Safeguarding Pipeline Integrity Through Stacked Ensemble Learning and Data Fusion

Abstract

This research presents a novel approach to pipeline Structure Health Monitoring (SHM) by utilizing frequency response function signals and integrating advanced data-driven techniques to detect and evaluate vibration responses regarding loose bolts, scale deposits within pipelines, and cracks at pipeline supports, aiming to determine the effectiveness of utilizing artificial neural networks (ANN) and an ensemble learning approach in detecting the aforementioned damages through a data-driven approach. The research starts by recording 6500 samples captured by two accelerometers, related to 11 replicated pipeline structural scenarios. The research demonstrated the potential of principal component analysis (PCA) in dimensionality reduction, achieving approximately 81% reduction in data set 1 acquired by accelerometer 1 and around 79.5% in data set 2 acquired by accelerometer 2, without significant loss of information. Additionally, two ANN base models were employed for fault recognition and classification, achieving over 99.88% accuracy and mean squared error values ranging from 0.00006 to 0.00019. A significant innovation of this work lies in the implementation of an ensemble learning approach, which integrates the strengths of the base models, showcasing outstanding performance that was proved consistent across multiple iterations, effectively mitigating the weaknesses of the base models and providing a reliable fault classification and prediction system. This research underscores the effectiveness of combining PCA, ANN, k-fold cross-validation, and ensemble learning techniques in pipeline SHM for improved reliability and safety. The findings highlight the potential for broader applications of this methodology in real-world scenarios, addressing urgent challenges faced by infrastructure owners and operators.