Time-Dependent Reliability Analysis of Degrading Structural Elements Using Stochastic FE and LSTM Learning

Time-dependent reliability assessment of complex structural systems that require finite element (FE) simulation to evaluate the limit state function by means of random fields is computationally demanding. The computational cost is frequently unaffordable unless reliable surrogate models of the performance function are developed. In this research, a novel approach is employed to conduct time-dependent reliability analysis of degrading concrete structures based on long short-term-memory (LSTM) technique and stochastic (S)FE simulation, where the structural material properties are simulated using random fields. LSTM is a neural technique that accounts for the sequence dependence among the input variables, and it is versatile for time series prediction. It is a type of recurrent neural network used in deep learning, and it is proven to be efficient in training for large architectures. The proposed analysis method was used to determine the time-dependent reliability indexes of two examples. The first example was related to a mathematical limit state function with time-independent random variables and time-dependent Gaussian processes. The second example was related to a degrading glass fiber reinforced polymer (GFRP) reinforced concrete beam modeled using SFE. Analysis results indicate the versatility of the proposed framework of analysis and the accuracy of the predicted reliability indexes.