Digital twin modeling for structural strength monitoring via transfer learning-based multi-source data fusion

Abstract

Experimental measurement and numerical simulation are two typical methods to monitor the strength variation of structures. However, the former method is difficult to lay sufficient sensors on structures with large sizes or complex curved surfaces, and the latter one suffers from low prediction accuracy due to the simplification and idealization of the physical entity. How to efficiently utilize and combine these two kinds of methods remains a challenging problem. In this study, a novel digital twin modeling method via transfer learning-based multi-source data fusion (DTM-TL-MSDF) is proposed to make full use of the experiment data and the simulation data, aiming to establish an accurate digital twin model for structural strength monitoring in real-time. In the off-line stage, the clustering algorithm is used to pre-process the huge simulation data to relieve the computational burden, and the deep neural network (DNN) model is then pre-trained using the pre-processed simulation data. In the on-line stage, the pre-trained DNN model is fine-tuned using the experimental data to carry out transfer learning. Moreover, the bagging algorithm is employed in the fine-tuning process to improve the robustness and prediction accuracy due to its ability to address the dataset with only a small number of training points. To illustrate the effectiveness of the DTM-TL-MSDF method, a one-dimensional test function and an experimental study of a rectangular plate with hole under axial tension are studied. Results indicate that the DTM-TL-MSDF method can build an accurate digital twin model by integrating the simulation data and the experimental data with excellent global and local accuracy, providing a novel solution to monitor the variations of the structural full-field strength in real-time.