Learning complex nonlinear dynamics of a 3D translational parallel manipulator using neural network

This study investigates modeling the dynamics of a 3D translational parallel manipulator with closed chains using feedforward neural networks (FFNNs). The dataset exceeds 50,000 samples, incorporating experimental data collected from a robot prototype using MATLAB® real-time workshop and the National Instruments^TM DAQ toolbox, as well as CAD simulation data from MSC ADAMS software. While achieving satisfactory mean squared error (MSE), some predictions did not fully capture the manipulator's dynamics, with small overfitting observed. A Deep Neural Network (DNN) was tested but faced overfitting and high computational costs, despite being trained on a subset of the dataset. This highlighted the limitations of DNNs for modeling such complicated parallel robots with closed chains and parallelograms. FFNNs were preferred for their simplicity and lower overfitting risk. L2 regularization and k-fold validation were applied to improve performance. Transfer learning (TL) was also employed, fine-tuning a new network with weights from pre-trained FFNNs using a smaller, unseen dataset. This approach significantly reduced MSE and completely eliminated overfitting, demonstrating the effectiveness of TL in refining model performance for forward and inverse dynamics. These findings suggest that FFNNs, combined with TL, L2 regularization, and k-fold validation, offer a robust method for accurately modeling complex robotic dynamics, enhancing control and optimization strategies for complicated robotic systems. Training for all networks was conducted within the MATLAB® environment.