978-1-4673-7678-5 ©2015 IEEE 883
2015 11th International Conference on Natural Computation (ICNC'15)
Comparison of sEMG-Based Feature Extraction and
Hand Motion Classification Methods
Lili Dai; Feng Duan
College of Computer and Control Engineering
Nankai University
Tianjin, China
Abstract—The myoelectric prosthetic hand is regard as a useful
tool to provide convenience for the upper amputees. There are
two key challenges for the control of myoelectric prosthetic hand,
one is the surface electromyogram (sEMG) feature extraction,
the other is the identification of hand motions. In this paper, we
analyzed the influence of feature selection from four feature sets
and determined the most appropriate feature in time-frequency
domain. Furthermore, we utilized two methods of wavelet neural
network (WNN) and support vector machines (SVMs) to identify
six kinds of hand motions. We trained the WNN using a hybrid
method which consists of back-propagation (BP) and least mean
square (LMS), and trained SVMs with grid search (GS) and
cross validation (CV) for getting the prediction model. The
classification results show that the training time of WNN for
hand motion classification is longer than that of SVMs. However,
comparing with SVMs, the classifier of WNN has the following
significant performance: 1) less identification time; 2) more
robustness; 3) higher accuracy rate.
Keywords-surface electromyogram signal; feature selection;
wavelet neural network; support vector machines; pattern
recogntion; prosthetic hand
I. INTRODUCTION
The amputees are usually caused by trauma, congenital
anomaly, and aging. They have the need to ensure a high
quality of life, so a number of researchers are actively
studying and developing assistive devices like prosthetic
devices [1, 2], orthotic devices, wheelchairs [3, 4],
exoskeleton [5] and so on. The prosthetic hand is regard as a
useful way to provide convenience for the upper amputees.
Using bio-signals from the upper amputees is a main method
to capture the amputees’ intentions and then control the
prosthetic hand effectively. Because of the noninvasive,
safety, and easy to collect, surface electromyogram (sEMG)
plays an important role in the field of human machine
interface [6, 7]. Many researchers have investigated the study
of prosthetic hand based on sEMG [8, 9]. In the study of
myoelectric prosthetic hand, the main challenges are feature
extraction and classification of hand motions.
Feature set is the most distinctive feature of sEMG, it
avoids the redundancy cased by the complexity of sEMG, as
well as it avoids dimension disaster because of the advantage
of feature set’s limited dimensions. Feature selection
influences the input of classifier, if the feature’s dimension is
too large, it may cause the classifier more complex and
classification time too long to response rapidly. However, if
the feature of sEMG is too simple, it will cannot recognize
hand motions accurately. Based on above reasons, we can
know that feature extraction is a significant work. Commonly
used methods of feature extraction mainly involved in time
domain, frequency domain, and time-frequency domain. The
features of time domain, such as autoregressive (AR) model
[10], root mean square (RMS), mean absolute value (MAV)
[11] and so on, are usually used. In frequency domain, there
are many algorithms like frequency median, frequency ratio,
and fast Fourier transform (FFT). In time-frequency domain,
wavelet transform (WT) [12-14] and wavelet packet transform
(WPT) [15] are commonly used. The analysis method of time-
frequency domain has the advantages of the time domain and
frequency domain, so it has gotten widely attention and
application. In this study, we compared the effects of different
feature sets to determine an optimal set. The process of feature
selection is conducive to the recognition of hand motions,
moreover it will promote the study of prosthetic hand
ultimately.
Many researchers have investigated the methods of pattern
recognition, for instance, K-Nearest Neighbor, Bayesian
Model [16], and Discriminant Analysis [17, 18]. Support
Vector Machines (SVMs) [17, 19] provides many unique
advantages in solving small sampling set, nonlinear, and high
dimensional pattern recognition, so it can be applied to
classification [20], regression analysis, and other problems in
the field of machine learning. Artificial Neural Network (ANN)
is also a popular way in classifying the movements based on
sEMG [21]. In addition, Wavelet Neural Network (WNN) and
Fuzzy Neural Network, as the improvement of ANN, have
been used in pattern recognition [22-24]. In this paper, we
utilized WNN and SVMs as the classifiers to identify six hand
motions.
The effects of hand motion classification need to be
evaluated by some performance benchmarks. The main
benchmark is accuracy rate, but only using accuracy rate is
neither complete nor objective. Classification time is an
important benchmark, taking into account the real-time control
of prosthetic hand. In addition, robustness is an essential factor
that influences the stability of the system of prosthetic hand
[25], so it is usually used to evaluate the availability of pattern
recognition of hand motions. Furthermore, training time is
also often used. Therefore, we verified the performance of
This work is supported by the National Natural Science Foundation o