Motion classification of EMG signals based on wavelet packet transform and LS-SVMs ensemble

Abstract

Abstract：This paper presents an effective method for motion classification using the surface electromyographic (sEMG) signal collected from the forearm. Given the nonlinear and time-varying nature of EMG signal, the wavelet packet transform (WPT) is introduced to extract time-frequency joint information. Then the multi-class classifier based on the least squares support vector machine (LS-SVM) is constructed and verified in the various motion classification tasks. The results of contrastive experiments show that different motions can be identified with high accuracy by the presented method. Furthermore, compared with other classifiers with different features, the performance indicates the potential of the SVM techniques combined with WPT in motion classification. Keywords：pattern recognition; wavelet packet transform; least squares support vector machine; surface electromyographic signal; neural network; separability Recently, bio-signals have been paid much attention due to their potentiality to provide convenient control channels between the disabled men and rehabilitation engineering [1][2][3] . By analyzing the electromyographic (EMG) signals collected from remnant muscles, as control inputs for the artificial limb, the intelligent artificial limbs can implement the corresponding functions. Recognizing bio-signals correctly is a preliminary stage of controlling the prosthesis devices. Various methods of determining features of bio-signals were known for discerning predefined motions of human [4][5][6][7] . In this paper, what is concerned about is the motion classification of the forearm by studying the corresponding surface electromyographic (sEMG) signals. The EMG signal is a form of electric manifestation of neuromuscular activation associated with contracting muscles and there exist many factors affecting its waveform and amplitude [8] , which makes it very difficult to be analyzed.There are various methods to identify the motion modes via EMG signals for assistive devices such as powered prosthesis [1,8,9] . In the case of EMG classification, many efforts from two aspects have been made to seek the efficient and effective classification strategy. In the first category [7,9] , attempts are made to find a certain set of features with good separability to the predefined classes.In this category, the problem gets straightforward once one finds appropriate threshold values of features, which make the least overlapping among the predefined classes to be classified. Nevertheless, under many circumstances, the feature extraction is a subject-dependent problem [7] . An effective set of features used in one circumstance maybe does not work well when used in another circumstance. In the other category in Refs. [1,6,10,11], the researchers employed some specific classifiers with advisable features to cope with the classification task. This category includes the methods of using the probabilistic model, the fuzzy classifier, and the neural network classifier. In these methods, the errors in the boundaries often can...

Support vector machines to detect physiological patterns for EEG and EMG-based human–computer interaction: a review

“…Also LDA performed similarly to SVM; MLP-ANN with two hidden layers performed similarly to LDA and SVM and finally MLP-ANN with one hidden layer achieved accuracy inferior by 6% with respect to the other classifiers. In [147] eight arm positions were discriminated by means of a linear SVM, with accuracy within 92% and 98%, while in [148] a fourlevel wavelet transform was investigated as a novel kernel for an LS-SVM aiming at classifying four different limb motions. Accuracy > 90% was obtained with just ten features, better than the one found with a MLP, especially with small training sets.…”

Estimation of Time-Frequency Muscle Synergy in Wrist Movements

“…The resolution of the wavelet packet signal in the high frequency part is higher than that of the binary wavelet. Combined with the non-linear and time-varying characteristics of the sEMG signals, the wavelet packet transform can be introduced to extract the time-frequency joint information of the sEMG signals [28]. Since this method decomposes the signal bandwidth through a series of different center frequency but the same filter, some signal features can be extracted from the selected frequency band.…”