Machine Learning–Based Damage Pattern Discovery in Composite Pressure Vessels Using Acoustic Emission Signals

Abstract

Ensuring the safety and reliability of hydrogen storage systems requires a detailed understanding of damage evolution in composite pressure vessels (CPVs). These structures experience complex degradation processes during service, which makes damage assessment particularly challenging. Acoustic emission (AE) monitoring offers a non-destructive and powerful method to capture damage-related activity in composite materials. However, distinguishing individual damage mechanisms remains difficult due to overlapping signal characteristics. This work proposes a data-driven framework for damage mechanism identification in full-scale CPVs using AE data and machine learning (ML) methods. Since reliable ground-truth labels are unavailable, an unsupervised approach was applied to AE features to extract inherent damage-related groups. To strengthen physical interpretation, finite element (FE) simulations at the structural scale were performed using established composite failure criteria. The simulated damage evolution is correlated with clustered AE activity to validate the pseudo-labeling strategy. The resulting automatic labels were subsequently used to train an explainable ML classification model that assigns AE events to the corresponding candidate damage-mechanism classes. Feature-level interpretation are achieved using explainable ML approaches, enabling systematic analysis of how individual AE descriptors contribute to the classification of damage mechanisms. The proposed approach enables interpretable and scalable damage classification in CPVs without manual labeling, supporting reliable structural health monitoring of hydrogen storage systems.