Concept for an ML-supported structural health monitoring of GFRP-tanks using strain data: A case study on the storage of a liquid organic hydrogen carrier under operational-like conditions

The current work provides a model-based concept for a machine-learning supported structural health monitoring system for a glass-fibre-reinforced polymer tank for the storage of a liquid organic hydrogen carrier, focusing on damage classification using strain-field data and operational parameters. Finite element generated synthetic data based on a DIN EN 13121-3-conform tank geometry are used to train and evaluate models under operationallike conditions, incorporating both ideal and imperfect tank geometries derived from buckling analyses. A Deep Neural Network model achieves the highest accuracy, whereas for classic ML methods the LinearSVC (LinSVC) model shows the best performance. Fast Fourier Transform proves to be the most effective datatreatment method, especially for imperfect geometries. Strain data in the circumferential direction of the tank cylinder yield more accurate results than axial strain. Considering the DNN, 400–500 load scenarios sufficefor ML training for ideal geometries, whereas 600-700 scenarios are required for imperfect geometries. Additive White Gaussian Noise from 1 % onward impacts prediction accuracy, with 5 % noise causing up to a 25 % accuracy reduction. At the same time, Wavelet denoising recovers accuracy close to the original non-noise values. The DNN model also proves more robust against sensor malfunctions than the LinSVC model.