LSTM-augmented deep networks for time-variant reliability assessment of dynamic systems

Abstract

This paper presents a long short-term memory (LSTM)-augmented deep learning framework for time-dependent reliability analysis of dynamic systems. To capture the behavior of dynamic systems under time-dependent uncertainties, multiple LSTMs are trained to generate local surrogate models of dynamic systems in the time-independent system input space. With these local surrogate models, the time-dependent responses of dynamic systems at specific input configurations can be predicted as an augmented dataset accordingly. Then feedforward neural networks (FNN) can be trained as global surrogate models of dynamic systems based on the augmented data. To further enhance the performance of the global surrogate models, the Gaussian process regression technique is utilized to optimize the architecture of the FNNs by minimizing a validation loss. With the global surrogates, the time-dependent system reliability can be directly approximated by the Monte Carlo simulation (MCS). Three case studies are used to demonstrate the effectiveness of the proposed approach.