Multi-DoF Continuous Estimation for Wrist Torques Using Convolutional Neural Network

Using surface electromyography (sEMG) signals for simultaneous and proportional control (SPC) is appealing for natural control in human-machine interaction (HMI). Introducing spatial information across sEMG channels in the decoding process provides a more complete approach than methods that focus only on time and frequency domain features. In this article, we proposed an sEMG-based model for thesi-multaneous estimation of two degree-of-freedom (DoF) wrist torques, namely, Pro-Sup and Fle-Ext. Specifically, the 8 × 24 channels sEMG signals were transformed into images, which further served as inputs in a 29-layer convolutional neural network to estimate multi-DoF wrist torques. To prove the superiority of the CNN-based model, comparisons with other regression methods were performed on eight healthy subjects and two amputees in terms of the determination coefficient (R2) and normalized root mean square error (nRMSE). The results demonstrated that the CNN-based model (CNN-WAF) outperformed two conventional regression methods on estimation accuracy (R2: 0.851 ± 0.024 for Pro-Sup and 0.863 ± 0.016 for Fle-Ext; nRMSE: 6.933% ± 0.872% for Pro-Sup and 6.830% ± 0.856% for Fle-Ext). The average filtering in data preprocessing greatly improves (p < 0.05) the smoothness and performance of the estimation (CNN-WOAF: R2: 0.837 ± 0.015 for Pro-Sup and 0.853 ± 0.021 for Fle-Ext; nRMSE: 7.487% ± 0.870% for Pro-Sup and 7.370% ± 0.866% for Fle-Ext). Overall, the outcomes of this study provide meaningful methods to utilize spatiotemporal information from sEMG signals for SPC in practical applications.