Borehole Image Artifact Removal via Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing borehole imager tools face challenges in accurately characterizing reservoirs due to image artifacts caused by tool and environmental noise, calibration limitations, current leakage, geometric mismatches, tool eccentricity, and pad to pad offsets, which traditional signal processing methods struggle to effectively address in real-time.
Innovation Solution
A machine learning technique is employed to denoise and remove artifacts from borehole images by training a regression function using a machine learning algorithm. This approach utilizes synthetic data, high-quality image data, or data processed with existing algorithms, and introduces almost no delay once the model is trained, enabling real-time correction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional signal processing methods with statistical averaging are used to remove artifacts, then some noise reduction is achieved, but processing delay increases significantly and accuracy decreases when data is noisy within the window interval
Solution Approach 1:
The system performs preliminary action by training the machine learning model offline before real-time processing. During offline training, the model learns from large datasets containing various artifact patterns, enabling it to make rapid predictions during actual borehole imaging without requiring time-consuming statistical averaging windows. This separates the computationally intensive learning phase from the time-critical inference phase.
Solution Approach 2:
The patent replaces the mechanical signal processing system (statistical averaging, filtering algorithms) with an intelligent system based on machine learning. Instead of using fixed mathematical operations that require large data windows, the trained neural network model directly processes input data to predict and remove artifacts, dramatically reducing processing time while maintaining or improving accuracy.
2Ease of manufacture
If traditional signal processing methods are used, then processing can be performed with existing tools, but operator errors and inconsistencies occur due to manual parameter tuning requirements
Solution Approach 1:
The machine learning model performs self-service by automatically learning optimal processing parameters from training data. During the training phase, the model autonomously adjusts its internal parameters (weights and biases) to minimize prediction error, eliminating the need for operators to manually tune parameters. This self-learning capability ensures consistent performance across different users and conditions.
Solution Approach 2:
The system transforms fixed, manually-tuned parameters into adaptive parameters learned from data. The machine learning model automatically adapts its parameters based on the characteristics of the training data, enabling it to handle various noise patterns and artifact types without requiring operator intervention for parameter adjustment. This changes the nature of parameters from static and manual to dynamic and automated.
3Measurement precision
If larger data windows are used for statistical averaging, then more data points are available for correction, but processing delay increases and real-time capability is reduced
Solution Approach 1:
The system performs preliminary computation during the offline training phase, where the model learns from large datasets. This pre-learning enables the model to make accurate predictions with minimal input data during real-time operation, effectively decoupling the amount of training data needed from the amount of real-time data required for processing.
Solution Approach 2:
The machine learning model provides dynamic processing capability, adapting its behavior based on the input data characteristics. Unlike fixed-window statistical methods that require a predetermined data length, the trained model can process varying amounts of data dynamically, optimizing between accuracy and speed based on the specific conditions encountered during real-time borehole imaging.
Data Source
AI summary
A method for correcting borehole images may include acquiring a raw data image of a formation using a downhole tool that takes one or more measurements and processing the raw data image through a machine learning model to form a corrected image. The method may further include displaying the corrected image and identifying one or more formation properties based at least in part on the corrected image.