An optimized artificial neural network for human-force estimation: consequences for rehabilitation robotics

Abstract

Purpose This paper aims to present an application of design of experiments techniques to determine the optimized parameters of artificial neural networks (ANNs), which are used to estimate human force from Electromyogram (sEMG) signals for rehabilitation robotics. Physiotherapists believe, to make a precise therapeutic exercise, we need to design and perform therapeutic exercise base on patient muscle activity. Therefore, sEMG signals are the best tool for using in therapeutic robots because they are related to the muscle activity. Using sEMG signals as input for therapeutic robots need precise human force estimation from sEMG. Furthermore, the ANN estimator performance is highly dependent on the accuracy of the target date and setting parameters. Design/methodology/approach In the previous studies, the force data, which are collected from the force sensors or dynameters, has widely been used as target data in the training phase of learning ANN. However, force sensors or dynameters could measure only contact force. Therefore, the authors consider the contact force, limb’s dynamic and time in target data to increase the accuracy of target data. Findings There are plenty of algorithms that are used to obtain optimal ANN settings. However, to the best of our knowledge, they do not use regression analysis to model the effect of each parameter, as well as present the contribution percentage and significance level of the ANN parameters for force estimation. Originality/value In this paper, a new model to estimate the force from sEMG signals is presented. In this method, the sum of the limb’s dynamics and the contact force is used as target data in the training phase. To determine the limb’s dynamics, the patient’s body and the rehabilitation robot are modeled in OpenSim. Furthermore, in this paper, sEMG experimental data are collected and the ANN parameters based on an orthogonal array design table are regulated to train the ANN. Taguchi is used to find the optimal parameters settings. Next, analysis of variance technique is used to obtain significance level, as well as contribution percentage of each parameter, to optimize ANN’s modeling in human force estimation. The results indicate that the presented model can precisely estimate human force from sEMG signals.