Mode shape prediction based on convolutional neural network and autoencoder

Abstract

Mode shape is a dynamic characteristic that plays an important role in civil engineering. In this paper, an approach to predict the mode shape of a bridge is proposed using a convolutional neural network (CNN) and an autoencoder. First, a large mode shape database of a bridge is established by the finite element method for training networks. Second, a mode shape tensor is formed based on the mode-shape results. Then, an autoencoder is trained to encode the tensor to a three-dimensional latent-space representation and restore it from the representations. The CNN can output the representation directly rather than the mode shape to reduce the training difficulty and improve the accuracy. The CNN takes 18 bridge design parameters and an original shape tensor, which is constructed based on 16 geometric parameters. An evaluation of the test set shows that the approach can predict the first three order mode shapes well, with the accuracy of 0.92, 0.83 and 0.79, while performs poorly in the fourth and fifth orders, with the accuracy of 0.68 and 0.63. In addition, the spatial distribution of the latent space representation is explored. The necessity of an autoencoder and the original shape tensor is demonstrated.