An Automatic Modal Identification Framework for Civil Structures Based on Deep Learning and Frequency-Damping Heatmaps

Abstract

Real-time estimation of modal parameters from dynamic responses is important for structural health monitoring of civil structures, facilitating the rapid development of automatic modal identification algorithms. However, existing algorithms still involve human interaction and utilize image information to a limited extent. To fill this gap, this paper proposes a computer vision-based automatic modal identification framework combining stochastic subspace identification (SSI) and faster region-based convolutional network (Faster R-CNN), which can directly estimate modal parameters from images. Specifically, the SSI method is first used to generate modal parameter candidates (including physical and spurious modes), based on which the frequency-damping images containing rich visual information on modal parameters can be obtained. Then, the Faster R-CNN is employed to obtain physical modes from the images. Finally, the modal parameters of each structural mode are obtained by sorting the extracted physical modes by natural frequencies. The proposed framework is trained and validated through numerical simulation studies. Besides, the trained framework is applied to automatically identify modal parameters of a 600-m-tall supertall building during a typhoon event. This paper aims to develop an automatic algorithm for estimating modal parameters of civil structures and to promote the application of computer vision in the field of automatic modal identification.