| The human surface EMG signal is an important physiological signal that contains a wealth of information about the human movement state.In recent years,the gesture recognition of human surface EMG signals has been a hot topic in the field of human-computer interaction,and it has important applications in many fields such as military and medical.Traditional recognition methods mainly use machine learning methods,such as k-nearest neighbour algorithms and support vector machines.These methods do not dig deeper into the features of the EMG signal and have a lower accuracy rate.In recent years,the theory of deep learning has been improved and its practice on the field of EMG signal recognition has become more and more popular.To address the above problems,the following work is done in this thesis.(1)When using traditional convolutional neural networks for EMG gesture recognition,the use of a single network architecture,which only utilises the spatial features of the EMG signal,leads to a low accuracy rate of recognition.To address the above problems,this thesis optimises the traditional recognition model by combining the temporal features of the EMG signal and proposes a parallel CNN-RNN EMG gesture recognition model.The model first extracts the spatial features of surface EMG signals using a parallel convolutional neural network,then uses the spatial features obtained by training with a serial recurrent neural network,extracts its temporal features on the basis of the spatial features,and finally uses the Soft Max layer in the neural network to classify the fused features to obtain the final gesture classification results.Simulation experiments on the Nina Pro EMG dataset show that the gesture recognition rate is improved compared to a single-row CNN network or a parallel CNN network.(2)The traditional neural network model has deep layers and many parameters,which often results in slow training and overfitting during the training iteration,leading to a low accuracy rate of recognition.To address the above problems,this thesis proposes a myoelectric gesture recognition model based on deep separable networks,combining the characteristics of deep separable networks.The main architecture of the model uses a parallel hybrid network architecture,where each layer of the depth-separable network is followed by a recurrent neural network to extract temporal features.In addition,an auxiliary residual network is introduced to further extract surface EMG features,and the fused features are finally classified to obtain the final gesture recognition results.Simulation experiments on the Nina Pro EMG dataset show that the algorithm effectively reduces the number of parameters in the hybrid network and achieves a certain improvement in gesture recognition rate. |
