Rock Cuttings Imaging with Deep Learning Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for rock type identification in oilfield drilling are subjective and time-consuming, and fail to accurately analyze mixtures of rock types and quantify their proportions, which are crucial for reconstructing geological layers.
Innovation Solution
A deep learning approach using a convolutional neural network for pixel-level classification of rock cuttings images, trained on annotated datasets to segment and quantify rock types, allowing for automated and semi-automated identification and quantification of rock types in drilling cuttings.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual rock type identification is performed by mud loggers, then geological background expertise can be applied, but the process becomes subjective and time-consuming
Solution Approach 1:
The patent replaces the manual mechanical examination process with an automated optical imaging system coupled with machine learning algorithms. The system captures images of drill cuttings and uses trained neural networks to automatically classify rock types, eliminating the need for manual visual inspection while maintaining or improving identification accuracy.
Solution Approach 2:
The system enables self-service automation where the imaging system and machine learning model work together without human intervention. The trained model autonomously processes images, identifies rock types, and quantifies proportions, allowing the system to serve itself in performing tasks that previously required expert mud loggers.
2Measurement precision
If manual examination of cutting samples is performed, then rock type recognition can be attempted, but quantification of rock type proportions becomes subjective
Solution Approach 1:
The patent replaces subjective manual quantification with automated image analysis. The system processes images through trained machine learning models that objectively determine rock type proportions based on pixel classification, eliminating human bias and subjectivity from the quantification process.
Solution Approach 2:
The system incorporates feedback mechanisms where the machine learning model continuously improves its classification accuracy based on training data. The model learns from annotated examples and provides feedback-driven improvements, ensuring consistent and reliable objective quantification across different samples and conditions.
3Adaptability or versatility
If traditional imaging methods are used, then simple rock types can be identified, but mixtures of rock types and variations in texture, color, and grain size cannot be accurately analyzed
Solution Approach 1:
The patent employs machine learning models that can dynamically adjust to various parameters including texture, color, grain size, and lithology composition. The system transforms images through multiple processing stages and uses learned parameters to accurately classify and quantify rock types despite variations in these physical properties.
Solution Approach 2:
The imaging system is designed with universal capabilities to handle diverse rock types and conditions. The machine learning model is trained on varied datasets encompassing different lithologies, textures, colors, and grain sizes, enabling the single system to universally process and accurately analyze any rock sample within its training scope.
Data Source
Figure 1
Figure 2
Figure 3~5
AI summary
Apparatus and methods for ascribing one of multiple predetermined sub-classes to multiple pixels of an image of an unknown rock sample retrieved from a geological formation. The ascription utilizes a deep learning model trained with an annotated training dataset. The annotated training dataset includes multi-pixel images of known rock samples and, for each known rock sample image, which sub-class corresponds to at least a subset of pixels of that image. For each pixel of the unknown rock sample image having an ascribed sub-class, which one of predetermined meta-classes is associated with that pixel is derived based on the sub-class ascribed to that pixel. The meta-classes represent different predetermined rock types. At least one property of the formation is predicted utilizing the ascription-derived meta-classes, including which rock type(s) are present in the formation.