DEVELOPING A MACHINE LEARNING-BASED DIGITAL TWIN FOR REAL-TIME MULTIPHASE LEAK DETECTION IN OFFSHORE GAS PIPELINES

Abstract

Leak detection in offshore gas pipelines is critical for operational safety, environmental sustainability, and economic efficiency [1]. Conventional methods suffer from poor accuracy, low sensitivity, delayed response, and high false alarm rates, particularly under multiphase flow conditions. This thesis develops a novel digital twin framework for real-time and robust leak detection and localization. The framework utilizes OLGA-generated synthetic data, validated against experimental multiphase flow data, to optimize ML models such as Random Forest, SVM, XGBoost, and stacked ensembles. The final model achieves R2 > 0.96 for single leaks and >90% accuracy for multiple leaks with zero false alarm rate. Also, this study introduces a self-learning model to enhance the adaptability of the digital twin across different pipeline conditions, minimizing retraining efforts. This comprehensive approach advances AI-driven pipeline monitoring, with applications in automated integrity management, predictive maintenance, and real-time risk assessment, setting the foundation for next-generation offshore monitoring systems and improved leak response strategies.