Feature mapping multi-fidelity neural network for damage assessment of aeronautical structures

Multi-fidelity methods offer a practical approach to integrating low-fidelity data with limited high-fidelity data, effectively balancing prediction accuracy and modeling costs. Existing multi-fidelity approaches in forward modeling rely on low-fidelity models refined with high-fidelity data. However, when applied to inverse problems in structural damage assessment, these methods overlook intrinsic discrepancies in the feature distribution of low- and high-fidelity data. This neglect introduces bias into the baseline predictions of low-fidelity model and degrades the final accuracy of multi-fidelity models. To address this, a novel feature mapping multi-fidelity neural network architecture is proposed, tailored for structural damage assessment. In this approach, a surrogate model is constructed using low-fidelity data via a neural network, while a feature mapping network establishes effective relationships between high- and low-fidelity data features. Additionally, a customized loss function is designed to simultaneously minimize feature mapping and damage assessment errors, ensuring efficient model training. This architecture is validated through two structural damage assessment cases. Results demonstrate that the method outperforms existing multi-fidelity approaches in prediction accuracy, requires fewer training samples, and exhibits superior generalization capabilities. These advantages contribute to enhancing the structural life management of individual aircraft and promoting the development of the airframe digital twin. Furthermore, this architecture also shows potential for broader applications to other inverse problems.