ABSTRACT
Time-variant reliability analysis is an essential computational procedure to quantify the failure probability and risk throughout the life-cycle of structures. A major challenge of time-variant reliability analysis is the computational effort and expense, especially when sampling-based methods are involved. Artificial Intelligence inspires a new perspective for developing advanced life-cycle computational tools in this field. This paper presents a meta-learning method to provide an alternative computational scheme that can achieve higher efficiency in terms of utilizing the evaluations from earlier point-in-time reliability analysis. The proposed meta-learning method uses the deep neural network (DNN) as the surrogate for reliability analysis. The model-agnostic meta-learning (MAML) method is adopted as the meta-learning algorithm to learn a set of adaptable parameters for subsequent DNN training. A corrosion problem is used to illustrate the applicability and efficiency of the proposed method.
