Vibrational modal shape interpolation through convolutional auto encoder

In this paper we propose a first data-driven approach to the super-resolution of vibrational modal shapes. In order to reconstruct a high resolution modal shape from the subsampled data, we adopt a convolutional autoencoder with additional max pooling residual connections, inspired by solutions proposed in the literature of computer vision. We tested the proposed architecture on a modal shapes dataset of isotropic rectangular plates with simply supported boundary conditions. The network performance is analyzed in terms of reconstruction with respect to a baseline given by a similar network and bicubic interpolation. Moreover the robustness of the architecture with respect to noisy input and missing data has been investigated. Results suggest that the proposed network outperformed model based interpolation and it is able to deal with noisy and scaling of the input signal.