第1页 / 共10页
第2页 / 共10页
第3页 / 共10页
第4页 / 共10页
第5页 / 共10页
第6页 / 共10页
第7页 / 共10页
第8页 / 共10页
A_Novel_Feature_Extraction_forRobust_EMG_Pattern_Recognition___..pdf
JOURNAL OF COMPUTING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
A Novel Feature Extraction for
Robust EMG Pattern Recognition
Angkoon Phinyomark, Chusak Limsakul, and Pornchai Phukpattaranont
71
Abstract—Varieties of noises are major problem in recognition of Electromyography (EMG) signal. Hence, methods to remove
noise become most significant in EMG signal analysis. White Gaussian noise (WGN) is used to represent interference in this
paper. Generally, WGN is difficult to be removed using typical filtering and solutions to remove WGN are limited. In addition,
noise removal is an important step before performing feature extraction, which is used in EMG-based recognition. This research
is aimed to present a novel feature that tolerate with WGN. As a result, noise removal algorithm is not needed. Two novel mean
and median frequencies (MMNF and MMDF) are presented for robust feature extraction. Sixteen existing features and two
novelties are evaluated in a noisy environment. WGN with various signal-to-noise ratios (SNRs), i.e. 20-0 dB, was added to the
original EMG signal. The results showed that MMNF performed very well especially in weak EMG signal compared with others.
The error of MMNF in weak EMG signal with very high noise, 0 dB SNR, is about 5-10% and closed by MMDF and Histogram,
whereas the error of other features is more than 20%. While in strong EMG signal, the error of MMNF is better than those from
other features. Moreover, the combination of MMNF, Histrogram of EMG and Willison amplitude is used as feature vector in
classification task. The experimental result shows the better recognition result in noisy environment than other success feature
candidates. From the above results demonstrate that MMNF can be used for new robust feature extraction.
Index Terms—Electromyography (EMG), Feature extraction, Pattern recognition, Robustness, Man-machine interfaces.
—————————— ——————————
1 INTRODUCTION
S
URFACE Electromyography (sEMG) signal is one of
the electrophysiological signals, which is extensively
studied and applied in clinic and engineering. In this
research, the application of sEMG signal in assistive tech-
nology and rehabilitation engineering is paid attention.
Main application of these fields is the control of the pros-
thesis or other assistive devices using the different pat-
terns of sEMG signal [1-2]. Nevertheless, the major draw-
back of EMG pattern recognition is the poor recognition
results under conditions of existing noises especially
when the frequency characteristic of noise is random. Ma-
jor types of noise, artefact and interference in recorded
sEMG signal are electrode noise, electrode and cable mo-
tion artifact, alternating current power line interference,
and other noise sources such as a broad band noise from
electronic instrument [3-4]. The first three types of noise
can be removed using typical filtering procedures such as
band-pass filter, band-stop filter, or the use of well elec-
trode and instrument [3-4] but the interferences of ran-
dom noise that fall in EMG dominant frequency energy is
difficult to be removed using previous procedures. Gen-
erally, white Gaussian noise (WGN) is used to represent
————————————————
• Angkoon Phinyomark is with Department of Electrical Engineering,
Prince of Songkla University, Songkhla, Thailand 90112.
• Chusak Limsakul is with the Department of Electrical Engineering,
Prince of Songkla University, Songkhla, Thailand 90112.
• Pornchai Phukpattaranont is with the Department of Electrical Engi-
neering, Prince of Songkla University, Songkhla, Thailand 90112.
the random noise in sEMG signal analysis [5-6]. Adaptive
filter or wavelet denoising algorithm, advance digital sig-
nal filter, has been received considerable attention in the
removal of WGN [7-8]. However, WGN cannot be re-
moved one hundred percent and sometimes some impor-
tant part of sEMG signals are removed with noise even if
we use adaptive filter and wavelet denoising algorithm.
The broad band and random frequency characteristic of
noise in this group is a main reason that make it difficult
to be removed. Moreover, the amplitude of noise is bigger
than the sEMG signal amplitude; the amplitude of raw
signal is about 50 µV-100 mV [9].
In EMG-based pattern recognition, sEMG signal is
preprocessed the spectral frequency component of the
signal and extracted some features before performing
classification [1]. Normally, in preprocessing and signal
condition procedure, method to remove noise is a signifi-
cant step to reduce noises and improve some spectral
component part [3]. Next important step, feature extrac-
tion, is used for highlighting the relevant structures in the
sEMG signal and rejecting noise and unimportant sEMG
signal [5]. The success of EMG pattern recognition de-
pends on the selection of features that represent raw
sEMG signal for classification. This study is motivated by
the fact that the limitation of the solutions to remove
WGN in the preprocessing step and EMG-based gestures
classification need to do the extraction step. The selection
of feature that torelance of WGN and the modified of ex-
isting EMG feature to improve the robust property are
proposed. As a result, WGN removal algorithms in the
preprocessing step are not needed.
© 2009 Journal of Computing
http://sites.google.com/site/journalofcomputing/
JOURNAL OF COMPUTING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
72
Fig. 1. Estimated six upper limb motions (a) wrist flexion (b) wrist
extension (c) hand close (d) hand open (e) forearm pronation (d)
forearm supination.
From the literatures, the development of robust feature
extractions in speech, texture, and image are presented
[10-11] but there is no selection and modification of ro-
bust EMG feature extraction. There are some evaluations
about the effect of noise with EMG features [5, 12-14].
However, these literatures attend to the quality of EMG
features in maximum class separability point of view. The
description and discussion about the robustness are infe-
riority. Furthermore, features that used to evaluate in the
literatures are not fair with the available methods today.
In 1995, Zardoshti-Kermani et al. [5] evaluated seven fea-
tures in time domain and frequency domain. WGN with 0
to 50% of rms amplitude signal are used to test the effect
of noise. The cluster separability index and classification
result are presented that histrogram of EMG is the better
feature in very high noise (50% of rms amplitude signal).
Later, in 2003, thirteen features with combination and
various orders are tested the robustness property by
Boostani et al. [12]. One level, one tenth of sEMG peak-to-
peak amplitude, of 50 Hz interference and random noise
is considered and the sensitivity of feature is reported. In
addition, our previous work [13-14] compared the effect
of eight features and their relevant features with 50 Hz
interference and WGN. The results of mean square error
(MSE) criterion show that Willison amplitude with 5 mV
threshold parameter is the best feature compared to the
other features.
However, there is an increase in EMG feature methods
that is published in many literatures this day. In this pa-
per, sixteen features in time domain and frequency do-
main from the literatures [5, 12-17] are used to test the
robustness with the additive WGN at various signal-to-
noise ratios (SNRs). Moreover, the effect of the level of
signal amplitude was tested. Eighteen features that used
in this research represent most features in EMG pattern
recognition. Generally, most of the attempts to extract
features from sEMG signal can be classified into three
categories including time domain, frequency domain, and
time-frequency domain [1]. We considered only former
two categories because they have computational simplici-
ty and they have been widely used in research and in
Fig. 2. Strong sEMG signal (gray line) and weak sEMG signal (black
line) of (a) wrist extension motion. (b) hand open motion.
clinical practice. In addition, two novel feature calcula-
tions using frequency properties are presented. We mod-
ified the mean frequency and median frequency by calcu-
lating the mean and median of amplitude spectrum in-
stead of power spectrum that we called Modifed Mean
Frequency (MMNF) and Modified Median Frequency
(MMDF). This paper is organized as follows. Experiments
and data acquisition are presented in Section 2. Section 3
presents a description of EMG feature extraction methods
in time domain and frequency domain. In addition, the
evaluation criterion is introduced. Results and discussion
are reported in Section 4, and finally the conclusion is
drawn in Section 5.
2 EXPERIMENTS AND DATA ACQUISITION
In this section, we depict our experimental procedure for
recording sEMG signals. The sEMG signal was recorded
from flexor carpi radialis and extensor carpi radialis lon-
gus of a healthy male by two pairs of Ag-AgCl Red Dot
surface electrodes on the right forearm. Each electrode
was separated from the other by 2 cm. A band-pass filter
of 10-500 Hz bandwidth and an amplifier with 60 dB gain
was used. Sampling frequency was set at 1 kHz using a 16
bit analog-to-digital converter board (National Instru-
ments, DAQCard-6024E).
A volunteer performed four upper limb motions in-
cluding hand open, hand close, wrist extension, and wrist
flexion as shown in Fig. 1 (a-d). In this study, the effect of
signal strength was performed by divided the sEMG sig-
nal to two types: strong sEMG signal and weak sEMG
signal. Strong sEMG signals were collected from extensor
carpi radialis longus in hand close and wrist flexion and
JOURNAL OF COMPUTING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
73
Fig. 3. The eight electrode placements of the right forearm.
were collected from flexor carpi radialis in hand open and
wrist extension as shown in Fig. 2 (gray line). In addition,
the others motion and electrode channel are weak sEMG
signals as shown in Fig. 2 (black line). Ten datasets were
collected for each motion. The sample size of the sEMG
signals is 256 ms for the real-time constraint that the re-
sponse time should be less than 300 ms. This dataset was
used for the MSE criterion that represent the effect of
noise with the value of EMG features.
The second dataset is used to evaluate the performance
of the classification results of EMG features in noisy envi-
ronment. Seven upper limb motions including hand open,
hand close, wrist extension, wrist flexion, forearm prona-
tion, forearm supination and resting as shown in Fig. 1
and eight electrode positions on the right forearm as
shown in Fig. 3 were used in the classification procedure
to measure the performance of the EMG feature space
quality with WGN. This dataset was acquired by the Car-
leton University in Canada [17]. A duo-trode Ag-AgCl
surface electrode (Myotronics, 6140) was used and an Ag-
AgCl Red-Dot surface electrode (3M, 2237) was placed on
the wrist to provide a common ground reference. This
system set a bandpass filter with a 1–1000 Hz bandwidth
and amplifier with a 60 dB (Model 15, Grass Telefactor).
The sEMG signals were sampled by using an analog-to-
digital converter board (National Instruments, PCI-
6071E), and the sampling frequency was 3 kHz. However,
in pattern recognition, downsample of EMG data from 3
kHz to 1 kHz was done. Each trial of the database con-
sisted of four repetitions of each motion. There are six
trials in each subject. Three subjects were selected in this
study. More details of experimentals and data acquisition
are described in [17].
3 METHODOLOGY
Eighteen time domain features and frequency domain
features are described in this section. Thirteen time do-
main variables are measured as a function of time. Be-
cause of their computational simplicity, time domain fea-
tures or linear techniques are the most popular in EMG
pattern recognition. Integrated EMG, Mean absolute val-
ue, Modified mean absolute value 1, Modified mean abso-
lute value 2, Mean absolute value slope, Simple square
integral, Variance of EMG, Root mean square, Waveform
length, Zero crossing, Slope sign change, Willison ampli-
tude, and Histogram of EMG are used to test the perfor-
mance. All of them can be done in real-time and electron-
ically and it is simple for implementation. Features in this
group are normally used for onset detection, muscle con-
traction and muscle activity detection. Moreover, features
in frequency domain are used to represent the detect
muscle fatigue and neural abnormalities, and sometime
are used in EMG pattern recognition. Three traditional
and two modified features in frequency spectrum are per-
formed namely autoregressive coefficients, mean and
median frequencies, modified mean and median frequen-
cies. Afterward, the evaluation methods of two criterions
that used to measure the robustness property of the
whole features are introduced.
3.1 Time Domain Feature Extraction
3.1.1 Integrated EMG
Integrated EMG (IEMG) is calculated as the summation of
the absolute values of the sEMG signal amplitude. Gener-
ally, IEMG is used as an onset index to detect the muscle
activity that used to oncoming the control command of
assistive control device. It is related to the sEMG signal
sequence firing point, which can be expressed as
IEMG
N
= ∑ ,
x
n
n
=
1
(1)
where N denotes the length of the signal and xn
represents the sEMG signal in a segment.
3.1.2 Mean Absolute Value
Mean Absolute Value (MAV) is similar to average recti-
fied value (ARV). It can be calculated using the moving
average of full-wave rectified EMG. In other words, it is
calculated by taking the average of the absolute value of
sEMG signal. It is an easy way for detection of muscle
contraction levels and it is a popular feature used in
myoelectric control application. It is defined as
MAV
1
N
= ∑ .
N =
n
x
n
1
(2)
3.1.3 Modified Mean Absolute Value 1
Modified Mean Absolute Value 1 (MMAV1) is an exten-
sion of MAV using weighting window function wn. It is
shown as
MMAV1
=
1
N
N
∑
n
=
1
w x
n
n
,
w
n
{
=
0.25
1,
0.5,
if
N n
0.75
≤ ≤
otherwise
.
N
(3)
3.1.4 Modified Mean Absolute Value 2
Modified Mean Absolute Value 2 (MMAV2) is similar to
MMAV1. However, the smooth window is improved in this
method using continuous weighting window function wn. It
is given by
MMAV2
=
1
N
∑
N =
n
1
w x
n
n
,
(4)
JOURNAL OF COMPUTING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
74
if
1,
0.25
n N
if
4 /
,
0.25
n N N
if
) /
,
0.75
N n
N
0.75
≤ ≤
N n
>
N n
<
−
.
w
n
{
=
4(
3.1.5 Mean Absolute Value Slope
Mean Absolute Value Slope (MAVSLP) is a modified ver-
sion of MAV. The differences between the MAVs of adja-
cent segments are determined. The equation can be de-
fined as
presented in Section 4.1.
3.1.10 Zero Crossing
Zero crossing (ZC) is the number of times that the ampli-
tude value of sEMG signal crosses the zero y-axis. In
EMG feature, the threshold condition is used to abstain
from the background noise. This feature provides an ap-
proximate estimation of frequency domain properties. It
can be formulated as
MAVSLP =MAV ‐MAV .
i+1
i
i
(5)
ZC
=
3.1.6 Simple Square Integral
Simple Square Integral (SSI) uses the energy of the sEMG
signal as a feature. It can be expressed as
SSI
N
2
= ∑ .
x
n
n
=
1
(6)
3.1.7 Variance of EMG
Variance of EMG (VAR) uses the power of the sEMG sig-
nal as a feature. Generally, the variance is the mean value
of the square of the deviation of that variable. However,
mean of EMG signal is close to zero. In consequence, va-
riance of EMG can be calculated by
N
−
1
∑
n
=
1
sgn
⎡
⎣
(
x
n
×
x
n
+
1
)
∩ −
x
n
x
n
+
1
≥
threshold
⎤
;
⎦
x
sgn( )
{
=
1,
0,
if x
≥
otherwise
threshold
.
(10)
3.1.11 Slope Sign Change
Slope Sign Change (SSC) is similar to ZC. It is another
method to represent the frequency information of sEMG
signal. The number of changes between positive and neg-
ative slope among three consecutive segments are per-
formed with the threshold function for avoiding the inter-
ference in sEMG signal. The calculation is defined as
SSC
=
N
−
1
∑
n
=
2
f
⎡
⎣
(
⎡
⎣
x
n
−
x
n
−
1
)
×
(
x
n
−
x
n
+
1
)
⎤
⎦
⎤
;
⎦
(11)
VAR
=
N
N
1
− ∑ .
1
x
2
n
=
n
1
(7)
f x
( )
{
=
1,
0,
if x
≥
otherwise
threshold
.
3.1.8 Root Mean Square
Root Mean Square (RMS) is modeled as amplitude mod-
ulated Gaussian random process whose RMS is related to
the constant force and non-fatiguing contraction. It relates
to standard deviation, which can be expressed as
RMS
= ∑ .
x
2
n
1
N
N =
n
1
3.1.12 Willison Amplitude
Willison amplitude (WAMP) is the number of times that
the difference between sEMG signal amplitude among
two adjacent segments that exceeds a predefined thre-
shold to reduce noise effects same as ZC and SSC. The
definition is as
(8)
WAMP
=
(
f x
n
−
x
n
+
1
)
;
N
−
1
∑
n
=
1
in
the
literatures
The comparison between RMS and MAV feature is
reported
[3, 18]. Clancy et al.
experimentally found that the processing of MAV feature
is equal to or better in theory and experiment than RMS
processing. Furthermore, the measured index of power
property that remained in RMS feature is more advantage
than MAV feature.
3.1.9 Waveform Length
Waveform length (WL) is the cumulative length of the
waveform over the time segment. WL is related to the
waveform amplitude, frequency and time. It is given by
WL
=
N
−
1
−∑
x
n
+
1
n
=
1
x
n
.
(9)
All of these features above, 3.1.1‐3.1.9, are computed
based on sEMG signal amplitude. From the experimental
results, the pattern of these features is similar. Hence, we
selected the robust feature representing for the other
features in this group. The results and discussion is
f x
( )
{
=
1,
0,
if x
≥
otherwise
threshold
.
(12)
WAMP is related to the firing of motor unit action po-
tentials (MUAP) and the muscle contraction level.
The suitable value of threshold parameter of features
in ZC, SSC, and WAMP is normally chosen between 10
and 100 mV that is dependent on the setting of gain value
of instrument. Nevertheless, the optimal threshold that
suitable for robustness in sEMG signal analysis is eva-
luated and discussed in Section 4.1.
3.1.13 Histogram of EMG
Histogram of EMG (HEMG) divides the elements in
sEMG signal into b equally spaced segments and returns
the number of elements in each segment. HEMG is an
extension version of the ZC and WAMP features. The
effect of various segments is tested and expressed in Sec-
tion 4.1.
JOURNAL OF COMPUTING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
75
Fig. 5. Original sEMG (gray line) signal and noisy sEMG signal at 5
dB SNR (black line) in six upper limb motions.
Modified Mean Frequency (MMNF) is the average fre-
quency. MMNF is calculated as the sum of the product of
the amplitude spectrum and the frequency, divided by
the total sum of spectrum intensity, as in
Fig. 4. Power spectrum (×10-5) (top) and amplitude spectrum (×10-3)
(bottom) of noisy sEMG signal at 20 dB SNR in hand close motion.
3.2 Frequency Domain Feature Extraction
3.2.1 Autoregressive Coefficients
Autoregressive (AR) model described each sample of
sEMG signal as a linear combination of previous samples
plus a white noise error term. AR coefficients are used as
features in EMG pattern recognition. The model is basi-
cally of the following form:
x
n
= −
p
∑
i
1
=
a x
i n i
w−
n
+
,
(13)
where xn is a sample of the model signal, αi is AR coeffi-
cients, wn is white noise or error sequence, and p is the
order of AR model.
The forth order AR was suggested from the previous
research [19]. However, the orders of AR between the first
order and the tenth order are found. The results are dis-
cussed in Section 4.1.
3.2.2 Modifed Median Frequency
Modified Median Frequency (MMDF) is the frequency at
which the spectrum is divided into two regions with
equal amplitude. It can be expressed as
MMNF
M
= ∑
j
1
=
f A
j
j
M
∑
j
1
=
A ,
j
(15)
where fj is the frequency of spectrum at frequency bin j.
3.2.4-3.2.5 Mean Frequency and Median Frequency
Traditional median frequency (MDF) and traditional
mean frequency (MNF) are calculated based on power
spectrum. We can calculate using the sEMG power spec-
trum Pj instead of amplitude spectrum Aj. They can be
expressed as
MDF
∑
j
1
=
P
j
=
j
M
∑
=
MDF
P
j
=
1
2
M
∑
j
1
=
P ,
j
MNF
M
= ∑
j
1
=
f P
j
j
M
∑
j
1
=
P .
j
(16)
(17)
The outline of amplitude spectrum and power spec-
trum is similar but the amplitude value of amplitude
spectrum is larger than amplitude value of power spec-
trum as shown in Fig. 4. Moreover, the variation of ampli-
tude spectrum is less than the power spectrum. For that
reason, variation of MMNF and MMDF is also less than
traditional MNF and MDF.
3.3 Evaluation methods
The percentage error (PE) is used to evaluate the quality
of the robust of WGN of EMG features, as in
MMDF
∑
j
1
=
A
j
=
j
=
M
∑
MMDF
A
j
=
1
2
M
∑
j
1
=
A ,
j
(14)
PE
=
feature
feature
−
clean
feature
clean
noise
×
100%
,
(18)
where Aj is the sEMG amplitude spectrum at frequency
bin j.
3.2.3 Modifed Mean Frequency
where featureclean denotes the feature vector of the original
sEMG signal and featurenoise represents the feature vector
of the noisy sEMG signal. WGN at different level is added
JOURNAL OF COMPUTING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
76
(a)
(c)
(b)
(d)
(e)
(f)
Fig. 6. Average PE of sEMG signals of (a) features in time domain based on signal amplitude, 3.1.1-3.1.9 (b) ZC with various threshold value
(10-50 mV) (c) SSC with various threshold value (10-50 mV) (d) WAMP with various threshold value (10-50 mV) (e) HEMG with various
segment parameters (3, 5, 7, 9, 11 segments) (f) AR coefficients with various orders (1-10) at various signal-to-noise ratios (20-0 dB) in four
motions.
to the original sEMG signal.
The performance of the methods is the best when PE is
the smallest value. We calculated average PE for each
motion with ten repeated datasets. Therefore, there are 80
datasets with four motions and two channels for each
feature and noise level was varied from 20 to 0 dB SNR
for each dataset. Moreover, WGN was added 10 times in
each noise level to confirm the results. SNR is calculated
by
SNR
P
= 10log clean
P
noise
,
(19)
where Pclean is power of the original sEMG signal and Pnoise
is power of WGN.
The classification rate (CR) is used to evaluate the qual-
ity of the recognition system with the noisy environment
of sEMG signal. The performance of the methods is the
JOURNAL OF COMPUTING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
77
(a)
(b)
Fig. 7. (a-b) Average PE of strong sEMG signals of nine selected features at various signal-to-noise ratios (20-0 dB SNRs) in four motions.
(c-d) Average PE of weak sEMG signals of nine selected features at various signal-to-noise ratios (20-0 dB SNRs) in four motions.
(c)
(d)
best when the CR values still have the same value with
the noisy sEMG signals. Original sEMG signal and noisy
sEMG signal were sent to hand movement recognition. In
this study, we evaluated the performance of robust fea-
tures in pattern recognition view point with Myoelectric
Control development (MEC) toolbox [17]. The window
size is 256 ms and window slide is 64 ms for the real-time
constraint that the response time should be less than 300
ms. The feature vector of selected robust features was
evaluated by linear discriminant anslysis classifier (LDA)
and majority vote (MV) post-processing was performed
in this study. In summary, the robust features should
have the small value of PE and still have maximum classi-
fication accuracy.
4 RESULTS AND DISCUSSION
4.1 The Quality of the Robustness of EMG Features
with WGN
To test the robustness of sixteen traditional features and
two novel features, we measured the PE with sEMG sig-
nal with additive WGN. The results of PE are plotted for
SNRs from 20 dB to 0 dB, as shown in Fig. 6-7, in practice;
we can select feature extraction to be suitable for each
application depend on the level of interference of sEMG
system. For the easy way to describe the results of a large
number of features, we discussed and evaluated the fea-
tures that have the same pattern in recognition and eva-
luated some parameter of each feature in Fig. 6. As a re-
sult, only nine representatives are discussed as the results
shown in Fig. 7.
In Fig. 6 (a), the PE of time domain features computed
using sEMG signal amplitude demonstrates that RMS
results in powerful performance in robust noise tolerance
than the other features. Hence, RMS feature is used to
represent the features in this group. Fig. 6 (b‐d) present
the evaluation of suitable value of threshold. Threshold
value was chosen between 10 and 50 mV. The optimal
threshold is 10 mV for ZC and WAMP but the suitable
threshold of SSC is 30 mV. However, the minimum PE of
SSC is higher than ZC and WAMP. ZC and WAMP with
10 mV threshold value are selected for the representative
features of this group. Afterward, the second bin of the
third segment HEMG was adopted from the result in Fig.
6 (e) and the first-order of AR is chosen because the PE of
the other AR orders are much bigger than the first one as
shown in Fig. 6 (f).
Therefore, we evaluated the performance of robustness
of nine representative features namely RMS, ZC and
WAMP with 10 mV threshold, HEMG with 2nd bin, AR
order 1, MNF, MDF, MMNF, and MMDF. Two types of
JOURNAL OF COMPUTING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
78
CLASSIFICATION RATE (%) OF 7 EMG FEATURES
USING LEAVE-ONE-OUT VALIDATION
Level of SNR noise
TABLE 1
20 dB
15 dB
Clean
10 dB
60.7835 49.1590 41.7926 34.6817
79.3059 34.1707 14.4347 12.5186
86.6298 92.2504 47.0087 21.6095
41.1326 36.3636 32.6804 17.1386
Method
HEMG
WL
WAMP
MMNF
MAV,WL,
ZC,SSC
RMS,AR2
HEMG,WA
MP,MMNF
95.6781 67.4260 22.5676
7.9838
96.4871 89.8872 64.8712 25.2714
93.0807 96.1891 64.0622 28.1243
(a)
sEMG signal, strong signals and weak signals, are used to
evaluate the robust of nine features. The weak sEMG sig-
nal has the effect of interference more than the strong
sEMG signal. In practice, we can select the robust features
to be suitable for each application. Fig. 7 (a) and Fig. 7 (b)
show the average PE of strong sEMG. For strong sEMG
signals and low noise, SNR more than 10 dB, MMNF has
the smallest average PE, followed closely by the MMDF,
MNF, and MND. For SNR less than 10 dB that showed
high noise, the PE of MNF and MDF rapidly increased
and SNR less than 3 dB that showed very high noise,
WAMP has the average PE close to MMNF. The average
error of MMNF in strong sEMG with very high noise, 0
dB SNR, is only 6%. Moreover in wrist extension and
hand open from extensor carpi radialis longus, it is only
3.5%. HEMG and ZC have slightly larger error compared
to the first group in Fig. 7 (b). The PE of RMS and AR1 are
large that they were expected to perform poorly.
The average PE of weak sEMG signals shown in Fig. 7
(c-d) clearly demonstrate that MMNF is the best robust-
ness feature and closed by MMDF and HEMG, whereas
the error of other features is more than 20%. In very high
noise, 0 dB SNR, it provides average PE about 10% and
the PE of wrist extension from flexor carpi radialis is only
5%. Other feature results are similar to the results of
strong sEMG signal but the results of PE of weak sEMG
signal are larger than the PE of strong sEMG signal. The
results above show that MMNF was the best feature
comparing with others in four motions. In summary,
MMDF and WAMP can be used with HEMG for multi-
feature. Hence, it is compared the classification results in
noisy environment with other successful individual fea-
ture and multi-feature sets from the literatures [5, 14-15,
20] in Section 4.2.
4.2 The Quality of the Recognition System of EMG
Features with the Noisy Environment
Four individual features and three multi-feature sets are
examined in this study. The classification results of seven
representative features are reported in Table 1. Leave-
one-out validation was used to guarantee an exact per-
formance measure for this dataset. The first single robust
(b)
Fig. 8. Plot of the classification results as a function of time (Top)
and plot of confusion matrix (Below) of our multi-feature set with (a)
original sEMG signal. (b) noisy sEMG signal, 15 dB SNR.
feature is HEMG that suggested by Zardoshti-Kermani et
al. [5] and is comfirmed with our result in Section 4.1. The
second feature, WL is recommended by Oskoei and H.
Hu [15] that it is the best single feature in their experi-
ment. Lastly, two
individual features, WAMP and
MMNF, are aprroved by our previous experiment [14]
and the experimental results in this paper. The recogni-
tion results of single feature did not perform well but it is
commonly used in EMG recognition system. The CR of
WAMP in clean and low noisy environment is good but
its CR is rapidly decreased in high noise. The CR of
HEMG is still stability even if noise increases. In addition,
no surprising that the CR of WL in noisy environment is
poor that comfired by the result in Section 4.1 and the CR
of MMNF is poor because of the limitation of their ability
to discriminate between classes. However, in practice, we
are usually combined this feature with other features to
get the useful information features. Because of only one
feature per channel of feature that provided from features
in time domain and frequency domain, it is effective and
small enough to combine with other features for a more
powerful feature vector and avoiding additional signifi-
cant computational burden.
相关推荐
2023年江西萍乡中考道德与法治真题及答案.doc
2012年重庆南川中考生物真题及答案.doc
2013年江西师范大学地理学综合及文艺理论基础考研真题.doc
2020年四川甘孜小升初语文真题及答案I卷.doc
2020年注册岩土工程师专业基础考试真题及答案.doc
2023-2024学年福建省厦门市九年级上学期数学月考试题及答案.doc
2021-2022学年辽宁省沈阳市大东区九年级上学期语文期末试题及答案.doc
2022-2023学年北京东城区初三第一学期物理期末试卷及答案.doc
2018上半年江西教师资格初中地理学科知识与教学能力真题及答案.doc
2012年河北国家公务员申论考试真题及答案-省级.doc
2020-2021学年江苏省扬州市江都区邵樊片九年级上学期数学第一次质量检测试题及答案.doc
2022下半年黑龙江教师资格证中学综合素质真题及答案.doc
