Myoelectric Control Algorithm Based On Deep Learning And Its Applications

With the deep integration of artificial intelligence,big data and precision instrument manufacturing,intelligent equipment has been widely used in family life and industrial manufacturing.Among them,the intelligent prosthetic equipment based on electromyographic signal control has important research significance and use value in the application fields of rehabilitation medicine and intelligent prosthetic control.The EMG signal is the superposition of the action potential generated by the muscle during exercise in time and space,which can reflect the functional state of nerve and muscle movement,and it is produced ahead of body’s action.The EMG control technology accurately decodes the human motion intention by extracting the characteristic parameters of the surface EMG signal,and realizes the classification of discrete actions and the estimation of continuous motion.This control method can improve the initiative of the human in the human-machine coordination system.This paper starts with the algorithm of EMG control,designs data preprocessing and feature extraction schemes under different EMG control tasks,and focuses on the application of deep learning algorithms in discrete action classification and continuous motion estimation.In conclusion,the impact of electrodes’ movement and damage on the procedure of discrete action classification is decreased and the multi-degree of freedom;In the existing continuous motion estimation,there are few degrees of freedom and great difference between individuals,the continuous and stable proportional control of the upper limb based on deep learning are realized by us.The specific research content is as follows:(1)High-density electrodes are prone to relative movement between the electrodes during the acquisition process,and due to irregular use and electrode aging,part of the electrodes can be damaged.These problems can affect the accuracy of gesture recognition and the stability of control.In this article,by simulating the movement and damage of the original array electrode with a size of 8*24,a time and space characteristic EMG which fuses high-density EMG signal is proposed to reduce the impact of high-density electrodes’ movement on gesture recognition accuracy.This method is designed by designing a deep and fast convolutional neural network based on convolutional operation,and using the translation invariance of the convolutional layer and parameter sharing to achieve accurate as well as stable classification of gestures.Compared with the spatial feature-based Kriging Variogram,local sensitivity discriminant analysis model,and long-short-term memory network,the results show that the deep learning algorithm proposed in this paper can reach an average accuracy of 96.3% in the case of movement or damage of electrodes.A single amputee can maintain about 71.4%,which has obvious advantages over pattern recognition algorithms.Moreover,in a variety of combinations of electrode movements and damages in different individuals,the inter-individual difference is less than 3%,which shows the best control stability.(2)At present,among the commercialized electromyographic prostheses,the control method of discrete action classification based on pattern recognition algorithm is mostly adopted.This method achieves fewer degrees of freedom and its control is sequential,which make it difficult to achieve multiple degrees of freedom of anthropomorphic or continuous and flexible control.In response to the above problems,this research proposes a convolutional neural network based on time-frequency diagram input.By designing the synchronous acquisition,recalibration and feature extraction of surface EMG signal and joint angle data under the Delsys EMG acquisition system,the continuous estimation of the shoulder and elbow joints in space motion based on deep learning is realized.What’s more,compared with the multi-layer perceptron model,the algorithm proposed in this article have 83.5%,84%,and 83.5% of the three-channel angle prediction accuracy,which is 8% in total higher than that of the multi-layer perceptron model,and the difference between individuals is only 1.3 %.The result proves the feasibility of the deep learning algorithm in decoding surface EMG signals to achieve continuous control tasks.(3)The focus of the application of smart prosthesis control lies in the dexterity of hand functions,such as the flexibility of control and multi-degree-of-freedom grasping.This paper uses the public data set Nina Pro to realize the continuous motion estimation of the three algorithms under different grasping actions,and proposes a long and shortterm memory network algorithm based on the time series prediction model to establish the regression relationship between the surface EMG signal and the angle of the finger joints.From the analysis of statistical parameters,the fitting accuracy of the deep learning algorithm proposed in this paper is significantly higher than other methods.It has a better fitting effect and has higher stability between different movements.