Training Strategy and sEMG Sensor Positioning for Finger Force Estimation at Various Elbow Angles

Abstract

Recently, several simultaneous proportional control techniques using surface electromyography (sEMG) have been developed to control protheses. However, there is a gap between research and clinical applications. This gap is due to estimation error in interposition and sEMG sensor parameters. Many previous studies have shown that the estimation accuracy of the model decreases at the interposition. Several studies have also found that the number and placement of sEMG sensors affect the performance of the models. In this paper, we propose a training strategy and guidelines for positioning sEMG sensors to reduce this gap. In this study, the finger force estimation model was trained by combining training data from various elbow angles into one feature vector. This training protocol reduced the root mean square error (RMSE) by 12.40% and increased the interdependency ratio (IR) by 3.40% in interposition. We also used the correlation coefficient between the finger force and sEMG signal as an evaluation index to determine the optimal placement of the sEMG sensor. The models trained using the channels with high correlation coefficients achieved better estimation performance, and the number of channels was reduced to four. When the model was retrained with the proposed training protocol using these 4 channels, the RMSE decreased by 10.73% and the IR increased by 1.87% in interposition. We expect the training strategy to close the researchapplication gap by reducing the number of sensors, finding the optimal placement of the sensors, and reducing the estimated error in interposition.