Accurate Recognition Of Hand Movements Based On EMG Signal Decoding

With the improvement of people's living standards,it is of great significance to design a functional upper limb rehabilitation robot to meet the actual needs of patients with amputation and upper limb dysfunction.In order to realize the precise control of hand movement,the core of rehabilitation robot is the accurate decoding and recognition based on surface electromyography(s EMG).Due to the weakness and instability of s EMG,s EMG is easily affected by the acquisition equipment and the surrounding environment.There are still many problems and challenges to achieve accurate action analysis,such as accurate interception of s EMG activity points,optimal combination of time-domain features,frequency-domain features,time-frequency features and non-linear features,few types of action recognition,complex algorithm and low real-time performance can not meet the actual needs.In view of the above problems,this paper mainly carried out the following aspects of research:(1)Aiming at the problem of inaccurate signal extraction of gesture activity segment,an improved multi threshold energy judgment method is proposed.On the basis of the window function method,the average energy value of each channel signal,the energy domain value of the rest segment,the length of the action data and the prior knowledge were taken as the judgment standard of the start and end point of the EMG signal,which overcomes the problem of inaccurate cutting between the signal start and end po int detection point and the silent state signal existing in the traditional short-term energy detection method.The experiment shows that the accuracy of action capture in active segment can be significantly improved by continuous judgment of multi field values;(2)According to the problem that there are many kinds of EMG eigenvalues,and the recognition effect of each eigenvalue combination is different,this paper constructs the combination feature optimization model.That was to say,all kinds of eigenvalues were preliminarily screened according to Pearson correlation coefficient method,then the high-quality feature set was extracted for permutation and combination calculation,and then each feature combination scheme was input into the algorithm for comparison and verification round by round until the optimal subset was obtained.The experimental results show that the recognition results of the combined feature optimization model are better than other feature optimization methods;(3)Aiming at the problems of few kinds of gesture recognition,low recognition accuracy and poor stability of multi-source EMG signals,a gesture recognition method based on fish swarm optimization extreme learning machine is proposed.After feature extraction and optimization,artificial fish swarm algorithm was used to iteratively select the random parameters of extreme learning machine.The experimental results show that the average recognition rate of FA-ELM algorithm model combined with combined feature model and optimal sliding window design is 96.6%,and the results of super BP,SVM and ELM are 35.6%,5.3% and 1.3% respectively,which effectively proves the accuracy of the proposed method;(4)In order to reduce the complexity of the original signal and explore the single source s EMG gesture recognition method,this paper proposes a single source EMG multi gesture recognition method based on deep learning decoding.In the experiment,the information of single flexor superficialis muscle was collected,and the selected feature data information was transformed into picture information,and then trained and tested in the convolution model framework designed in this paper.The experimental results show that the CNN network model designed in this paper achieves an average accuracy of 94.4%in decoding and recognition of single source EMG signals of 8 gestures,which is much higher than the four classification models of FA-ELM,BP,ELM and SVM,which fully proves the effectiveness of the method.