Professional Documents
Culture Documents
Simple and Fast Compensation of sEMG Interface Rotation For Robust Hand Motion Recognition
Simple and Fast Compensation of sEMG Interface Rotation For Robust Hand Motion Recognition
Ch 3 Ch 1 Ch 8
Abstract—Surface electromyography (sEMG) measurements Ch 2
Ch 4
have demonstrated the potential to recognize complex hand Ch 7
Ch 2
motions. Additionally, sEMG enables natural recognition without Ch 5 Ch 1 Rotation
disturbing movements, and thus, can be used in various fields Ch 6
Sources Initial Sources Rotated
Ch 3
such as teleoperation, assistant robots, and prosthetic hands. Ch 6 Ch 8 electrode electrode
placement Ch 4 Ch 5 placement
However, sEMG signals highly depend on electrode placements Ch 7
Initial configuration
Rotation Compensation
Source Source Source
Initial Rotated Compensated
electrode electrode electrode
Known parameters placement placement placement
Unknown parameters
Identified parameters
Source Electrode
sEMG signal
parameter parameter
compensation
identification identification
Compensated
Electrode parameters Electrode parameters Electrode parameters
sEMG signals
Several methods have been proposed for robust motion parameters are changed according to the interface rotation,
recognition against electrode displacement. Methods of en- the source parameters remain unchanged because they depend
suring robustness by analyzing high-density sEMG signals on anatomical factors. Therefore, once source parameters
[21], [22] have been proposed. However, these methods require are identified before rotation, the electrode parameters after
many sEMG channels. Additionally, Muceli et al. [23] demon- rotation can be identified via a simple procedure. The scheme
strated that activation signals extracted using non-negative for the model-based rotation compensation is shown in Fig. 2.
matrix factorization (NMF) [24] are robust against electrode Heuristically, it is sufficient to recognize the rotation by per-
displacement. forming only a single motion for approximately two seconds.
However, these conventional methods were demonstrated The motion for compensation is dependent on the anatomical
for small displacements of only about 1 or 2 cm. Although factors of the user, which is selected as the ulnar deviation for
in the case of a prosthetic hand, compensation of this level of able-bodied subjects in this study.
displacement might be sufficient due to a user-specific socket After rotation, when the electrode parameters are identified,
[19], the user should carefully wear the interface for consistent the sEMG signals can be converted into those with respect to
performance. In the case of a teleoperation, the effect of the initial electrode parameters. Therefore, this compensation
electrode displacement is dominant. enables sEMG-based motion recognition without additional
In this study, simple and fast rotation compensation for training to determine the relationship between motions and
a wide range is proposed. The overview of the motion sEMG signals after rotation.
recognition with compensation is described in Fig. 1. This Additionally, the compensation process is separated from
research is based on our previous work on sEMG signal the motion recognition process. Therefore, any recognition
decomposition as muscle activation sources [25]. We assumed method, such as linear discriminant analysis classifier [26],
that the contraction of a muscle involves activation sources. support vector machine classifier [27], or deep learning [28],
The sEMG signal measured from an electrode is a combina- can be integrated.
tion of source activations attenuated by the conductivity of In conclusion, the proposed method provides robust sEMG-
each source and the distance between each source and the based motion recognition, which is useful in assistant robots
electrode. If the electrode parameters, including the three- and prosthetic hands that require long-term and repeated usage
dimensional placement, size, and shape, are known, then the in everyday activities. The users do not require additional
source parameters, including the three-dimensional placement, motion training each time they put on and take off the
conductivity, and source amplitude, can be identified using the interface.
measured root-mean-square (RMS) sEMG signals. The remainder of the paper describes the details of the
In our previous study, rotation of the sEMG interface model and compensation method. Then, hand motion recog-
was recognized by comparing the source positions identified nition will be described to demonstrate the rotation compen-
before and after rotation. However, this is time-consuming and sation.
impractical.
In this study, our proposed method is improved for simple II. M ETHODS
and fast compensation. If either of the electrode parameters
or source parameters are known, the remaining parameters A. Subjects
can be identified using sEMG signals. Electrode parameters Ten able-bodied subjects and one transradial amputee partic-
are known before the interface rotation; thus, the source pa- ipated in the study. The conducted experiments were approved
rameters can be identified. Theoretically, even if the electrode by the institutional review board (PIRB-2018-E079) of the
3
Pohang University of Science and Technology (POSTECH), signals measured at the two active electrodes relative to that
Pohang, Korea. of the reference electrode. Active electrodes are placed on the
skin surface near the target muscle, and a common reference
B. Experimental Setup and Data Acquisition electrode is placed on the bony region or between the two
active electrodes.
For the able-bodied subjects, the motions selected included The proposed model assumes that the signal measured
wrist extension/flexion, hand close/open, and ulnar deviation. from each electrode is a summation of the attenuated source
A transradial amputee has a very limited muscle contraction amplitudes. The corresponding relationship is expressed as
pattern, and thus, was asked to contract muscles while intend- follows:
ing four motions (wrist extension/flexion, ulnar deviation, and
hand close). X
An MYO armband (Thalmic Labs Inc., Kitchener, Canada) Vit = dit k Vsk ,
was used to measure the sEMG signals from the forearm. The k
interface consists of eight bipolar channels, which are arranged i = 1, ..., m, k = 1, ...n, t = 1, 2, r (1)
at constant intervals. Each channel consists of two active where Vit denotes the summation of the attenuated source
electrodes and one reference electrode between the active amplitudes measured at electrode t consisting the ith channel;
electrodes. Additionally, an inertial measurement unit (IMU) t denotes the index of the electrodes, where 1 and 2 denote
was built into the MYO armband. sEMG signals and IMU two active electrodes and r denotes the reference electrode; m
signals were simultaneously measured at sampling frequencies denotes the number of overall channels; n denotes the number
of 200 Hz and 50 Hz, respectively. of sources; Vsk denotes the kth non-negative source amplitude;
and dit k denotes the function value that decreases with the
C. Experimental Procedures and Signal Analysis conductivities and distance between the source position and
The subjects were comfortably seated on a chair and in- electrode position as follows:
structed to perform muscle contractions following a specific s
motion displayed on a monitor. For each motion, the subjects (xit − xk )2 (yit − yk )2 (zi − zk )2
were asked to maintain the pose with moderate and constant dit k = 1/ + + t (2)
σxk σyk σzk
force according to their convenience.
Each motion was contracted for 3 s and repeated three times. where xit , yit , and zit denote the three-dimensional po-
Between contractions, the subjects rested for at least 1 s. In sition of the electrode, σxk , σyk , and σzk denote the three-
the first contraction, the middle two seconds of contractions dimensional conductivity of the kth source, and xk , yk , and
were used for model identification, reconstruction, and motion zk denote the three-dimensional position of the kth source.
training. The remaining contractions were used for the tests. For the bipolar configuration of the sEMG channel, the
After all motions were performed, the subjects removed the measured sEMG signal can be represented as follows:
interface, rotated it, and put it back on again.
The non-negative sEMG signals are required to identify
X
Vi = ||di1 k − dir k | − |di2 k − dir k ||Vsk
the model parameters and to compensate for the rotation. k
Therefore, both parameter identification and motion training X
= Dik Vsk (3)
were based on the RMS sEMG signals. The signals were
k
extracted from a window of 200 ms every 100 ms. From
the contraction of 2 s, 19 samples for each motion can be where Vi denotes the RMS sEMG signal measured from the
extracted. A large number of samples might provide a more ith channel.
convincing performance; however, the number of samples The electrode parameters, including the electrode placement
used in this study might guarantee at least the minimum and the channel configuration, are known according to the
performance. sEMG interface specification. Therefore, the source parameters
The relationship between the RMS sEMG signal and motion for the motion can be identified using the sEMG signals
was trained via a shallow neural network that contained 10 measured during static muscle contractions.
neurons. The network was trained using the scaled conjugate The source parameters for each motion were independently
gradient backpropagation method by using MATLAB 2017a identified. There were 19 RMS sEMG signal samples from a
(MathWorks, Inc., Natick, MA, USA). static contraction of 2 s, 8 equations for a number of channels,
To investigate statistical significance, a one-sample and and 7 unknown source parameters (x, y, z, σx , σy , σz , Vs ) per
paired-sample t-test, and a two-sample t-test were performed each source; all 19 samples were used to find best-fit pa-
according to the data. All analyses were performed using rameters. After identification, the first 6 source parameters
MATLAB 2017a. were used for further analysis, and the last parameter, Vs ,
only represented the identified source amplitude during the
model identification. To guarantee a unique solution, a number
D. Model Configuration of channels should be larger than the unknown parameters.
Generally, an sEMG channel consists as a bipolar configu- To identify the parameters, the lsqcurvefit function in the
ration in which the sEMG signal represents the difference in MATLAB 2017a Optimization Toolbox was used.
4
∗ ∗
)VS0
PNp
p dpit k VM = f (PS , CS , PM
dit k = (6)
Np = f (PS , CS , T PM )VS0 (8)
where dpit k denotes the function value between the kth where VM ∗
and PM∗
denote the measured sEMG signals and
source and the pth point, which is comprised of electrode t electrode positions after rotation, respectively, and T denotes
∗
consisting of the ith channel, and Np denotes the number of the rotation matrix considering only roll rotation. PM can be
points selected for the electrode construction. represented as T PM because the electrodes comprising the
The target limb, which is the forearm, was considered to interface are simultaneously rotated. Therefore, the rotation
be a rigid cylinder. The forearm thickness varies depending can be identified by estimating T .
5
0.8
(d) Hand close (e) Hand open (f) Ulnar deviation Fig. 6. Signal reconstruction using two sources for hand close from a
representative subject: individual source positions (blue and red points) and
Fig. 4. Normalized sEMG signal distribution and its reconstruction from a reconstructed signal distributions according to each source (blue and red lines).
representative subject: The measured signal distribution for each electrode Additionally, the merged reconstructed signal distribution (green line) and
arrangement (red line), source position (black point), and reconstructed signal measured signal distribution (black line) are described. This indicated that
distribution (blue line). two sources are sufficient for reconstructing the signal.
6
100
With compensation 350
80 Without compensation
300
150
20
100
45°
0 50
Small Large Measured
rotation rotation Recognized
0
1 2 3 4 5 6 7 8 9
Fig. 10. Motion recognition accuracy according to level of rotation. There was Session
no significant difference (p = 0.17) between the small and large rotations.
Fig. 11. Measured and recognized angular differences based on the first
session from a representative subject. The shape of the actual forearm is not
in Fig. 9. The case in which compensation was not performed cylindrical. Therefore, the angular increment in each trial was not exact 45◦ .
Furthermore, there were differences between the measured and recognized
demonstrated poor recognition performance (37.5%). How- angular differences due to the model assumption.
ever, the proposed model-based compensation provided con-
sistent motion recognition performance after rotation (99.4%). 360
315
100
90
50
40
30
20
(a) Recognition for session 2.
10
0
0 50 100 150 200 250 300
Time [min]
Fig. 14. Motion recognition accuracy with and without compensation for
long-term and repeated usage from a representative subject. Rotation com-
pensation produced a consistently high performance.
Session 2
(b) Recognition for session 3.
Difference of 𝟒𝟓° Fig. 17. Motion recognition with and without compensation. The performance
Session 1 was significantly increased from 51.1% to 79.5% and from 41.6% to 83.6%
for sessions 2 and 3, respectively.
Wrist Wrist Ulnar Hand
extensions flexions deviations close
Session 3
F. Validation with a Transradial Amputee
Wrist Wrist Ulnar Hand
extensions flexions deviations close
The measured sEMG signals during muscle contractions are
Difference of 𝟏𝟐°
described in Fig. 15. In each session, the subject was asked
Wrist Wrist Ulnar Hand
to take off the interface, randomly rotate it, and put it on
extensions flexions deviations close
again. Therefore, very different sEMG patterns were obtained
Fig. 15. Measured sEMG signals for different sessions. Due to the rotation, depending on the session.
different patterns were obtained. The average patterns during static muscle contractions
are described in Fig. 16. Unlike the able-bodied subjects,
wrist extension produced a consistent pattern. Therefore, wrist
extension-based compensation was performed. This indicates
that the optimal motion for rotation compensation depends on
the muscle properties of the user.
From the first session, model parameters were identified and
the relationship between the motion and sEMG signal was
trained. The recognition with and without compensation are
(a) Wrist extension (b) Wrist flexion presented in Fig. 17. The compensation significantly increased
the recognition accuracy from 51.1% to 79.5% and from
41.6% to 83.6% for sessions 2 and 3, respectively.
IV. D ISCUSSION
A. Summary of Experiments
(c) Ulnar deviation (d) Hand close With known electrode parameters, the source parameters can
Session 1 Session 2 Session 3 be identified for an individual motion. Each source contains
information about the anatomical factors that are related to
Fig. 16. Normalized RMS sEMG signal distributions from all sessions. Wrist
extension produced a consistent distribution compared to other motions.
the corresponding motion. Once the source parameters are
identified, the rotation can be compensated for by performing
a single motion for two seconds.
To demonstrate both rotation recognition and its compen-
pensation for different sessions with a representative subject sation, motion recognition was conducted by rotating the
are shown in Fig. 14. The results confirmed that the proposed sEMG interface. Ten able-bodied subjects and one transradial
model-based compensation guarantees long-term and repeated amputee participated in the experiments. The results indicated
usage without additional training (p < 0.005). The overall that the proposed model guarantees the recognition of any
accuracy was increased from 40.8% to 97.1%. rotation.
9
B. Muscle Property Variation degradation for the actual angle-based compensation was
In this paper, ulnar deviation was consistently selected for caused by the difference between the actual forearm and the
the able-bodied subjects to compensate rotation. Additionally, model assumption. As mentioned above, the current version of
the number of sources for each motion was heuristically the proposed model does not accurately represent the elliptical
chosen by observing the sEMG signal distribution and recon- shape of an actual forearm. Furthermore, skin deformation can
struction ratio. However, the sEMG spatial patterns depend occur due to the muscle contraction and the electrode pressing
on the involved anatomical factors. There might be many against the skin. Enhanced compensation based only on the
variations depending on the subject. Especially, that might be angle measured using an IMU could be achieved using a more
dominant for a person who has an amputation or a muscular accurate model that considers the actual forearm shape and
disorder. As an example, in the case of the transradial amputee skin deformation.
who participated in this experiment, wrist extension was the Additionally, during donning and doffing, not only rotation
best motion for compensation. Therefore, to achieve accurate but also translation can occur. By including the translation
compensation, optimal selection of subject-specific parameters matrix in (8), the translation could be compensated. However,
is required by analyzing the muscle properties. a band-type interface has a longitudinal limitation; thus, a large
translation cannot be compensated for. A sleeve-type high-
C. Inter-Subject Variability density sEMG interface could help to provide a more reliable
compensation that considers translation due to a high spatial
The proposed model facilitates long-term and repeated
resolution and a large skin contact area.
usage by identifying the parameters for the muscle activa-
One of the advantages of the proposed compensation is that
tion source that depend on the anatomical factors. However,
the process is separate from the motion recognition process.
anatomical factors vary from user to user. Thus, the source
Therefore, any motion recognition method can be integrated.
parameters determined for one user cannot be simply applied
In this study, a simple shallow neural network was integrated.
to a different user. We conducted several experiments in which
Thus, the motion recognition performance can be improved
the model parameters obtained for one user were applied
using a state-of-the-art algorithm.
to another user (not described in this paper). However, the
performance was inconsistent and even significantly degraded
in several cases. Therefore, the proposed model-based com- E. Limitation of Source Extraction
pensation is suitable for applications that are for only a single The proposed model requires non-negative sEMG signals
user. to identify the model parameters. Therefore, other types
of features widely used in sEMG signal analysis, such as
D. Scope for Improvement wavelength, slope sign change, and autoregressive coefficients,
Fig. 18 describes a comparison of the performance between cannot be applied. However, many deep learning-based motion
ulnar deviation-based and actual angle-based compensations estimations have been proposed. By integrating a state-of-
for a representative subject. The actual angle was measured the-art motion-estimating method, the proposed compensation
using IMU. The accuracy of actual angle-based compensation might guarantee high performance using only RMS features.
(94.8%) was slightly lower (p = 0.14) than that of ul-
nar deviation-based compensation (98.6%). The performance F. sEMG Signal Reconstruction
The proposed model-based source extraction is a dimension
100
reduction technique. In this study, only eight channels were
used due to the specification of the MYO armband. Therefore,
95 there might be no significant differences between source-
Accuracy of recognition [%]
can be accomplished by performing only a single motion. [17] H. J. Hermens, B. Freriks, C. Disselhorst-Klug, and G. Rau, “Devel-
The proposed method can be particularly useful in assistant opment of recommendations for semg sensors and sensor placement
procedures,” Journal of electromyography and Kinesiology, vol. 10,
robots and prosthetic hands. These applications require long- no. 5, pp. 361–374, 2000.
term and repeated usage in everyday activities. By applying [18] L. Hargrove, K. Englehart, and B. Hudgins, “The effect of electrode
rotation compensation, their reliability and functionality might displacements on pattern recognition based myoelectric control,” in
Engineering in Medicine and Biology Society, 2006. EMBS’06. 28th
be improved. Annual International Conference of the IEEE. IEEE, 2006, pp. 2203–
2206.
[19] A. J. Young, L. J. Hargrove, and T. A. Kuiken, “The effects of electrode
size and orientation on the sensitivity of myoelectric pattern recognition
R EFERENCES systems to electrode shift,” IEEE Transactions on Biomedical Engineer-
ing, vol. 58, no. 9, pp. 2537–2544, 2011.
[1] M. T. Wolf, C. Assad, M. T. Vernacchia, J. Fromm, and H. L. Jethani, [20] C. Pylatiuk, S. Schulz, and L. Döderlein, “Results of an internet
“Gesture-based robot control with variable autonomy from the jpl survey of myoelectric prosthetic hand users,” Prosthetics and orthotics
biosleeve,” in Robotics and Automation (ICRA), 2013 IEEE International international, vol. 31, no. 4, pp. 362–370, 2007.
Conference on. IEEE, 2013, pp. 1160–1165. [21] A. Boschmann and M. Platzner, “Towards robust hd emg pattern
[2] J. Vogel, C. Castellini, and P. van der Smagt, “Emg-based teleoperation recognition: Reducing electrode displacement effect using structural
and manipulation with the dlr lwr-iii,” in Intelligent Robots and Systems similarity,” in Engineering in Medicine and Biology Society (EMBC),
(IROS), 2011 IEEE/RSJ International Conference on. IEEE, 2011, pp. 2014 36th Annual International Conference of the IEEE. IEEE, 2014,
672–678. pp. 4547–4550.
[3] Z. O. Khokhar, Z. G. Xiao, and C. Menon, “Surface emg pattern [22] L. Pan, D. Zhang, N. Jiang, X. Sheng, and X. Zhu, “Improving
recognition for real-time control of a wrist exoskeleton,” Biomedical robustness against electrode shift of high density emg for myoelectric
engineering online, vol. 9, no. 1, p. 41, 2010. control through common spatial patterns,” Journal of neuroengineering
[4] Y. Yun, S. Dancausse, P. Esmatloo, A. Serrato, C. A. Merring, P. Agar- and rehabilitation, vol. 12, no. 1, p. 110, 2015.
wal, and A. D. Deshpande, “Maestro: an emg-driven assistive hand [23] S. Muceli, N. Jiang, and D. Farina, “Extracting signals robust to
exoskeleton for spinal cord injury patients,” in Robotics and Automation electrode number and shift for online simultaneous and proportional
(ICRA), 2017 IEEE International Conference on. IEEE, 2017, pp. myoelectric control by factorization algorithms,” IEEE Transactions on
2904–2910. Neural Systems and Rehabilitation Engineering, vol. 22, no. 3, pp. 623–
[5] D. Kim, O. Hilliges, S. Izadi, A. D. Butler, J. Chen, I. Oikonomidis, and 633, 2014.
P. Olivier, “Digits: freehand 3d interactions anywhere using a wrist-worn [24] D. D. Lee and H. S. Seung, “Algorithms for non-negative matrix
gloveless sensor,” in Proceedings of the 25th annual ACM symposium factorization,” in Advances in neural information processing systems,
on User interface software and technology. ACM, 2012, pp. 167–176. 2001, pp. 556–562.
[6] G. Marin, F. Dominio, and P. Zanuttigh, “Hand gesture recognition with [25] M. Kim and W. K. Chung, “Muscle activation source model-based semg
leap motion and kinect devices,” in Image Processing (ICIP), 2014 IEEE signal decomposition and recognition of interface rotation,” in Intelli-
International Conference on. IEEE, 2014, pp. 1565–1569. gent Robots and Systems, 2018. IROS 2018. IEEE/RSJ International
[7] M. Amjadi, A. Pichitpajongkit, S. Lee, S. Ryu, and I. Park, “Highly Conference on. IEEE, 2018, pp. 2780–2786.
stretchable and sensitive strain sensor based on silver nanowire– [26] H. Zhang, Y. Zhao, F. Yao, L. Xu, P. Shang, and G. Li, “An adaptation
elastomer nanocomposite,” ACS nano, vol. 8, no. 5, pp. 5154–5163, strategy of using lda classifier for emg pattern recognition,” in Engi-
2014. neering in Medicine and Biology Society (EMBC), 2013 35th Annual
International Conference of the IEEE. IEEE, 2013, pp. 4267–4270.
[8] J. Kim, M. Lee, H. J. Shim, R. Ghaffari, H. R. Cho, D. Son, Y. H.
[27] M. Yoshikawa, M. Mikawa, and K. Tanaka, “A myoelectric interface for
Jung, M. Soh, C. Choi, S. Jung et al., “Stretchable silicon nanoribbon
robotic hand control using support vector machine,” in Intelligent Robots
electronics for skin prosthesis,” Nature communications, vol. 5, p. 5747,
and Systems, 2007. IROS 2007. IEEE/RSJ International Conference on.
2014.
IEEE, 2007, pp. 2723–2728.
[9] J. G. Ngeo, T. Tamei, and T. Shibata, “Continuous and simultaneous
[28] M. Atzori, M. Cognolato, and H. Müller, “Deep learning with convo-
estimation of finger kinematics using inputs from an emg-to-muscle ac-
lutional neural networks applied to electromyography data: a resource
tivation model,” Journal of neuroengineering and rehabilitation, vol. 11,
for the classification of movements for prosthetic hands,” Frontiers in
no. 1, p. 122, 2014.
neurorobotics, vol. 10, p. 9, 2016.
[10] M. V. Liarokapis, P. K. Artemiadis, P. T. Katsiaris, K. J. Kyriakopou- [29] T. O. Kvålseth, “Cautionary note about r 2,” The American Statistician,
los, and E. S. Manolakos, “Learning human reach-to-grasp strategies: vol. 39, no. 4, pp. 279–285, 1985.
Towards emg-based control of robotic arm-hand systems,” in Robotics [30] M. Gazzoni, N. Celadon, D. Mastrapasqua, M. Paleari, V. Margaria, and
and Automation (ICRA), 2012 IEEE International Conference on. IEEE, P. Ariano, “Quantifying forearm muscle activity during wrist and finger
2012, pp. 2287–2292. movements by means of multi-channel electromyography,” PloS one,
[11] W. Geng, Y. Du, W. Jin, W. Wei, Y. Hu, and J. Li, “Gesture recognition vol. 9, no. 10, p. e109943, 2014.
by instantaneous surface emg images,” Scientific reports, vol. 6, p. [31] M. Rojas-Martı́nez, M. A. Mañanas, and J. F. Alonso, “High-density
36571, 2016. surface emg maps from upper-arm and forearm muscles,” Journal of
[12] M. Carozza, G. Cappiello, G. Stellin, F. Zaccone, F. Vecchi, S. Micera, neuroengineering and rehabilitation, vol. 9, no. 1, p. 85, 2012.
and P. Dario, “On the development of a novel adaptive prosthetic
hand with compliant joints: experimental platform and emg control,”
in Intelligent Robots and Systems, 2005.(IROS 2005). 2005 IEEE/RSJ
International Conference on. IEEE, 2005, pp. 1271–1276.
[13] A. Fougner, Ø. Stavdahl, P. J. Kyberd, Y. G. Losier, and P. A. Parker,
“Control of upper limb prostheses: terminology and proportional myo-
electric control—a review,” IEEE Transactions on neural systems and
rehabilitation engineering, vol. 20, no. 5, pp. 663–677, 2012.
[14] A. D. Roche, H. Rehbaum, D. Farina, and O. C. Aszmann, “Prosthetic
myoelectric control strategies: a clinical perspective,” Current Surgery
Reports, vol. 2, no. 3, p. 44, 2014.
[15] A. J. Young, L. J. Hargrove, and T. A. Kuiken, “Improving myo-
electric pattern recognition robustness to electrode shift by changing
interelectrode distance and electrode configuration,” IEEE Transactions
on Biomedical Engineering, vol. 59, no. 3, pp. 645–652, 2012.
[16] L. Hargrove, K. Englehart, and B. Hudgins, “A training strategy to re-
duce classification degradation due to electrode displacements in pattern
recognition based myoelectric control,” Biomedical signal processing
and control, vol. 3, no. 2, pp. 175–180, 2008.