Cross-Comparison of EMG-to-Force Methods for Multi-DoF Finger Force Prediction Using One-DoF Training

Abstract

Surface electromyography (sEMG) signal is one of the widely applied biological signals in the research field of the force intention prediction. However, due to the severe cross-talk issue of sEMG signals during fine hand contractions, few studies have related sEMG to multiple degree-of-freedom (DoF) force prediction of individual fingers simultanously. Accordingly, this study proposed methods mainly based on neural networks: Convolutional neural Network (CNN) and Recurrent Neural Network (RNN) to achieve better prediction results. Several improvements on traditional methods are also proposed in this article such as: Common Spatial Pattern (CSP), Softmax function and several new channel selection standards to solve the cross-talk issues for the estimation of EMG-force during multiple finger contractions. High-density sEMG signals of forearm extensor muscles were obtained, and experimental data from seven able-bodied subjects were analyzed. Subjects produced 1-DoF and Multi-DoF forces up to 30% maximum voluntary contraction (MVC). Then, the root-mean-square values of sEMG were related to joint force. To realize a better practical use, the EMG-to-force models were trained with minimal numbers of trials (using 1-DoF trials only), then assessed on multi-DoF trials. Our results showed that the proposed modifications on traditional method also made an improvement on the prediction results. Our findings suggest that Multi-DoF control for individual fingers with minimal training procedure (using 1-DoF trials only) may be feasible for practical use. Furthermore, methods based on neural networks greatly outperform traditional methods and the combination of CNN and LSTM showed the best performance. INDEX TERMS Neural networks, EMG-force prediction, EMG signal processing.

MuscleNET: mapping electromyography to kinematic and dynamic biomechanical variables by machine learning

“…The initial architecture of machine learning models was adopted from prior researches [13, 28, 33, 35–37, 55]. The number of hidden layers, neurons in each hidden layer, feedback delay, input delay, convolution filtering size, pooling size, and strides were selected by trial and error (Figure 4) to have the highest regression accuracy.…”

Optimal strategy of sEMG feature and measurement position for grasp force estimation

“…Huang [ 10 ], Hu [ 11 ], and Zhang [ 12 ] employed a high-density array with 128 channels of sEMG electrodes to predict the muscle force, elbow-flexion force, and joint force, respectively. Chen et al performed a cross-comparison of sEMG to a force method based on CNN and recurrent neural network (RNN) for multi degrees of freedom finger force prediction with high-density sEMG signals (160 channels) [ 13 ].…”

“…That is, the input dimension of GRNN equals to the number of features multiplied by the number of measurement positions. For different FSs and MPSs, the input dimension may be the same, such as regions 4, 6 and 13 in Fig10.Regions 4,6,13 are the result of one feature and four measurement positions, two features and two measurement positions, four features and one measurement position, respectively. Although the input dimension of these three regions is 4, the numbers of features and measurement positions are different, and the results of these regions are different in all evaluation indexes.…”

A novel sEMG-based force estimation method using deep-learning algorithm

“…Results demonstrated that the combination of CNN and long short-term memory (LSTM) network is an effective method. Apart from hand grasping, Chen et al [32] delved degrees of freedom force prediction on hand with the similar method. While both of them obtained sEMG signals with high-density electrodes, which are expensive and unnecessary [23].…”