Perpustakaan Digital ITB

Pipeline leakage remains one of the major challenges in oil and gas transportation systems due to its potential to cause economic losses, environmental contamination, operational disruptions, and safety hazards. Conventional leak detection and localization methods, such as mass balance calculations and pressure-gradient analysis, often suffer from limited accuracy and long processing times, particularly under dynamic operating conditions and multiple-leak scenarios. With the increasing adoption of real-time monitoring systems equipped with pressure, flow rate, and temperature sensors, deep learning has emerged as a promising approach for improving leak detection and localization performance. Therefore, this study investigates the performance of deep learning-based leak detection and localization systems under both single-leak and dual-leak conditions using USAD (UnSupervised Anomaly Detection) and Transformer architectures. The proposed methodology consists of two main modules. The first module employs the USAD architecture for leak detection by learning normal operational patterns from multivariate time-series data. Anomaly scores generated by the model are evaluated using ROC-AUC analysis to determine the optimal detection threshold. The second module utilizes Transformer-based architecture to predict leak locations from segmented leak-event data. The dataset consists of pressure, flow rate, temperature, and leak-rate measurements collected from a controlled pipeline experimental setup equipped with multiple sensors and leak valves. Data preprocessing includes cleaning, missing-value handling, normalization, feature selection, and event segmentation. Hyperparameter tuning is performed using Optuna with a Tree-structured Parzen Estimator (TPE) sampler and Median Pruner strategy to improve model performance. The results demonstrate that the USAD model is capable of effectively detecting leak events in both singleleak and dual-leak scenarios, achieving high classification performance with strong sensitivity to abnormal operating conditions while maintaining low false-alarm rates. Furthermore, the Transformer model provides accurate leak localization results in the single-leak case, with low prediction errors and satisfactory regression metrics. However, localization performance decreases in the dual-leak case due to the complex interaction between multiple leak points and the similarity of pressure-drop patterns generated by different leak combinations. Comparative analysis indicates that leak multiplicity has a more significant impact on localization accuracy than on detection capability. The novelty of this study lies in the integration of an unsupervised anomaly detection model (USAD) with a Transformer-based localization framework for comparative evaluation under both single- and dual-leak conditions. Unlike previous studies that primarily focus on single-leak detection or localization separately, this study proposes a unified deep learning workflow capable of handling leak detection and localization simultaneously while assessing the influence of leak multiplicity on model performance. The findings provide valuable insights for developing intelligent real-time pipeline monitoring systems that enhance operational safety, environmental protection, and pipeline integrity management.