Prediction of externally applied forces to human hands using frequency content of surface EMG signals

In this work, a new signal processing method was proposed in order to predict externally applied forces to human hands by deriving a relationship between the surface electromyographic (SEMG) signals and experimentally known forces. This relationship was investigated by analyzing the spectral features of the SEMG signals. SEMG signals were recorded from three subjects during isometric contraction and from another three subjects during anisometric contraction. In order to determine force-SEMG signal relationship, higher order frequency moments (HOFMs) of the signals were calculated and used as characterizing features of SEMG signals. Subsequently, artificial neural networks (ANN) with backpropagation algorithm were trained by using the HOFMs. Root mean square difference (RMSD) between the actual and predicted forces was calculated to evaluate force prediction performance of the ANN. In addition to RMSD, cross-correlation coefficients between actual and predicted force time histories were also calculated for anisometric experiment results. The RMSD values ranged from 0.34 and 0.02 in the isometric contraction experiments. In the anisometric contraction tests, RMSD results were between 0.23 and 0.09 and cross-correlation coefficients ranged from 0.91 to 0.98. In order to compare the performance of the HOFMs with a widely used EMG signal processing technique, root-mean-squared (RMS) values of the EMG signals were also calculated and used to train the ANN as another characterizing feature of the signal. Predicted forces using HOFMs technique were in general closer to the actual forces than those of obtained by using RMS values. The results indicated that the proposed signal processing method showed an encouraging performance for predicting the forces applied to the human hands, and the spectral features of the EMG signal might be used as input parameter for the myoelectric controlled prostheses.