For the reliability analysis problem characterized by a single failure mode， cognitive uncertainty， and “black-box” models， an active learning reliability analysis method based on evidence theory is proposed. This method efficiently and accurately determines the credibility and verisimilitude of structures. It handles cognitive uncertain variables using evidence theory， initiates initial training sample construction for a Kriging model， and combines optimization methods with active learning to search for optimal training samples across the entire input variable space. This approach refines the Kriging model chronically with optimal training samples， replacing the functional function with the Kriging model to predict unknown points for credibility and verisimilitude calculation of the structure. By integrating optimization methods with active learning， the method relaxes constraints on candidate sample locations during traditional training sample search， thereby identifying training samples that better enhance the Kriging model’s correction effects and improve the efficiency and success rate of Kriging model construction. Numerical examples demonstrate the method’s computational effectiveness and its application to the reliability analysis of vehicle frontal collisions.