Co-Optimization of Vibration Suppression and Data Efficiency in Robotic Manipulator Dynamic Modeling

Abstract

In response to the limitations of vibration suppression performance caused by the difficulty in accurately modeling nonlinear friction during robotic manipulator dynamics parameter identification, this paper proposes a hybrid identification method based on a Broad Learning System (BLS) optimized by Particle Swarm Optimization (PSO). First, a joint excitation trajectory is designed using a fifth-order Fourier series with zero boundary conditions to ensure sufficient excitation of system dynamics. Then, a linear regression formulation of the manipulator’s structural dynamics is established, and the BLS network is employed to model the unstructured residuals—primarily arising from nonlinear friction—with high precision. Finally, the PSO algorithm is applied to optimize the hyperparameters of the BLS network, achieving global model optimality. Simulation results demonstrate that under typical motion conditions of the manipulator, the proposed method exhibits excellent capability in capturing nonlinear disturbances, maintaining joint prediction errors below 6 × 10−12 N·m. This significantly improves the accuracy and robustness of the feedforward vibration suppression control. Moreover, by integrating PSO-based hyperparameter optimization and trajectory design with sufficient excitation, the proposed method enhances data efficiency during the identification process, offering a novel and practical identification strategy for precise modeling and control of complex mechanical systems.