Due to scarcity of the evaluation index of rubber asphalt mixture fatigue performance, and the difficulty of determination of fatigue failure criterion, on the basis of research of fatigue damage evolution process of rubber asphalt mixture, the characteristic value of damage curve was taken as the evaluation index of fatigue performance, and the prediction model of BP neural network full cycle fatigue life was established. The results show that the deterioration rate of the material during the fatigue damage of the rubber asphalt mixture can be maintained in a stable state without obvious fatigue failure. The lossless stiffness modulus S0, instability rate V, fatigue stability K and transformation stiffness modulus St can be used as the evaluation index of fatigue performance with clear physical meanings. To obtain high accuracy, the BP neural network model is used to predict the full cycle fatigue life. The Levenberg Marquardt training algorithm has high convergence speed and great generalization ability with the maximum relative error between 170% and 823%.