Unsupervised ML Model for Sanding Event Prediction in Deepwater Wells
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deepwater wells face significant risks of sand contamination, leading to equipment damage and operational inhibition due to sand migration from surrounding regions, which existing technologies fail to predict effectively, resulting in costly maintenance and potential well damage.
Innovation Solution
A scalable unsupervised machine-learning model is trained using historical and real-time sensor data from wells to predict future sanding events by reconstructing operation characteristics and determining anomaly scores, allowing for timely preventative measures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional monitoring methods are used to detect sanding events, then equipment damage can be detected after occurrence, but the detection precision and timing are insufficient to prevent damage
Solution Approach 1:
The patent applies preliminary action by training the machine learning model on historical data to establish baseline patterns before actual sanding events occur. The system continuously monitors well operation data and compares it against learned patterns to predict sanding events before they cause equipment damage, enabling preventive rather than reactive detection.
Solution Approach 2:
The patent replaces traditional mechanical or rule-based monitoring systems with a machine learning-based predictive system. Instead of using fixed thresholds or simple anomaly detection, the system uses trained models to analyze complex patterns in well operation data, achieving higher detection precision and reliability for predicting sanding events.
2Measurement precision
If machine learning models are implemented to predict sanding events, then prediction accuracy improves, but system complexity increases
Solution Approach 1:
The patent applies universality by designing a machine learning system that can handle multiple types of well operation data (pressure, temperature, flow rate, etc.) and predict different aspects of sanding events using a unified model framework. This multi-functional approach improves prediction accuracy while managing system complexity through consolidation rather than multiple separate systems.
Solution Approach 2:
The system applies self-service through automated model training and deployment pipelines. The machine learning models are trained on historical data and automatically updated as new data becomes available, reducing the need for manual intervention and simplifying the operational complexity of maintaining high prediction accuracy.
3Reliability
If historical data is collected and processed for model training, then prediction capability improves, but data processing time and computational resources increase
Solution Approach 1:
The patent applies preliminary action by pre-processing and storing historical well operation data in structured formats before training is needed. Data cleaning, feature extraction, and labeling are performed in advance, so when model training is required, the system can quickly process the prepared data without time-consuming preprocessing steps, reducing overall processing time while maintaining prediction reliability.
Data Source
AI summary
An unsupervised machine-learning model is trained using historical operation characteristics of a well. Operation characteristics of the well for a duration of time is reconstructed by the unsupervised machine-learning model. Whether a sanding event will occur in the future at the well is predicted based on the difference between the operation characteristics of the well and the reconstructed operation characteristics of the well.