Perpustakaan Digital ITB
This study presents an intelligent framework for oilfield predictive maintenance that integrates machine
learning to cluster tubing corrosion risk and forecast Electrical Submersible Pump (ESP) failures in the
mature Offshore South East Sumatra fields, where 119 tubing leak incidents have been recorded across 92
wells, most occurring before the ESP's targeted run life. Principal Component Analysis (PCA) was applied
to nine corrosion-related parameters to reduce dimensionality, followed by K-Means and Gaussian Mixture
Model (GMM) clustering to segment wells into Low, Moderate, and High corrosion risk regimes. The K
Means model achieved a silhouette score of 0.417 and a Davies-Bouldin Index of 1.077, outperforming
GMM. For ESP failure prediction, six baseline classification algorithms were evaluated, together with a
seventh soft-voting ensemble that combined them, and a hyperparameter-tuned XGBoost model achieved
the best performance with a PR-AUC of 0.64 and ROC-AUC of 0.704 against unseen test data, addressing
the dataset's 75:25 class imbalance through cost-sensitive learning. SHapley Additive exPlanations (SHAP)
were used to interpret model outputs, identifying bottom-hole temperature, differential pressure, fluid
velocity, and CO2 partial pressure as the dominant drivers of ESP failure. The clustering, classification, and
SHAP outputs were integrated into a rule-based, web-based application that automatically generates well
level corrosion risk diagnostics, ESP success probability, and optimized tubing material and workover
recommendations. Validation on a blind test set of two wells withheld from training demonstrated the
framework's capability to deliver actionable, well-specific intervention strategies, though this small
validation sample reflects workflow feasibility rather than statistically robust field performance; the
approach nonetheless offers a practical, data-driven pathway to enhance tubing and ESP reliability and
reduce unnecessary workover costs in mature oilfield operations.