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.

翻译：研究机器人机械臂的动力学特性为基于模型的控制提供了有力支持。通常，严格的基于第一性原理的动力学建模以及机构参数的精确辨识对于实现基于模型控制的高精度至关重要，而数据驱动的模型重构则为上述过程提供了替代方法。本文以数据激活水平为指标，通过K-means聚类算法对采集的机器人机械臂数据进行分类。借助基础先验知识，我们分别挖掘分类数据背后的相应动力学特性。随后，利用稀疏非线性动力学辨识方法，根据分类数据的激活水平逐步重构机器人机械臂的动力学模型。仿真结果表明，所提方法不仅降低了基函数库的复杂度，使SINDy方法得以应用于多自由度机器人机械臂，还减少了数据噪声对回归结果的影响。最后，基于重构模型的动力学控制在实验平台上进行部署，实验结果证明了所提方法的有效性。