Research On Gesture EMG Recognition Algorithm Based On Deep Residual Neural Network

The contraction of human muscles may produce weak bioelectrical signals during exercise.The additive superposition of these signals on the skin surface becomes the surface EMG signal,which is closely related to human actions.Hence,it is possible to analyze and recognize the action and intention of the human body by analyzing the s EMG signal.Gesture recognition based on the s EMG signal thus becomes the frontier of human-computer interaction.With the development of deep learning technology in recent years,gesture recognition based on deep learning has gradually become a research hotspot and has made some progress.This article also researches this and does the following work.(1)This thesis introduces a deep residual network into EMG-based gesture recognition.The original deep residual network has defects such as many model parameters,long training time,and long algorithm delay,and the recognition performance on one-dimensional signal data is also poor.To solve these problems,this thesis proposes a residual pooling model.This model modifies the original residual network algorithm,adds a pooling layer and identity mapping,which reduces the overall parameters of the model,and enables more effective learning of signal data.Experimental results show that the residual pooling model improves the accuracy of gesture classification compared with the original residual network and the traditional convolutional network.(2)For some applications with high accuracy requirements,this thesis proposes an EMG-based gesture recognition algorithm based on a hybrid fusion strategy.The algorithm first obtains multiple data sets by controlling the length of the data overlap,reducing the inherent error of a single data set.Secondly,this algorithm introduces a decision-level model combination method.By using multiple data sets and multiple model classification methods,this algorithm increases the difference between the models,generates multiple complementary probability models.Then the algorithm synthesizes the characteristics and results of multiple models,so as to overcome the deviation feature extracted from the single model.Experimental results show that the algorithm in this thesis effectively reduces the error rate of gesture recognition.