An accurate and stable data-driven model is proposed in this work for fatigue life prediction in presence of mean stresses. Multiple independent extreme learning machines are integrated into the model with distinct neural network configurations to simulate the complex correlations among mean stress levels, material properties, and fatigue lives. Meanwhile, the theoretical prediction, as a representation of domain knowledge, is used to optimize the data-driven processes of model training and prediction. Extensive experimental data of 13 metallic materials with different mean stress levels are collected from the open literatures for model training and evaluation. The results demonstrate that the proposed model can achieve high accuracy and good stability in fatigue life prediction under mean stress loading conditions, even with a small training dataset, showing great applicability for fatigue life prediction.