SEMG-Based Continuous Estimation Of Multijoint Motion Under Complex Tasks

Surface EMG signal is an electrophysiological signal that can reflect the intention of human movement.It is widely used in clinical medicine,rehabilitation medicine and human-robot interaction.Compared with pattern recognition,continuous motion estimation based on surface EMG signals can provide a more flexible,natural and efficient way of human-robot interaction.This study analyzes the existing problems and summarizes three types of factors that complicate the task which will bring challenges to estimation,that is multiple-degrees-of-freedom motion,multi-state motion and multi-mode motion.This paper proposes different estimation algorithms to evaluate the estimation effect from the aspects of accuracy,real-time performance and robustness aiming at estimation of complex tasks.The main work of this paper is as follows:1)The sparse pseudo-input Gaussian process regression algorithm is used to construct the mapping model of the multi-degrees-of-freedom motion of the hand and the surface EMG signal.In this study,we quantitatively analyze the delay between the estimated angle and the measured angle,and the delay between the activation of the s EMG signal and the generation of the estimated angle.The average CC of the SPGP model estimation effect is0.89,the average RMSE is 13.7°,and the estimation delay is less than 150 ms.The results show that the SPGP model can achieve high precision,high real-time online estimation.2)Aiming at the influence of the grasping speed and grasping force on the s EMG state during the grasping motion,this research proposes a simplified long-short term memory networks(s LSTM)to achieve continuous estimation of the five-degree-of-freedom finger movement under a variety of electromyographic states.The experimental results of five healthy subjects and one amputee subject showed that the s LSTM model was significantly better than the SPGP model in the estimation of multi-state s EMG signals(p = 0.003).Compared with the amputee subject,the estimation accuracy of healthy subjects will also be improved,because they can stably control the activation of s EMG signals in various states.The average CC of healthy subject and amputee subjects were 0.92 and 0.91,the average RMSE were 12.9° and 20.4°,respectively.Experimental results prove that both healthy subjects and the amputation subject can realize the decoding of s EMG and then control the manipulator in a variety of s EMG states through s LSTM method.3)The human hand adjusts the grasp mode for objects of different shapes,which called multi-mode motion.For such multi-mode motion of the human hand,this study proposes an estimation framework using real-time pattern recognition combined with continuous estimation.Specifically,a hybrid estimation model of random forest combined with s LSTM is used to achieve four-modes of hand movement.The feature extraction method of Bayesian filtering is adopted to extract stable EMG feature so that improve the accuracy of real-time pattern recognition.The proposed hybrid estimation model effectively improves the estimation accuracy compared with the traditional estimation model,especially RMSE dropped from 13.1°to 12.6°.The result proves that the random forest(RF)+ s LSTM hybrid estimation model can make the continuous estimation of multi-mode motion with high precision and high robustness.