Unsupervised transfer learning for monitoring CFRP responses using discrete strains

Abstract

A deep learning (DL) model based on an unsupervised domain adaptation (UDA) approach is developed to learn shared features from labeled simulation datasets and transfer them to unlabeled experimental data for predicting CFRP displacement response and delamination growth. Different from traditional transfer learning methods based on fine-tuning strategies, the UDA-DL model focuses on the unlabeled target domain, aiming to learn prior knowledge in the source domain data for transfer. Specifically, a DL model with an encoder-decoder architecture is first built to construct an inverse mapping between discrete strains and displacement responses. The model is verified to efficiently and accurately predict the displacement field based on strains. Furthermore, the impact of the number of strain points and data type on the prediction of the out-of-plane displacement field is discussed. Subsequently, the UDA strategy is introduced into the DL model, which realizes the transfer of simulated data to experimental data based on shared features extracted by domain separation. The comparison with experimental results confirms the potential of the UDA-DL model in the prediction of displacement fields and delamination growth. This study provides a promising solution to the challenge of state sensing with unlabeled monitoring data in structural health monitoring.