A new active learning method for system reliability analysis with multiple failure modes

For practical system problems, all response values of the component performance function can be obtained by running an expensive computational model once. In current adaptive Kriging-based methods for system reliability analysis, only one or all the component Kriging models are updated in each iteration. The former may waste computational resources, whereas the latter has the problem of overfitting. To improve the efficiency and stability of these problems, this study proposes a new active learning method for system reliability analysis, where a specific number of component Kriging models are refined in each iteration, rather than updating only one or all the component kriging models, as in the existing methods. First, a new learning function based on an approximate estimation of the error probability of the system is proposed. Subsequently, two strategies are proposed to stabilize the adaptive Kriging-based algorithms. Finally, five examples are used to compare the proposed approach with the other existing active Kriging-based methods. Practical validations show that the proposed method outperforms the other methods in terms of accuracy, efficiency, and stability.