rLSTM-AE for dimension reduction and its application to active learning-based dynamic reliability analysis

A novel method termed rLSTM-AE is developed for the low-dimensional latent space identification of the stochastic dynamic systems with more than 1000 input random variables and the active learning-based dynamic reliability analysis. First, the long short-term memory network considers both the time-variant stochastic excitation and the time-invariant random variables is developed (rLSTM), which adopts the time-series excitation as the pertinent input feature and makes it available for the metamodeling of the high-dimensional stochastic dynamic systems. To circumvent the insufficient accuracy of deep neural networks for reliability analysis results from the limited observations, autoencoder (AE) is incorporated with the rLSTM (rLSTM-AE) and utilized to decompose the approximate extreme value space found by rLSTM onto a low-dimensional latent space. The dimension of the latent space is adaptively determined by a Gaussian process regression reconstruction error, which enables the Gaussian process regression with the similar accuracy as rLSTM regarding the extreme responses prediction. The proposed rLSTM-AE conducts the low-dimensional features extraction from the perspective of the output space decomposition and considers the time-dependent property of the dynamic systems. Finally, the detected latent variables can be combined with the active learning-based Gaussian process regression for the high-dimensional dynamic reliability analysis. One single-degree-of-freedom system and a reinforced concrete frame structure subjected to the stochastic excitation are investigated to validate the performance of the proposed method.