Force regression from EMG signals under different grasping patterns

To implement the force control of a prosthetic hand when grasping objects, a method of support vector regression (epsilon-SVR) is adopted to extract the force information from multi-channel myoelectric (eletromyography, EMG) signals. Six surface EMG electrodes are attached on the forearm for recording EMG signals. A six-dimensional force sensor is used for collecting the force data. The regression accuracy between these two signals is studied under several hand grasping modes, i.e., one random grasping mode and three standardized grasping modes. The experimental results show that the epsilon-SVR can achieve better cross-session regression accuracy. Under the random mode, the mean squared error (MSE) is (6.31±1.20) N, and the squared correlation coefficient (SCC) is 0.85±0.05. While under the standardized modes, the mean MSE and SCC can arrive at (5.04±0.67) N and 0.90±0.03, respectively. Companying with pattern recognition, the online force regression can acquire an error around 5 N, which is bellow 10% of the full force range.