Hand Motion Recognition Based On Deep Learning And EMG Feature Selection

Bioelectrical signals can provide a more natural,convenient,and effective connection for human-computer interaction.However,in the research of EMG pattern recognition based on traditional machine learning,the increase in the types of EMG features and the dimension of the feature set will decrease the generalization of the classification model.Moreover,considering that the recognition effect of the intelligent model relies too much on the advantages and disadvantages of the EMG feature extraction method,In practical applications,due to the diversity of gesture categories,the existing manual feature extraction methods combined with traditional machine learning models are still difficult to fully distinguish the subtle differences between similar gestures.Therefore,it is of great significance to carry out research on EMG pattern recognition combined with reduction of high-dimensional feature sets.For the above problems,this paper takes the Nina Pro DB1 EMG data set as the research object,uses 32 EMG feature extraction methods,and uses the EMG feature selection algorithm and deep learning as the research basis,and carry out research on EMG pattern recognition framework combining the best EMG feature set and deep convolutional neural network.The main research work of this paper includes:(1)Aiming at the problem of different sources of original data for EMG pattern recognition and the wide variety of EMG feature extraction.This paper used the Nina Pro DB1 EMG set as the original EMG signal data set.And performed signal preprocessing.Finally,extracted 32 time-domain and frequency-domain EMG features in the above data set,and formed the original EMG feature set for subsequent feature selection.(2)Aiming at the current problem of higher dimension of EMG feature set.This paper proposed a two-way recursive feature selection algorithm,determine the candidate EMG feature set with the greatest correlation through the forward dynamic recursive feature selection algorithm.And combine the backward dynamic filtering feature selection algorithm to further remove the redundant features.Finally,the best EMG feature set is constructed for deep learning model training below.(3)Aiming at the limitations of traditional machine learning gradually reflected in EMG pattern recognition.This paper used a deep convolutional neural network model to replace the traditional machine learning model for EMG image gesture recognition.And used a one-dimensional convolution kernel to replace the conventional multidimensional convolution kernel to extract deep abstract features of EMG and EMG feature images.Finally,simulation analysis shows that the deep convolutional neural network combined with multiple EMG feature images can achieve the highest gesture recognition rate of 86.19%,which is 8.65% higher than the best recognition rate of machine learning model.