Fatigue failure is the main failure mode of mechanical structure, and the premise of the fatigue life pre-diction on mechanical structure is to obtain S-N curves of high precision based on test data. However, the fatigue test data often show large dispersion, so it is more appropriate to use probabilistic fatigue life curve (P-S-N) to describe the stress cycle relationship. To overcomes the limitations of the existing classical machine learning model which can only give the determined value of fatigue life but cannot quantify its dispersion in the analysis of material fatigue data, this study proposed a P-S-N curve estimation method for metal materials based on Bayesian neural net-work (BNN) model. Firstly, the weight parameters of the traditional neural network model were regarded as random variables, and the posterior distribution of the weight parameters was estimated by the BNN model according to the training samples. Then, considering that the existing posterior distribution estimation methods are easy to underestimate the uncertainty of weight parameters, this paper introduced a black-box alpha (BB-α) algorithm based on α divergence to estimate the posterior distribution of the weight parameters. Finally, the BNN model was used to analyze the fatigue data of 2524-T3 aluminum alloy, 2024-T4 aluminum alloy and S420MC steel. The results show that BNN can be used as an effective and robust model for fatigue data fitting and uncertainty quantification. Mean-while, the BNN model based on BB-α algorithm can give a more accurate uncertainty quantification results of fatigue data.