•  
  •  
 

Corresponding Author

Gnana Allan Whinney Gnanapragasam

Keywords

Artificial intelligence, Machine learning, Lost circulation, Drilling analytics, Real-time prediction, Time-series anomaly detection, Class imbalance, Upstream oil and gas, Explainable AI, Digital transformation

Document Type

Original Research Article

Abstract

Artificial intelligence (AI) and machine learning are reshaping upstream oil and gas operations, yet their value depends less on algorithmic sophistication than on converting heterogeneous sensor streams into timely operational decisions. This study develops and evaluates an AI-enabled framework for early detection of lost circulation incidents (LCIs) during drilling, positioning this safety-critical use case within the broader innovation architecture of upstream operations. Uncontrolled drilling-fluid losses generate non-productive time, well-control risk, and substantial cost. The empirical component uses historical surface drilling data from approximately 200 wells that experienced severe or total losses, recorded at an average frequency of 0.2 Hz (one data vector every five seconds). Ten principal variables — weight on bit, hook height, hook load, torque, standpipe pressure, flow-in, flow-out, rate of penetration, RPM, and total mud-system volume — together with trip-tank volume were used to construct time-series windows. The workflow combines data cleansing, window extraction, feature engineering, window normalisation, aggressive sampling, and focal loss to address severe class imbalance. Critically, wells rather than individual records were partitioned 80/10/10 into training, validation, and test sets to prevent information leakage and overly optimistic validation. Each well contributed 60 LCI and 180 normal-operation samples. The model produces a continuous real-time LCI probability curve; in the reported severe-loss event it exceeded 95% probability before the rig crew's manual pump shutdown, demonstrating an operational lead over conventional human monitoring. The results are framed as event-level proof of operational concept rather than a generalisable accuracy claim. The paper concludes with an AI maturity roadmap spanning data readiness, interoperability, explainability, human-in-the-loop oversight, continuous validation, digital twins, and edge computing.

Download

Share

COinS