The Reduction And Classification Of Nonlinear Features For Semg Signals Based On Rough Set And Support Vector Machines

The surface electromyography signal (SEMG) is a kind of physiological signal produced by the neuromuscular system with voluntary contraction. It represents the function and status of the neuro-muscle system. The SEMG is used widely in clinic diagnosis, sport medicine, rehabilitation engineering, electrical physiology, neuron-physiology and the work efficiency estimation of human-machine etc.The booming application urges the researchers to study the inherent feature deeply and to figure out the fine characteristics of the SEMG. The novel and emerging signal-processing strategy play the important role on the recognition and classification of the SEMG. This paper focus on the two key points of the pattern recognition with respect to the action SEMG: feature selection-reduction and classifier design. We exploit theoretically the novel strategy for action SEMG recognition problem. The main contributions we executed are described as follows:Firstly, to the used-widely multi-channel data acquisition technique, we point out that due to the wide coverage of the surface electrode, the coupling among the different channels will be inevitable. With deep understand of the physiological basic of the surface SEMG, we introduce the independent component analysis (ICA) to analyze the surface EMG and to realize the decoupling of the multi-channel EMG signals. Through the analysis on signal complexity, we prove that after decoupling by ICA the complexity of SEMG signal at each channel descends and in some sense this phenomena proves the effect of ICA in multi-channel data acquisition system.To the feature extraction, we give the general description about the nonlinear characteristics of the SEMG. From the overall viewpoint, we introduce briefly the applications of nonlinear entropy, fractal dimension, phase-space reconstruction and chaos dynamics analysis in bio-medicine signal processing field. We execute the comparison and analysis in detail among the several usual-used classical pure time-domain and pure frequency-domain algorithms. We point out the limitations of these algorithms. Then, we propose two kinds of feature-extraction algorithm based on joint time-frequency domain. They are the sub-band energy based on wavelet decomposition and the AR coefficients derived from the IMFs of the EMD (abbreviated as EMD-AR). Furthermore, we introduce the class-separability index to evaluate the feature sets extracted by the above-mentioned two methods. At the same time, we also take the computing time into account and balance the efficiency and performance of the two feature-extraction methods. The experiments prove that the normalized sub-band energy is effective with little subject-dependence. The experiments also demonstrate the effect of the EMD-AR. But this method has two fundamental drawbacks. In one hand, it has heavy computation load. In the other hand, its performance varies with the order of AR model and the segment number of the analyzed SEMG signal. Moreover, we also explore the multi-fractal dimension of the SEMG and find that it is not appropriate as the discriminative feature for the SEMG-based motion classification task.From the viewpoint of reducing the redundant information and executing real-time online recognition, we need to perform the reduction for the extracted feature set so as to speed up the computation and strength the system robustness. In this paper, we introduce the rough set theory to realize the feature reduction and give comparison with the PCA-based feature reduction algorithm. Both the two reduction methods are classifier-independent and we call them"filter-type"reduction strategy. For the convenience of evaluation, we still introduce the neural network as the estimator, taking into account of its excellent performance of nonlinear global convergence, for the separability of the reduced feature set. The contrastive experiments prove that as a kind of linear transform, the PCA is not competent for extracting the fine and discriminative characteristic. Considering that the rough set theory is only applicable to the discrete decision system, we must perform the feature discretion firstly before using it to simplify the feature set. In our study, we propose the K-means to discretize the continuous-value feature.With respect to the classifier design, our interest is focused on the support vector machine based on the fuzzy logic. Before training the fuzzy LS-SVM based multi-class classifier, we firstly adopt the particle swarm optimization technique to cluster the training dataset. For those samples from training dataset which are clustered and assigned correctly, we figure out the edge samples which are at the edge of the different clusters and the central samples which are located at the center of the different clusters. Then we select these two kinds of samples as the support vectors and assign them different weights to train the support vector machine. Through this selection mechanism, by avoiding the effect of those malicious samples, we can improve the generalization and accuracy of the support vector machine. Because only adopting the central samples and the edge samples of each cluster as the support vectors, we can quicken the training process of the classifier significantly. Furthermore, compared with the neural network, the contrastive experiments show the superiority of the fuzzy support vector machines in avoiding the local convergence, over-training and insufficient training. In other words, the fuzzy LS-SVM is insensitive to the over-training and insufficient training and has good generalization.Relatively speaking, contrasted to the multi-classifier configuration, the single classifier can't utilize completely all the effective information and to some extent the decision it make will be arbitrary. For the further verification the performance of fuzzy LS-SVM and avoiding the arbitrary decision made by a single-classifier, we combine the Fuzzy LS-SVM, ANN, ANFIS and CART together and adopt the fuzzy integral strategy to acquire the final trade-off decision after the result fusion. The experiments demonstrate that to those samples classified correctly by the fusion strategy, the Fuzzy LS-SVM can classify them correctly over 98% while other three classifiers are relatively of poor performance. This result demonstrates that the Fuzzy LS-SVM burdens the oriented function during the process of decision fusion and has superior performance in contrast to other three peers.Our research indicates that the multi-classifier fusion algorithm can effectively make up the deficiency of the single classifier in the case of parameter drifting and construct damage. At the same time, we also point out, as the trade-off, the computing time will increase slightly when adopt the fusion mechanism. Among the four classifiers, the result given by Fuzzy LS-SVM is more reliable and stable. All the conclusions are based on the isometric SEMG signal and in the experiments we avoid the affect of fatigue facto r.