Last modified: 2021-06-15
Abstract
The computational inverse techniques provide an effective solution for many engineering problems to inverse the internal characteristics and external environmental parameters which are difficult to be determine in complex structures. However, engineering problems are very complicated and the calculation scale is large, many high dimensional response data usually contains a lot of redundancy. Due to the existence of these high-dimensional responses with many redundancies, the computational difficulty of parameter inversion is seriously increased and the computational efficiency is reduced. We propose a dimensionality reduction method called “bound–autoencoder” which combines standard autoencoder and PCA by introducing some constraints are applied on weights of autoencoder. This method can conduct linear and nonlinear dimensionality reduction. Based on the two-way TubeNets and DWI method, the bound–autoencoder is then applied to identify material constants of composite and make comparison with PCA and standard autoencoder. Exemplification demonstrates bound-autoencoder not only overcome the shortcomings of PCA and standard autoencoder, but also, cooperating with two-way TubeNets and DWI method, both linear and nonlinear bound-autoencoder can achieve high accuracy in identifying parameters. The proposed bound-autoencoder denote a novel way for dimensionality reduction in inverse problem.