Control of Underactuated Systems Based on Machine Learning Model: Case Studies

The rapid pace of industrial development has significantly increased the demand for virtual testing of machines and control algorithms. Besides traditional simulations, machine learning has become widespread: certain subsystems or even the entire simulation model is replaced by neural network models.Besides existing concepts, this work explores the feasibility of employing neural network models to replace dynamic equations in multibody systems.The neural network model is developed using the MLP Regressor from Scikit-learn Python package in three levels: 1) forward dynamics of minimum-coordinate models and 2) constrained models and 3) the inverse dynamics control of underactuated multibody systems. The neural network model is utilized to represent the inverse dynamics model, for which the train data are generated using forward dynamic simulations. The study demonstrates that low degree-of-freedom planar systems can be approximated by middle-scale neural network models up to six hidden layers and 166 perceptrons, requiring training time of minutes on a personal computer.