An active weight learning method for efficient reliability assessment with small failure probability

In current years, the metamodel-based reliability analysis method has been developed to assess the failure probability for engineering problems involving time-consuming computational model. Despite the fact that some sequential metamodel-based reliability analysis methods have improved the computational efficiency, there still exists a certain possibility to further reduce the computational effort without loss of accuracy. In this study, an active weight learning method based upon the Kriging model is well proposed for reliability analysis. An active weight learning function based on the optimization theory is built to replace the traditional learning function, in which the important degrees of sampling points on the limit state function are assigned as different weight indices. The Kriging surrogate model is updated according to the proposed active weight learning function. In addition, the proposed strategy is extended to solve the system reliability problem, which can effectively avoid the nonlinearity of composite function in the traditional approach. A novel stopping criterion is also exploited to guarantee the convergence of the proposed method. Five numerical examples are provided to verify the effectiveness of the proposed method and convergence strategy. Results indicate that the proposed method can significantly improve the computational efficiency of reliability analysis without sacrificing computational accuracy.