Feature Extraction from sEMG of Forearm Muscles, Performance

Analysis of Neural Networks and Support Vector Machines for

Movement Classification

Luis Morales and Jaime Cepeda

Departamento de Automatización y Control Industrial, Escuela Politécnica Nacional Ladrón de Guevara,

E11-253, Quito, Ecuador

Keywords: Support Vector Machines, Feedforward Neural Networks, Pattern Recognition, EMG Signals, Feature

Extraction.

Abstract: The propose of this work is to extract different features from surface EMG signals of forearm muscles such

as MAV, RMS, NZC, VAR, STD, PSD, and EOF's. Signals are acquired through 8 channels from "Myo

Armband" sensor that is placed in the forearm of the human being. Then, identification and classification of

5 types of movements are done, including open hand, closed hand, hand flexed inwards, out and relax position.

Classification of the movement is performed through machine learning and data mining techniques, using two

methods such as Feedforward Neural Networks and Support Vector Machines. Finally, an analysis is done to

identify which features extracted from the sEMG signals and which classification method present the best

results.

1 INTRODUCTION

Nowadays advances in robotics have made life easier

for human beings, both domestically and industrially.

An application of the first one, is to assist people with

different types of disabilities, helping them to lead

their lives in the most normal way possible.

Specifically, in the case of people who have suffered

the loss of a superior member such as the amputation

of a hand, it is indispensable that the disabled person

recovers the ability to take or manipulate objects. The

muscular groups present in the forearm of the human

being are directly related to the different states of the

hand (Khushaba, Al-Timemy, Kodagoda, and

Nazarpour, 2016), for example, completely open,

closed, flexed inwards, flexed out, relax position, etc.

The surface EMG can be measured easily and

non-invasively (Nakajima, Keeratihattayakorn,

Yoshinari, and Tadano, 2014), through the use of dry

sensors, which measure the potentials generated by

muscle contractions. EMG signals are widely used to

perform medical diagnoses (Abel, Zacharia, Forster,

and Farrow, 1996), as well as to determine

movements of the upper limbs and thus control hand

prosthetics (Kawano and Koganezawa, 2016). With

multisensory information is possible identify human

hand motion via feature extraction and classification

(Ju and Liu, 2014)(Ju, Ouyang, Wilamowska-Korsak,

and Liu, 2013). There are different studies that have

allowed the estimation of mathematical models that

establish the generation of potentials in muscle

groups as in the case of those belonging to the

forearm, to study its behavior and its mechanism

which may be potentially used for assessment or

neuromuscular rehabilitation (Nakajima et al., 2014).

Other studies have focused on identifying different

states of the hand through myoelectric sensors placed

in the forearm, to control robotic prostheses

establishing states of supination, pronation, open and

closed hand, this through the classification of the

signals through the harmonic wavelet packet

transform (Wang, Zhiguo, Xiao, Hongbo, and

Zhizhong, 2006), and detection of the angle of the

hand, considering the position of relax, semi-flexed

and flexed to replicate those movements in an

orthopedic hand that may be useful for rehabilitation

(Kavya, Dhatri, Sushma, and Krupa, 2015), the

classification of these states is done through Support

Vector Machines (SVM). The force generated

between each of the fingers and the thumb is also

considered to determine the behavior of EMG signals

of the forearm (16 channels) and its relation to these

254

Morales, L. and Cepeda, J.

Feature Extraction from sEMG of Forearm Muscles, Performance Analysis of Neural Networks and Support Vector Machines for Movement Classiﬁcation.

DOI: 10.5220/0006429402540261

In Proceedings of the 14th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2017) - Volume 1, pages 254-261

ISBN: 978-989-758-263-9

movements which may be useful to identify precise

movements (Fang, Ju, Zhu, and Liu, 2014), and may

help identify which muscle group interacts in greater

proportion with each of these movements. There are

a lot of features that can be obtained from the signals

in time, frequency and wavelet domain. (Boostani and

Moradi, 2003), studying which of them improves the

results in the classification process.

This work proposes to identify and classify 5

types of movements, including

open hand, closed hand,

hand flexed inwards and flexed out and relax position,

through different patterns such as: Mean Absolute

Value “MAV”, Root Mean Square “RMS”, Number

Zero Crossing “NZC”, Variance “VAR”, Standard

Deviation “STD”, Power Spectral Entropy “PSD”,

and Empirical Orthogonal Functions “EOF's”,

extracted from the surface myoelectric signals in 8

channels from "Myo Armband" sensor that is placed

in the forearm of the human being. Classification of

the movement is performed through machine learning

and data mining techniques, using two methods such

as Neural Networks and Support Vector Machines,

finally an analysis is done to identify which pattern

extracted from the signal and which method of

classification present the best results, through

subjects/users of training and test groups.

2 METHODOLOGY

The system to be tested consists of the steps shown in

Figure 1, for the proposed analysis using different

features extracted for the sEMG signals, and

classified by two methods such as Feedforward

Neural Networks and Support Vector Machines.

Figure 1: Scheme of the proposed approach.

2.1 Signal Acquisition

In this work is performed feature extraction and

pattern recognition of a sEMG signals from the

forearm to identify different class of movements.

Signals are acquired using “Myo Armband” sensor,

which is a gesture recognition device worn on the

forearm and manufactured by Thalmic Labs. It uses a

set of electromyographic sensors that sense electrical

activity in the forearm muscles, combined with a

gyroscope, accelerometer and magnetometer to

recognize gestures.

The sensor consists of 8 channels (Figure 2) to

acquire the myoelectric signals of muscles of the

forearm from users/subjects.

Figure 2: Myo Armband Sensor and its channels.

This sensor allows to know the myoelectric

signals of the muscular groups present in the forearm

(Figure 3), the signals are acquired at a sampling

frequency of 200 Hz, and are normalized to values

between -1 and 1 as shown in Figure 4.

Figure 3: Positioning the sensor for signals acquisition.

Feature Extraction from sEMG of Forearm Muscles, Performance Analysis of Neural Networks and Support Vector Machines for

Movement Classiﬁcation

255

Figure 4: EMG signal acquired by Myo Armband Sensor.

The electrodes are in contact with the different

muscles around the forearm as shown in Table 1.

Table 1: Electrodes distribution in muscles of the forearm.

Myo

Armband

Muscle

Channel 1 Extensor Di

itorium Cummunis

Channel 2 Extensor Car

i Radialis

Channel 3 Brachioradialis

Channel 4 Pronator Teres

Channel 5 Flexor Digitorum Sublimas

Channel 6 Flexor Car

i Ulnaris

Channel 7 Flexor Di

itorum Profundus

Channel 8 Extensor Carpi Ulnaris

The signals acquired by the 8 channels of sensor

(S1, S2, ... S8) allow detecting 5 hand movements

(M1, M2, ... M5) which are produced by the

combination of contraction and relaxation of different

muscles in the forearm. Figure 5.

Figure 5: Multi-channel sensor and its relation with the

movements of the hand scheme.

Signal acquisition is made to fifteen user/subjects

who were separated in seven subjects for training and

eight subjects for test. They were asked to perform the

movements that are to be detected. Figure 6.

Figure 6: Proposed movements to identify a) closed hand,

b) open hand, c) hand flexed inwards d) hand flexed out e)

relax position.

2.2 Feature Extraction

Different features can be extracted from the obtained

signals (Boostani and Moradi, 2003), in this work has

been calculated parameters in time domain such as:

MAV, NZC, RMS, VAR, STD, and EOF, in the

frequency domain is calculated PSD. With these

parameters, is desired to reduce the number of signal

data to facilitate pattern recognition and movement

classification.

 MAV (Mean Absolute Value): This feature

determines the mean value of the difference in

amplitudes of consecutive samples in a time

segment.















(1)

Where: x

is the value of -th sample, is the number

of samples.

 RMS (Root Mean Square): This feature

determines the root mean square of consecutive

samples in a time segment.

















(2)

 NZC (Number Zero Crossing): is the number of

times that signal passes the zero-amplitude axis.

ICINCO 2017 - 14th International Conference on Informatics in Control, Automation and Robotics

256



















(3)











10

0

 VAR (Variance): This feature measures the

spread of data from the mean (̅) of consecutive

samples in a time segment.



1













(4)

Where ̅ is the mean of data in a time segment.

 STD (Standard Deviation): this feature measures

the data dispersion of consecutive samples in a

time segment from its mean.





1















(5)

 EOF (Empirical Orthogonal Function): is a time

series data mining technique that allows

decomposing time series into a sum of a set of

discrete functions namely EOF’s (Cepeda and

Colome, 2014).

In this work time series correspond to the signals

obtained from eight channels of the forearm F.









⋯



⋮⋱⋮





⋯





(6)

The SVD of F matrix has the form:



U







/





(7)

Where Λ



/

is a diagonal matrix with the square roots

of eigenvalues from U



or V



. U



and V



are an

orthogonal matrix whose columns are the

orthonormal eigenvectors of FF′ and F′F

respectively.

The number of selected eigenvectors of V



defined by eigenvalues of Λ, which allow a measure

of the corresponding explained variability (), and

whose elements are known as EOF’s.











∑









100%

(8)

Then the EOF’s will be used for training of

classifiers.

 PSD (Power Spectral Entropy): is used to extract

information content in a discrete signal. To

calculate PSD is necessary to apply the FFT to the

signals in a finite time. The algorithm to calculate

this parameter is summarized to the following

steps (Zhang, Yang, and Huang, 2008):

- The discrete Fourier Transform 



 can be

computed by FFT. Considering 



is the i-th

frequency of the spectrum.

- The Power Spectral Density is computed by:























|



(9)

- In this work, the sum of the 



 corresponding

to each frequency is the pattern that will be used.















(10)

2.3 Classification

In this work two methods of pattern recognition are

presented: Feedforward Neural Networks and

Support Vector Machines, then it is done a

comparison between the performance of the

classifiers when they are tested with different patterns

that are obtained from the acquired signals.

2.3.1 Feedforward Neural Networks (FNN)

Neural Networks are processing algorithms whose

operation simulates a biological brain. They can

process large parallel amounts of information, even if

it is partial and diffuse. This method can learn and

memorize very varied information and formalize it

and, of course, predictions from the data with which

it has been trained. They provide a powerful non-

linear interpolation tool and multidimensional. So,

they have been used mainly in identification and

prediction of patterns.

Figure 7: Feedforward Neural Network Scheme.

The feedforward topology is characterized by

running the processing in one direction only (Russell

and Norvig, 2009). Distinguishing three layers of

computation called neurons: input layer where the

Feature Extraction from sEMG of Forearm Muscles, Performance Analysis of Neural Networks and Support Vector Machines for

Movement Classiﬁcation

257

data to be processed is received; the intermediate

layer or layers, where is the processing itself and the

output layer. (See Figure 7).

2.3.2 Support Vector Machines (SVM)

SVM is a supervised machine learning algorithm

which can be used for both classification or

regression challenges. However, it is mostly

used in classification problems. In this algorithm, is

plotted each data item as a point in n-dimensional

space (where n is number of features) with the value

of each feature being the value of a particular

coordinate. Then, is performed classification by

finding the hyper-plane that differentiate the two

classes. In Figure 8, is shown the representation of

SVM’s and the optimal hyperplane.

Figure 8: Representation of SVM’s.

Support Vector Classifiers (SVC) needs a priori

an off-line learning stage, in which the classifier must

be trained using a training set of data. Each element

in the training set contains one ¨target value¨ and

several attributes (Cepeda, 2013). Training involves

the minimization of the error function:













(11)

Subject to the constraints:

























1







0

(12)

Where:  is the margin parameter,  is the vector of

coefficients, is the bias constant, and 



represents

parameters for handling nonseparable data (inputs).

The index i labels the  training cases. 



represents

the class labels and 



represents the independent

variables. The kernel  is used to transform data from

the input (independent) to the feature space. Radial

Basis Kernel function used in this work, it has the

form (Chih-Wei Hsu, Chih-Chung Chang, 2008):





,





















,0

(13)

In this work, the inputs are the different features

extracted of the signals of the forearm muscles and

the targets are the hand movements.

2.3.3 K-Fold Cross Validation

It is a technique used to measure how accurate is the

classification algorithm. In this method, the sample

data are divided into K subsets. One of the subsets is

used as test data and the rest (K-1) as training data.

The cross-validation process is repeated during K

iterations, with each of the possible subsets of test

data. Finally, the arithmetic mean of the results of

each iteration is performed to obtain a single result.

This method is very accurate since we evaluate from

K combinations of training and test data. In practice,

the choice of the number of iterations depends of the

measurement of the data set.

2.3.4 Reduction of Dimensionality of the

Data

In this work, it is identified the clusters belonging to

the 5 movements of the hand to verify its separation.

Data have been obtained through the 8 channels of the

sensor, then it necessarily to reduce their

dimensionality to show the information in the space.

Through the principal component analysis (PCA)

is proposed to reduce the dimensionality of elements

to represent the clusters in the plane (2-dimensions)

or space (3-dimensions) without excessive loss of

accuracy.

Principal component analysis (PCA) is a data

mining technique that allows transform the original

data into a new set of variables which are uncorrelated

(Cepeda, 2013), and can be obtained through

calculation of singular value decomposition (SVD) of

the covariance matrix (S).

Considering a data matrix X:

X





⋯



⋮⋱⋮





⋯



 (14)

The covariance matrix is calculated from data matrix

(X):

S















(15)

where:

I: identity matrix

1: all ones vector

Applying spectral decomposition (SVD) to S matrix:

SUΛU´

(16)

ICINCO 2017 - 14th International Conference on Informatics in Control, Automation and Robotics

258

Where Λ is a diagonal matrix containing eigenvalues

(



) of S, U is an orthonormal matrix containing the

eigenvector of S.

The number of selected eigenvectors of U is

defined by eigenvalues of Λ, which allow a measure

of the corresponding explained variability Ec (8).

Thus, the new data matrix Y is the projection of

original data X on the hiper-plane defined by U.

YXU

(17)

3 RESULTS

Feature extraction has been for time window of 100

milliseconds as shown Figure 9. Once the data for

these segments are obtained, it is calculated: MAV,

NZC, RMS, VAR, STD, and EOF, and PSD.

Figure 9: Time window used for feature extraction

(example).

Figure 10, shows clusters obtained with 3 PC´s

which have an Explained Variability 



=93.3%.

Figure 10: Clusters identified for 5 hand movement.

In Table 2, the results obtained to measure the

robustness of the implemented classifiers are

presented, taking the signals of 7 subjects/users (4

men, 3 women) for off-line training of Feedforward

Neural Networks (10 hidden layers and

backpropagation method for training phase) and

Support Vector Machines with 350 time windows of

signals corresponding to the 5 movements of the

hand, that is: 70 open hand, 70 closed hand, 70 hand

flexed inwards, 70 hand flexed out and 70 hand

relaxed, the accuracy of the classifiers is done using

K-fold cross validation algorithm for K = 10.

Tests were performed with independent features

and with the best combinations of them to find the

best accuracy.

Table 2: Classification accuracies throw K-fold cross

validation method.

Pattern

Feedforward

Neural

Network

Support Vector

Machines

MAV 79.2 % 63.2 %

NZC 34.5 % 29.5 %

RMS 87.5 % 66.0 %

VAR 77.0 % 55.5 %

STD 86.0 % 69.0 %

PSD 45.0 % 35.5 %

EOF 22.5 % 39.0 %

MAV-STD 86.0 % 67.0 %

RMS-STD 86.5 % 67.5 %

MAV-RMS-STD 78.8 % 62.4 %

MAV-RMS-VAR 82.4 % 62.4 %

MAV-RMS-VAR-STD 78.0 % 63.2 %

It is recommendable to identify the best

parameters C and γ of SVM classifier to obtain better

results (Chih-Wei Hsu, Chih-Chung Chang, 2008).

Differential evolution algorithm has been used to

identify the constants, obtaining the accuracy results

shown in Table 3.

Table 3: Classification accuracies throw K-fold cross

validation method with identified parameters of SVM.

Pattern

Feedforward

Neural

Network

Support

Vector

Machines

MAV 79.2 % 100 %

NZC 34.5 % 61.0 %

RMS 87.5 % 99.5 %

VAR 77.0 % 98.5%

STD 86.0 % 100 %

PSD 45.0 % 65.5 %

EOF 22.5 % 56.0 %

MAV-STD 86.0 % 99.5 %

RMS-STD 86.5 % 100 %

MAV-RMS-STD 78.8 % 99.6 %

MAV-RMS-VAR 82.4 % 99.2%

MAV-RMS-VAR-STD 78.0 % 99.6 %

Feature Extraction from sEMG of Forearm Muscles, Performance Analysis of Neural Networks and Support Vector Machines for

Movement Classiﬁcation

259

In Table 3, it can be evidenced when using a

method of optimization for calculation the SVM

constants, the performance of the classifier can be

greatly improved.

Table 4: Classification accuracies in the test group.

Classification accuracy with different features

(%)

Test CLA MAV STD

RMS

STD

MAVR

STD

MAV

RMS

VAR

STD

User 1

79.0 84.0 74.0 40.0 50.0

SVN 94.0 92.0 92.0 82.0 94.0

User 2

FNN 80.0 38.0 50.0 74.0 74.0

SVN 90.0 84.0 84.0 90.0 92.0

User 3

FNN 70.0 58.0 68.0 72.0 42.0

SVN 88.0 84.0 84.0 76.0 76.0

User 4

FNN 50.0 34.0 28.0 36.0 22.0

SVN 64.0 62.0 62.0 78.0 90.0

User 5

FNN 74.0 84.0 44.0 84.0 96.0

SVN 92.0 92.0 100 100 100

User 6

FNN 80.0 74.0 56.0 52.0 46.0

SVN 99.0 98.0 100 100 100

User 7

FNN 46.0 20.0 18.0 82.0 52.0

SVN 88.0 86.0 84.0 84.0 88.0

User 8

FNN 58.0 30.0 36.0 52.0 50.0

SVN 82.0 96.0 78.5 92.0 96.0

After the training of the FNN and SVM

algorithms, accuracy measurements, were performed

taking the signals of 8 subjects/users (6 men, 2

women) of the test group, with 400 time windows of

signals corresponding to the 5 movements of the

hand, that is: 80 open hand, 80 closed hand, 80 hand

flexed inwards, 80 hand flexed out and 80 hand

relaxed. It has been done with the better features such

as: MAV, STD, RMS-STD (R-S),

MAV-RMS-STD

(M-R-S) and MAV-RMS-VAR-STD

(M-R-V-S) which

exceed 99.5% accuracy in the training phase of SVM.

This results are shown in Table 4.

Table 5, shows the average accuracies for each

classifier, and Figure 11 a comparative graph of these

values.

Table 5: Average accuracies in the test group.

Feedforward

Neural Networ

Support Vector

Machines

MAV 67.1 % 87.1 %

STD 52.8 % 86.8 %

R-S 46.8 % 85.6 %

M-R-S 61.5 % 87.8 %

M-R-V-S 54.0 % 92.0 %

Results presented in Figure 11, show the best

feature for the classification of movements in users

who did not participate in the training phase, this is

the combination of the features M-R-V-S, with the

SVM classifier, which have an average accuracy of

92.0%.

Figure 11: Average accuracies of classifiers test.

4 CONCLUSIONS

The present work has allowed to determine the

identification of different movements of the hand

through the acquisition of sEMG signals of the

forearm. In this work, it has been shown that SVM

presents better accuracy regarding the FNN for

classification, and the feature that is considered the

best for this aim is the combination MAV-RMS-

VAR-STD with 92% of accuracy.

In future works, it is proposed to detect other

movements, especially including the fingers of the

hand and to verify other classification techniques

such as Linear Discriminant Analysis (LDA) among

others to transfer these movements ton robotic

orthotic prostheses.

REFERENCES

Abel, E., Zacharia, P., Forster, A., and Farrow, T. (1996).

Neural network analysis of the EMG interference

pattern. Medical Engineering and Physics, (1), 12–17.

Retrieved from http://www.sciencedirect.com/science/

article/pii/1350453395000321

Boostani, R., and Moradi, M. H. (2003). Evaluation of the

forearm EMG signal features for the control of a

prosthetic hand. Physiological Measurement, 24(2),

309. https://doi.org/10.1088/0967-3334/24/2/307

Cepeda, J. C. (2013). Real Time Vulnerability Assessment

of Electric Power Systems Using Synchronized Phasor

ICINCO 2017 - 14th International Conference on Informatics in Control, Automation and Robotics

260

Measurement Technology. Facultad de Ingeniería de la

Universidad Nacional de San Juan.

Cepeda, J. C., and Colome, D. G. (2014). Benefits of

empirical orthogonal functions in pattern recognition

applied to vulnerability assessment. 2014 IEEE PES

Transmission and Distribution Conference and

Exposition, PES T and D-LA 2014 - Conference

Proceedings, 2014–Octob. https://doi.org/10.1109/

TDC-LA.2014.6955254

Chih-Wei Hsu, Chih-Chung Chang, and C.-J. L. (2008). A

Practical Guide to Support Vector Classification. BJU

International, 101(1), 1396–400. https://doi.org/

10.1177/02632760022050997

Fang, Y., Ju, Z., Zhu, X., and Liu, H. (2014). Finger pinch

force estimation through muscle activations using a

surface EMG sleeve on the forearm. IEEE International

Conference on Fuzzy Systems, 1449–1455.

https://doi.org/10.1109/FUZZ-IEEE.2014.6891790

Ju, Z., and Liu, H. (2014). Human hand motion analysis

with multisensory information. IEEE/ASME

Transactions on Mechatronics, 19(2), 456–466.

https://doi.org/10.1109/TMECH.2013.2240312

Ju, Z., Ouyang, G., Wilamowska-Korsak, M., and Liu, H.

(2013). Surface EMG based hand manipulation

identification via nonlinear feature extraction and

classification. IEEE Sensors Journal, 13(9), 3302–

3311. https://doi.org/10.1109/JSEN.2013.2259051

Kavya, S., Dhatri, M. P., Sushma, R., and Krupa, B. N.

(2015). Controlling the hand and forearm movements

of an orthotic arm using surface EMG signals, 1–6.

https://doi.org/10.1109/INDICON.2015.7443749

Kawano, T., and Koganezawa, K. (2016). A Method of

Discriminating Fingers and Wrist Action from Surface

EMG Signals for Controlling Robotic or Prosthetic

Forearm Hand. https://doi.org/10.1109/AIM.2016.

7576736

Khushaba, R. N., Al-Timemy, A., Kodagoda, S., and

Nazarpour, K. (2016). Combined influence of forearm

orientation and muscular contraction on EMG pattern

recognition. Expert Systems with Applications, 61, 154–

161. https://doi.org/10.1016/j.eswa.2016.05.031

Nakajima, Y., Keeratihattayakorn, S., Yoshinari, S., and

Tadano, S. (2014). An EMG-CT method using multiple

surface electrodes in the forearm. Journal of

Electromyography and Kinesiology, 24(6), 875–880.

https://doi.org/10.1016/j.jelekin.2014.07.010

Russell, S., and Norvig, P. (2009). Artifical Intelligence: A

Modern Approach (Third Edition). Prentice Hall.

https://doi.org/10.1017/S0269888900007724

Wang, G., Zhiguo, Y., Xiao, H., Hongbo, X., and Zhizhong,

W. (2006). Classification of surface EMG signals using

harmonic wavelet packet transform. Medical and

Biological Engineering and Computing, 44(10), 865–

872. https://doi.org/10.1007/s11517-006-0100-y

Zhang, A. Z. A., Yang, B. Y. Bin, and Huang, L. H. L.

(2008). Feature Extraction of EEG Signals Using

Power Spectral Entropy.

2008 International

Conference on BioMedical Engineering and

Informatics, 2, 1359–1363. https://doi.org/10.1109/

BMEI.2008.254

Feature Extraction from sEMG of Forearm Muscles, Performance Analysis of Neural Networks and Support Vector Machines for

Movement Classiﬁcation

261