Structural reliability analysis using a neural network

Abstract

In estimating reliability of a structural system, a limit-state function is needed to relate the structural state (failure or safety) to random variables of the system. However, it is not easy to obtain such an explicit function for complex structures. As a consequence, structural analysis must be performed repeatedly to check the structural state, which is very expensive. We develop an approximate limit-state function by using a neural network. Orthogonal factorial designs are selected as learning data for the network. An "active learning algorithm" is proposed to enable the network to determine important failure regions by itself and also to do further learning at those regions to achieve a good fitness with the real structural state there. The validity of the method is illustrated through numerical examples.