In order to improve the performance of switched reluctance driving system, it is necessary to build an accurate switched reluctance motor (SRM) model. In this paper, a nonlinear flux-linkage model and a torque model of SRM are presented by using the measured accurate flux-linkage data, torque data and nonlinear mapping ability of BP neural network, which is based on fast self-configuring algorithm. In contrast with the traditional models, these two models have the abilities of fast convergence in training, good learning generalization, small network scale and easy real-time control. An experiment is carried out to demonstrate the accuracy and feasibility of the presented models. The result shows that the models have a better accuracy than the previous ones and are good for further optimization of the energy conversion and reducing the torque ripple.