Pipeline transportation remains one of the most efficient methods for moving oil and gas across long distances; however, leaks present ongoing operational and environmental risks. Traditional detection approaches—such as pressure-difference monitoring, mass balance methods, or threshold-based residual analysis—often struggle in multiphase flow environments where operating conditions are highly transient and variable. Such limitations can lead to delayed leak diagnosis or increased false alarm rates.

This study introduces a data-driven leakage detection framework based on supervised machine learning, trained using OLGA-simulated two-phase oil–gas flow data. Key process variables, including pressure, temperature, and mass flow rates, were processed through feature engineering to capture non-linear and transient patterns linked with leakage events. Two classification models, Random Forest and XGBoost, were developed and evaluated. Both models achieved robust performance in distinguishing leak from no-leak conditions.

To enable near real-time applicability, a sliding-window monitoring approach was implemented, allowing the models to perform continuous leak assessment while accounting for detection latency and false alarm control. Results indicate that the machine-learning framework provides improved responsiveness and stability under dynamic flow conditions compared to conventional threshold-based techniques.

This work demonstrates the potential for advanced data-driven monitoring systems to support safer and more reliable pipeline operation, particularly in challenging multiphase flow environments.