micromachines

Article
Design of an Effective Prosthetic Hand System for Adaptive
Grasping with the Control of Myoelectric Pattern
Recognition Approach
Yanchao Wang 1 , Ye Tian 1,2, *, Haotian She 1,2 , Yinlai Jiang 2,3 , Hiroshi Yokoi 2,3 and Yunhui Liu 1,4

1 School of Mechatronical Engineering, Beijing Institute of Technology, Beijing 100081, China;

[email protected] (Y.W.); [email protected] (H.S.); [email protected] (Y.L.)
2 Beijing Advanced Innovation Center for Intelligent Robot and System, Beijing 100081, China;
[email protected] (Y.J.); [email protected] (H.Y.)
3 Graduate School of Informatics and Engineering, The University of Electro-Communications,
Tokyo 1828585, Japan
4 Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong,
Hong Kong 999077, China
* Correspondence: [email protected]

Abstract: In this paper, we develop a prosthetic bionic hand system to realize adaptive gripping with
two closed-loop control loops by using a linear discriminant analysis algorithm (LDA). The prosthetic
hand contains five fingers and each finger is driven by a linear servo motor. When grasping objects,
four fingers except the thumb would adjust automatically and bend with an appropriate gesture,
while the thumb is stretched and bent by the linear servo motor. Since the change of the surface
electromechanical signal (sEMG) occurs before human movement, the recognition of sEMG signal
with LDA algorithm can help to obtain people’s action intention in advance, and then timely send





 control instructions to assist people to grasp. For activity intention recognition, we extract three
Citation: Wang, Y.; Tian, Y.; She, H.; features, Variance (VAR), Root Mean Square (RMS) and Minimum (MIN) for recognition. As the
Jiang, Y.; Yokoi, H.; Liu, Y. Design of results show, it can achieve an average accuracy of 96.59%. This helps our system perform well
an Effective Prosthetic Hand System for disabilities to grasp objects of different sizes and shapes adaptively. Finally, a test of the people
for Adaptive Grasping with the with disabilities grasping 15 objects of different sizes and shapes was carried out and achieved good
Control of Myoelectric Pattern experimental results.
Recognition Approach.
Micromachines 2022, 13, 219. https:// Keywords: prosthetic bionic hand; sEMG signal; LDA; motion recognition
doi.org/10.3390/mi13020219

Academic Editor: Aiqun Liu

Received: 4 November 2021

1. Introduction
Accepted: 27 January 2022
Published: 29 January 2022
As an important part for individuals with upper-limb differences, a prosthetic bionic
hand is indispensable to keep the structural integrity of the human body. For the study of
Publisher’s Note: MDPI stays neutral
prosthetic bionic hands, many studies contribute greatly. For example, Dr Paul Chappell
with regard to jurisdictional claims in
from the University of Southampton in the UK developed the Southampton Hand [1] series
published maps and institutional affil-
based on the observation of human hands and fingers when grasping. In this scheme,
iations.
each finger has three joints except the thumb, and the motion of two joints at the end are
correlatively coupled. However, the hand only has three fingers, which is too small for
people to act like ordinary people. Dalley et al. [2] developed a hand with 16 joints driven by
Copyright: © 2022 by the authors.
five independent actuators, which can provide eight hand postures. However, the specific
Licensee MDPI, Basel, Switzerland. control process is too complicated, which is not practical for individuals with upper-limb
This article is an open access article difference. Another study is the i-Limb hand, which has five independently controlled
distributed under the terms and fingers that is controlled by action signals from two electrodes [3]. There are also some
conditions of the Creative Commons other classic prosthetic bionic hands, including Okada hand [4], Stanford/JPL [5], Shadow
Attribution (CC BY) license (https:// hand [6], Utah/MIThand [7–9], Hitachi hand [10], etc. In all, the research on intelligent
creativecommons.org/licenses/by/ prosthetic hands is relatively too complicated, resulting in various inconveniences for
4.0/). the individuals with upper-limb difference in practical application. However, due to

Micromachines 2022, 13, 219. https://doi.org/10.3390/mi13020219 https://www.mdpi.com/journal/micromachines

Micromachines 2022, 13, 219 2 of 15

seeking more freedom, the motor driving system of these prosthetic hands are relatively
too complicated in hardware structure and control system, leading to the weak strengthen
of hand structure and poor real-time performance. Therefore, improving prosthesis control
is a key requirement for enhancing prosthesis satisfaction. By improving intuitive control
prosthesis, users may experience decrease cognitive effort and decrease dependency on
vision to guide and monitor actions.
Among them, the Ottobock hand [11] produced by Otto Bock Company in Germany
is one of the most successful and widely used prosthetic hands in the world. Although this
kind of hand can achieve good accuracy, its control movement is fixed, which could not
realize the adaptive grasp of objects of different shapes. In order to realize the adaptive
grasping function of the prosthetic bionic hand, many studies begin to concentrate on the
feature extraction method and advanced intelligent algorithms.
The surface electromechanical signal (sEMG) has a peak value of 0–6 mv and a fre-
quency range of 0–500 hz is one-dimensional time-series signal recorded by electrodes
on the surface of the muscle during neuromuscular system activity [12,13] have the prop-
erties of non-stationary, non-linear, complexity, and large variation. At the same time,
when the handicapped wear the prosthetic bionic hand, there will also be sweating of
the skin, muscle fatigue [14–17], ECG noise, device noise [18,19] and other environmental
conditions. These characteristics make it so difficult to analyze surface electromechanical
signals under multiple action modes well. Generally, the recognition of human movement
intention through electromyographic signals is mainly divided into three steps [20]: signal
preprocessing, feature extraction and feature classification [21–23]. The signal preprocess-
ing process generally includes signal amplification, filtering and notch. This process is a
significant step, which directly affects the reliability of signals [24]. Feature extraction is
used for extracting relevant information from sEMG signal, and rejecting noise and other
unimportant components. Feature extraction of sEMG signals mainly includes time domain
analysis, frequency domain analysis and time-frequency domain analysis. Time domain
analysis includes Integrated EMG, Mean Absolute Value (MAV), Root Mean Square (RMS),
Zero Crossing, etc. [25]. Frequency domain analysis includes Autoregressive coefficient
(AR), Modified Median Frequency (MMDF), etc. [26], and time-frequency domain analysis
includes Wavelet Transform (WT), Wigner–Ville distribution (WVD), etc. [27].
Among the feature extraction methods, time domain analysis is computational sim-
plicity and effective, which is widely used in many studies. For example, Hudgins et al.
recognized four forearm motions by using five time domain features including MAV, Mean
Absolute Value Slope (MAVS), Zero crossing (ZC), Slope Sign Changes (SSC) and Waveform
Length (WL), and gained an average accuracy rate of 91% [28]. Kim et al. successfully
classified four wrist movements by using Integrated Absolute Value (IAV) and Root Mean
Square (RMS) [29].
For sEMG-based control of prosthetic hands, there are many classification algorithms,
such as linear discriminant analysis (LDA) [30], fuzzy logic, artificial neural networks
(ANN), and support vector machine (SVM) [31]. For example, S.M. Mane [32] used the
ANN algorithm for hand gesture recognition; H. Zhang [33] adopted LDA for sEMG
pattern recognition; K. Xing [34] applied the support vector machine (SVM) into the real-
time sEMG pattern recognition system for the control of the prosthetic hand. However, for
the control of prosthetic hands, we need to consider two indicators, running speed and
recognition accuracy.
To select a suitable algorithm for the control of prosthetic hand, Sumit A [35] identify
different hand movements sEMG signals of prosthesis hand with LDA, SVM and ANN
algorithm for comparison, and proved that the LDA algorithm can achieve a high accuracy,
as well as faster running behavior than other algorithms. Besides, D. Zhang [36] also
claimed that LDA can achieve better control performance in EMG classification for a
prosthetic hand. The same work can also be seen in the work from C. Antuvan [37].
Therefore, we can conclude that the LDA algorithm can perform better in the classification
Micromachines 2022, 13, 219 3 of 15

work in sEMG signals for the control of prosthesis hand, whether in recognition accuracy
or in running speed. So, we choose LDA algorithms for intension recognition in our work.
Based on the analysis above, this paper designed an adaptive prosthetic bionic hand
with an LDA algorithm. We present the mechanical structure design of our prosthetic
bionic hand and also analyze the movement track of the finger. Then the control strategy
is designed based on the sEMG signal. In the control process, we firstly collect the sEMG
data and segment it with a sliding window for feature extraction. Then the LDA algorithm
is applied to recognize the motion intension through these features. Once the algorithm
recognizes a certain movement, the system will send the command to the motor of BIT
hand C so as to realize the grasping and opening action.
The contribution of this study can be summarized as four aspects: (1) The bionic
hand adopts the connecting rod drive design so that the bionic hand can better adapt to
the grasping action of objects of different shapes and sizes. This kind of design is more
anthropomorphic than the Ottobock bionic hand on the market. (2) In the grasp control
strategy of the bionic hand, we adopt two closed-loop control strategies of the current loop
and position loop, which makes the control of the bionic hand safer and more reliable. (3) In
order to make the hand light, cheap and durable as much as possible, we design it with
aluminum alloy, nylon and rubber, which is only 332 g. This kind of design can help people
grasp objects more freely than other designs. (4) We adopt LDA linear classifier for the
signal detection from the people, which makes the control process faster and more accurate.

2. Method
2.1. System Components of Hardware System
The prosthetic bionic hand designed in this paper is named BIT Hand C. In order to
make the hand as light and convenient as possible, we design it with aluminum alloy, nylon
and rubber. The total mass of BIT Hand C is only 332 g. It is driven by five electric motors,
which has the characteristics of small size, high accuracy and self-locking. Four linear
actuators (LA10-21D) control the fingers and one linear actuator (LAS10-23D) control the
thumb. The linear servo driver, produced by Beijing Inspire Robots Technology Co., Ltd.,
adopts DC brushed micro motor. Its reducer adopts a two-stage planetary gear reduction
device to form a large reduction ratio and realizes low speed and large torque.
The BIT Hand C developed in our study is shown in Figure 1. It has 5 fingers with
a total of 11 joints and 6 degrees of freedom. Each finger has its own independent motor
drive so that each finger can realize independent movement. The micro linear servo drives
of this prosthetic bionic hand are all placed inside the palm. Placing all actuators inside
the palm makes the design simpler and control more efficient. On the premise of great
power, the weight of the whole prosthetic bionic hand is also light, which is acceptable to
individuals with upper-limb differences. The hand, BIT Hand C, has a strong grasping
ability, it can lift up to 5 KG of weight, and the maximum pinching force of the fingertips is
up to 12 N.
BIT Hand C has the function of wrist rotation, but the wrist rotation movement is
passive, and there is no rotating motor. The wrist is controlled by the damping force of the
rubber apron in the process of rotating motion. The damping force setting of wrist motion
is realized by adjusting the fastening force of bolts. A slip ring mechanism is arranged in
the structure of the wrist so that the wrist movement of the bionic hand will not cause wear
and damage to the connection power line of the motor inside the bionic hand palm.
In order to make the hand movement simpler and more effective, each finger is
designed with one less joint, namely one degree of freedom. The finger trajectory analysis
model is composed of three links, which can be calculated by vector equation, as shown in
Figure 2a. The vector diagram of the finger kinematics is presented in Figure 2b.
Micromachines
Micromachines2022,
2022,13,
13,x 219
FOR PEER REVIEW 4 of 1615
4 of

(a) (b)

Figure 1. The prosthetic bionic hand. (a) 3D modeling of BIT Hand C. (b) Photograph of BIT Hand
C.

BIT Hand C has the function of wrist rotation, but the wrist rotation movement is
passive, and there is no rotating motor. The wrist is controlled by the damping force of
the rubber apron in the process of rotating motion. The damping force setting of wrist
motion is realized by adjusting the fastening force of bolts. A slip ring mechanism is
arranged in the structure of the wrist so that the wrist movement of the bionic hand will
not cause wear and damage to the connection power line of the motor inside the bionic
hand palm.
In order to make the hand movement simpler and more effective, each finger is
designed with one less joint, namely one degree of freedom. The finger trajectory analysis
(a) links, which can be calculated (b)
model is composed of three by vector equation, as shown
in Figure 2a. The vector diagram of the finger kinematics is presented in Figure 2b.
Theprosthetic
Figure1.1.The
Figure prostheticbionic
bionichand.
hand.(a)
(a)3D
3Dmodeling
modelingof
ofBIT
BITHand
Hand C.
C. (b)
(b) Photograph
Photograph of BIT Hand
Hand C.
C.

BIT Hand C has the function of wrist rotation, but the wrist rotation movement is
passive, and there is no rotating motor. The wrist is controlled by the damping force of
the rubber apron in the process of rotating motion. The damping force setting of wrist
motion is realized by adjusting the fastening force of bolts. A slip ring mechanism is
arranged in the structure of the wrist so that the wrist movement of the bionic hand will
not cause wear and damage to the connection power line of the motor inside the bionic
hand palm.
In order to make the hand movement simpler and more effective, each finger is
designed with one less joint, namely one degree of freedom. The finger trajectory analysis
model is composed of three links, which can be calculated by vector equation, as shown
in Figure 2a. The vector diagram of the finger kinematics is presented in Figure 2b.

(a) (b)
Figure 2. Design of connecting rod of finger mechanism. (a) The connecting rod structure of the
finger, (b) Finger connecting rod structure schematic diagram.

The closed vector equation of the finger linkage shown in Figure 2b can be written as:
→ → → →
r4 + r6 = r3 + r5 (1)
→ → → → → →
r1 + r2 + r7 = r3 + r5 + r8 (2)
With XY components for each closed vector equation, the two vector Equations (1) and (2)
can rewrite with four XY scalar component equations:

r4 sin θ4a + r6 sin θ5 = r3 sin θ1 + r5 sin θ3 ; (3)

(a) r cos θ − r cos θ = r cos θ − r cos(b)
θ3 (4)
4 4a 6 5 3 1 5
 With XY components for each closed vector equation, the two vector Equations (1)
and (2) can rewrite with four XY scalar component equations:
r sinθ + r sin θ = r sinθ + r sinθ ; (3)
Micromachines 2022, 13, 219 5 of 15
r cosθ − r cos θ = r cosθ − r cosθ (4)
r sinθ + r sin θ + r sin θ = r sinθ + r sinθ + r sinθ ; (5)
1 sin θ2−+
r rcosθ sinθθ4++r r7cos
r r2cos sinθθ5a ==−r
r3 sin θ1 ++rr5 sin
cosθ θ3 +
sinθ + rr8 sin
sinθθ7 ; (6)(5)
where θ = θ r+ 1 cos θ2 − θr2 cos
δ and =θ θ4 +
+ βr7 ,cos
δ and
θ5a = −r3constants.
β are cos θ1 + r5 sin θ3 + r8 sin θ7 (6)
Thenθ the=system can θbe solved
θ4 + δ and with these equations by using MATLAB for finger
where 4a 5a = θ5 + β, δ and β are constants.
positionThen
analysis. Figure 3 shows
the system can be solved withthe fingerthese
movement
equationstrajectories
by usingofMATLAB
BIT Hand forCfinger
in
different
positionpostures
analysis.andFigure
manipulable
3 shows space. As canmovement
the finger be seen from the figure,
trajectories ofthe
BITmaximum
Hand C in
movement angle of the first knuckle can reach 84.5 degrees, and the maximum
different postures and manipulable space. As can be seen from the figure, the maximum movement
angle of the second
movement angle ofknuckle
the firstisknuckle
135 degrees.
can reach It can
84.5be seen that
degrees, andBIT
the Hand
maximum C can fit the
movement
maneuverability
angle of the second knuckle is 135 degrees. It can be seen that BIT Hand C can fitisthe
space of the normal hand well. While the finger of the Attobock hand
only composed of aspace
maneuverability solidofstructure,
the normal andhandthewell.
grasping
Whilefinger has only
the finger of theone degree hand
Attobock of
freedom,
is only when the hand
composed of agrasps an object, and
solid structure, the contact surfacefinger
the grasping betweenhasthe
onlyfinger
one and
degreethe of
object mostly
freedom, has the
when only onegrasps
hand fulcrum. Furthermore,
an object, BIT surface
the contact Hand adopts
betweenan the
underactuated
finger and the
connecting rod mechanism
object mostly has only one design, whichFurthermore,
fulcrum. not only ensures BITthe
Handdriving motor
adopts an pulling force
underactuated
butconnecting
also enables fingers to realize underactuated grasping ability, so
rod mechanism design, which not only ensures the driving motor pullingas to ensure at least
force
two surfaces contact between fingers and objects, which can better ensure the
but also enables fingers to realize underactuated grasping ability, so as to ensure at least stability of
grasping action.contact between fingers and objects, which can better ensure the stability of
two surfaces
grasping action.

Micromachines 2022, 13, x FOR PEER REVIEW 6 of 16

Figure 3. Design of connecting rod of finger mechanism.

Figure 3. Design of connecting rod of finger mechanism.
Thefour
The fourfingers
fingersofofBIT
BITHand
HandCCarearedriven
driven
byby the
the connecting
connecting rodrod mechanism,
mechanism, which
which is
is designed
designed to suit
to suit the driving
the driving principle.
principle. Take
Take the theindex
right rightfinger,
indexfor
finger, for example,
example, we
we describe
describe
the fingerthe finger in
structure structure in shown
detail, as detail, as
in shown in Figure 4.
Figure 4.

Figure 4. The exploded view of the finger.

Figure 4. The exploded view of the finger.
The finger is mainly composed of the distal phalanx, proximal phalanx, fingertip,
The finger
connecting is mainly
rod, coil springcomposed of the distal
and pin. Nylon phalanx,
is selected proximal
for distal phalanx,
phalanx fingertip,
and proximal
connecting rod, coil spring and pin. Nylon is selected for distal phalanx and proximal
phalanx, making it lightweight. Fingertip is made of rubber for easier grasping for
different objects. The connecting rod is made of aluminum alloy, making it strong.
The finger of BIT Hand C is a kind of compliant linkage mechanism. It only needs a
micro linear private server driver (LA10-21D) to complete the basic movements of
grasping. Its maximum push force is 70 N, and the maximum stroke of these two drives
 Figure 4. The exploded view of the finger.

The finger is mainly composed of the distal phalanx, proximal phalanx, fingertip,
Micromachines 2022, 13, 219 6 of 15
connecting rod, coil spring and pin. Nylon is selected for distal phalanx and proximal
phalanx, making it lightweight. Fingertip is made of rubber for easier grasping for
different objects. The connecting rod is made of aluminum alloy, making it strong.
phalanx, making
The finger of itBIT
lightweight.
Hand C isFingertip
a kind ofiscompliant
made of rubber
linkage formechanism.
easier grasping for needs
It only different
a
objects. The connecting rod is made of aluminum alloy, making
micro linear private server driver (LA10-21D) to complete the basic movements of it strong.
The Its
grasping. finger of BIT Hand
maximum C is a is
push force kind
70 N,of and
compliant linkage mechanism.
the maximum stroke of theseIt only
two needs
drivesa
micro linear private server driver (LA10-21D) to complete the basic movements
is 10 mm. When the push rod of the drive moves to the right, the finger will be bent under of grasping.
Itspull
the maximum push
of the link. force
The is 70 N,spring
extension and the maximum
will bend to a stroke
certainof these showing
extent, two drives is 10 mm.
a relatively
large displacement without any permanent deformation or torsion. The extension the
When the push rod of the drive moves to the right, the finger will be bent under pull
spring
isofarranged
the link.atThe
theextension spring joint
second knuckle will bend
of thetofinger.
a certain
Theextent,
bionic showing
hand hasaarelatively large
fingertip grip
displacement
of 12 N. without any permanent deformation or torsion. The extension spring is
arranged at the second knuckle joint of the finger. The bionic hand has a fingertip grip
The thumb part, the maximum angle of rotation of which is 75 degrees consists of the
of 12 N.
proximal knuckle, distal knuckle, and fingertip, as shown in Figure 5. A linear servo drive
The thumb part, the maximum angle of rotation of which is 75 degrees consists of
placed in the palm of the hand enables it to rotate. However, the opening and closing
the proximal knuckle, distal knuckle, and fingertip, as shown in Figure 5. A linear servo
movement of the thumb is passive, which requires individuals with upper-limb
drive placed in the palm of the hand enables it to rotate. However, the opening and closing
differences to adjust the motion position through their own healthy hands. A linear servo
movement of the thumb is passive, which requires individuals with upper-limb differences
drive is LAS10-023D, with a maximum push force of 105 N, which is used for rotation
to adjust the motion position through their own healthy hands. A linear servo drive is
drive.
LAS10-023D, with a maximum push force of 105 N, which is used for rotation drive.

Thedesign
Figure5.5.The
Figure designofofthe
thethumb.
thumb.

2.2. Control Strategy for Prosthetic Bionic Hand

2.2. Control Strategy for Prosthetic Bionic Hand
2.2.1. Control Process of BIT Hand C
2.2.1. Control Process of BIT Hand C
The BIT hand C is controlled by surface electromechanical signals(sEMG). The control
The isBIT
system handdivided
mainly C is controlled by surface
into four parts, electromechanical
including signal source (EMG signals(sEMG).
signal fromThethe
control system is mainly divided into four parts, including
forearm), data collection, data processing, and decision and control. Figure signal source (EMG signal
6 shows the
Micromachines 2022, 13, x FOR PEER REVIEW 7 of 16
from
flow the forearm),
of the controldata collection,
system. The sEMGdata processing,
signals are and decision
collected fromand control.electrodes.
2 channel Figure 6
shows the flow of the control system. The sEMG signals are
Data processing includes signal preprocessing, window processing, feature extraction collected from 2 channel
and
electrodes. Data processing
feature classification. The finalincludes signal preprocessing, window processing, feature
extraction and feature classification. Theclassification results
final classification resultsare
are sent
sent totothe
the control system,
control which
controls
system, thecontrols
which actiontheof action
the prosthetic bionic
of the prosthetic hand
bionic through
hand through the actuator.
the actuator.

Figure
Figure 6. Control
6. Control flowofchart
flow chart of BIT
BIT hand C. hand C.

2.2.2. Control Strategy of BIT Hand C

The control strategy of BIT hand C is shown in Figure 7. Firstly, the sliding window
is applied to process real-time data into frames, and the window size and frameshift are
250 ms and 100 ms, respectively. After data amplification and filtering, feature extraction
Micromachines 2022, 13, 219 7 of 15

2.2.2. Control Strategy of BIT Hand C

The control strategy of BIT hand C is shown in Figure 7. Firstly, the sliding window
is applied to process real-time data into frames, and the window size and frameshift are
250 ms and 100 ms, respectively. After data amplification and filtering, feature extraction
and action pattern recognition are carried out on the processed data, then the final classifica-
tion results are imported into the control system. Next, we adopt a classification algorithm
to classify the mode of it, that of grasping or opening. Once the algorithm detects the mode
successfully, it can be then be judged as corresponding motion. Finally, the system
Micromachines 2022, 13, x FOR PEER REVIEW will
8 of 16
send the command to the motor of BIT hand C and then it will be driven to move, so as to
realize the grasping and opening action.

Theworking
Figure7.7.The
Figure workingdiagram
diagramof
ofthe
thewhole
wholecontrol
control system.
system.

The highlight of this study is that our control system sets up two closed-loop controls,
3. Data Processing
the current loop and the position loop to ensure the adaptively grasping of objects of
3.1. Data Collection
different shapes and sizes. When the BIT hand C is working, the system will monitor the
current
Theflow through
prosthetic each hand
bionic motor.designed
Once theiscurrent flow isunder-actuated
an adaptive recognized as aprosthetic
certain intention,
hand,
which only needs two electrode channels sEMG signal for high control accuracy. In thisC
the motor will be pushed to grasp or open. In this way, the adaptive grasp of BIT hand
can be
way, it realized,
only needswhichlittlecan better control
calculation workvarious
for datadaily grasp movements
processing and thus of individuals
improves the
with upper-limb
running speed to differences.
make sure the real-time performance.
Before electrodes are placed, the skin paste should be treated, including removing
3. Data
hair, Processing
removing sweat, and wiping clean with alcohol. As experiment shows, we select
3.1. Data Collection
brachioradialis muscle and extensor digitalis muscle as objects to obtain the sEMG signals,
whichThe prosthetic
show obviousbionic
change hand designed
during is anhand
different adaptive under-actuated
movements. prosthetic
The placement hand,
position
which
of only needs
the electrodes is two
shown electrode channels
in Figure sEMG
8. In this paper,signal forthe
we set high control frequency
sampling accuracy. In of this
the
way, it only
surface needs little calculation
electromechanical signal atwork
1000 for
Hz.data
Theprocessing
amplitudeand thus
of the improves
original EMGthesignal
runningis
speed
only 0–6to Mv,
make sosure the real-time
we design performance.
the amplifier circuit to amplify it, and we also perform a filter
Beforethe
to remove electrodes are placed,
noise. Finally, the skin
the usable EMGpaste should
signal canbebetreated, including
obtained. removing
In Figure 9, we
hair, removing sweat, and wiping clean
present the raw sEMG signal from two electrodes. with alcohol. As experiment shows, we select
brachioradialis muscle and extensor digitalis muscle as objects to obtain the sEMG signals,
which show obvious change during different hand movements. The placement position
of the electrodes is shown in Figure 8. In this paper, we set the sampling frequency of the
surface electromechanical signal at 1000 Hz. The amplitude of the original EMG signal is
only 0–6 Mv, so we design the amplifier circuit to amplify it, and we also perform a filter to
 Before electrodes are placed, the skin paste should be treated, including removing
hair, removing sweat, and wiping clean with alcohol. As experiment shows, we select
brachioradialis muscle and extensor digitalis muscle as objects to obtain the sEMG signals,
which show obvious change during different hand movements. The placement position
Micromachines 2022, 13, 219 of the electrodes is shown in Figure 8. In this paper, we set the sampling frequency of8 the
of 15
surface electromechanical signal at 1000 Hz. The amplitude of the original EMG signal is
only 0–6 Mv, so we design the amplifier circuit to amplify it, and we also perform a filter
toremove
removethethe noise.
noise. Finally,
Finally, the the usable
usable EMGEMG signal
signal can
can be be obtained.
obtained. In Figure
In Figure 9, we
9, we present
present the raw sEMG signal from two
the raw sEMG signal from two electrodes.electrodes.

Micromachines 2022, 13, x FOR PEER REVIEW 9 of 16

Micromachines 2022, 13, x FOR PEER REVIEW 9 of 16

Theposition
Figure8.8.The
Figure positionofoftwo
twochannel
channelelectrodes.
electrodes.

Figure 9. SEMG signals from two electrodes.

3.2. Feature
Figure
Figure 9. Extraction
9. SEMG
SEMG signals
signals from
from two
two electrodes.
electrodes.
In the feature
3.2. Feature
Feature Extraction extraction process, we adopt the sliding window to segment sEMG
3.2.
signals. TheExtraction
detail of the method is shown in Figure 10. To make sure real-time detection
In the
In the feature
feature extraction
extraction process,
process, wewe adopt
adopt the
the sliding
sliding window
window to to segment
segment sEMG
sEMG
and decision, the sliding window needs to be overlapped. It is known that the higher the
signals. The detail
signals. The detail of the method is shown in Figure 10. To make sure real-time detection
overlapping radio of is, the
the method is shownspeed
faster response in Figure 10.However,
it has. To make sure it is real-time
not to seedetection
that we
and decision,
and decision, the sliding window needs to be overlapped. It is known that the
the higher
higher the
the
should set it asthe sliding
large windowwithout
as possible needs to be overlapped.
limitation. We shouldIt is known
also takethat
recognition and
overlapping radio is, the faster response speed it has. However, it is not to see that we
overlapping
decision radiointo
accuracy is, the faster response
consideration. speed it has.
Additionally, However,
a suitable size it
of is not towindow
sliding see thatcan
we
should set it as large as possible without limitation. We should also take recognition and
should
not onlyset it as large
improve the as possible speed
algorithm withoutbutlimitation.
also improveWe should also takeaccuracy.
the recognition recognition and
In this
decision accuracy into consideration. Additionally, a suitable size of sliding window can not
decision
paper, weaccuracy
set the into consideration.
window length atAdditionally,
250 ms and athe suitable size ofatsliding
increment 100 ms window can
for better
only improve the algorithm speed but also improve the recognition accuracy. In this paper,
not only
performance.improve the algorithm speed but also improve the recognition accuracy. In this
we set the window length at 250 ms and the increment at 100 ms for better performance.
paper, we set the window length at 250 ms and the increment at 100 ms for better
performance.

Figure 10. Sliding window scheme.

Figure 10. Sliding window scheme.

In10.
Figure this paper,window
Sliding we adopt time domain information for feature extraction. Among them,
scheme.
the integral myoelectric value (iEMG), denoted in Equation (7), the root mean square
valueIn(RMS),
this paper, we adopt
denoted time domain
in Equation (8), information for feature
and mean absolute extraction.
value (MAV),Among them,
denoted in
Micromachines 2022, 13, 219 9 of 15

In this paper, we adopt time domain information for feature extraction. Among them,
the integral myoelectric value (iEMG), denoted in Equation (7), the root mean square value
(RMS), denoted in Equation (8), and mean absolute value (MAV), denoted in Equation (9),
can reflect the variation of sEMG signal amplitude.
Z t+ T
iEMG = | EMG (t)|dt (7)
t
r
1 N
RMS =
N ∑ n =1 x n 2 (8)

1 N
N ∑ n =1
MAV = | xn | (9)

where N is the number of samples in a certain window, and xn is the nth sample in the
EMG signal.
Zero crossing (ZC) is the frequency sEMG signals that cross the x-axis. In order to
avoid the influence of low voltage fluctuation and background noise, a threshold is applied
to the signal amplitude, as shown in the following formula:

N −1 1, i f x ≥ threshold
ZC = ∑i=1 [sgn( xn × xn−1 ) ∩ | xn − xn−1 | ≥ threshold]; sgn( x ) = . (10)
0, otherwise
Variance (Var) is a measure of dispersion degree for a set of data. The equation is
shown below:
2
∑ni=1 Xi − X
Var = (11)
n−1
where Xi represents the ith sample and X represents the mean value of sample data, and n
represents the total number of samples.

3.3. Feature Classification

There are many algorithms for feature classification, which are mature and commonly
used at present, such as Linear Discrimination Analysis (LDA), Support Vector Machine
(SVM), K-Nearest Neighbor (KNN), Artificial Neural Network (ANN), decision tree (DT)
and so on. Linear discriminant analysis is a simple but effective method for classification,
which can achieve fast recognition. The basic idea of LDA classifier is to find an optimal
discriminant vector space to maximize the ratio of the inter-class dispersion and intra-class
dispersion of samples projected.
The calculation steps of the LDA algorithm are shown as follows:
(a) Firstly, calculate the mean vector of each class using Equation (12):

1
ui =
ni ∑x∈classi x (12)

where ni is the number of samples of each class, and x denotes the original feature
vectors of each class.
(b) Secondly, calculate the mean of samples using Equation (13):

1 m
u=
m ∑ i =1 x i (13)

where m is the total number of all classes.

(c) Thirdly, calculated the inter-class divergence matrix and the in-class divergence matrix
using Equations (14) and (15).
c
Sb = ∑i=1 ni (ui − u)(ui − u)T (14)
 Micromachines 2022, 13, 219 10 of 15

c
Sω = ∑i=1 ∑x∈classi (ui − xk )(ui − xk )T (15)

Here, it is important to note that a weighted average is needed when calculating the
total Sb and Sω , because the number of samples for each class may be different.
(d) Finally, we can express the Fisher criterion in terms of Sb and Sω as:

ωT S b ω
J(ω) = (16)
cromachines 2022, 13, x FOR PEER REVIEW ωT Sω ω 11 o
By maximizing the generalized Rayleigh quotient, we use the Lagrange multiplier
method to find the required eigenvectors. Then the original feature matrix can be projected by:

4. Experiments Results y = ωT x (17)

4.1. Intention Recognition Results
The matrix y represents the projection eigenvector of original data, through which we
We,
can in total,
obtain asked
the best 10 subjects
recognition to conduct the experiment, of which 9 subjects
accuracy. w
able-bodied, between 28 and 38 years old, and one individual with upper-limb differen
4. Experiments Results
aged4.1.
52.Intention
All subjects provided informed consent before the experiment. The experime
Recognition Results
procedures were reviewed and approved by the Ethics Committee of The University
We, in total, asked 10 subjects to conduct the experiment, of which 9 subjects were
Electro-Communications
able-bodied, between 28 and (No.
38 10006(5)).
years old, and one individual with upper-limb difference,
For 52.
aged theAllclassification of informed
subjects provided LDA classifier, we finally
consent before collected
the experiment. The 128 samples for e
experimental
procedures
subject. Then, were reviewed
64 were and approved
randomly selectedbyfor
thetraining,
Ethics Committee
and the of The University
other 64 samples of were u
Electro-Communications (No. 10006(5)).
for testing. The subjects were required to perform hand opening and grasp
For the classification of LDA classifier, we finally collected 128 samples for each
movements, and64each
subject. Then, weremovement hold for
randomly selected fortraining,
5 s. Every hand
and the gesture
other waswere
64 samples repeated
used 10 tim
To avoid muscle
for testing. The fatigue, the required
subjects were subjectstowere
performgiven
handaopening
five-minute breakmovements,
and grasping before the next t
and each movement hold for 5 s. Every hand gesture was
Then, VAR, RMS and MIN features were extracted and the LDA algorithm repeated 10 times. To avoid
was app
muscle fatigue, the subjects were given
for action recognition, as shown in Figure 11. a five-minute break before the next test. Then,
VAR, RMS and MIN features were extracted and the LDA algorithm was applied for action
recognition, as shown in Figure 11.

Figure 11. Opening and gripping action recognition with LDA algorithm.

Figure 11.Green
Opening and
points gripping
represent theaction recognition
eigenvalue when with LDAisalgorithm.
the hand open, while pink points
represent the eigenvalue when the hand is grasped. It can be seen from Figure 11 that the
three
Greenfeatures,
points VAR, RMS, MIN,
represent thecan be well distinguished
eigenvalue when theusing
handtheisLDA classification
open, while pink po
represent the eigenvalue when the hand is grasped. It can be seen from accuracy
algorithm. After training, we tested the new action data and calculated its Figure 11 that
with Equation (18).
three features, VAR, RMS, MIN, can be well distinguished using the LDA classificat
algorithm. After training, we tested the new action data and calculated its accuracy w
Equation (18).
𝑁𝑂. 𝑜𝑓 𝑎𝑙𝑙 𝑡ℎ𝑒 𝑓𝑒𝑎𝑡𝑢𝑟𝑒𝑠 − 𝑁𝑂. 𝑜𝑓 𝑓𝑒𝑎𝑡𝑢𝑟𝑒 𝑐𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑐𝑎𝑡𝑖𝑜𝑛 𝑒𝑟𝑟𝑜𝑟𝑠
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = × 100%
𝑁𝑂. 𝑜𝑓 𝑎𝑙𝑙 𝑡ℎ𝑒 𝑓𝑒𝑎𝑡𝑢𝑟𝑒𝑠
 Micromachines 2022, 13, 219 11 of 15

NO. o f all the f eatures − NO. o f f eature classi f ication errors

Accuracy = × 100% (18)
NO. o f all the f eatures

As the results show, we can achieve an average accuracy of 96.59% with the LDA algorithm.
To prove the effectiveness of the LDA algorithm, we compare the recognition accuracy
with other recognition algorithms, such as native bayes (NB), support vector machine
(SVM), decision tree (DT), nearest neighbor (KNN) and linear discriminant analysis (LDA),
as shown in Table 1. For convenience, we only present the results of six subjects.

Table 1. Recognition accuracy of different classifiers for 6 subjects (S1–S6).

Micromachines
Accuracy2022,
(%) 13, x FOR
S1PEER REVIEWS2 S3 S4 S5 S6 Average 12 of 16

NB 87.5 95.31 84.38 93.75 98.44 92.19 91.93

SVM 95.31 96.88 96.13 96.88 97.88 95.31 96.40
DT 87.94 92.19 91.38 93.75 98.44 90.63 92.39
KNN 86.94 KNN
92.19 86.94
86.39 92.1992.19 86.39 92.19
98.44 98.44
92.19 92.19
91.39 91.39
LDA 94.31 LDA
97.31 94.31
97.13 97.3197.43 97.13 97.43
96.92 96.92
96.45 96.45
96.59 96.59

Among
Among thethealgorithms,
algorithms,LDA LDA shows
shows high
high accuracy.
accuracy. Furthermore,
Furthermore, we compare
we compare the the
computation time of different classifiers, as shown in Figure
computation time of different classifiers, as shown in Figure 12. 12.

Figure
Figure 12.
12. Comparison
Comparison ofofdifferent
different classifiers
classifiers in terms
in terms of accuracy
of accuracy and running
and running time. time.

From Figure
From Figure12,12,we
wecan
canseesee
that LDA
that LDAcan can
achieve an accuracy
achieve of 96.59%,
an accuracy which is
of 96.59%, which is
greatly higher by 4.2% and 4.66% than that of DT and NB respectively. Furthermore, for
greatly higher by 4.2% and 4.66% than that of DT and NB respectively. Furthermore, for
running speed, LDA only needs 6.33 ms, which is only slightly slower by 0.5 ms and
running speed, LDA only needs 6.33 ms, which is only slightly slower by 0.5 ms and 0.16
0.16 ms than that of DT and NB respectively. Weighing these factors, we conclude that LDA
ms than that
performs ofthan
better DT and
otherNB respectively. Weighing these factors, we conclude that LDA
algorithms.
performs better
For the than other
observation of algorithms.
the classification results of the LDA algorithm clearly, we
For the observation
present the confusion matrix ofofthe classification
subject results
1 for certain 6 timesofexperiments,
the LDA algorithm clearly,
for example, as we
present
shown inthe confusion
Figure 13. matrix of subject 1 for certain 6 times experiments, for example, as
shown in Figure 13.
 greatly higher by 4.2% and 4.66% than that of DT and NB respectively. Furthermore, for
running speed, LDA only needs 6.33 ms, which is only slightly slower by 0.5 ms and 0.16
ms than that of DT and NB respectively. Weighing these factors, we conclude that LDA
performs better than other algorithms.
For the observation of the classification results of the LDA algorithm clearly, we
Micromachines 2022, 13, 219 12 of 15
present the confusion matrix of subject 1 for certain 6 times experiments, for example, as
shown in Figure 13.

Micromachines 2022, 13, x FOR PEER REVIEW 13 of 16

Figure 13.
Figure 13. Confusion
Confusion matrix
matrixof
ofsubject
subject11for
forcertain
certain6 6times experiments
times (E1–E6).
experiments 1—grasping,
(E1–E6). 2—
1—grasping,
opening.
2—opening.

4.2.
4.2.Grasp
GraspTest
TestResults
Results
In
Inthis
thispart, wewe
part, verify the reliability
verify of the of
the reliability control strategystrategy
the control based on based
the LDAonalgorithm.
the LDA
During tests,
algorithm. subjects
During aresubjects
tests, asked to rest
are for five
asked minutes
to rest for fiveinstead
minutesof instead
conducting five grip
of conducting
tests to avoid
five grip testsmuscle
to avoidfatigue.
muscle The test content
fatigue. The testis content
to control
is BIT hand CBIT
to control to hand
grasp C
objects of
to grasp
different shapes and sizes. In this experiment, five kinds of objects with different
objects of different shapes and sizes. In this experiment, five kinds of objects with different shapes
were
shapesselected, includingincluding
were selected, cylinder, cube, sphere,
cylinder, torus
cube, and pentagonal
sphere, torus and column. Each
pentagonal shape
column.
comes in three sizes: large, medium and small, as shown in Figure
Each shape comes in three sizes: large, medium and small, as shown in Figure 14.14.

Figure14.
Figure 14.Tests
Testson
onindividuals
individualswith
withupper-limb
upper-limbdifference
differencegrasping
graspingdifferent
differentobjects.
objects.

Figure
Figure1515shows
showsthe
thehand
handposture
posturefor forgrasping
graspingobjects
objectsof
ofdifferent
differentsizes
sizesand
andshapes
shapes
with
withour
ourprosthetic
prostheticbionic
bionichand.
hand.ItItcan
canbe
beeasily
easilyfound
foundthat
thatour
ourhands
handscan
can adaptively
adaptivelygrasp
grasp
various
variousobjects.
objects.
In this part, we invited 10 volunteers to participate in these trials, including 9 able-
bodied volunteers, aged between 28 and 36, and one volunteer with upper limb disability.
Participants were all required to wear BIT hand C to grab different objects and move them
from a table into a plastic basket. For each participant, they were asked to perform 6 trials
 Micromachines 2022, 13, x FOR PEER REVIEW 14
Micromachines 2022, 13, 219 13 of 15

Figure 15. Grasp test of different objects of BIT hand C.

Figure 15. Grasp test of different objects of BIT hand C.
In this part, we invited 10 volunteers to participate in these trials, including 9 able-
Tableaged
bodied volunteers, 2. Grasping
betweenresults of different
28 and 36, andobjects with 10 participants.
one volunteer with upper limb disability.
Participants were all required
Shape to wear BIT hand C
Size to grab different objects
Success Rateand move them
Time-Consuming
from a table into a plastic basket. For each participant,
Small they were asked
90.74% to perform 6 trials4.84
for each object. Finally, we obtain grasping time data and success rate data from 60 trials
Sphere Middle 96.30% 3.94
for each object.
Big 100.00% 4.08
Volunteers were required to grasp a single object from initial status and then drop it in
the basket. A trial was regarded a success when Small 98.15%
the objects were transported successfully3.91
Cube
without being dropped in this process. We defined the time this process spends as time-3.91
Middle 100.00%
consuming. The results are shown in Table 2. Big 100.00% 3.73
Small 96.30% 4.36
Table 2. Grasping resultsTorus
of different objects with 10 participants.
Middle 98.15% 4.20
Big 100.00% 4.21
Shape Size Success Rate Time-Consuming(s)
Small 92.59% 5.11
Small 90.74% 4.84
Pentagonal Column Middle 98.15% 4.31
Sphere Middle 96.30% 3.94
Big Big 100.00% 100.00% 4.08 4.33
Small
Small 98.15%
90.74% 3.91
5.08
Cube Cylinder Middle Middle 100.00% 96.30% 3.91 4.06
Big Big 100.00% 100.00% 3.73 4.16
Small 96.30% 4.36
Torus 5. Conclusions Middle 98.15% 4.20
In this paper,Big 100.00% prosthetic bionic,4.21
we developed a humanoid hand BIT hand C, wh
is controlled by two electrodes placed on
Small the skin surface. This 5.11
92.59% system can identify h
Pentagonal Column Middle movements with
opening and grasping 98.15%
a high accuracy of 4.3196.59% based on
Big 100.00% 4.33
experiments. The hand can realize adaptive grasping through connecting rod mechan
At the same time,Small
two control loops, including
90.74% the current loop 5.08
and the position loop,
Cylinder Middle strategy to ensure
adopted in the control 96.30%
the adaptive gripping4.06
of the artificial hand
Big
results show, our system performs well100.00% 4.16
in grasping objects of different sizes and sha
which has the potential to be applied for individuals with upper-limb differences. A th
As the results show, our system performs better in grasping objects, no matter what
shape and size it is, and the success rate can be all above 90%. Besides, people only need a
Micromachines 2022, 13, 219 14 of 15

little time to grasp the objects, almost no more than 5 s, which shows that our prosthetic
hand has a better real-time performance. We can also conclude that with the size increase,
the success rate can be improved, and the grasping time is shorter. This means that if the
object is too small, the system should adjust their body posture repeatedly and then take
more time for adaptive grasping. Therefore, the prosthetic bionic hand can meet the daily
life application in a variety of tasks.

5. Conclusions
In this paper, we developed a humanoid prosthetic bionic, hand BIT hand C, which
is controlled by two electrodes placed on the skin surface. This system can identify hand
opening and grasping movements with a high accuracy of 96.59% based on our experiments.
The hand can realize adaptive grasping through connecting rod mechanism. At the same
time, two control loops, including the current loop and the position loop, are adopted in
the control strategy to ensure the adaptive gripping of the artificial hand. As results show,
our system performs well in grasping objects of different sizes and shapes, which has the
potential to be applied for individuals with upper-limb differences. A thing to be noticed is
that our prosthetic hand can only help restore some functionality to grasping tasks so far. It
is now a proof of concept for a new type of prosthetic hand. However, this functional and
reliable hand might improve prosthesis satisfaction and decrease the cognitive effort and
visual attention required to control the device. In future work, we will further increase the
function and availability of the hand and make it more accessible to people.

Author Contributions: Conceptualization, Y.W. and Y.L.; methodology, Y.W.; formal analysis, Y.J.;
data curation, H.S.; writing—original draft preparation, Y.W.; writing—review and editing, Y.T.;
supervision, H.Y. All authors have read and agreed to the published version of the manuscript.
Funding: This work was supported by the Beijing Advanced Innovation Center of Intelligent Robots
and Systems under Grant 2016IRS23.
Institutional Review Board Statement: The study was conducted according to the guidelines of
the Declaration of Helsinki and approved by the Institutional Review Board of The University of
Electro-Communications (10006(5) approved on 8 June 2018).
Informed Consent Statement: Informed consent was obtained from all subjects involved in the study.
Data Availability Statement: The data presented in this study are available on request from the
corresponding author.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Kyberd, P.J.; Light, C.; Chappell, P.H.; Nightingale, J.M.; Whatley, D.; Evans, M. The design of anthropomorphic prosthetic hands:
A study of the Southampton Hand. Robotica 2001, 19, 593–600. [CrossRef]
2. Dalley, S.A.; Wiste, T.E.; Withrow, T.J.; Goldfrab, M. Design of a Multifunctional Anthropomorphic Prosthetic Hand with Extrinsic
Actuation. IEEE ASME Trans. Mechatron. 2009, 14, 699–706. [CrossRef]
3. Connolly, C. Prosthetic hands from Touch Bionics. Ind. Robot 2008, 35, 290–293. [CrossRef]
4. Okada, T. Computer control of multi-jointed finger system for precise object-handling. IEEE Trans. Syst. Man Cybern. 1982, 12,
289–299. [CrossRef]
5. Salisbury, J.K.; Mason, M.T. Robot Hands and the Mechanics of Manipulation; MIT Press: Cambridge, MA, USA, 1985.
6. Kochan, A. Shadow delivers first hand. Ind. Robot. Int. J. 2005, 32, 15–16. [CrossRef]
7. Jacobsen, S.C.; Wood, J.E.; Knutti, D.F.; Biggers, K.B. The UTAH/MIT dextrous hand: Work in progress. Int. J. Robot. Res. 1984, 3,
21–50. [CrossRef]
8. Iversen, E.; Sears, H.H.; Jacobsen, S.C. Artificial arms evolve from robots, or vice versa. IEEE Control Syst. Mag. 2005, 25, 16–18.
9. Jacobsen, S.C. Design of the Utan/MIT Dextrous Gripper. In Proceedings of the 1986 IEEE International Conference on Robotics
and Automation, San Francisco, CA, USA, 7–10 April 1986; Volume 4, pp. 1520–1523.
10. Nakano, Y.; Fujie, M.; Hosada, Y. Hitachi robot Hand. Robot. Age 1984, 6, 18–20.
11. Tawfik, M.A.; Baqer, I.A.; Abdulsahib, A.D. Grasping Force Controlling by Slip Detection for Specific Artificial Hand (ottobock
8E37). Eng. Technol. J. 2018, 36, 979–984. [CrossRef]
Micromachines 2022, 13, 219 15 of 15

12. Chu, J.U.; Moon, I.; Lee, Y.J.; Kim, S.-K.; Mun, M.-S. A Supervised Feature-Projection-Based Real-Time EMG Pattern Recognition
or Multifunction Myoelectric Hand Control. IEEE ASME Trans. Mechatron. 2007, 12, 282–290. [CrossRef]
13. Nam, Y.; Koo, B.; Cichocki, A.; Choi, S. GOM-Face: GKP, EOG, and EMG-Based Multimodal Interface with Application to
Humanoid Robot Control. IEEE Trans. Biomed. Eng. 2014, 61, 453–462. [CrossRef] [PubMed]
14. Burgess-Limerick, R. Squat, stoop, or something in between? Int. J. Ind. Ergon. 2003, 31, 143–148. [CrossRef]
15. Hsiang, S.M.; Brogmus, G.E.; Courtney, T.K. Low back pain (LBP) and lifting technique—A review. Int. J. Ind. Ergon. 1997, 19,
59–74. [CrossRef]
16. Brad, S.; Sherry, S. The extraction of neural strategies from the surface EMG. J. Appl. Physiol. 2004, 96, 1486–1495.
17. Scheme, E.J.; Englehart, K.B.; Hudgins, B.S. Selective Classification for Improved Robustness of Myoelectric Control Under
Nonideal Conditions. IEEE Trans. Biomed. Eng. 2011, 58, 1698–1705. [CrossRef] [PubMed]
18. Clancy, E.A.; Morin, E.L.; Merletti, R. Sampling, noise-reduction and amplitude estimation issues in surface electromyography.
J. Electromyogr. Kinesiol. 2002, 12, 1–16. [CrossRef]
19. Reaz, M.; Hussain, M.S.; Mohd-Yasin, F. Techniques of EMG signal analysis: Detection, processing, classification and applications.
Biol. Proced. Online 2006, 8, 11–35. [CrossRef] [PubMed]
20. Shair, E.F.; Ahmad, S.A.; Marhaban, M.H.; Mohd Tamrin, S.B.; Abdullah, A.R. EMG Processing Based Measures of Fatigue
Assessment during Manual Lifting. BioMed Res. Int. 2017, 2017, 3937254. [CrossRef] [PubMed]
21. Nsugbe, E.; Phillips, C.; Fraser, M.; McIntosh, J. Gesture Recognition for Trans-humeral Prosthesis Control Using EMG and NIR.
IET Cyber. Syst. Robot. 2020, 2, 122–131. [CrossRef]
22. Phinyomark, A.; Hirunviriya, S.; Limsakul, C.; Phukpattaranont, P. Evaluation of EMG feature extraction for hand move-
ment recognition based on Euclidean distance and standard deviation. In Proceedings of the International Conference on
Electrical Engineering/Electronics Computer Telecommunications and Information Technology IEEE, Chiang Mai, Thailand,
19–21 May 2020.
23. Wang, N.; Lao, K.; Zhang, X. Design and Myoelectric Control of an Anthropomorphic Prosthetic Hand. J. Bionic Eng. 2017, 14,
47–59. [CrossRef]
24. Farina, D.; Lucas, M.-F.; Doncarli, C. Optimized wavelets for blind separation of non-stationnary surface myoelectric signals.
IEEE Trans. Biomed. Eng. 2007, 55, 78–86. [CrossRef] [PubMed]
25. Tkach, D.; Huang, H.; Kuiken, T.A. Study of stability of time-domain features for electromyographic pattern recognition.
J. Neuroeng. Rehabil. 2010, 7, 21. [CrossRef] [PubMed]
26. Potluri, C.; Kumar, P.; Anugolu, M.; Urfer, A.; Chiu, S.; Subbaram Naidu, D.; Schoen, M.P. Frequency Domain Surface EMG
Sensor Fusion for Estimating Finger Forces. In Proceedings of the 2010 Annual International Conference of the IEEE Engineering
in Medicine and Biology, Buenos Aires, Argentina, 31 August–4 September 2010; pp. 5975–5978.
27. Sapsanis, C.; Georgoulas, G.; Tzes, A. EMG Based Classification of Basic Hand Movements Based on Time-Frequency Features. In
Proceedings of the 21st Mediterranean Conference on Control and Automation, Platanias, Greece, 25–28 June 2013.
28. Hudgins, B.; Parker, P.; Scott, R.N. A new strategy for multifunction myoelectric control. IEEE Trans. Biomed. Eng. 1993, 40, 82–94.
[CrossRef] [PubMed]
29. Kim, K.S.; Choi, H.H.; Moon, C.S.; Mun, C.W. Comparison of k-nearest neighbor, quadratic discriminant and linear discriminant
analysis in classification of electromyogram signals based on the wrist-motion directions. Curr. Appl. Phys. 2011, 11, 740–745.
[CrossRef]
30. Zecca, M.; Micera, S.; Carrozza, M.C.; Dario, P. Control of multifunctional prosthetic hands by processing the electromyographic
signal. Crit. Rev. Biomed. Eng. 2002, 30, 459–485. [CrossRef]
31. Oskoei, M.A.; Hu, H. Support Vector Machine-Based Classification Scheme for Myoelectric Control Applied to Upper Limb. IEEE
Trans. Biomed. Eng. 2008, 55, 1956–1965. [CrossRef]
32. Mane, S.M.; Kambli, R.A.; Kazi, F.S.; Singh, N.M. Hand Motion Recognition from Single Channel Surface EMG Using Wavelet &
Artificial Neural Network. Procedia Comput. Sci. 2015, 49, 58–65.
33. Zhang, H.; Zhao, Y.; Yao, F.; Xu, L.; Shang, P.; Li, G. An adaptation strategy of using LDA classifier for EMG pattern recognition.
In Proceedings of the 35th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC),
Osaka, Japan, 3–7 July 2013; pp. 4267–4270.
34. Xing, K.; Yang, P.; Huang, J.; Wang, J.; Zhu, Q. A real-time EMG pattern recognition method for virtual myoelectric hand control.
Neurocomputing 2014, 136, 345–355. [CrossRef]
35. Raurale, S.A.; Chatur, P.N. Identification of Real-Time Active Hand Movements EMG Signals for Control of Prosthesis Robotic
Hand. In Proceedings of the International Conference on Computation of Power, Chennai, India, 16–17 April 2014.
36. Zhang, D.; Zhao, X.; Han, J.; Zhao, Y. A comparative study on PCA and LDA based EMG pattern recognition for anthropomorphic
robotic hand. In Proceedings of the 2014 IEEE International Conference on Robotics and Automation (ICRA), Hong Kong, China,
31 May–7 June 2014.
37. Antuvan, C.W.; Masia, L. An LDA-Based Approach for Real-Time Simultaneous Classification of Movements Using Surface
Electromyography. IEEE Trans. Neural Syst. Rehabil. Eng. 2019, 27, 552–561. [CrossRef]