Dynamics Parameter Identification of Torque Compensation Based on Neural Networks

To address the issue of low fitting accuracy of the rigid flexible coupling dynamic model of a mechanical arm，a dynamics parameter identification method based on neural network torque compensation was proposed by analyzing the effect of dynamic parameters on joint torque. First，the dynamical model was linearized，the minimum inertial parameter was established，the mechanical arm excitation trajectory was optimized，and data sets for each joint were collected. Second，a multilayer neural network architecture was built，and the training effects of the neural network models with different numbers of hidden layers were compared，confirming the accuracy of the proposed model. The joint position，velocity and acceleration data sets were used as the input of the network architecture，the torque was calculated after neural network. Finally，the identification model was verified. The root-mean-square error of the predicted moment was taken as the evaluation standard. The findings of joint fitting torque reveal that the proposed method has better fitting accuracy than inverse dynamics and control effect on the mechanical arm system. Further，the model reduces the impact of nonlinear factors，such as joint friction，on the identification experiments，obtains a more accurate dynamical model，and enhances precision in the manipulation of the robotic arm. © 2024 Tianjin University. All rights reserved.