Research On SEMG Gesture Recognition Algorithm Based On Lightweight Neural Network

In recent years,various researches based on bioelectric signals have been widely used in many fields such as human-computer interaction,prosthesis control,clinical diagnosis,and medical rehabilitation.Among them,gesture recognition based on surface EMG signals,as one of the important research branches,is gaining more and more attention from researchers due to its advantages of easy acquisition,real-time and non-invasive,and has great development potential and prospect.But at the same time,due to the weak signal strength,non-smoothness,complex structure,low robustness,and noise interference of surface EMG signals,the final recognition results are still unsatisfactory.While deep neural networks,which have been widely used in this field in recent years,can improve the gesture recognition accuracy to a certain extent,they also make the model more and more demanding on the hardware equipment deployed.This thesis investigates this issue,and the main work is as follows.The sEMG gesture recognition algorithm based on convolutional neural network is investigated.The existing convolutional neural network evolves with the continuous development of the model,and its increasing demand for memory and operation volume contradicts the application scenarios of various mobile,embedded and other resource-constrained devices.In this thesis,we address this problem by introducing the idea of network compression and propose a lightweight CNN-based myoelectric gesture recognition algorithm from the perspective of architecture optimization.The algorithm builds a network model based on depthwise separable convolution and cross-stage modules,and adds three shortcut connections at different locations of the network with label smoothing operations to ensure the lightweight nature and recognition accuracy of the network model.Simulation results on the Ninapro public database show the effectiveness of the modules in this network architecture and the accuracy of the algorithm in reducing the computation and the number of parameters while ensuring the accuracy of gesture recognition.Research on s EMG gesture recognition algorithms based on recurrent neural networks is carried out.Existing EMG gesture recognition algorithms based on convolutional networks are limited by the size of the convolutional kernel and the operational mechanism of convolution itself,which cannot establish long-term dependence on the timing information in the EMG signal and cannot fully utilize the timing characteristics in the EMG signal.In this thesis,a lightweight CNN-RNN based EMG gesture recognition algorithm is proposed to address this problem,introducing a recurrent neural network based on bi-directional long short-term memory to model the features extracted from the depthwise separable convolution in a temporal order,which effectively ensures that the classification results of EMG signals at each time point depend on the whole time series of the input.In addition,considering the feature differences in the signal channels,this thesis also applies an attention module to the extracted temporal features to achieve weight reassignment on the temporal dimension.Simulation results on the Ninapro public database show that the algorithm further improves the EMG signal gesture classification accuracy with the help of bidirectional LSTM and attention mechanism.