Lithology Segmentation Model Updating for Unseen Wellbore Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automated lithology systems face challenges in integrating diverse datasets, failing to generalize effectively to new environments, and lacking adaptability, leading to performance degradation when faced with unseen data.
Innovation Solution
Implement an automated lithology instance segmentation machine learning model that is continuously monitored and updated through performance evaluation, retraining, or replacement based on performance metrics and ground truth data, using techniques like Mask-R-CNN, YOLACT, and DETR for improved segmentation and classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a machine learning model is deployed in production to perform lithology segmentation, then automated analysis speed and productivity are improved, but the model encounters performance degradation when faced with new, unseen data
Solution Approach 1:
The system continuously monitors model performance metrics (precision, recall, F1-score) on new data and uses this feedback to trigger retraining when performance degrades below thresholds, ensuring the model adapts to new formations while maintaining high productivity
Solution Approach 2:
The model transitions from a static deployed state to a dynamic system that automatically retrain s when performance degradation is detected, allowing the segmentation model to adapt to new formations and maintain reliability without sacrificing speed
2Adaptability or versatility
If the model is retrained continuously to adapt to new data, then adaptability to new formations is improved, but computational resources and processing time are consumed
Solution Approach 1:
The system performs preliminary evaluation of new data against performance thresholds before initiating retraining, and uses efficient training techniques to minimize computational resources while ensuring the model learns from new formations effectively
Solution Approach 2:
The system changes training parameters such as learning rate, batch size, and epochs dynamically based on the amount and complexity of new data, optimizing the balance between adaptability and computational resource consumption
3Measurement precision
If manual labeling of images is performed to improve model accuracy, then measurement precision of lithology identification is improved, but loss of time and labor requirements increase
Solution Approach 1:
The system automatically evaluates its own performance metrics and triggers retraining without human intervention, using automated performance monitoring and threshold-based decision-making to reduce manual labor while maintaining high precision
Solution Approach 2:
Automated performance feedback loops continuously monitor model output against ground truth data and trigger retraining when precision degrades, eliminating the need for manual labeling of every image while maintaining high measurement precision
Data Source
AI summary
Aspects provide a method of automated lithology at a wellbore. The method includes imaging cuttings collected while drilling the wellbore. The method includes one or more processors: obtaining, from an instance segmentation machine learning model, one or more lithology predictions associated with the wellbore based on the images; determining one or more performance metric scores for the instance segmentation machine learning model based on the one or more lithology predictions and reference data; and comparing the one or more performance metrics scores to one or more performance metric score thresholds and/or one or more performance metrics scores of a second model. The method includes based on the comparison: labeling the images; and (i) retraining the model using the labeled images; or (ii) replacing the model with the second model. The method includes obtaining, from the retrained or second instance segmentation machine learning model, one or more second lithology predictions.