<p>Accurate and efficient estimation of structural failure probability often requires balancing predictive accuracy with computational cost, particularly when high-fidelity models are involved. To address this challenge, this study develops an adaptive ensemble of surrogates with approximate upper bound function (AES-AUBF) for reliability analysis. The proposed framework integrates multiple polynomial chaos Kriging (PCK) surrogates through a novel weighting scheme, enabling dynamic adjustment of model importance according to both global accuracy and local predictive uncertainty. An approximate upper bound function (AUBF) is introduced within a Bayesian active learning framework to guide the sequential selection of new informative samples, effectively reducing epistemic uncertainty in failure probability estimation. Furthermore, a reward-based learning function allocation strategy is proposed to adaptively select the most effective learning function from a portfolio, while a parallel enrichment mechanism accelerates convergence by adding multiple samples per iteration. A hybrid error-based stopping criterion ensures termination at an optimal balance between accuracy and efficiency. Three numerical examples, including a nonlinear oscillator, a multi-branch series system, and the fatigue reliability assessment of a monopile-supported offshore wind turbine, are employed to investigate the performance of AES-AUBF. Results show that AES-AUBF achieves accuracy comparable to direct Monte Carlo simulation while significantly reducing the number of function evaluations. The proposed framework provides a flexible and efficient tool for reliability analysis, and its modular structure allows seamless integration with dimension-reduction and advanced simulation techniques for future extension to high-dimensional or rare-event problems.</p>