Black-box Identification with Static Neural Networks of Nonlinearities of an Elastomer-based Elastic Joint Manipulator

Abstract

Through the developments of human interactive robots, Series Elastic Actuators (SEA) have played an important role in relation to safety. However, using a compliant element adds new nonlinearities to the system. Considering this, the work presented aims to develop models to describe and characterize the elastomer-based elastic joint (eSEA). Three models are proposed: a linear, a nonlinear, and a combined model. The parametric linear model is implemented using the AutoRegressive Moving Average model with the eXogenous (ARMAX) model, and the nonlinear model used is the Nonlinear Autoregressive eXogenous (NARX) model using an Artificial Neural Network (ANN). The last model combines the linear model and an ANN. The models’ predictions are analyzed and compared with the experimental data obtained from an original assembly. The ANNs show an important ability to characterize the nonlinearities involved in flexible joint dynamics. The nonlinear model composed by an ANN makes adequate predictions compared to the linear model. The best model is obtained using the linear model and ANN, decreasing the Mean Absolute Error (MAE) by 93%.