A Deep Learning Model for Leakage Identification in Multiphase Flow Systems

Abstract The secure and dependable transportation of energy via offshore pipelines relies heavily on precise leakage detection, especially in multiphase flow scenarios where conventional detection techniques are susceptible to false alarms. Although machine learning (ML) approaches have demonstrated potential in flow regime categorization and pressure drop prediction, their utilization for leakage localization in multiphase pipeline systems is yet insufficiently investigated. This study fills the gap by creating a machine learning framework to categorize leakage scenarios—no leakage, single leakage, and double leakage—in a horizontal pipeline conveying plug-type gas-liquid flow. Synthetic pressure time series data were produced utilizing a validated transient numerical model in ANSYS-Fluent, simulating 24 unique operating scenarios. The model outputs were juxtaposed with experimental data, revealing an average error of 10%, so affirming its reliability. A suite of ML algorithms, including Support Vector Machines (SVM), k-Nearest Neighbors (k-NN), and Random Forest (RF), were applied to the generated pressure signals. Statistical and spectral features were extracted from the time series, and a moving window technique was introduced to preserve dynamic flow information. Without the moving window, classification accuracies peaked at 71.43%, with spectral features outperforming statistical ones. Incorporating the moving window approach significantly enhanced performance: the RF model achieved 100% classification accuracy across all leakage scenarios, including correct identification of no-leakage cases, thereby eliminating false alarms. The results indicate that the suggested moving-window-based machine learning framework can efficiently identify transient patterns in multiphase flows and provides an accurate option for leakage detection in offshore pipelines. Future work will be extended to encompass more intricate flow regimes, including slug flow, utilizing experimental, numerical, and field data.