Abstract: As the most widely used connection method in large-scale equipment, bolted connections exhibit diverse failure modes and significant tension-bending coupling effects, posing severe challenges such as low computational efficiency and insufficient accuracy of failure criteria for traditional numerical methods (e.g., finite element method) when performing failure behavior prediction and bearing capacity uncertainty quantification. In this paper, a Bayesian-driven feature learning (Bayesian-driven feature learning, BDFL) method is proposed. Specifically, it is found that the inflection points of the bolt axial force curve are critical for predicting failure behavior, so a Bayesian-based method is used for inflection point recognition. Then, a dimension reduction process combining the supervised surrogate model and unsupervised principal component analysis is proposed. Taking the above variables and failure loads as input data, the nested stochastic Kriging model is applied for uncertainty analysis and prediction. Compared with the traditional static model and Johnson-Cook model, the results demonstrate that the proposed method can provide the most accurate results in terms of bearing capacity and uncertainty quantification. This method offers a promising approach that leverages machine learning to integrate simulation and experimental data, thereby improving large-scale equipment simulation without increasing computational costs.