Research on dynamics of robotic manipulators provides promising support for model-based control. In general, rigorous first-principles-based dynamics modeling and accurate identification of mechanism parameters are critical to achieving high precision in model-based control, while data-driven model reconstruction provides alternative approaches of the above process. Taking the level of activation of data as an indicator, this paper classifies the collected robotic manipulator data by means of K-means clustering algorithm. With the fundamental prior knowledge, we find the corresponding dynamical properties behind the classified data separately. Afterwards, the sparse identification of nonlinear dynamics (SINDy) method is used to reconstruct the dynamics model of the robotic manipulator step by step according to the activation level of the classified data. The simulation results show that the proposed method not only reduces the complexity of the basis function library, enabling the application of SINDy method to multi-degree-of-freedom robotic manipulators, but also decreases the influence of data noise on the regression results. Finally, the dynamic control based on the reconfigured model is deployed on the experimental platform, and the experimental results prove the effectiveness of the proposed method.