0% found this document useful (0 votes)
8 views4 pages

A Machine-Learning-Based Blind Detection On Interference Modulation Order in NOMA Systems

Uploaded by

chiranjeevit
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
8 views4 pages

A Machine-Learning-Based Blind Detection On Interference Modulation Order in NOMA Systems

Uploaded by

chiranjeevit
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 4

IEEE COMMUNICATIONS LETTERS, VOL. 22, NO.

12, DECEMBER 2018 2463

A Machine-Learning-Based Blind Detection on Interference


Modulation Order in NOMA Systems
Ningbo Zhang, Member, IEEE, Kai Cheng , and Guixia Kang

Abstract— In order to blindly detect the modulation order interference user in power-domain NOMA. Anderson-Darling
of interference signals in downlink non-orthogonal multiple test (AD) extracts features from received constellation points
access systems, a machine learning (ML) algorithm based on and its introduction makes machine learning algorithm feasible
Anderson–Darling test is proposed in this letter. The proposed
algorithm adopts ML to determine the modulation order of in blind detection area.
interference user equipment from the raw received constellation The technical contributions of this letter include:
points automatically. In feature extraction, a novel feature is • Proposing a new feature of received constellation points
introduced to improve the accuracy of blind detection. To eval- after equalization with Anderson-Darling test.
uate the performance of blind detection, the detection rate • Introducing machine learning algorithm to blind detection
and the throughput are simulated under different scenarios.
Simulation results show that the proposed algorithm outperforms
area in communication discipline.
conventional algorithm on modulation order detection. • Proposing a new blind detection algorithm based on
Anderson-Darling test and machine learning algorithm.
Index Terms— Blind detection, machine learning, Anderson-
Darling test, modulation order, NOMA.
II. S YSTEM M ODEL
This letter considers a downlink single-cell scenario which
I. I NTRODUCTION
consists one base station (BS) and N user equipments (UEs).

N ON-ORTHOGONAL multiple access (NOMA) has


become one of the key technologies in the fifth genera-
tion (5G) systems [1]–[5]. One of candidate receiver schemes
The UEs are denoted as Ui , i = 1 · · · N . And the sys-
tem model considers a multiple carrier channel scenario.
The available NOMA scheme can broadly be divided into
is codeword level successive interference cancellation (CWIC) power-domain NOMA and code-domain NOMA [1]. This
receiver, which uses successive interference cancellation (SIC) letter focuses on the power-domain NOMA. Based on the
technique [6]. Assistance information, including modulation concept of power-domain NOMA, signals of two or more
order of interference user, is required for CWIC receiver to users are superposed together with specified power ratios and
cancel the inter-superposition-layer interference [6]. Methods, are transmitted on the same time-frequency resource. This
which may be considered for obtaining assistance information, process is called superposition coding (SC) and can be written
include blind detection (BD), signaling, etc [6]. However, as [1], [6]
signaling will consume extra resources, which would increase
Nc 

the burden of control channel. Thus, BD is a practical method
t= αi Pr x(i) , Nc ≤ N (1)
in real systems, and has been discussed by researchers [7], [8].
i=1
Generally, the solutions of BD can be divided into like-
lihood based (LB) and feature based (FB) [9]. In LB area, where Nc is the number of superposed UEs, and Pr denotes
exact or approximated likelihood function is used to solve the the total radiation power.Ncαi is the portion of Pr assigned
classification problem. Lee et al. propose an algorithm based to Ui , which satisfies i=1 αi = 1. x(i) is the signal of Ui
(i) (i) (i)
on max-log approximation [10]. In FB area, high-order cumu- and can be written as x(i) = [x1 , x2 , · · · , xK ]T , where
(i)
lants are widely used. Huang et al. propose a cumulant-based i = 1, · · · , Nc . (·)T denotes the transpose of a vector, xk
maximum likelihood algorithm [11]. Existing blind detection and K denote the k-th symbol for user Ui and the number of
(i)
algorithms do not use any prior experience and cause perfor- symbols, respectively. Symbol xk is chosen from a constel-
mance degradation. This letter proposes a machine learning lation set Cp(i) , whose cardinality is denoted by |Cp(i) | and
algorithm based on Anderson-Darling test (MLAD), which p(i) represents the modulation order of Ui . The k-th symbol
uses prior experience to blindly detect the modulation order of in t, denoted by tk , is chosen from a composite constellation
set Cpc ,
Manuscript received August 15, 2018; revised September 28, 2018; accepted
September 29, 2018. Date of publication October 8, 2018; date of current (i)
 Nc |C
  p | 
version December 10, 2018. This work was supported by the Fundamen-
tal Research Funds for the Central Universities (2017RC06) and National Cpc = c(i,j) |c(i,j) = C(p(i) ,j) , (2)
Natural Science Foundation of China (61501056). (Corresponding author: i=1 j=1
Kai Cheng.)  c (i)
The authors are with the Key Laboratory of Universal Wireless Com- where pc = N i=1 p and C(p(i) ,j) is the j-th point in Cp(i) .
munications, Ministry of Education, Beijing University of Posts and
Telecommunications, Beijing 100876, China (e-mail: [email protected]; pc is termed as the composite modulation order of the super-
[email protected]; [email protected]). posed transmission. A schematic diagram is given in Fig. 1 to
Digital Object Identifier 10.1109/LCOMM.2018.2874218 show how superposition coding is performed.
1558-2558 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: VNR Vignana Jyothi Inst of Eng & Tech. Downloaded on October 18,2023 at 04:10:20 UTC from IEEE Xplore. Restrictions apply.
2464 IEEE COMMUNICATIONS LETTERS, VOL. 22, NO. 12, DECEMBER 2018

Fig. 1. An example of composite constellation generated from two QPSK


constellations with different power ratios.

Fig. 3. The features of QPSK and 16QAM when modulation order of target
user is 64QAM on different SNRs.

Fig. 2. The diagram of MLAD, including two phases: training phase and constellation points. For the sake of simplicity, the set of
blind detection phase. clusters are denoted by
Dpc = {· · · , Sj , · · · }, j = 1, · · · , 2pc . (5)
(i)
Defining r as the received constellation points vector after
equalization at the user Ui . Then, r (i) can be written as 2) Feature Extraction: Feature extraction is the most impor-
tant part in the proposed algorithm. It reduces the complexity
r (i) = H (i) t + n(i) , (3) of received constellation points. The goal of this step is
to extract accurate features of received constellation points.
where H (i) denotes the channel matrix after equalization, and AD test is used to extract features which can be written as [12]
n(i) is the additive noise vector, whose elements are inde-
K

pendent and identically-distributed (i.i.d.) complex Gaussian. 2k − 1 (i) (i)
A2 = −K − [ln(F (rk )) + ln(1 − F (rK+1−k ))],
E[|n(i) |2 ] = σi2 , where E[·] denotes the expectation operator, K
k=1
and |·| represents the absolute value of a complex number. The
(6)
aim of blind detection is to determine the modulation order of
interference user from r (i) . where F (·) is hypothesized distribution function and {r1 <
(i)
(i) (i)
r2 < · · · < rK } is the ordered data. Stephens shows that for
III. P ROPOSED A LGORITHM the case where the mean (μ) and variance (σ) are estimated
from data, the statistic must be corrected according to [12]
In this section, we introduce a new feature of received
constellation points r (i) and propose a MLAD algorithm, 4 25
A2∗ = A2 (1 + − 2 ). (7)
including clustering, feature extraction, classification training n n
and model parameter selection criteria. Its block diagram is The formula for p-value depends on the value of AD
shown in Fig.2. Without loss of generality, two users are statistic. There are two methods to calculate p-value, which
considered in the remainder of this letter. can be found in [13]. Assuming that the p-value of Sj in Dpc
(p )
is pj c . Then, the feature of received signals is
A. Training Phase 2 pc
1  (pc )
1) Clustering: Assuming that the noise at the receiver is fi = p i = 1, · · · , w, (8)
additive white Gaussian noise (AWGN). Defining Sj as a 2pc j=1 j
cluster of received points. Sj has the minimum Euclidean norm
where w is the number of candidate modulation orders for
to the j-th composite constellation point from Cpc , which can
interference user. The physical meaning of fi is the average
be written as
extent that each cluster obeys Gaussian distribution when we
(i)
Sj = {rk }, split received constellation into 2pc clusters. After performing
(i) (i) feature extraction, we can get a feature vector
s.t. ||rk − cj || < ||rk − cz ||, cj , cz ∈ Cpc , cj = cz . (4)
f = [· · · , fi , · · · ]T , i = 1, · · · , w. (9)
where j is in the range of [1 . . . 2pc ] and pc is one of candidate
composite modulation orders. To make it easier to understand, the features extracted from
In this step, all candidates of modulation order for different combinations of modulation orders and SNRs are
interference user should be used to cluster received shown in Fig. 3.

Authorized licensed use limited to: VNR Vignana Jyothi Inst of Eng & Tech. Downloaded on October 18,2023 at 04:10:20 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: ML-BASED BLIND DETECTION ON INTERFERENCE MODULATION ORDER IN NOMA SYSTEMS 2465

3) Classification Training: The goal of classification train- TABLE I


ing is to find a model parameter of logistic regression model. S IMULATION PARAMETERS
Based on the features extracted from previous step, the model
estimates the probability of modulation order that interference
user may use. From Fig. 3, we can find that clusters are well-
separated. But some cases are non-linear separable. A better
way to fit the data well is to create more features. We choose
a simple method to map features and it can be written as
f˜ = [1, f1 , f2 , · · · , fw , f12 , f22 , · · · , fw2 , · · · , f1l , f2l , · · · , fwl ],
(10)
where fi is the i-the feature in the feature vector f . l is the
TABLE II
highest degree and is also one of parameters in logistic regres-
M ODULATION C OMBINATIONS OF TARGET AND I NTERFERENCE U SER
sion. Based on this higher-dimension feature vector, the trained
model will have a more complex decision boundary. To avoid
overfitting, a regularized logistic regression is used in this step.
Defining TSNR,ptar as the training set. TSNR,ptar is made up of
mapped features and labels. The labels are actual modulation
order of interference user, which are explicitly known in the
training phase. TSNR,ptar can be written as 2) Blind Detection: Inputing the feature vector fe to the
⎡ ˜(1) T ⎤
(f ) y (1) logistic regression model, and the output is the modulation
⎢ (f˜(2) )T y (2) ⎥ order of interference user.
⎢ ⎥
TSNR,ptar = ⎢ .. .. ⎥, (11)
⎣ . . ⎦ IV. S IMULATIONS AND C OMPARISON
(f˜(m) )T y (m)
In this section, we provide a series of simulation results to
where m is the number of training examples and y (i) is the verify the efficacy and accuracy of the proposed algorithm.
label. The subscript SNR and ptar represent Signal-to-Noise We use the blind detection rate and throughput as the mea-
ratio and the modulation of target user, respectively. Then, surements of pros and cons of the proposed algorithm. The
conventional logistic regression process, such as cost func- definition of blind detection rate is R = v/Z, where R is the
tion computing and gradient descent, should be executed to detection rate, v is the number of transmissions which blindly
find optimal model parameters. Moreover, receiver operating detect the modulation order of interference user successfully
characteristic curve (ROC), learning curve and precision-recall and Z is the total number of transmissions. As the performance
(P-R) curve should be used to select the best model parameter of blind detection only affects target user, the throughputs
in training phase. of target user are compared. Moreover, the throughput com-
parison of orthogonal multiple access (OMA) and NOMA
B. Blind Detection Phase is also considered, and the definition of total throughput of
NOMA is the throughput on the same time-frequency resource
1) Model Parameter Selection Criteria: Model parameter
block. To demonstrate its effectiveness, the performance of
selection criteria is a method to choose one model para-
max-log likelihood based blind detection algorithm is used as
meter for the classification model. After the training phase,
benchmark [10]. The legends, which contain “ideal” in Fig. 5
a series of model parameters for classification model have
and Fig. 6, are best performance curves where receiver always
been established on different SNRs and ptar s. Those SNRs
knows the modulation order of interference user. Simulation
are denoted by Sl , l = 1, · · · , L, where L is the number of
parameters and modulation combinations are listed in Table I
SNRs. As shown in blind detection phase in Fig.2, the common
and Table II, respectively. The word “None” in Table II
steps in dashed block should be performed repeatedly on
represents that there is no interference user.
each candidate modulation order of interference user. After
In the training phase, the range of SNR is 0-30 dB with
that, a feature vector of received constellation points fe is
an interval of 1 dB and different types of modulation and
generated.
coding scheme (MCS) are performed to embody the superi-
Assuming that the estimated SNR is Se , and the modulation
ority of proposed algorithm. The blind detection rate curves
of target user is ptar
e . Se is generated in channel estimation of different algorithms are shown in Fig. 4, and throughput
process. Then, the model parameter whose SNR has the
comparisons are shown in Fig. 5 and Fig. 6.
minimum distance to Se and satisfies ptar = ptar e is chosen. From Fig. 4, it can be seen that the proposed algorithm
The selection criteria can be written as
is more accuracy than the conventional max-log algorithm on
ptar = ptar
e , modulation order blind detection. There are two reasons for
SNR = arg min |Se − Si |, i = 1, · · · , L (12) this phenomenon. One is that the features of received con-
i stellation points are extracted properly, and they can represent
Then model parameter can be uniquely determined by (12). the modulation order parameter. Another is that the proposed

Authorized licensed use limited to: VNR Vignana Jyothi Inst of Eng & Tech. Downloaded on October 18,2023 at 04:10:20 UTC from IEEE Xplore. Restrictions apply.
2466 IEEE COMMUNICATIONS LETTERS, VOL. 22, NO. 12, DECEMBER 2018

max-log algorithm. The throughputs of NOMA and OMA are


also addressed in Fig. 5. Due to the failure blind detection of
modulation order and imperfect SIC receiver, the throughput
for target user in NOMA systems may be worse than that of
OMA, but the total throughput of NOMA is better than that
of OMA in most SNR regions.
Fig. 6 shows the throughput comparison with different MCS
levels for both algorithms. It can be seen that the proposed
algorithm brings a mild throughput improvement with the
increasing of MCS level. This is because high MCS levels
are usually scheduled in high SNR region where the blind
detection rates for MLAD and max-log benchmark are both
optimistic.
Fig. 4. Blind detection rate comparisons on different detection algorithms.
V. C ONCLUSION
In this letter, we introduce new features to represent the
modulation order parameter, and propose an MLAD algorithm
to blindly detect that parameter of interference user in down-
link NOMA systems. The algorithm consists of two phases
and four main steps, including clustering, feature extraction,
classification training and model parameter selection criteria.
Simulation results have proved that the performance of the
proposed method outperforms max-log algorithm.

R EFERENCES
[1] S. M. R. Islam, N. Avazov, O. A. Dobre, and K.-S. Kwak, “Power-
domain non-orthogonal multiple access (NOMA) in 5G systems: Poten-
tials and challenges,” IEEE Commun. Surveys Tuts., vol. 19, no. 2,
Fig. 5. Throughput comparisons on different detection algorithms. Target pp. 721–742, 2nd Quart., 2017.
user: 16QAM (MCS=11), Interference user: QPSK (MCS=5). [2] Z. Ding, X. Lei, G. K. Karagiannidis, R. Schober, J. Yuan, and V. Bhar-
gava, “A survey on non-orthogonal multiple access for 5G networks:
Research challenges and future trends,” IEEE J. Sel. Areas Commun.,
vol. 35, no. 10, pp. 2181–2195, Oct. 2017.
[3] L. Xiao, Y. Li, C. Dai, H. Dai, and H. V. Poor, “Reinforcement learning-
based NOMA power allocation in the presence of smart jamming,” IEEE
Trans. Veh. Technol., vol. 67, no. 4, pp. 3377–3389, Apr. 2018.
[4] Y. Du et al., “Efficient multi-user detection for uplink grant-free NOMA:
Prior-information aided adaptive compressive sensing perspective,” IEEE
J. Sel. Areas Commun., vol. 35, no. 12, pp. 2812–2828, Dec. 2017.
[5] N. Zhang, J. Wang, G. Kang, and Y. Liu, “Uplink nonorthogonal multiple
access in 5G systems,” IEEE Commun. Lett., vol. 20, no. 3, pp. 458–461,
Mar. 2016.
[6] Study on Downlink Multiuser Superposition Transmission (MUST) for
LTE (Release 13), document 3GPP Tr 36.859 v13.0.0, 3GPP, Dec. 2015.
[7] LS on Potential Parameters for Blind Detection in Must, document R1-
163836, 3GPP RAN1 #84bis, Busan, Korea, Apr. 2016.
[8] A. Bayesteh, E. Yi, H. Nikopour, and H. Baligh, “Blind detection of
SCMA for uplink grant-free multiple-access,” in Proc. IEEE 11th Int.
Fig. 6. Throughput comparison on different detection algorithms. Target Symp. Wireless Commun. Syst. (ISWCS), Aug. 2014, pp. 853–857.
user: 64QAM, Interference user: QPSK. [9] O. A. Dobre, A. Abdi, Y. Bar-Ness, and W. Su, “Survey of automatic
modulation classification techniques: Classical approaches and new
trends,” IET Commun., vol. 1, no. 2, pp. 137–156, Apr. 2007.
algorithm has more prior experience than benchmark scheme. [10] H. Lee, J.-H. Lim, S. Cho, and S. Kim, “Interference cancellation based
The prior experience is learned from numbers of training on blindly-detected interference parameters for LTE-advanced UE,” in
examples in the training phase, and it is saved in model Proc. IEEE Int. Conf. Commun. (ICC), Jun. 2015, pp. 3143–3148.
[11] S. Huang, Y. Yao, Y. Xiao, and Z. Feng, “Cumulant based maximum
parameter. likelihood classification for overlapped signals,” Electron. Lett., vol. 52,
The accuracy improvement of blind detection improves no. 21, pp. 1761–1763, 2016.
the throughput performance. From Fig. 5, it is clear that [12] M. A. Stephens, “EDF statistics for goodness of fit and some compar-
isons,” J. Amer. Statist. Assoc., vol. 69, no. 347, pp. 730–737, 1974.
the throughput of the proposed algorithm is very closed [13] R. B. D’Agostino, Goodness-of-fit-Techniques. Evanston, IL, USA:
to the ideal curve, and achieves larger gains compared to Routledge, 2017.

Authorized licensed use limited to: VNR Vignana Jyothi Inst of Eng & Tech. Downloaded on October 18,2023 at 04:10:20 UTC from IEEE Xplore. Restrictions apply.

You might also like