The Study Of Classification For Different Motions In Lower Limb SEMG

With the intensification of the aging and the aggravation of the traffic accidents, the demand for the prosthetic or assisted rehabilitation platform is growing. The EMG signal contains a wide variety of control information. Using an appropriate signal processing methods, the user’s intentions can be detected. Therefore, the EMG signals can be used as command signals for controlling artificial devices.Surface electromyography (sEMG) is able to repeat obtained in a given collection point in a given motion state. And the feature set is different in different motion states. Compared to using the injected EMG, the accuracy rate of identification will not be reduced when using sEMG in analysis and process. And the collection of sEMG is simple, convenient and non-invasive.In this paper, the sEMG is collected and then used to identify the movement intent of the users to help the users rehabilitation. The feature extraction for continuous signal is studied firstly in this paper, and then classifier design is discussed, at last the post-processing is considered. The main work includes the following four aspects:(1) Using the linear discriminant analysis (LDA) and combined feature vectors to identify the pattern of the sEMG. At first, the signals collected from five normal subjects are pre-processing with slip analysis windows. Then the feature vectors are extracted which are waveform length (WL) characteristics, autoregression (AR) coefficients and the wavelet coefficients. And a fusion feature vector is got by the separation information of the feature vectors. At last, using the improved LDA classifier to identify the patterns. And the performance is very well.(2) A novel algorithm is proposed based on fuzzy theory. The algorithm is a fuzzy pattern recognition algorithm based on Gaussian radial basis function. Using different feature vectors, the recognition accuracy rate of the novel algorithm is high. And the performance is related to the feature vector’s Fisher’s index (FI). The larger the value, the recognition error rate is lower. Compared with LDA and back propagation neural network, the recognition is lower when using WL as the feature.(3) Post-processing after pattern recognition. The post-processing is used to improve the recognition accuracy on condition not to affect real-time. The post-processing methods of majority vote (MV) and the improved method of waiting for the next window (WNW) are used here. The methods can reduce the error rate to0.04%.(4) Design and implement the sEMG recognition software. The algorithms which have higher recognition accuracy and computational efficiency are used to design the recognition software to identify the motion intent by using the sEMG.In this paper, overlapping analysis window are used to process the collected sEMG, which can meet the real-time requirements. Six different limb motions are identified accurately which are useful to help the users for rehabilitation.