| With the continuous advancement of technology,the intelligent bionic hand has brought great convenience to the amputee disabled patients.Although the intelligent bionic hand has made great research progress in mechanical devices,its precise control,especially the use of non-invasive bio signal precise control,needs to be further improved.In the field of biological signal control of intelligent prosthetic hands,the use of surface EMG signals to achieve precise control of intelligent prosthetic hands has become a research hotspot.The problem of low accuracy of gesture recognition based on surface EMG signals restricts the application of intelligent prosthetic hands and becomes an urgent problem to be solved.In this paper,the recognition accuracy of surface EMG signals is improved by setting multiple parallel channels with different convolution kernel sizes and time-frequency feature fusion.The specific research work is mainly reflected in the following aspects:(1)In view of the low accuracy of gesture recognition,a gesture recognition model based on the Improved Multi-Channels Convolutional Neural Network(IMC-CNN)structure was constructed to improve the recognition accuracy of gesture actions.The model constructs a CNN network with 3-channel convolution kernels of different sizes for spectral features by combining spectral features in three dimensions: time domain,energy domain and time energy to perform feature fusion to avoid the problem of feature loss caused by the same convolution kernel size.The classification and recognition of surface EMG signals of10 types of gesture actions was realized.The model has high accuracy,sensitivity and specificity when tested on the dataset collected by MYO and the public dataset Nina Pro DB5.(2)Aiming at the problem of inaccurate recognition of continuous hand movements,a Long Short-Term Memory Convolution Neural Network based on Multiple Features Fusion(LSTMCNN-MFF)is constructed to improve the recognition accuracy of continuous gestures.The model performs feature fusion by constructing the RMS time domain feature CNN network channel and the Hilbert-Huang spectrum frequency domain feature CNN network channel.The fused high-dimensional features have both time domain and frequency domain feature information,avoiding the low recognition rate caused by a single feature.The fused high-dimensional features are input into the LSTM network to achieve accurate classification and recognition of 10 types of continuous gestures.The model has high accuracy when tested on the public datasets Nina Pro DB2,Nina Pro DB8,UCI HAR and MYO datasets.(3)Aiming at the problem that the existing 3D deformation model has a large amount of calculation and low real-time efficiency,the grouping calculation method and 3D skin calculation method of the projection distance in the bone-driven 3D model are given.By initializing the bone calculation and projection distance,the time overhead problem caused by a large number of complex multi-dimensional vector calculations can be solved,and efficient real-time 3D deformation are realized.This method is not only applicable to the deformation of the human body geometric model but also to the bone-driven deformation of animals and hands.(4)In order to verify the accuracy,timeliness and generality of the model,a multimodal real-time control system for the robot arm was constructed,and the IMC-CNN model and the LSTMCNN-MFF model were applied to the EEG under the SSVEP experimental paradigm and EMG decoding.Real-time control of 6 categories of EEG and 4 categories of surface EMG signals of the robotic arm was realized.In addition,an intelligent prosthetic hand was designed and 3D printed,and the LSTMCNN-MFF model was used to simulate and control multiple gestures of the simulated robot arm. |
