Variational Neural Network Embedded with Digital Twins for Probabilistic Structural Damage Quantification

Abstract

Quantifying structural damage using online monitoring data is crucial for condition-based maintenance to ensure aviation safety. However, most data-driven methods hardly use accumulated domain knowledge, making it difficult to address parameter variability across different structures due to manufacturing as well as compromising result interpretability. To address these challenges, this study proposes a physics-decoded variational neural network for structural damage quantification and model parameter calibration. The innovation of this method lies in seamlessly integrating a reduced-order digital twin containing damage states and influencing parameters as a decoder within the variational neural network and training a data-driven physical feature extraction model using the variational inference. This architecture enables the individualized, real-time structural damage quantification and parameter calibration across an entire fleet, while accounting for uncertainties. Validation on typical damaged aeronautical panels demonstrates that the proposed method accurately predicts structural damage states and quantifies associated uncertainties, thereby ensuring high interpretability and accuracy. This approach is expected to be integrated into the airframe digital twin framework to enable condition-based maintenance across a fleet.