A Hybrid Machine Learning Framework for AE-Based Anomaly Detection in Pressure Vessels

Abstract

In the transportation of hazardous materials, ensuring the structural integrity of pressure vessels is essential for safety and regulatory compliance. This paper proposes a hybrid machine learning framework for acoustic emission (AE)–based anomaly analysis in pressure vessels, designed to enable early detection and characterisation of internal damage under static and fatigue loading conditions. The proposed methodology follows a multi-phase workflow comprising data preparation and validation, anomaly detection using an attention-enhanced LSTM autoencoder, adaptive threshold determination, damage type classification using a Light Gradient Boosting Machine (LightGBM), and spatial localisation of anomalies through a distributed sensor network. This hybrid architecture combines unsupervised and supervised learning to address the complexity, non-linearity, and noise inherent in AE signals. Experimental results obtained from both simulated and real-world datasets demonstrate that the proposed framework can reliably identify and differentiate multiple degradation mechanisms, including matrix cracking, fibre/matrix debonding, delamination, and fibre rupture. Furthermore, the use of attention mechanisms and interpretable outputs supports explainability, making the framework suitable for predictive maintenance and informed decision-making in safety-critical applications.