Estimating Simultaneous and Proportional Finger Force Intention Based on sEMG Using a Constrained Autoencoder

To boost the usability of a robotic prosthetic hand, providing degrees of freedom to every single finger is inevitable. Under the name of simultaneous proportional control (SPC), many studies have proposed methods to achieve this goal. In this paper, we propose a method to generate a regression model of a neuromuscular system called the Constrained AutoEncoder Network (CAEN) that estimates finger forces using a surface electromyogram (sEMG). Modifying the autoencoder from deep learning, the model is generated in a semi-unsupervised manner where only sEMG data and finger labels are used. In the learning process, the finger labels are used at the central layer of the network, where the three finger forces are estimated, to prevent penetration of other finger signals to each finger node and the network is trained in the constrained manner. This process results in independence among estimated finger forces such that the manipulability of multiple fingers is highly improved. The proposed model was compared with four previously reported SPC models in two tests: offline and online tests. In the offline test, the CAEN showed good results but not the best results. However, in the online test, which involved reaching target positions for three fingers simultaneously and proportionally, the proposed model showed the best results for three of six online performance indices (the completion rate, completion time, and throughput). Emphasizing the independence among estimated finger forces in the training process is the key point of the proposed method distinct from previous studies and the results showed that it was effective in the online control.