Perpustakaan Digital ITB

This study aims to develop a real-time drilling anomaly detection framework by combining unsupervised machine learning methods with a rule-based approach. The primary objectives are to determine the most effective and efficient machine learning model for real-time anomaly detection, develop feature engineering as model inputs, identify early signs of drilling anomalies, and validate detection results by comparing them against historical data and field reports. The case study was conducted on three wells in South Sumatra, Indonesia (Well A, Well B, and Well C) using historical drilling parameter data simulated as if it were streaming in real-time. Well A was focused on kick detection validation, Well B on lost circulation, and Well C on stuck pipe, while bit balling cases were validated using the Drilling Anomaly Index (DAI) and lithology due to the absence of specific incident records. Data pre-processing included separating data by drilling section and handling missing values, while model preparation involved tuning the Isolation Forest algorithm hyperparameters and determining the severity threshold to balance sensitivity and noise tolerance. The results indicate that the developed framework can detect anomalies earlier than conventional observations recorded in the Daily Drilling Report (DDR) or Final Well Report (FWR) and can identify early signs such as kick potential, lost circulation potential, differential sticking, and mechanical sticking. Validation using the DAI demonstrated consistent detection of abnormal drilling behaviors across the wells. Integration with a rule-based system also enhanced decision support capabilities by providing recommended actions before anomalies escalate. This study emphasizes the importance of high-quality real-time drilling parameter data for prediction accuracy and operational safety. The developed framework provides a robust platform for proactive drilling anomaly management, with potential for further development in non-rotary operations and expanded rulebased determination. Overall, the combination of feature engineering, unsupervised learning, and a rulebased system proved effective in improving safety, efficiency, and decision-making in drilling operations.