To analyze high-sigma yields of complex circuits, surrogate model methods based on Bayesian optimization are widely used. Although Bayesian optimization using Gaussian Process regression accurately samples from failure regions through uncertainty quantification, it faces computational challenges in high-dimensional spaces, requiring dimension reduction or separate nonlinear modeling that may lead to loss of information. To address this limitation, the Deep Ensemble was adopted as a Bayesian neural network approximation, which efficiently quantifies uncertainty and estimates yield in high dimensions. For improved training efficiency, combined batch Bayesian optimization and Bayesian active learning are used for adaptive sampling to exploit worst-case and high-uncertainty failure boundary regions, with feature sensitivity-based data augmentation to increase the number of samples in rare-failure regions. Experiments on circuits across different design domains confirm the effectiveness of the proposed yield analysis framework, achieving 2.6–4.6 times speed-up, compared to existing approaches with similar accuracy rates.