You are on page 1of 6

A Deep learning approach for the Estimation of

Middleton Class-A Impulsive Noise Parameters


Bassant Selim∗ , Md Sahabul Alam∗ , Georges Kaddoum∗ , Mohammad T. AlKhodary∗ , and Basile L. Agba∗,§

∗ Departmentof Electrical Engineering, École de Technologie Supérieure, Montreal, Canada,


(emails: bassant.selim.1@ens.etsmtl.ca; md-sahabul.alam.1@ens.etsmtl.ca; georges.kaddoum@etsmtl.ca;
mohammad.alkhodary.1@ens.etsmtl.ca)
§ Hydro-Quebec Research Institute (IREQ), Montreal, Canada,

(email: agba.basilel@ireq.ca)

Abstract—Impulsive noise is a common impediment in class-A model [1] is very popular and widely accepted
many wireless, power line communication (PLC), and due to its canonical property, tractable probability dis-
smart grid communication systems that prevents the system tribution function (PDF), and capability of capturing
from achieving error-free transmission. To overcome the
detrimental effects of such impulsive interference, knowl- the statistical behaviour of a wide variety of impulsive
edge of impulsive noise parameters is generally required interference.
by the available mitigation techniques. This work considers
a machine learning perspective for the estimation of the Numerous methods have been investigated to im-
impulsive noise parameters in communication systems un- prove the performance of wireless communications in
der the influence of Middleton class-A noise. Precisely, we the presence of impulsive noise. A simple and efficient
consider a deep learning approach and design a deep neural approach is to precede the receiver with a non-linear
network (DNN) that classifies a set of received symbols
according to the parameters of the impulsive noise affecting preprocessor such as clipping, blanking, or combined
them. It is sown that the classification accuracy greatly clipping/blanking. Ndo et al. [11] have considered an
depends on the number of symbols fed into the neural adaptive clipping-based impulsive noise mitigation tech-
network as well as the number of considered states in the nique to overcome the noxious effect of impulsive noise
classification, where the proposed approach can reach a in orthogonal frequency division multiplexing (OFDM)-
testing accuracy of more than 99%.
Index Terms—Impulsive noise mitigation, Middleton based PLC channels. Since it was clearly shown that
class-A noise, Parameter estimation, Deep neural networks, the value of the clipping/blanking threshold selected is
Smart grid communication. of paramount importance in the mitigation performance,
they have determined an optimized clipping threshold
I. I NTRODUCTION based on the well-known false alarm and good detection
Interference and noise with non-Gaussian impulsive trade-off. The authors in [12] have derived a closed-
behaviors are widespread phenomena encountered in form optimal blanking threshold for OFDM receivers
communication systems that can severely degrade the employing blanking non-linearity to cancel the effect
system’s performance. Indeed, several studies show suf- of impulsive noise. The obtained results show that the
ficient evidences that impulse man-made noise is en- optimized threshold maximizes the SNR at the output of
countered in various metropolitan, high voltage, manu- the blanking nonlinearity. A comprehensive study of the
facturing plants, and indoor environments [1], [2]. It was performance of clipping and blanking methods for the
shown in [3]–[7] that communication systems designed mitigation of the performance degradation in impulsive
under the additive white Gaussian noise (AWGN) as- noise environments was considered in [13]. Recently,
sumption typically experience severe performance degra- advanced algorithms have been proposed to determine
dation when subjected to impulsive noise. This elevates the optimal threshold to maximize the impulsive noise
the need for the performance analysis of communication reduction after clipping/blanking, see [14], [15] and the
systems, which are not only disturbed by background references therein. Meanwhile, for single carrier commu-
Gaussian noise, but also by impulsive (non-Gaussian) nication systems, an effective impulsive noise mitigation
noise, in order to provide pragmatic information for technique is to calculate the exact log-likelihood ratio
the system designer. Consequently, over the last few (LLR) of each symbol by taking into account the exact
decades, several statistical models and their canonical impulsive noise statistics [4], [6], [16]–[18] where it
parameters were suggested to model different impulsive was demonstrated that significant bit error rate (BER)
behaviour [1], [8]–[10]. Out of them, the Middleton performance gain can be achieved through this scheme.

978-1-7281-5089-5/20/$31.00 ©2020 IEEE

Authorized licensed use limited to: Carleton University. Downloaded on August 03,2020 at 09:01:49 UTC from IEEE Xplore. Restrictions apply.
However, these works rely on the impractical assump- with
tion of perfect knowledge of the impulsive noise statistics exp−A Am
pm = (2)
at the receiver in order to evaluate the clipping/blanking m!
threshold or the LLR. To estimate these parameters, for
and
Middleton class-A noise, the works in [19], [20] resorted
2 m/A + T
to Markov Chain Monte Carlo which requires a large σm = σ2 . (3)
1+T
number of iterations in order to converge. More recently,
2
Sacuto et al. [21] used fuzzy C-means clustering to where pm and σm are the steady state probability and
estimate the partitioned Markov chain impulse noise the variance of the mth impulsive source, respectively.
parameters in PLC. This work only considers three For m = 0, the model generates the traditional AWGN
possible impulsive noise states. component.
On the other hand, machine learning approaches, in- The parameters A, T , and σ 2 are called global pa-
cluding deep learning, are becoming increasingly popu- rameters [1] as they characterize the PDF. The physical
lar in numerous applications such as signal detection and significance of these parameters are [1]:
resource allocation in wireless communication systems.
• A is the impulsive index which is defined as the
Their position in the spotlight can be attributed to their
average number of impulses per unit time (λ)
capacity to seamlessly perform complicated tasks, which
impinging on the receiver times the impulse mean
are generally impossible to achieve or would require
duration (T̄ )
substantial efforts if traditional algorithm were used.
In the context of impulsive noise mitigation, recently, A = λT̄ . (4)
Barazideh et al. [22] proposed to use a deep neural
A smaller A implies a lower number of impulsive
network (DNN) to detect the instances of impulsive
events and/or their duration. In such cases, the
noise before a blanking nonlinearity. Although instruc-
impulses are not dominant compared to the AWGN
tive, the proposed work is limited to OFDM systems
[1]. Higher values of A imply that the impulsive
with clipping/blanking nonlinearity. Thus, in this work,
noise is more dominant compared to the AWGN
we seek to estimate the impulsive noise parameters on
and by increasing A, the impulsive noise becomes
Middleton class-A noise, which can thereafter be used
closer to the Gaussian noise.
in the implementation of various mitigation techniques.
• T is the Gaussian to impulsive noise power ratio. It
To do so, inspired by the exceptional ability of neural
provides information on how strong the impulsive
networks to classify images based on their content, see
noise is compared to the independent AWGN noise
[23] and the references therein, this work makes use
and is defined as
of DNNs to determine the parameters of the impulsive
noise in communication systems. The proposed DNN 2
T = σG /σI2 . (5)
can accurately classify the impulsive noise parameters
in real time and with a very small number of samples The lower T is, the stronger the impulsive noise is,
required. It is shown that the proposed approach can compared to the background AWGN noise.
be used with classical mitigation techniques in order to • σ 2 represents the total power of the noise nk and
accurately eliminate the effect of impulsive noise on the is given by
communication system.
σ 2 = σG
2
+ σI2 . (6)
The remainder of this paper is organized as follow.
Section II presents an overview of the Middleton class-A The class-A model can be simplified and made more
noise model while Section III introduces the considered manageable by considering its truncated version with the
DNN design. Moreover, simulation results are provided first M terms as
in Section IV. Finally, Section V concludes this work.
M
 −1  
ṕ n2
II. M IDDLETON C LASS -A N OISE M ODEL f˜(nk ) = √ m exp − k2 , (7)
m=0
2πσm 2σm
The Middleton class-A model can be seen as a su-
perposition of statistically independent impulsive source with
emissions where the sources are Poisson distributed pm
ṕm = M −1 . (8)
and the amplitude of the sources follow a Gaussian m=0 pm
distribution. The PDF of a real-valued Middleton class-
A noise sample nk , where k represents the discrete-time It was shown in [24] that such a truncation with the
index, is given by [1] first four terms (M = 4) approximates the theoretical
∞   PDF with sufficient accuracy. Therefore, in this paper, we
 pm n2k consider this accurate approximation in order to simplify
f (nk ) = √ exp − 2 , (1)
m=0
2πσm 2σm the simulation of the Middleton class-A impulsive noise.

Authorized licensed use limited to: Carleton University. Downloaded on August 03,2020 at 09:01:49 UTC from IEEE Xplore. Restrictions apply.
III. I MPULSIVE N OISE PARAMETER E STIMATION of the network. Therefore, we propose to sort the input
Without loss of generality, let us consider an informa- vector in an ascending, or descending, order such that
tion source transmitting M -ary phase shift keying (PSK) the affected symbols are always fed to specific neurons.
modulated data to a destination over a flat fading channel Let y = [y1 , y2 , ..., yN ] be the input vector consisting of
in the presence of impulsive noise. The received signal N received symbols. Here, yn represents the magnitude
at each time epoch k is therefore given by of the k largest magnitude of the N symbols considered.
Thus, the DNN input layer consists of N nodes. Mean-
rk = x k + n k , (9) while, the soft outputs of the network θ = [θ1 , θ2 , ...., θI ]
where xk is the transmitted symbol from the source and indicate to which state θi , from the I states considered
nk = nw I in the training phase, the input vector is mapped. Here,
k + bnk denotes the total noise which is the
Middleton class-A model. Let r = {r1 , r2 , ..., rK }, be the term state represents the set of parameters of the
a sequence of received symbols impeded by impulsive impulsive noise θi = {Ai , Ti }. Here, the relationship
Middleton class-A noise with parameters A and T , between the layers of the network are given by
 
which we seek to estimate. H[1] = φ[1] W[1] y + b[1] , (10)
A. Deep Neural Network Design  
Deep learning is a supervised learning method, which H[2] = φ[2] W[2] H[1] + b[2] , (11)
learns to map the input to an output based on training and  
performed on labeled input-output pairs. For classifiers, θ = φ[3] W[3] H[2] + b[3] (12)
where the aim is to determine to which of the k out-
puts the input belongs, deep neural networks seek to where φ[l] , b[l] , and W[l] denote the lth layer’s vec-
approximate some function f , which maps the input y tors of activation functions and biases, and matrices,
to the output θ. To do so, these networks use the labeled respectively. Here the output of a layer is also the input
training data to determine the most suitable mathematical of the subsequent one. We use the rectified linear unit
manipulation to map the input to the output. We consider (ReLU) function for the hidden layers and a sigmoid for
a DNN, depicted in Fig. 1, with N input neurons and the output layer. The considered DNN uses backward
2 dense layers with Q neurons each. Here, it is well propagation and the Adam algorithm for its training [25].
known that there is no rule of thumb to follow n order The Adam optimizer has been recently proposed as an
to determine the optimal number of layers and neurons; extension to the classical stochastic gradient descent. The
hence, the characteristics of DNNs are usually defined main feature of this optimizer is its ability to adapt the
based on trial and error. learning rate.

Hidden Hidden IV. S IMULATION RESULTS


Layer 1 Layer 2 In this section, we assess the performance of the pro-
Output
Input ReLU ReLU Layer posed approach for the estimation of the impulsive noise
Layer 1 1 parameters in a point-to-point communication system
y 1
1 under the influence of Middleton class-A noise. Do to
ReLU so, we start by evaluating the accuracy of the machine
ReLU
2 2 learning approach proposed for the parameter estimation.
y 2
2
In addition, we evaluate the performance of the LLR
mitigation technique, when the estimated impulsive noise
3 parameters are considered.
A. Impulsive noise parameter estimation
y N
Without loss of generality, in this section, we consider
ReLU ReLU
I a single carrier system with low-density parity-check
Q Q (LDPC) coded transmission under the effect of Middle-
ton class-A noise. It is worth mentioning that the con-
Fig. 1: Considered DNN block diagram sidered modulation is binary phase shift keying (BPSK);
however. the DNN was tested with different modulation
It is recalled that impulsive noise, is characterized orders where it was shown that this parameter does not
by random occurrences in time. Thus, out of the N affect the classification performance of the network. The
input symbols of the DNN, only a fraction of them are number of neurons considered in the hidden layers of
affected by impulsive noise. Moreover, these symbols are the DNN is N + 70, where N denotes the number of
randomly distributed in the input sequence. Intuitively, input neurons, i.e., the number of symbols considered.
this randomness would negatively affect the performance Moreover, unless mentioned otherwise, the number of

Authorized licensed use limited to: Carleton University. Downloaded on August 03,2020 at 09:01:49 UTC from IEEE Xplore. Restrictions apply.
outputs of the DNN is 25, from which 24 correspond over, it is observed that increasing the size of the input
to the different impulsive noise states obtained form all vector y positively affects the classification accuracy.
possible combinations of A = [0.5 0.1 0.05 0.01] and This result is expected since, as the number of input
T = [0.1 0.02 0.01 0.005 0.002 0.001], while the last symbols increases, the number of samples affected by
state corresponds to the AWGN case. The training and impulsive noise in the input is increased, which increases
testing set included 9 × 105 split between training and the classification performance.
testing with a 2 : 1 ratio. Moreover, Table I shows the test accuracy of the
considered DNN when the number of impulsive states
considered is varied. Here, we consider a DNN with
90 N = 200 input samples and Q = 270 neurons in
Sorted Inputs each hidden layer. As expected, the test accuracy of the
Unsorted Inputs
85 network is greatly improved when the considered states
are more easily distinguishable. This is the case when I
80
is decreased and when the considered parameters greatly
DNN Test Accuracy (%)

differ from each other. Precisely, for the considered


states, the DNN achieves an exceptional performance
75
when the number of states considered is I ≤ 10.

70 B. Impulsive Noise Mitigation

65
100
60
10-1
55
100 200 300 400 500
Number of input samples (N) 10-2

Fig. 2: DNN test accuracy versus number of input symbols (N) the
considered DNN with 2 layers where Q = N + 70 neurons in each 10-3
BER

layer and I = 25 distinct impulsive noise states.


10-4
Here, we consider a practical communication system
where we seek to determine the impulsive noise pa-
10-5
rameters from the received signal, on the fly. Since the
activation function of each neuron is the ReLU function
for all hidden layers, the computational complexity of 10-6 LLR (Genie)
LLR (DNN)
the neural network during the prediction, as a forward No Mitigation
feeding, is linear of an order of O(Q) [26]. 10-7
0 0.5 1 1.5 2 2.5 3 3.5
By definition, impulsive noise is characterized by a
SNR (dB)
high power and random occurrence where the impul-
sive index determines the average number of impulses Fig. 3: BER performance of LLR-based mitigation assuming the Genie
case with perfect knowledge of the impulsive noise parameters, the
per unit time. To train the network, we generate data estimated parameters using DNN, and the no mitigation case for A =
according to (9), where nk come from the Middleton 0.0083 and T = 0.02.
class-A distribution in (1). For each state considered, we
generate a sequence of 5×104 labeled vectors of N input Here, we evaluate the performance of the LLR based
samples. Here, 70% of the data is used for the training mitigation with the estimated parameters. To do so,
while the remainder 30% of the datasets are considered we consider the case where the actual impulsive noise
in the testing phase. parameters that we seek to estimate, i.e., A = 0.0083
As detailed in Section III-A, we ordered the input and T = 0.02, do not correspond to any of the consid-
vector in the training and testing datasets in order to ered states. In this context, the DNN network estimates
increase the learning efficiency. Fig. 2 shows the neural A = 0.02 and T = 0.01. Thus, we assume that a
network testing accuracy versus the considered number sequence of equally likely information bits of length
of input samples N for both ordered and non ordered 32, 400 is encoded using LDPC channel coding based on
input vectors. Interestingly, it is observed that ordering the DVB-S2 standard [27] with a code rate of 1/2. The
the data is associated with an increased accuracy and coded sequence is then mapped into a BPSK modulation
consequently a better classification performance. More- sequence. The number of iterations is set to 50.

Authorized licensed use limited to: Carleton University. Downloaded on August 03,2020 at 09:01:49 UTC from IEEE Xplore. Restrictions apply.
TABLE I: Test Accuracy versus Number of States (I)

Number of Test Accuracy States

States (I) (%)

5 99.62 θ5 = [(A = 0.5, T = 0.001) (A = 0.01, T = 0.1) (A = 0.5, T = 0.1) (A = 0.01, T = 0.001)

(A = 0, T = 1)]

7 95.54 θ7 = [θ5 (A = 0.5, T = 0.01) (A = 0.01, T = 0.01)]

10 96.77 θ10 = [(θ7 (A = 0.1, T = 0.1) (A = 0.1, T = 0.01) (A = 0.1, T = 0.001)]

13 82.32 θ13 = [θ10 (A = 0.1, T = 0.005) (A = 0.01, T = 0.005) (A = 0.5, T = 0.005)]

17 80.20 θ17 = [θ13 (A = 0.05, T = 0.005) (A = 0.05, T = 0.001) (A = 0.05, T = 0.01) (A = 0.05, T = 0.1)]

21 95.54 θ21 = [θ17 (A = 0.5, T = 0.002) (A = 0.1, T = 0.002) (A = 0.01, T = 0.002) (A = 0.05, T = 0.002)]

25 73.84 θ25 = [θ21 (A = 0.5, T = 0.05) (A = 0.1, T = 0.05) (A = 0.01, T = 0.05) (A = 0.05, T = 0.05)]

Fig. 3 figure presents the BER of the LLR based [2] K. L. Blackard, T. S. Rappaport, and C. W. Bostian, “Measure-
mitigation assuming the Genie case where perfect knowl- ments and models of radio frequency impulsive noise for indoor
wireless communications,” IEEE Journal on Sel. Areas in Comm.,
edge of the impulsive noise parameters are available at vol. 11, no. 7, pp. 991–1001, Sept 1993.
the receiver as well as the case where the parameters [3] A. Spaulding and D. Middleton, “Optimum reception in an
are estimated by the proposed DNN. As a benchmark, impulsive interference environment-part i: Coherent detection,”
IEEE Trans. Commun., vol. 25, no. 9, pp. 910–923, 1977.
we also include the performance of the system without [4] M. S. Alam, F. Labeau, and G. Kaddoum, “Performance analysis
any mitigation at the receiver side. Interestingly, it is of DF cooperative relaying over bursty impulsive noise channel,”
observed that, assuming the worse case where the actual IEEE Trans. Commun., vol. 64, no. 7, pp. 2848–2859, 2016.
[5] M. S. Alam, G. Kaddoum, and B. Agba, “Performance analysis
impulsive noise parameters are not in the set of con- of distributed wireless sensor networks for Gaussian source
sidered states, the performance of the Genie based LLR estimation in the presence of impulsive noise,” IEEE Sig. Proc.
detection is similar to the of the proposed estimation. Letters, vol. 25, no. 6, pp. 803–807, 2018.
[6] M. S. Alam, G. Kaddoum, and B. L. Agba, “Bayesian MMSE es-
This is because the DNN selects the parameters that most timation of a Gaussian source in the presence of bursty impulsive
accurately mimic the exact parameters of the system. noise,” IEEE Commun. Letters, vol. 22, no. 9, pp. 1846–1849,
2018.
V. C ONCLUSION [7] ——, “A novel relay selection strategy of cooperative network
This work proposed a Deep learning approach for impaired by bursty impulsive noise,” IEEE Trans. Veh. Technol.,
vol. 68, no. 7, pp. 6622–6635, 2019.
the estimation of the impulsive noise parameters in [8] M. Ghosh, “Analysis of the effect of impulse noise on multicarrier
communication systems under the effect of Middleton and single carrier QAM systems,” IEEE Trans. Commun., vol. 44,
class-A impulsive interference. Precisely, we designed no. 2, pp. 145–147, 1996.
[9] D. Fertonani and G. Colavolpe, “On reliable communications
a DNN consisting of two hidden layers to classify the over channels impaired by bursty impulse noise,” IEEE Trans.
impulsive noise parameters of an input vector of N sam- Commun., vol. 57, no. 7, pp. 2024–2030, 2009.
ples between I states. It was shown that the considered [10] G. Ndo, F. Labeau, and M. Kassouf, “A Markov-Middleton model
for bursty impulsive noise: modeling and receiver design,” IEEE
approach maintains a high classification accuracy, even Trans. Power Del., vol. 28, no. 4, pp. 2317–2325, 2013.
when the number of inputs is decreased. Thus, it can [11] G. Ndo, P. Siohan, and M.-H. Hamon, “Adaptive noise mitigation
be applied in realistic communication systems in order in impulsive environment: Application to power-line communica-
tions,” IEEE Trans. Power Delivery, vol. 25, no. 2, pp. 647–656,
to estimate the impulsive noise parameters from the 2010.
received signal. In this context, it was shown that the [12] S. V. Zhidkov, “Performance analysis and optimization of OFDM
classification accuracy obtained greatly depends on the receiver with blanking nonlinearity in impulsive noise environ-
ment,” IEEE Trans. Veh. Technol., vol. 55, no. 1, pp. 234–242,
number of considered states and the length of the input 2006.
vector where a higher accuracy can always be obtained [13] H. Oh and H. Nam, “Design and performance analysis of
by increasing the length of the input vector used for the nonlinearity preprocessors in an impulsive noise environment,”
IEEE Trans. Veh. Technol., vol. 66, no. 1, pp. 364–376, 2017.
classification. [14] B. Adebisi, K. Anoh, K. M. Rabie, A. Ikpehai, M. Fernando,
R EFERENCES and A. Wells, “A new approach to peak threshold estimation for
impulsive noise reduction over power line fading channels,” IEEE
[1] D. Middleton, “Statistical-physical models of electromagnetic Systems J., vol. 13, no. 2, pp. 1682–1693, 2018.
interference,” IEEE Trans. Electromagn. Compat., vol. EMC-19, [15] N. Rožić, P. Banelli, D. Begušić, and J. Radić, “Multiple-
no. 3, pp. 106–127, 1977. threshold estimators for impulsive noise suppression in multicar-

Authorized licensed use limited to: Carleton University. Downloaded on August 03,2020 at 09:01:49 UTC from IEEE Xplore. Restrictions apply.
rier communications,” IEEE Trans. Sig. Process., vol. 66, no. 6, for parameter estimation of partitioned markov chain impulsive
pp. 1619–1633, 2018. noise model,” in IEEE Int. Conf. Smart Grid Commun. IEEE,
[16] M. S. Alam, B. Selim, and G. Kaddoum, “Analysis and com- 2013, pp. 348–353.
parison of several mitigation techniques for middleton Class-A [22] R. Barazideh, S. Niknam, and B. Natarajan, “Impulsive noise
noise,” in Proc. Latin-American Conf. Commun. (LATINCOM). detection in OFDM-based systems: A deep learning perspective,”
IEEE, 2019, pp. 1–6. in IEEE 9th Annual Computing and Communication Workshop
[17] H. Nakagawa, D. Umehara, S. Denno, and Y. Morihiro, “A and Conference (CCWC), Jan 2019, pp. 0937–0942.
decoding for low density parity check codes over impulsive noise [23] V. Sze, Y. Chen, T. Yang, and J. S. Emer, “Efficient processing
channels,” in Proc. Int. Symposium Power Line Communications of deep neural networks: A tutorial and survey,” Proceedings of
and Its Applications. IEEE, 2005, pp. 85–89. the IEEE, vol. 105, no. 12, pp. 2295–2329, Dec 2017.
[18] D. Umehara, H. Yamaguchi, and Y. Morihiro, “Turbo decoding [24] K. Vastola, “Threshold detection in narrow-band non-gaussian
in impulsive noise environment,” in Proc. IEEE Global Telecom. noise,” IEEE Trans. Commun., vol. 32, no. 2, pp. 134–139, 1984.
Conf., vol. 1. IEEE, 2004, pp. 194–198. [25] D. P. Kingma and J. Ba, “Adam: A method for stochastic
[19] Y.-Z. Jiang, X.-l. Hu, W.-L. Li, and S.-X. Zhang, “Estimation optimization,” arXiv preprint arXiv:1412.6980, 2014.
of two-dimensional class A noise model parameters by markov [26] “Complexity analysis of multilayer perceptron neural network
chain monte carlo,” in 2nd IEEE International Workshop on embedded into a wireless sensor network,” Procedia Computer
Computational Advances in Multi-Sensor Adaptive Processing. Science, vol. 36, pp. 192 – 197, 2014, complex Adaptive Systems
IEEE, 2007, pp. 249–252. Philadelphia, PA November 3-5, 2014.
[20] Y.-Z. Jiang, X.-l. Hu, X. Kai, and Z. Qi, “Bayesian estimation [27] D. J. C. Mackay, “Digital Video Broadcasting (DVB), Second
of class A noise parameters with hidden channel states,” in 2007 generation framing structure, channel coding and modulation
IEEE International Symposium on Power Line Communications systems for Broadcasting, Interactive Services, News Gathering
and Its Applications. IEEE, 2007, pp. 2–4. and other broadband satellite applications (DVB-S2),” ETSI Std.
[21] F. Sacuto, F. Labeau, and B. L. Agba, “Fuzzy C-means algorithm EN, vol. 302, no. 307, p. V1.2.1, 2009.

Authorized licensed use limited to: Carleton University. Downloaded on August 03,2020 at 09:01:49 UTC from IEEE Xplore. Restrictions apply.

You might also like