FWI Quality Control Using Unsupervised Cycle-Skipping Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional quality control (QC) methods for Full Waveform Inversion (FWI) in seismic data processing are time-consuming, labor-intensive, and subjective, with existing supervised machine learning approaches being unsuitable due to the laborious nature of data labeling, and current probabilistic methods fail to reliably identify cycle skipping issues.
Innovation Solution
Employing unsupervised machine learning (UML) with a k-means clustering algorithm to generate feature images and cycle-skipping risk maps, flagging cycle skipping in seismic data, and generating class maps to assess the quality of candidate velocity models without the need for training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional quality control methods are used for FWI, then human evaluation can identify cycle skipping issues, but the process consumes significant time and human effort, increasing turnaround time
Solution Approach 1:
The patent replaces the mechanical human evaluation system with an automated computational system using machine learning algorithms. The system automatically identifies cycle skipping issues in velocity models by comparing synthetic seismic data with observed data, eliminating the need for manual quality control while maintaining or improving detection reliability.
Solution Approach 2:
The quality control system performs self-evaluation by automatically detecting cycle skipping issues without human intervention. The machine learning model trains on labeled data and independently assesses velocity models, enabling the system to serve itself in identifying and flagging problematic areas.
2Ease of operation
If human evaluation is used for quality control, then subjective assessment can be performed, but consistency varies over time and between evaluators
Solution Approach 1:
The patent transforms subjective human evaluation into objective quantitative assessment by changing the parameter from human judgment to algorithmic computation. The system uses consistent mathematical criteria and thresholds to evaluate cycle skipping, ensuring identical results regardless of when or by whom the evaluation is performed.
Solution Approach 2:
The patent replaces the variable human judgment mechanism with a consistent computational algorithm. The machine learning model applies the same evaluation criteria uniformly across all velocity models, eliminating inter-evaluator and intra-evaluator variability while maintaining operational simplicity.
3Extent of automation
If supervised machine learning is used for quality control, then automated detection can be achieved, but laborious training data labeling is required
Solution Approach 1:
The patent extracts and separates the training data preparation step from the core automated detection process. By using pre-labeled reference data and focusing the unsupervised learning on new velocity models, the system minimizes the labor-intensive labeling requirement while maintaining high automation in the actual quality control operation.
Solution Approach 2:
The patent performs preliminary action by preparing reference velocity models and synthetic data in advance. These pre-prepared materials serve as a foundation for the unsupervised learning algorithm, reducing the need for extensive real-time labeling and simplifying the training process.
4Extent of automation
If existing machine learning approaches are used, then automated quality control can be implemented, but reliability is insufficient due to cycle skipping detection limitations
Solution Approach 1:
The patent enhances detection reliability by adding spatial and temporal dimensions to the analysis. Instead of evaluating single traces in isolation, the system analyzes entire velocity models and seismic gathers, examining cycle skipping patterns across multiple dimensions to improve detection accuracy and reduce false positives.
Solution Approach 2:
The patent combines multiple evaluation approaches into a composite quality control system. It integrates synthetic data comparison, feature image analysis, and machine learning classification to create a robust multi-faceted detection method that overcomes the limitations of individual approaches.
Data Source
AI summary
A quality control method and apparatus generate maps for cycle skipping analysis of velocity models that may be the input or the output of a full waveform inversion. A comparison of observed data and corresponding synthetic data generated using a candidate velocity model yields feature images of shots. The feature images are then grouped in clusters using unsupervised machine learning. Deviation of the feature images from an ideal non-cycle skipped image allows to measure the quality of the candidate velocity model.