final ANN model output to a target-data set utilising said ANN, generating a desired output data.