-
Electromyographic pattern analysis andclassification for a
robotic prosthetic arm
doi:10.1533/abbi.2005.0039
M. José H. Erazo Macias1 and S. Alejandro Vega21Department of
Electric and Electronic Engineering, Technological Institute of
Reynosa, Reynosa Tamaulipas, México2Technological Institute of
Superior Studies of Monterrey, Campus Queretaro, Querétaro,
México
Abstract: This paper deals with the statistical analysis and
pattern classification of electromyographicsignals from the biceps
of a person with amputation below the humerus. Such signals
collected from anamputation simulator are synergistically generated
to produce discrete elbow movements. The purposeof this study is to
utilise these signals to control an electrically driven prosthetic
or orthotic elbowwith minimum extra mental effort on the part of
the subject. The results show very good separabilityof classes of
movements when a learning pattern classification scheme is used,
and a superpositionof any composite motion to the three basic
primitive motions—humeral rotation in and out, flexionand
extension, and pronation and supination. Since no synergy was
detected for the wrist movement,different inputs have to be
provided for a grip. In addition, the method described is not
limited by thelocation of the electrodes. For amputees with shorter
stumps, synergistic signals could be obtained fromthe shoulder
muscles. However, the presentation in this paper is limited to
biceps signal classificationonly.
Key words: Electromyographic, amputated, orthotic, pronation,
supination.
INTRODUCTION
Since World War II, many attempts have been made to useexternal
power to operate artificial prosthetic or orthoticarms (Alderson
1954). Electromyographic (EMG) signalsfrom the body’s intact
musculature have been suggestedand utilised by many researchers as
an effective noninvasivemethod to provide commands to control an
electricallypowered artificial limb (Graupe et al. 1978; Lyman et
al.1974; Saidis and Stephenou 1977). Most of this researchhas
produced single motion controls of the bang-bang type,which require
special training and effort on the part of thesubject.
The introduction of microcomputers has made it pos-sible to
experiment with more sophisticated signal detec-tion and motion
control of human prostheses. Graupe andSalahi (1978) and Graupe et
al. (1978) proposed and suc-cessfully implemented a time series
identification to re-cover control information from synergistic EMG
signals
Corresponding Author:M. José H. Erazo MaciasDepartment of
Electric and Electronic EngineeringTechnological Institute of
ReynosaReynosa TamaulipasMéxicoEmail: erazo [email protected]
produced by the subject’s intact musculature. However,this
project was terminated before any conclusive designprocedures were
developed.
This paper presents an approach to detect, analyse, andclassify
synergistic EMG signals generated by the biceps ofan
above-the-elbow amputee in an attempt to move the ar-tificial limb
without extra mental effort. The method is byno means limited to
this location of the electrodes (Graupeand Salahi 1979). However,
since stronger synergies areexpected from the arm muscles, they
were prepared in thisstudy. An elbow of three degrees of freedom is
considered.It was built using a newly designed product (made in
theDepartment of Electric and Electronic Engineering,
Tech-nological Institute of Reynosa, Tamaulipas, Mexico), whilethe
mechanism and all its parts were made in the Techno-logical
Institute of Reynosa (Figure 1). The elbow could bedriven to
perform all possible simultaneous combinationsin the following six
primitive motions:
• Humeral rotation inside• Humeral rotation outside• Elbow
flexion• Elbow extension• Elbow pronation of wrist• Elbow
supination of wrist
The wrist and hand grasp was not included in the
finalclassification because no synergistic signals were
obtained
C© Woodhead Publishing Ltd 113 ABBI 2006 Vol. 3 No. 2 pp.
113–119
-
M. José H. Erazo Macias and S. Alejandro Vega
Figure 1 Prosthetic elbow.
experimentally on the biceps. The grasp could be
generatedthrough other muscles of the body.
A complete statistical analysis of the EMG signals
corre-sponding to the above motions was first reported by
Saridisand Newman (1978), who used simulated amputation
mea-surements as described in the next section. The results
in-dicated that the generated signals—owing to their nonsta-tionary
properties—are unreliable for time series analysisand physiological
interpretation, but their true averagesmay serve as feature vectors
for pattern classification.
Subsequently, statistical pattern recognition algorithmswere
used to classify the signals to classes correspondingto each of the
possible signal and combined primitive mo-tions of the arm, and
their misclassification error was stud-ied. Loading and velocity
conditions found to be linearlyrelated to the latter’s relations to
the features associatedwith the classification problem, for
example, zero crossing(ZC) and signal variance, were established
(Bigland andLippold 1954). Using envelopes of the classes for
allloads and velocities, a classification algorithm independentof
varying loading and velocity conditions was obtained.Finally, by
discovering certain linear superposition prop-erties on the
statistics of the features, it was possible todecompose combined
motion to their relative velocities.
Load information was obtained through the integral ofthe
muscular activity variable evaluated from the ZC andvariance
measurements. The generation of the control al-gorithms, as well as
further elaboration on the muscular ac-tivity variable, will be
reported in a follow-up paper aimedat the direct control of a
prosthetic arm by EMG signalswith minimum effort or training on the
part of the amputee.
EXPERIMENTAL PROCEDURE
The EMG signals are easily gathered using skin electrodes.In
this study, two electrode sites were used. At each site,one
differential silver–silver chloride electrode is separatedby a
ground electrode. They are separated from the skinby
gel-impregnated foam, and were attached by an adhe-sive foam pad.
The electrodes are placed parallel to thehumeral axis, slightly off
the center of the bulge of each ofthe biceps. The locations are the
lowestmost placementsthat are undamaged in most above-the-elbow
amputees forwhom this work will be applicable (Newman and
Saridis1978; Rimroth et al. 1978; Saridis and Newman 1979;
Wirta
Figure 2 Electrode. Test on real muscle.
et al. 1968). An on-board microcomputer should sample thetwo
signals and sequentially update its decision as to whichmotion is
meant to be in progress. The stream of decisionsis fed into a
coordinator of the motions of the arm. Thisis done at intervals of
several milliseconds. The amputee’svisual feedback of the arm’s
position and velocity shouldthus provide accurate enough correction
of the motion.We note that the coordinator is designed with a
mem-ory to provide trend characteristics of the decision streamto
eliminate inadvertent errors and improve
classificationaccuracy.
Experimental data were obtained from an amputationsimulator
instead of a real amputee, where the subjectwould completely
immobilise his arm in a cast at and belowthe humerus. The device
electrode is shown in Figure 2(Saridis and Newman 1979). The reason
for the simulationwas, first, availability of more subjects from a
student pop-ulation, for example, three different subjects have
alreadybeen used, and second, the flexibility to use the results
fordesigning either upper limb prostheses or orthoses. Theresults
are claimed to be representative of EMG signalsfrom actual amputees
with nonatrophised musculature be-cause with the cast, the subject
would generate the samesynergistic signals over his biceps when he
would attemptto move his immobilised arm.
The experiments were performed with the arm in avertical
position. Small-angle variations of the arm positionaround the
perpendicular tested showed no effect in theclassification. Over
30◦ angle changes in the position of thearm from the vertical, as
well as a test on real amputees,are planned in a later phase of
this project.
The amputation simulator is constructed of cast alu-minum and
extends from just proximal to the elbow to thefingertips. The cast
holds the elbow at a mid-elbow flex-ion of 90◦, humerus to forearm,
and pronation–supination.
114ABBI 2006 Vol. 3 No. 2 doi:10.1533/abbi.2005.0039 C© Woodhead
Publishing Ltd
-
EMG pattern analysis and classification for a robotic prosthetic
arm
Differential Amplifier
Vs
GND
Input
Input
R9
R8R7
R6
R5
R4
R3
R2
R1
Low-pass filter
Input X
Output m
R18 R19
R20
R21
C6
C5C4
C3
R17
R16
R15R14
High-pass filter
InputOutput
C2C1
R13R12
R11R10
−+
−+
−+
−+
−+
−+
Figure 3 EMG amplifier and filters.
When the shoulder has normal freedom of motion, its posi-tion
may be monitored through potentiometers mounted ateach pin joint.
The aluminium elbow mounting bracket forthe two high-gain, high
input impedance EMG amplifiersis shown in Figure 3.
Twenty-nine motions were defined, along with onecontrol (zero)
motion, each consisting of either graspopen/close, or one to three
of the wrist, elbow andhumeral primitives. A more detailed
description is given inNewman and Saridis (1978) and Saridis and
Newman(1979). Since the biceps waveforms apparently containedvery
little information from the grasp primitives, classifica-tion of
the grasp primitives was not pursued further in thisstudy.
Figure 4 Outline of the arm with the weight, showing thevectors
that intervene in the taking of the signals.
STATISTICAL EMG ANALYSIS
These signals will be registered in a team cyberamp 100,with
gain (100), for conditioning of the signals, withjoining of AC of
0.01 Hz using another cyberamp 380,with gain (50), with joining AC
of 10 Hz. After filtra-tion it passes the first floor of 400 Hz. We
also have afilter that eliminates the noise generated in the band
of60 Hz.
A card digidata 1322A with 16 resolution bites occu-pies a
frequency of sampling of 2000 per second. At firsta weight is
placed in the palm of a person for 5 s in aseated position, and
this is repeated with different weights(Figure 4). The weights used
are as follows:
Weights
P1 240 gP2 610 gP3 650 gP4 940 gP5 1.74 kgP6 2.98 kgP7 3.24
kg
Sequences of EMG data were obtained with a student ofnormal
constitution, 27 years old, making two series, onein the morning
and the other in the afternoon.
Morning
Name of the file: Sitting position Event
EMG19 0000 P5+P6+P7EMG19 0001 P5EMG19 0002 P5+P1EMG19 0003
P5+P1+P2EMG19 0004 P5+P1+P2+P3EMG19 0005 P5+P1+P2+P3+P4EMG19 0006
P5+P6EMG19 0007 P5+P7
115C© Woodhead Publishing Ltd doi:10.1533/abbi.2005.0039 ABBI
2006 Vol. 3 No. 2
-
M. José H. Erazo Macias and S. Alejandro Vega
Name of the file: Foot position Event
EMG19P 0000 P5+P6+P7EMG19P 0001 P5EMG19P 0002 P5+P1EMG19P 0003
P5+P1+P2EMG19P 0004 P5+P1+P2+P3EMG19P 0005 P5+P1+P2+P3+P4EMG19P
0006 P5+P6EMG19P 0007 P5+P7
Afternoon
Name of the file: Sitting position Event
EMG19V 0000 P5+P6+P7EMG19V 0001 P5EMG19V 0002 P5+P1EMG19V 0003
P5+P1+P2EMG19V 0004 P5+P1+P2+P3EMG19V 0005 P5+P1+P2+P3+P4EMG19V
0006 P5+P6EMG19V 0007 P5+P7
Name of the file: Foot position Event
EMG19VP 0000 P5+P6+P7EMG19VP 0001 P5EMG19VP 0002 P5+P1EMG19VP
0003 P5+P1+P2EMG19VP 0004 P5+P1+P2+P3EMG19VP 0005
P5+P1+P2+P3+P4EMG19VP 0006 P5+P6EMG19VP 0007 P5+P7
These registrations were introduced into a program, tobe run in
the MatLab software, to obtain the correspondinggraphs for each one
of the cases, some of which are shownhere.
Once the data are registered, MatLab provides the meanvalues and
the securities RMS or the combinations that arewanted.
The mean of the signal, composed of a high number of(more than
500) sample points, is verified by computationin MatLab (Figure 5).
The variance σ 2 of the EMG signalv(t) as given by
σ 2 = 1800
800∑t=1
v(t)2
It is used as one variable. The absolute value of the moment|σ
3| given by
|σ 3| =∣∣∣∣∣ 1800
800∑t=1
v(t)3∣∣∣∣∣
is another variable. The absolute value is taken to
greatlyreduce within-class separation. Also computed is the
fourth
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Signal’s half of absolute value
Signal’s half of absolute value
0.2
0.15
0.1
0.05
0
0.1
0
0.02
0.04
0.08
0.06
Sig
nal’s
Var
ianc
e
EMG Signal
Figure 5 Electromyograph of the variance against thestocking of
the absolute value of the signal.
moment σ 4, the absolute value of the fifth moment |σ 5|,and the
number of times the signal potential passes throughor touches zero
(i.e., ZCs).
Also computed for each signal were the cross correla-tions
between the bicep signal Rb(τ ), the power spectraSb(ω), computed
by fast Fourier transform, and the signalvoltage density Db(v).
Density and correlation were laterdiscarded because they did not
apparently contribute classseparation, that is, in distinguishing
between motions.
Sb(ω) and Rb(τ ) have both 800 components, each ofwhich is a
measurement variable. Although the informationcontent of an
autocorrelation and a power spectrum is in asense redundant, and
although R(0) = σ 2, all of these arecomputed to simplify the
process of feature selection. Themeasurement space, therefore, has
2058 components.
[σ 2 + |σ 3| + σ 4 + |σ 5| + ZC + R(τ ) + S(ω)].
The final conclusions from the statistical analysis are
asfollows.
• The EMG signals are not stationary and are band limitedto
virtually 1200 Hz.
• No synergistic signals are generated from the hand grasp.•
Pattern information of the arm is contained in the ZC and
depends on the electrical activity of the muscles.• Loading and
velocity information of the arm is contained
in the ZCs and variance, and depends on the internal mus-cular
electrical activity.
• Motion information is concentrated mostly in the variancesand
less in the higher-order moments.
• In most motions, the variance and ZC information aregrouped in
separable clusters that may be approximatedby Gaussian densities,
and their moments represent theclasses.
116ABBI 2006 Vol. 3 No. 2 doi:10.1533/abbi.2005.0039 C© Woodhead
Publishing Ltd
-
EMG pattern analysis and classification for a robotic prosthetic
arm
Trajectory Segmentation Tasks
Subtasks
Motion Blender
Learning Automaton
Subgoal1
Subgoalp
Controller1
Controllerp
Subprocess1
Subprocessp
EMG Command
ElementaryMotions
Level 1Organization
Level 2Coordination
Level 3Control
ProcessArm
Figure 6 Hierarchically intelligent control of a
prostheticarm.
All of the above indicate that pattern recognition methodsmay be
applicable to classify and recognize the variousmotions under
consideration.
PATTERN CLASSIFICATION FOR EMG SIGNALS
Hierarchical intelligent control methods (Saridis andStephenou
1977) may be applicable to drive the pros-thetic arm if movement
and loading information is prop-erty decoded and applied to a
microcomputer system, pro-grammed accordingly. Pattern recognition
methods arequite adequate to provide such information as will be
ob-vious from this study. The final control configuration
ispresented in Figure 6 for an arm incremental motion.
In most pattern recognition problems, it is necessary togreatly
reduce the measurement space into a feature spaceof suitable
dimension so that the final pattern recognitionalgorithm can be
managed using an on-board microcom-puter. This will be accomplished
at the end of this study byselecting as features the variables that
contain maximuminformation on class separation.
The variance of the signal was selected as the measureof scatter
to determine the importance of each variable. Ifa variable has
large between-class and small within-classscatter, that is, the
criterion for class separation, for anytwo classes, then it is
considered important. All unimpor-tant variables are immediately
discarded. The ranking ofeach of the remaining features is
determined by the totalnumber of classes separated and the relative
amount ofseparation of the variable. When the feature count can
be
reduced to 50 or less, a computer can handle the com-putations
necessary for a rigorous mathematical analysisto further minimize
the feature vector, while consider-ing any resulting classification
error that may be expected(Fukunaga 1972). The learning algorithm
(Appendix I)makes it possible to determine the number of iterations
foreach pair of movements in the elbow.
THE PROPERTY OF SUPERPOSITION OF PRIMITIVES
Owing to the physical nature of the problem, it seemsreasonable,
perhaps even more straightforward, to con-sider only six classes,
each defined as the presence of oneof the primitives. In doing so,
it was noted that usingthe variances, the primitives exhibit
superposition proper-ties, which might provide the means to develop
a six-classscheme. That is, linear combinations of the variance
dataare taken for more than one primitive, the resulting
distri-bution approximates that of the combined motion involv-ing
the primitives used. Thus, distributions of combinedmotions should
be decomposed into the distributions oftheir primitives, possibly
permitting primitive identifica-tion from the variances of unknown
motions.
For each pair of primitive motions, coefficients of alinear
combination of points from the two variance distri-butions were
calculated, which resulted in the mean of thederived distribution
being equal to the mean of the actualcombined motion’s distribution
(see Appendix II).
CONCLUSIONS
Considerable data are available regarding communicat-ing
information from a below-the-humerus incapacitatedsubject and a
microcomputer system programmed to hi-erarchically implement
intelligent control of a prostheticarm.
Learning pattern classification methods and
primitivedecomposition may yield fairly accurate commands on
gen-erating effects on this subject for a variety of loading
andvelocity conditions. Memory of the previous movementand visual
feedback may be utilised to interrupt a motionthat was incorrectly
understood. The features needed areonly ZCs and variances, as in
simple speech recognitionproblems (Rabiner and Schafer 1978). The
integrated mus-cular activity may be used to evaluate the load and
ve-locity parameters involved. Further research is aimed
atexperimental verification of the actual movement and con-trol of
the prosthetic arm. The final test of the methodwill be with actual
amputees from the local rehabilitationcenter.
REFERENCES
Alderson S. 1954. The electric arm. In Klopsteg PE, Wilson
PD,eds. Human Limbs and Their Substitutes. New York:McGraw-Hill,
Chap. 13. (reprinted by Hafner, 1969).
117C© Woodhead Publishing Ltd doi:10.1533/abbi.2005.0039 ABBI
2006 Vol. 3 No. 2
-
M. José H. Erazo Macias and S. Alejandro Vega
Bigland B, Lippold OCJ. 1954. The relation between force,
velocityand integrated electrical activity in human muscles. J
Physiol,123:214–24.
Fukunaga K. 1972. Introduction to Statistical Pattern
Recognition.New York: Academic Press.
Graupe D, Salahi J. 1979. Multifunctional Artificial Limb
Controlvia EMG Temporal Analysis—Background and AmputeeTests [PhD
thesis]. Department of Electrical Engineering,Illinois Institute of
Technology, Chicago.
Graupe D, Magnussen J, Bees A. 1978. Microprocessor system
formultifunctional control of upper-limb prostheses. IEEE
TransAutomat Contr, AC-23:538–44.
Lyman J, Freedy A, Zadaca H. 1974. Studies and development
ofheuristic end-point control for artificial upper limbs.
UCLABiotechnology Laboratory. Technical report 54.
Newman MA, Saridis GN. 1978. Development of prosthetic
arm.Purdue University, West Lafayette, IN. AARL Memo 30.
Rabiner LR, Schafer RN. 1978. Digital Processing of
SpeechSignals. Englewood Cliffs, NJ: Prentice-Hall.
Rimroth P, Newman MA, Saridis GN. 1978. A study
ofelectromyography for above-elbow amputation. PurdueUniversity,
West Lafayette, IN. AARL Memo 34.
Saridis GN, Newman MA. 1979. Upper limb EMG statisticalanalysis.
In Proceedings of MIDCON’79, Chicago, IL.
Saridis GN, Stephenou HE. 1977. A hierarchical approach to
thecontrol of a prosthetic arm. IEEE Trans Syst Man
Cybernet,SMC-7:407–20.
Wirta RW, Taylor DR, Finley FR. 1968. Engineering principles
inthe control of external power by myoelectric signals. Arch
PhysMed, 49:294–6.
118ABBI 2006 Vol. 3 No. 2 doi:10.1533/abbi.2005.0039 C© Woodhead
Publishing Ltd
-
EMG pattern analysis and classification for a robotic prosthetic
arm
APPENDIX I: THE STOCHASTIC APPROXIMATIONALGORITHM FOR
COMPUTATION OF DISCRIMINANTCOEFFICIENTS
The general linear-discriminant classifier for two classesω1 and
ω2, V can be written as
h = VT X + v0 >< 0 ⇒ X ∈{
ω1ω2
. (AI.1)
This can be rewritten as
h = WTY ><
0 (AI.2)
h =n∑
i=0wi yi , (AI.3)
where y0 ≡ 1, y1 = x1, that is, w0 = v0.Now, if we let Z = Y for
Y∈ ω1, and Z = −Y for Y ∈
ω2, then WT Z > 0 for correct classifications.The stochastic
approximation can be expressed as
Wl+1 = Wl − al[WTl Z1 − γ (Zl )
]Zl , (AI.4)
where γ (Z1) is the desired classifier output (>0) for Z1,and
a1 is required to satisfy
liml→∞
a1 = 0,∞∑
l=1a1 = ∞, and
∞∑l=1
a2/ < ∞
To guarantee convergence, for example, a1 = 1/l. In orderto
speed up the convergence, the number a1 = 1/k wasadvanced to 1/k +
1 only when the term [WTl Zl − γ (Z1)]changed sign.
APPENDIX II: SUPERPOSITION OF PRIMITIVES TOMATCH THE MEAN OF
MULTIPLE PRIMITIVE MOTIONS
Let the mean of the double-primitive motion being con-sidered
be[
σ 2B
σ 2T
].
Let the data of the component primitives be[σ 2Bikσ 2Tik
]and
[σBjk
σ 2T jk
], i = 1, 27, j = 1, 20.
The data are used with the equation[σ 2Bnew
σ 2Tnew
]= c1 ·
[σ 2Bik
σ 2Tik
]+ c2 ·
[σ 2Bjk
σ 2T jk
]
with coefficients c1 and c2 such that the mean of the(σ 2Bnew,
σ
2Tnew
)distribution is equal to[
σ 2B
σ 2T
].
c1 and c2 are easily found by
c1σ̄ 2Bi + c2σ̄ 2T j = 2σ̄ 2B. (AII.1)And
c1σ̄ 2Ti + c2σ̄ 2T j = 2σ̄ 2T . (AII.2)The solution of (AII.1)
and (AII.2) is
c2 =(2σ̄ 2T − 2σ̄ 2Bσ̄ 2Ti /σ̄ 2Bi
)(σ̄ 2T j − σ̄ 2Bj σ̄ 2Ti /σ̄ 2Bi
) (AII.3)c1 =
(2σ̄ 2B − c2σ̄ 2Bi
)σ̄ 2Bi
. (AII.4)
119C© Woodhead Publishing Ltd doi:10.1533/abbi.2005.0039 ABBI
2006 Vol. 3 No. 2
-
120ABBI 2006 Vol. 3 No. 2 C© Woodhead Publishing Ltd
-
International Journal of
AerospaceEngineeringHindawi Publishing
Corporationhttp://www.hindawi.com Volume 2010
RoboticsJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume
2014
Hindawi Publishing Corporationhttp://www.hindawi.com Volume
2014
Active and Passive Electronic Components
Control Scienceand Engineering
Journal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume
2014
International Journal of
RotatingMachinery
Hindawi Publishing Corporationhttp://www.hindawi.com Volume
2014
Hindawi Publishing Corporation http://www.hindawi.com
Journal ofEngineeringVolume 2014
Submit your manuscripts athttp://www.hindawi.com
VLSI Design
Hindawi Publishing Corporationhttp://www.hindawi.com Volume
2014
Hindawi Publishing Corporationhttp://www.hindawi.com Volume
2014
Shock and Vibration
Hindawi Publishing Corporationhttp://www.hindawi.com Volume
2014
Civil EngineeringAdvances in
Acoustics and VibrationAdvances in
Hindawi Publishing Corporationhttp://www.hindawi.com Volume
2014
Hindawi Publishing Corporationhttp://www.hindawi.com Volume
2014
Electrical and Computer Engineering
Journal of
Advances inOptoElectronics
Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
The Scientific World JournalHindawi Publishing Corporation
http://www.hindawi.com Volume 2014
SensorsJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume
2014
Modelling & Simulation in EngineeringHindawi Publishing
Corporation http://www.hindawi.com Volume 2014
Hindawi Publishing Corporationhttp://www.hindawi.com Volume
2014
Chemical EngineeringInternational Journal of Antennas and
Propagation
International Journal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume
2014
Hindawi Publishing Corporationhttp://www.hindawi.com Volume
2014
Navigation and Observation
International Journal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume
2014
DistributedSensor Networks
International Journal of