Available via license: CC BY 4.0
Content may be subject to copyright.
Abstract— The use of deep neural networks in
electromyogram (EMG) based prostheses control provides a
promising alternative to the hand-crafted features by
automatically learning muscle activation patterns from the
EMG signals. Meanwhile, the use of raw EMG signals as input
to convolution neural networks (CNN) offers a simple, fast, and
ideal scheme for effective control of prostheses. Therefore, this
study investigates the relationship between window length and
overlap, which may influence the generation of robust raw
EMG 2-dimensional (2D) signals for application in CNN. And a
rule of thumb for a proper combination of these parameters that
could guarantee optimal network performance was derived.
Moreover, we investigate the relationship between the CNN
receptive window size and the raw EMG signal size.
Experimental results show that the performance of the CNN
increases with the increase in overlap within the generated
signals, with the highest improvement of 9.49% accuracy and
23.33% F1-score realized when the overlap is 75% of the
window length. Similarly, the network performance increases
with the increase in receptive window (kernel) size. Findings
from this study suggest that a combination of 75% overlap in 2D
EMG signals and wider network kernels may provide ideal
motor intents classification for adequate EMG-CNN based
prostheses control scheme.
Keyword: Convolution Neural Networks, EMG Pattern
Recognition, Prostheses, Window Size, Window Overlap
I. INTRODUCTION
The loss of upper limb leads to difficulty in performing
simple and complex daily activities which can cause
emotional and psychological problems that decrease the
quality of life of amputees. In order to restore limb
functionality in amputees, researchers have focused on
developing prostheses based on pattern recognition of motion
intents from surface electromyogram (sEMG) signals. The
sEMG signals present the sum of underlying motor action
This work was supported in part by the National Natural Science
Foundation of China (#81927804, #82050410452, #62150410439),
Shenzhen Governmental Basic Research Grant
(#JCYJ20180507182508857), Shenzhen Institute of Artificial Intelligence
and Robotics for Society, Shenzhen Governmental Collaborative Innovation
Program (#SGLH20180625142402055).
F. Kulwa, O.W. Samuel, M.G. Asogbon, and G. Li are with the CAS Key
Laboratory of Human-Machine Intelligence-Synergy Systems, Shenzhen
Institute of Advanced Technology (SIAT), Chinese Academy of Sciences
(CAS), Shenzhen, Guangdong 518055, China. (Correspondence: Dr.
Oluwarotimi Williams Samuel, e-mail: samuel@siat.ac.cn and Dr. Guanglin
Li, e-mail: gl.li@siat.ac.cn).
F. Kulwa is also with the Shenzhen College of Advanced Technology,
University of Chinese Academy of Sciences, Shenzhen, Guangdong 518055,
China.
O.O. Obe is with the Department of Computer Science, Federal
University of Technology, Akure, Nigeria.
potential produced from the limb muscle contraction [1] that
could serve as a potential control source for intelligent
prostheses. The stochastic and non-linearity nature of sEMG
have motivated the use of engineered features towards
adequate characterization of hidden patterns associated with
the motion intents. In that respect, Li et al. used a combination
of time domain features to train a regulator-based classifier in
mitigating the effects of muscle contraction force variation
when decoding motion intents from sEMG signals [2].
Likewise, to resolve the coexisting impact of subject mobility
and muscle contraction force variation on motion intents from
sEMG, Asogbon et al. proposed five invariant time-domain
features [3]. To improve the recognition of limb movements,
both time and spatial characteristics of the sEMG signals,
Samuel et al. introduced engineered spatio-temporal
descriptors which are robust to additive Gaussian noise [4].
Similar features were applied in [5] for optimal decoding of
movement intents in upper limb of the stroke survivors. While
these studies considered instantaneous (raw) sEMG signals
not suitable for prostheses control [6] and required feature
engineering, deep learning techniques have shown that feature
learning methods can be applied directly to such signals [7].
The sEMG-based upper limb gesture classification task
can be modeled as a two-dimensional (2D) signal recognition
problem using convolution neural networks (CNN), where the
input signal has a size of K ×T× 1 (width x length x depth).
Various methods have been deployed for constructing 2D
signals such as; construction of 2D signal from HD-sEMG
electrode array shape where the 2D depicts the array
placement during data collection [8], the use of spectrograms
generated from the short-time Fourier transforms of sEMG
segments [9], and generation of 2D signals from chunks of
sEMG signals using overlapping time windows (where the
width matches the number of electrodes and the length is equal
to the window size). Although recent research suggests that
focusing on the temporal information of the signals can help
CNN extract long and short-term patterns [8], [10], [11], they
are well-suited for extraction of spatial features. With regards
to sparse EMG collection configurations, the studies show that
the choice on the size of the analysis window and overlap
(during creation of 2D signals) has been made empirically to
trade-off the classification performance and computation time
[12] and there is no rule of thumb to follow. For example, [13]
used window length of 150ms, [14] used 150ms window
length and overlap of 90ms, [15] applies 100ms and increment
of 10ms which is equivalent to 200 samples per window and
overlap of 180 samples, [16] used window of 150 data points
and overlap of 30 data points, and [17] applied window of 300
ms and stride of 10 ms.
Although there are some studies that have analyzed the
effects of window parameters [8, 18-21], they specifically
Frank Kulwa,
Oluwarotimi Williams Samuel
*
,
Senior Member
,
IEEE
,
Mojisola Grace Asogbon
,
Member, IEEE, Olumide Olayinka Obe, and Guanglin Li*, Senior Member, IEEE
Analyzing the Impact of Varied Window Hyper-parameters on Deep
CNN for sEMG based Motion Intent Classification
focused on the extraction of engineered features and only
utilized conventional machine learning methods without
considering deep neural networks. To the best of the authors’
knowledge, no study has analyzed the effects of window
parameters (window length and overlap) on 2D EMG signals
for deep neural networks particularly CNNs to date. Thus, this
constitutes a research gap that the current study seeks to
address. Hence, this paper investigates the relationship
between window parameters (such as window size and
increment) on the generation of 2D raw EMG signals towards
yielding optimal classification results for the deep neural
networks. Moreover, we examine the impacts of receptive
window sizes (kernels) for varied configurations of CNNs and
how they impact the feature learning of sEMG signal.
II.METHODS
A. Subjects
A total of four transhumeral amputee subjects took part in
the sEMG data collection experiment in this research. Their
ages ranged from 35 to 49, and two of them were dominant in
their right hand. The subjects were briefed about the study's
goal and purpose prior to data collection, and they all agreed to
participate. Following that, the subjects signed a written
informed consent form and agreed to their data being
published for scientific and educational purposes. The
experimental protocol of the study was approved by the
Institutional Review Board of the Shenzhen Institutes of
Advanced Technology, Chinese Academy of Sciences, China.
B. Equipment setup and data acquisition
Five motion classes of wrist supination (WS), hand open
(HO), wrist pronation (WP), hand close (HC), and no
movement (NM) were investigated. A computer screen was
set up in front of the participants during the experiment. A
motion picture would be shown for each of the motion classes
engaged in the study. When a picture of the target movement
is shown on the screen, the subjects perform it. And when it
disappeared, the subject stopped doing the movement. The
five motion classes were performed at random with a
comfortable force level decided by the participants to avoid
muscular and mental tiredness.
Each motion lasted 5 seconds, with a 5-second rest period
(NM) in between two adjacent motions. Each subject had five
data recording sessions, each of which comprised of 40 active
motions (each of HO, HC, WP, and WS appeared 10 times at
random) and 40 repetitions of NM. A high-density sEMG
system (REFA 128, TMS International, the Netherlands) was
used to acquire the sEMG signals, with 32 monopolar
electrodes implanted on the skin surface of the residual arm
for each amputee.
C. Data preprocessing
The sEMG data corresponding to each of the indicated
classes of limb movement was preprocessed and evaluated
using the MATLAB programming tool. To be more specific,
the data was sampled at a rate of 1024Hz, with a 50Hz notch
filter used to reduce power-line interference. Variations in
EMG signals caused by motion artifacts and electrode
displacement were reduced using a band pass filter with
cut-off frequencies of 10 and 500 Hz.
D. Generation of 2D signal
In order to simplify the network’s learning process during
training, the data was first normalized by Z-score
normalization technique with unit standard deviation and zero
mean. During pilot experiment, Z-score and Mean-Max
normalization methods were evaluated on two types of
dataset configurations. Firstly, the dataset was normalized
across a particular class of movement. That is, the data
corresponding to HC movement of a particular subject is
normalized as a whole. Secondly, the normalization was done
segment-wise (2D signals). Finally, we considered using
Z-score on class-wise normalization due to the stable and
high results obtained for this method.
Then the 2D signals were generated using segmentation
window of size K ×T× 1, where K is the number of electrodes
which is 32 for this study and T is the window length. To
investigate the effects of window lengths and overlap on the
robustness of the generated 2D signals, three different
window lengths were examined, window length 125ms (data
points), 150ms and 175ms. These values were selected
because they have been suggested to give optimal results and
used by number of studies [20]. To set a rule of thumb we
vary the overlaps in percentage of the window length.
Therefore, for each window length the overlaps are varied at
75% of T, 50%, 25% and 0% of T (No overlap).
E. Convolutional neural network (CNN) architecture
We propose a model which consists of four feature learning
blocks, feature reduction layer and one feature selection layer
as shown in Figure 1.
The first block consists of convolution layer with 32 nodes,
ReLu activation function which is followed by 10% dropout
layer so as to reduce the overfitting and improve
generalization [22].
To allow kernel to cover more space within the feature
maps we keep the spatial dimension of the signal features the
same before and after each convolution by applying padding.
The other three blocks have similar layers packing with
only difference in the number of nodes, which are 32, 64 and
64, respectively. Each block consists of a convolution layer
followed by ReLu and dropout of 10%. To make the network
more robust to changes of the input signals such as translation
and shifts, we apply max pooling layer of size 2x2 at the end
of each block. After feature learning by the convolution
blocks, the feature maps are passed through global
maxpooling for dimension reduction. We use global max
maxpooling because it is not prone to over-fitting and
improves the model generalization.
The bottom part of the network consists of a dense layer of
128 nodes for more feature selection that is relevant to
classes. Lastly, the classification of the 5 limb motion intents
using a softmax layer.
To investigate the impacts of receptive window sizes
(kernels) of the network while learning general or fine
features within the EMG signal, we have designed three
networks of similar architecture but with different kernel
sizes. For each network the same size of convolution kernel is
applied in all layers. We use 3x3, 5x5, and 7x7 sizes because
they are the fundamental kernel sizes which have been
applied in many studies and give optimal network
performances [9, 13–15].
F. Experimental setup
All the models were trained using the Adam optimizer with
the learning rate of 0.0001 as it showed optimal network
performance during our preliminary experiments compared to
SGD optimizer. Moreover, we used a categorical cross
entropy as the loss function due to its better performance in
multiclass problems [23].
Observing the training and validation curves of the pilot
experiment in Figure 2, the curves tend to flatten at 30 epochs.
Thus, we used 35 epochs in all training experiments to allow
more confident results. The outcomes of all models are
presented using the fundamental metrics for classification
problems, accuracy and F1-score.
Figure 2: Training and validation curves to test the possible number of
epochs
III. RESULTS
A. Evaluating the impact of window overlap and size on the
robustness of the generated 2D EMG signal.
To analyze the effect of data overlap with respect to
segmentation (analysis) window, the window length (T) was
kept constant while varying the overlap per experiment.
Then, the generated 2D signals for each variation were used
to train the three different CNNs (with kernels sizes of 3, 5,
and 7) across each subject’s data. As suggested by previous
studies [16], 70% of the data was used for training the
networks and the remaining 30% for testing. The overall
average classification results obtained across all subjects were
shown in Figure 3. Besides, we only presented results for
kernel sizes of 3 and 5.
By carefully observing the results in Figure 3, it can be
seen that the classification performance increases with the
increase in percentage of overlap regardless of the 2D signal
size (window length (T)) in all proposed deep neural
networks. For instance, in Figure 3(b) when the network with
kernel =5 and window length of 150 is used, the accuracy of
97.80% is achieved at the overlap of 0.75T and 89.04%
accuracy at the overlap of 0T. Similarly, F1-score of 94.46%
and 72.80% are observed at overlaps of 0.75T and 0T,
respectively. This indicates an increment of approximately
8.76% accuracy and 21.66% F1-score. Meanwhile, the highest
improvement of 9.49% accuracy and 23.33% F1-score is
achieved when kernel=5 and T=175 in Figure 3(c). The results
above show that the 2D signals with an overlap of 75% can
achieve acceptable results for classification of upper limb
motion intents based on sEMG signals regardless of the
window length.
Figure 1. The architecture of the proposed CNN model. Where the output size of Block 1 and 2 is 32*T*32, Block 3 is 16*T/2 * 64,
and Block 4 is 8*T/4 *64. T represents the window length. The output classes at the final layer are no movement (NM), wrist
supination (WS), wrist pronation (WP), hand open (HO) and hand close (HC).
Figure 3. Limb motion intent classification accuracy (Blue bars) and F1-score
(Red bars) across two networks (kernel=3, 5) at window size (T), the overlap
is applied at 75% of T (0.75T), 50% (0.5T), 25% (0.25T) and No overlap (0T).
(a) When the 2D signals are of size 32x125 (b) the signals are of size 32x150
and (c) when the size is 32 x 175.
B.Evaluating the effects of network receptive fields against
the generated 2D EMG signal sizes
In this section the performance of the proposed networks
(kernel of 3, 5, and 7) are examined on different dimensions
of the 2D signals (at window length (T) of 125, 150, and 175).
Because of the relatively higher results obtained using signals
with overlap of 75%, only this overlap was considered. Thus,
the classification results based on F1-score metric for the
three CNNs at different window sizes are presented in Figure
5.
Observing the results in Figure 5, it can be seen that the
increase in kernel size led to increase in performance of the
networks. For example, at signal window length of 125 the
three networks with corresponding kernels of 3, 5, and 7
achieved 89.49%, 95.30% and 95.93% F1-score,
respectively. And this resulted to an increase of about 6.44%
in F1-score between kernel 3 and 7. Similarly, an increment
of 5.40% F1-score is observed when T=150, and the highest
increment of 7.79% F1-score is attained when the window
size is 175. Although high impact is seen in the variation of
kernel sizes, there is no much effect when varying the signal
window length on the performance of the networks.
Figure 5: The classification results of different networks (at kernels=3,4,5)
with variation in 2D window length size (T=125, 150, and 175). The results
are in percentages F1-score.
C.Analysis of individual motion intention across window
overlaps and kernel sizes
The analysis in the previous sections only present rather
general view of window sizes, overlaps, and kernel sizes
impact on motion intent classification. Therefore, there is need
to examine the impact of these parameters on the decoding
performance of individual class of movement. Thus, the effect
of kernel size and overlap on the characterization of individual
class of motion is presented in Figure 6. The effects of kernel
sizes and signal window overlap were investigated on each
MI. It is noteworthy that only the highest performing kernel
(kernel=7) and the lowest (kernel=3), likewise for the overlap
windows (0.75T and 0T) across individual class of motion
were considered as shown in the confusion matrices in Figure
6. This is because the window length only exhibits slight
impact on the classification performance. Thus, we used
T=175 for representation only and the confusion matrices are
presented as an average across all subjects.
From the classification results indicated in Figure 6, it can
be seen that the lowest accuracy is shown for Hand Open (HO)
motion intention by both networks (kernel equal 7 and 3)
when there is no overlap, and kernel =3 has lower results
compared to kernel=7. Classification accuracy higher that
95% was recorded for all motion intentions when the kernel=7
and the overlap is 0.75T as shown in Figure 6 (c). This shows
that the network of kernel =7 and 2D signals when overlap is
75% would be an ideal combination for characterizing the
EMG signal patterns related to these classes of hand
movements compared to others.
(a) Window size (T) = 125
(b) Window size (T) = 150
(c) Window
size (T) = 175
Figure 6: Classifiation performance of kernel 3 and 7 when the overlap is
75% and 0% of window length (T) across each motion intent.
IV. DISCUSSION
Feature learning through the use of deep neural networks
represent a promising alternative to hand crafted features
which consume a lot of time for preparation. To achieve
clinical use of DNN based prostheses, a relatively robust and
simple signal processing scheme (which is independent of the
structure of the classification network) is needed.
Therefore, in this study we explore the use of raw sEMG
signals in the identification of motor intents based on
convolution neural networks. Moreover, we investigate the
relationship between the 2D signal length and overlap that
can give optimal performance of the network. Additionally,
we examine the effects of kernel sizes of the CNNs with
respect to classification performance of the model when
applying raw EMG signals.
From the analyzed results of Figure 3, the percentage of
overlap within the generated 2D signal highly contributes to
the performance of the CNNs. The trend shows that 75% of
overlap yielded the highest performance for the networks
with an improvement of 9.49% in accuracy and 23.33% in
F1-score compared to without overlap. And the classification
performance increases with the increase in overlap. From
this, we can see that a positive correlation exists between the
performance of deep neural network and percentage of data
overlap, leading to the following rule of thumb: The
classification performance of Convolutional neural
networks increases with increase in the percentage of data
overlap at a constant window length.
Due to the non-stationary nature of the EMG signal, it is
possible to have 2D signals segmented at different time
window from the same class and subject to have different
probabilistic distributions [24]. This can be clearly observed
when the signals are generated without any overlap. The lack
of common distribution of data from the same class hinders
the optimal performance of the classification network [25].
The study presented in this paper show that a common
distribution between samples (segments) from the same class
can be achieved by increasing the percentage of overlap
which allows the model to learn easily the hidden distribution
(patterns) within the data and increase the performance. This
can be justified more by the analysis presented in Figure 6
where the classification results across individual motion class
show that 75% signal overlap has higher outcomes in all five
classes compared to when there is no overlap.
The kernel of the network can be considered as an eye of
the network, the wider the kernel, the wider the receptive field
of the network, meaning that it can have a broader range of
view over the signal and learn more general spatial features
(patterns) while a small kernel size learns fine features. [15].
The results obtained in Figure 5, show that the increase in
kernel size increases the performance by a considerable
margin regardless of the window size, with the highest
performance obtained by kernel 7. This implied that the
pattern of the EMG signals is present over a large spatial
dimension of the signal [26]. Moreover, a wide field of view
allows the network to model hidden temporal connectivity
within the signal [15]. Thus, it can learn both spatial and
temporal trend (patterns) of the signal.
It should be noted that in this study we focused only on
variation of three window lengths and kernel sizes which are
among the commonly used values. Since some studies that
focused on traditional machine learning and feature
engineering recommended that the window length can be as
lower as 32 and as higher as 300ms, in the future work,
further analysis will be carried out on these extreme ranges
and more kernel sizes will be considered with various CNNs
and other deep neural network architectures. Moreover, we
will investigate the robustness of the 2D signals at higher
percentage overlap in the presence of Gaussian noise which
are sometimes inherent in practical use of the prostheses.
REFERENCES
[1] Côté-Allard, U., Fall, C.L., Drouin, A., Campeau-Lecours, A., Gosselin, C.,
Glette, K., Laviolette, F. and Gosselin, B., 2019. Deep learning for
electromyographic hand gesture signal classification using transfer learning. IEEE
transactions on neural systems and rehabilitation engineering,27(4), pp.760-771.
[2] Li, X., Xu, R., Samuel, O.W., Tian, L., Zou, H., Zhang, X., Chen, S., Fang, P.
and Li, G., 2016, August. A new approach to mitigate the effect of force variation
on pattern recognition for myoelectric control. In 2016 38th Annual International
Conference of the IEEE Engineering in Medicine and Biology Society (EMBC)
(pp. 1684-1687). IEEE.
[3] Asogbon, M.G., Samuel, O.W., Geng, Y., Oluwagbemi, O., Ning, J., Chen, S.,
Ganesh, N., Feng, P. and Li, G., 2020. Towards resolving the co-existing impacts
of multiple dynamic factors on the performance of EMG-pattern recognition based
prostheses. Computer methods and programs in biomedicine,184, p.105278.
[4] Samuel, O.W., Asogbon, M.G., Geng, Y., Li, X., Pirbhu lal, S., Chen, S.,
Ganesh, N., Feng, P. and Li, G., 2019, July. Spatio-temporal based descriptor for
limb movement-intent characterization in EMG-pattern recognition system. In
2019 41st Annual International Conference of the IEEE Engineering in Medicine
and Biology Society (EMBC) (pp. 2637-2640). IEEE.
a) kernel=3, overlap=0.75T b) kernel=3, overlap=0T
c
)
kernel=7
, overlap=0.75T
d
)
kernel=7, overlap=0
T
[5] Samuel, O.W., Asogbon, M.G., Geng, Y., Rusydi, M.I., Mzurikwao, Z.B.,
Chen, S., Fang, P. and Li, G., 2020, Jul y. Characterizing Multiple Patterns of
Motor Intent Using Spatial-Temporal Information for Intuitively Active Motor
Training in Stroke Survivors. In 2020 42nd Annual International Conference of
the IEEE Engineering in Medicine & Biology Society (EMBC) (pp. 3831-3834).
IEEE.
[6] Hargrove, L.J., Scheme, E.J., Englehart, K.B. and Hudgins, B.S., 2010.
Multiple binary classifications via linear discriminant analysis for improved
controllability of a powered prosthesis. IEEE Transactions on Neural Systems and
Rehabilitation Engineering, 18(1), pp.49-57.
[7] Côté-Allard, U., Campbell, E., Phinyomark, A., Laviolette, F., Gosselin, B.
and Scheme, E., 2020. Interpreting deep learning features for myoelectric control:
A comparison with handcrafted features. Frontiers in Bioengineering and
Biotechnology, 8, p.158.
[8] Zanghieri, M., Benatti, S., Burrello, A., Kartsch, V., Conti, F. and Benini, L.,
2019. Robust real-time embedded EMG recognition framework using temporal
convolutional networks on a multicore IoT processor. IEEE transactions on
biomedical circuits and systems,14(2), pp.244-256.
[9] Zhai, X., Jelfs, B., Chan, R.H. and Tin, C., 2017. Self-recalibrating surface
EMG pattern recognition for neuroprosthesis control based on convolutional
neural network. Frontiers in neuroscience, 11, p.379.
[10] Olsson, A., Malešević, N., Björkman, A. and Antfolk, C., 2019, July.
Exploiting the intertemporal structure of the upper-limb sEMG: Comparisons
between an LSTM network and cross-sectional myoelectric pattern recognition
methods. In 2019 41st Annual International Conference of the IEEE Engineering
in Medicine and Biology Society (EMBC) (pp. 6611-6615). IEEE.
[11] Rahimian, E., Zabihi, S., Asif, A., Farina, D., Atashzar, S.F. and
Mohammadi, A., 2021. Fs-hgr: Few-shot learning for hand gesture recognition via
electromyography. IEEE transactions on neural systems and rehabilitation
engineering,29, pp.1004-1015.
[12] Khushaba, R.N. and Nazarpour, K., 2021. Decoding HD-EMG Signals for
Myoelectric Control-How Small Can the Analysis Window Size be?. IEEE
Robotics and Automation Letters, 6(4), pp.8569-8574.
[13] Atzori, M., Cognolato, M. and Müller, H., 2016 . Deep learning with
convolutional neural networks applied to electromyography data: A resource for
the classification of movements for prosthetic hands. Frontiers in
neurorobotics,10, p.9.
[14] Tsinganos, P., Cornelis, B., Cornelis, J., Jans en, B. and Skodras, A., 2018,
September. Deep Learning in EMG-based Gesture Recognition. In PhyCS (pp.
107-114).
[15] Rahimian, E., Zabihi, S., Atashzar, S.F., Asif, A. and Mohammadi, A., 2019,
November. Semg-based hand gesture recognition via dilated convolutional neural
networks. In 2019 IEEE Global Conference on Signal and Information Processing
(GlobalSIP) (pp. 1-5)..
[16] BAKIRCIOĞLU, K. and Özkurt, N., 2020. Classification of EMG signals
using convolution neural network. International Journal of Applied Mathematics
Electronics and Computers,8(4), pp.115-119.
[17] Gulati, P., Hu, Q. and Atashzar, S.F., 2021. Toward deep generalization of
peripheral EMG-Based Human-Robot interfacing: A hybrid explainable solution
for NeuroRobotic systems. IEEE Robotics and Automation Letters,6(2),
pp.2650-2657.
[18] Smith, L.H., Hargrove, L.J., Lock, B.A. and Kuiken, T.A., 2010. Determining
the optimal window length for pattern recognition-based myoelectric control:
balancing the competing effects of classification error and controller delay. IEEE
Transactions on Neural Systems an d Rehabilitation Engineering, 19(2),
pp.186-192.
[19] Wahyunggoro, O. and Nugroho, H.A., 2017, May. Effect of window length
on performance of the elbow-joint angle prediction based on electromyography. In
Journal of Physics: Conference Series (Vol. 853, No. 1, p. 012014). IOP
Publishing.
[20] Asogbon, M.G., Samuel, O.W., Jiang, Y., Wang, L., Geng, Y., Sangaiah,
A.K., Chen, S., Fang, P. and Li, G., 2020. Appropriate feature set and window
parameters selection for efficient motion intent characterization towards
intelligently smart EMG-PR system. Symmetry, 12(10), p.1710.
[21] Krasoulis, A., Vijayakumar, S. and Nazarpour, K., 2019. Multi-grip
classification-based prosthesis control with two EMG-IMU sensors. IEEE
Transactions on Neural Systems an d Rehabilitation Engineering, 28(2),
pp.508-518.
[22] Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I. and Salakhutdinov,
R., 2014. Dropout: a simple way to prevent neural networks from overfitting. The
journal of machine learning research,15(1), pp.1929-1958.
[23] Jayasinghe, S., Rambukkanage, L., Silva, A., de Silva, N. and Perera, A.S.,
Critical Sentence Identification in Legal Cases Using Multi-Class Classification.
In 2021 IEEE 16th International Conference on Industrial and Information
Systems (ICIIS) (pp. 146-151)
[24] Zhai, X., Jelfs, B., Chan, R.H. and Tin, C., 2017. Self-recalibrating surface
EMG pattern recognition for neuroprosthesis control based on convolutional
neural network. Frontiers in neuroscience, 11, p.379.
[25] Wang, W., Huang, Y., Wang, Y. and Wang, L., 2014. Generalized
autoencoder: A neural network framework for dimensionality reduction.
In Proceedings of the IEEE conference on computer vision and pattern
recognition workshops (pp. 490-497).
[26] Stango, A., Negro, F. and Farina, D., 2014. Spatial correlation of high density
EMG signals provides features robust to electrode number and shift in pattern
recognition for myocontrol. IEEE Transactions on Neural Systems and
Rehabilitation Engineering, 23(2), pp.189-198.
