Research On Hand Movement Recognition Method Based On Deep Learning

Surface Electromyography(s EMG)is a kind of bioelectrical signal generated on the skin Surface during muscle contraction,which reflects the movement state of muscle to a certain extent.Thanks to artificial intelligence,with the rapid development of materials science and other fields,the scholars of gesture recognition based on s EMG research also more and more thorough,the research achievements in the field of rehabilitation medical treatment,the control of prosthesis has been widely used,but whether it can accurately identify all kinds of gestures is prerequisite to improve its application scope,so how to establish a set of high-precision gesture recognition system is the key point of this study.This paper experimented and explored the classification model construction,design features and channel reduction scheme,and finally designed a set of efficient gesture classification model.The specific research contents are as follows:The experimental data used in this paper are Capg Myo database published by Professor Geng Weidong’s research group in Zhejiang University.The s EMG in the database are high-density EMG signals collected by array electrodes.In the process of collecting s EMG,there are many non-action signals and a lot of noise signals,which have a negative effect on the final gesture classification and recognition.In this paper,on the basis of signal filtering,the signal data of the action section is extracted.RMS curve method can effectively draw action section of the signal interval,by setting the appropriate threshold,accurately extract the action section of signal data,with sliding window method is used to intercept signal data,in ensuring continuity of signal at the same time,the experimental data are filled again,meet the deep learning method is the need for a large sample size data.Although deep learning has a strong learning ability and can effectively extract the features hidden in s EMG,it is generally not necessary to directly input the original s EMG data into the classification model,but to extract the features of s EMG and then input the feature parameters into the classification model.In the process of experiment,it is found that the classification accuracy of multi-feature sets is higher than that of single feature,but at the same time,the problem of feature redundancy is introduced,which requires feature reduction of multi-feature sets.In this paper,the correlation coefficient method is used to complete the reduction of feature parameters.In addition,due to the excessive number of signal channels in Capg Myo data set,signal channels must be reduced to eliminate redundant channels.In this paper,Relief F algorithm is used to calculate the weight of each signal channel,and then the ratio of the weight of each channel is compared.The channel with low weight ratio is eliminated by the reverse ordering method to achieve the purpose of channel reduction.Finally,considering the length of the memory network(LSTM)in the treatment of the time-series data have the advantage of large this article selects the LSTM networking gestures classification model,but in a variety of gestures in the process of classification,the basis of long LSTM neural network classification accuracy is not high,this paper,by using PSO algorithm to optimize LSTM network,it has been verified by the experiment,the optimized PSO-LSTM network classification accuracy is higher,further verify the PSO-LSTM gesture to the rationality and validity of the classification model.