High-dimensional reliability analysis using deep neural networks
Department of Mechanical Engineering-Engineering Mechanics
The limit state functions are often specified in high-dimensional forms in structural reliability analysis of real engineering applications. This paper presents a new approach to actively integrate autoencoder, deep feedforward neural network, and Gaussian process modeling to tackle the curse of dimensionality. A low-dimensional latent space with distinguishable failure surface is first introduced by training an autoencoder neural network. Then a deep feedforward neural network is adopted to link the high-dimensional inputs with the latent variables, and the limit state function in latent space is approximated using Gaussian process modeling. Thus the high-dimensional reliability is estimated by applying the Monte Carlo simulation on the approximated low-dimensional limit state function. To enhance the model fidelity and accuracy of reliability assessment, a distance-based sampling strategy is developed to identify new training samples. The effectiveness of the proposed approach is demonstrated by two examples.
AIAA Scitech 2020 Forum
High-dimensional reliability analysis using deep neural networks.
AIAA Scitech 2020 Forum,
Retrieved from: https://digitalcommons.mtu.edu/michigantech-p/14461