Digital twin modeling Enabling structure full field data reconstruction by Variable fidelity data fusion

Abstract

Structural health monitoring (SHM) is crucial for ensuring that structures meet operational requirements. However, monitoring large structures often requires a complex and redundant sensor array, which can hinder efficient structural verification. The challenge lies in minimizing sensor redundancy while maintaining accuracy in capturing the structure’s full-field health, particularly in static experiments where sensor placement and data alignment are critical for precise structural monitoring. To address this issue, we propose a Multi-Task Learning-based Variable Fidelity Digital Twin Modeling (MTL-VF-DTM) approach. This method leverages abundant, low-cost simulation stress data in conjunction with a limited number of challenging sensor readings to train a neural network. Additionally, a multi-task learning mechanism is incorporated to achieve stable accuracy improvements when fusing these different fidelity data types. The proposed MTL-VF-DTM framework unfolds in two phases: 1) a Convolutional Autoencoder (CAE) is pre-trained through self-supervised learning to extract features from simulation data; 2) the decoder is co-trained using both measured and simulated data via uncertainty-based multi-task learning to optimize loss weights. Experimental results from the twin modeling of the liquid rocket motor frame and one of its components demonstrate that the proposed approach can rapidly capture full-field structural responses with fewer sensors while significantly enhancing model accuracy.