Dimensionality Reduction And Classification Of Semg Signals Based On Deep Neural Network

sEMG(surface electromyography)signal is a bioelectric signal of neuro muscular system conducted by electrodes from skin surface which can reflect the states of muscle activity in real time and is widely used in many fields such as helping the disabled,rehabilitation technical aids,sports medicine and so on.Among the researches related to pattern recognition of sEMG signal,the performance of classification depends heavily on feature extraction,since sEMG signal has a high dimension.We collect 7200 sEMG signal samples of 6 kinds of hand movements for our research.With the discussion of determining the number of hidden layer and hidden node,activation function and classification model,this paper proposes a dimensionality reduction method that establishes a neural-network model(5 layers)based on multi-layer perceptron to extract the fuzzy features of sEMG signal.To accommodate the high dimension data as the input of the network,we use some binary classifiers in the last layer of the model.And then,train the model to reach a high accuracy(96.21%)on classification to reduce the dimension of original data from 3000 dimensions to 500 dimensions,100 dimensions and 6 dimensions.In order to achieve good generalization ability,we add L2 regularization to the loss function.And set a series of reasonable parameters to get the prefer results in practice.Since there are some similar computational procedures and a lot of matrix calculations in the parameter solution,we compartmentalize the procedure into matrix multiplication and activating operation.Based on the parallel acceleration theory of matrix calculation,we segment data into appropriate compute nodes and then merge the results to accelerate the matrix operation.In the final part of this paper also expounds the details of dimensionality reduction by using neural network,and with inter and intra class distance as the criteria,verifies that as the layer forwards from the input to the output,the separability of each layer becomes better.