Structural damage detection based on transfer learning strategy using digital twins of bridges

Abstract

In this paper, a novel structural damage detection (SDD) method based on the digital twin (DT) and transfer learning (TL) was proposed. The SDD methods based on the convolutional neural network (CNN) have proved their effectiveness in the many civil structures (models). However, their application to damage detection of real structures still faces some unprecedented challenges. It was widely known that a CNN needs a large number of training samples. It was difficult or/and impossible to obtain the sufficient samples covering various damage scenarios for in-service structures, which will limit the application of the CNN in real structures. Therefore, in this paper, a large number of damage samples of the numerical models were obtained by using the DT technology, and used to train a CNN as a pre-trained network. Then, the pre-trained CNN was transferred to the experimentally tested structure and real bridge structure by using the TL technology. The results confirm that the CNN trained by a large number of DT models has strong compatibility, and the detection accuracy of numerical models was more than 90%; the combination with TL technology significantly improves the performance of the CNN for experimental structures (the convergence speed was increased by 40–70%, and the detection accuracy was also improved by 5–17%). Meanwhile, the accuracy of damage detection for the real bridge structure reached 97.3% (76.6% higher than that of existing methods (non-digital twin)) by TL technology. It is demonstrated that the proposed method facilitates the application of the CNN in real structures.