This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2017.2666157, IEEE
Communications Letters

RSS-Based Localization in WSNs Using Gaussian

Mixture Model via Semidefinite Relaxation
Yueyue Zhang, Student Member, IEEE, Song Xing, Member, IEEE, Yaping Zhu, Student Member, IEEE,
Feng Yan, Member, IEEE, and Lianfeng Shen, Senior Member, IEEE

Abstract—Energy-based source localization methods are nor- programming (SDP) estimator has been designed specifically
mally developed according to the channel path-loss models for the RSS-based localization problems [7].
in which the noise is generally assumed to follow Gaussian Additionally, it is worth noting that traditional localiza-
distributions. In this letter, we represent the practical additive
noise by the Gaussian Mixture Model (GMM), and develop a tion algorithm designs in WSNs are based on the classical
localization algorithm based on the received signal strength (RSS) channel path-loss model with the noise generally modeled by
to achieve a maximum likelihood (ML) location estimator. By Gaussian distribution. However, in reality the noise does not
using Jensen’s inequality and semidefinite (SD) relaxation, the always follow Gaussian distribution due to the heterogeneity
initially proposed nonlinear and nonconvex estimator is relaxed of multiple sources [8]. Hence, the conventional Gaussian
into a convex optimization problem which can be efficiently
solved to obtain the globally optimal solution. Besides, the model cannot properly represents the measured noise, leading
corresponding Cramer-Rao lower bound (CRLB) is derived for to improper localization algorithms in WSNs. To the best of
performance comparison. Simulation and experimental results our knowledge, no study has utilized a practical channel path-
show a substantial performance gain achieved by our proposed loss model with non-Gaussian noise in the node localization
localization algorithm in wireless sensor networks (WSNs). algorithms in WSNs.
Index Terms—received signal strength (RSS), Gaussian mix- Motivated by above observations, in this letter we proposed
ture model (GMM), semidefinite relaxation an improved RSS-based node localization algorithm, called
Gaussian mixture-semidefinite programming (GM-SDP) esti-
I. I NTRODUCTION mator, to achieve the ML estimation of the node positions
in WSNs. Specifically, the noise is represented by a non-
W IRELESS source localization has attracted consider-
able attentions over the past decades. Among different
localization methods, energy-based localization via received
Gaussian model followed by an empirical parameter estimate
of the noise. Then we formulate the localization algorithm
signal strength (RSS) or received signal strength difference as an optimization problem to achieve ML performance. The
(RSSD) enables a simple implementation compared to other subsequent nonlinear and nonconvex optimization problem is
conventional technologies such as time of arrival (TOA) [1], solved by semidefinite relaxation to obtain the suboptimal
time difference of arrival (TDOA) [2], and angle of arrival solution. Finally, both simulation and experimental results
(AOA) [3]. Recent progress has made it practical to realize illustrate the performance gain of our proposed GM-SDP
the energy-based localization in various networks including algorithm over the traditional localization algorithms.
wireless sensor networks (WSNs) [4], wireless local area net- Notation: In the paper, Rn and Sn denote the set of n
work (WLAN) [5], and vehicular ad-hoc networks (VANETs). vectors and the n × n symmetric matrix, respectively. In
Nevertheless, RSS-based localization to achieve maximum addition, for any symmetric matrix A, A  0 means that A is
likelihood (ML) estimator of the coordinates of target nodes positive semidefinite. k · k1 , k · k2 and k · k∞ denote the `1 , `2
leads to a nonlinear and nonconvex optimization problem. and `∞ vector norms, respectively.
Several methods have been proposed to address this issue. In
[6], for example, a weighed least squares (WLS) formulation II. S YSTEM M ODEL AND P RELIMINARIES
has been derived to jointly estimate the sensor node location Let’s denote the unknown coordinates of the jth target
and the transmit power, based on the unscented transformation node as ϕj =[ϕj1 , ϕj2 ]T (ϕϕj ∈ R2 , j = 1, · · · , M ) and the
(UT). In [4], a proposed estimator approximates the ML known coordinates of the ith anchor node as αi =[αi1 , αi2 ]T
estimate for low noise level first, and then is relaxed by αi ∈ R2 , i = 1, · · · , N ), where M and N are the total
(α
applying the efficient convex relaxations that are based on the number of targets and anchors, respectively. From [4] and
second-order cone programming (SOCP). And, a semidefinite [7], the power received at the jth target from the ith anchor
This work was supported by the National Natural Science Foundation of (or vice versa) is typically modeled as
China (No. 61471164, 61601122), the Innovation Project of Jiangsu Province
(KYLX16 0222), and the Research Fund of National Mobile Communications ϕj , α i )
d(ϕ
Research Laboratory (NCRL), Southeast University (SEU) (No. 2016B02). Pi,j = P0 − 10β log10 + ni,j , (1)
Corresponding author: Lianfeng Shen.
d0
Song Xing is with the Department of Information Systems, California State where P0 denotes the transmitted power at reference distance
University, Los Angeles, CA, USA (e-mail: sxing@exchange.calstatela.edu).
The other authors are with NCRL, SEU (e-mail: douleyue@seu.edu.cn; ϕj , α i ) = kϕ
d0 from the receiver, d(ϕ ϕj −α
αi k2 is the Euclidean
xyzzyp@seu.edu.cn; feng.yan@seu.edu.cn; lfshen@seu.edu.cn). distance between the jth target and the ith anchor node, β

1089-7798 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2017.2666157, IEEE
Communications Letters

-45

RSS Measurement
finding the parameter τi,s . Then the ML estimator can be
-50

RSS(dBm) -55
logarithmic fitting
formulated as
-60

-65 max Φ(ϕ ϕj , τ )

-70
τ
ϕ j ,τ

s.t. C1 : ΣSs=1 τi,s = 1, ∀i

-75

-80
0 2 4
d(m)
6 8 10
(3)
C2 : 0 ≤ τi,s ≤ 1, ∀s, i
(a) A comparison between the empirical path loss model without
noise and the real RSS measurements. C3 : d(ϕϕj , α i ) 6= 0, ∀i ,
60 0.20
where Φ(ϕ ϕj , τ ) is the log-likelihood function of the joint
S=2

P j |ϕ

Probability density
conditional pdf in (2), i.e., Φ(ϕ ϕj , τ ) = ln p(P ϕj ), and the
Frequency counts

RSS Measurement Noise

S=3
0.15
S=1
40

0.10
mixture weights τi,s is constrained to sum up to 1.
20
However, the combinatorial nature of the objective function
0.05
in (3) results in an NP-hard problem. To reduce the complexity,
0 0.00
-50 -40 -30 -20 -10 0 10 20 the optimization problem can be reformulated by finding a
Noise(dBm)
lower bound of the original non-convex objective function.
(b) Histogram of the experimental RSS measurement noise and Then, the reformulated problem is solved as an SDP problem
different distributive profiles.
via relaxation to obtain the globally optimal solution. To be
Fig. 1. Experimental measurements analysis. concrete, we start to solve the original problem in (3) by
obtaining a suboptimal solution via the following proposition.
Proposition 1: For a feasible pair (ϕ ϕj , τ ), if the pair is an
is a path-loss exponent with the common value between 2 optimal solution to the following optimization problem
and 6, and the additive noise ni,j following the Gaussian N X
S
distribution represents the log-normal shadowing effect in
X
max τi,s ln N (µs , σs2 )
multipath environments, respectively. ϕj
i=1 s=1
(4)
However, there are two related issues with (1). One is s.t. C1, C2, C3 ,
related to RSS level behavior which is illustrated in Fig. 1
(a). We adopt the logarithmic curve for the fitting of the the feasible pair (ϕϕj , τ ) is suboptimal for the primal problem
RSS measurements. Another is as mentioned in Section I in (3).
that ni,j may not always follow Gaussian distributions in real Proof : The optimal value of ϕ j in (3) is defined as Φ∗ =
environment. In this letter, we aim at developing an improved sup{Φ(ϕ ϕj , τ )|C1, C2, C3}, and a feasible solution (ϕ ϕj , τ )
channel path-loss model by capturing the characteristics of the is  − suboptimal if Φ(ϕ ϕj , τ ) ≥ Φ∗ − . Using Jensen’s
additive noise, which is shown by the histogram of the RSS inequality, the lower bound of the primal non-convex weighted
measurement noise in Fig. 1 (b). objective function of (3) can be obtained and given by
Here, we represent the noise with a Gaussian mixture N X
X S
model (GMM) and the resulting joint conditional probability ϕj , τ ) ≥
Φ(ϕ τi,s ln N (µs , σs2 ) = Φ1 (ϕ
ϕj , τ ). (5)
distribution function (pdf) of observed power vector P j = i=1 s=1
[P1,j , . . . , PN,j ] at the jth target is Obviously, Φ1 (ϕ ϕj , τ ) serves as a lower bound for the log-
S
N X likelihood function of the parameters (ϕ ϕj , τ ). Hence, there
always exists 0 that makes (ϕ ϕj , τ ) ≥ Φ∗ −0 .
Y
P j |ϕ
p(P ϕj ) = τi,s N (µs , σs2 ), (2) ϕj , τ ) satisfy Φ1 (ϕ
i=1 s=1 ϕj , τ ) is 0 −suboptimal for the problem in (3) and
Therefore, (ϕ
where τi,s is the corresponding weight of the cluster s with the proposition holds. Furthermore, the optimization problem
mean µs and variance σs2 of the noise, and S is the number in (3) can be reformed as
of mixture component. ϕj , τ )
min Φ2 (ϕ
τ
ϕ j ,τ
Remark: The path-loss exponent is known and fixed a-priori (6)
in following the simulations and experiments. By using prior s.t. C1, C2, C3 ,
experimental measurements, P0 and β can be estimated by the where
logarithmic fitting, and fixed as P0 = −55 dBm, β = 2. ϕj , τ ) = −Φ1 (ϕ
Φ2 (ϕ ϕj , τ )
N X S
√ (ni,j − µs )2 (7)
 
III. P ROBLEM F ORMULATION AND S OLUTION
X
= τi,s ln 2πσs + 2
.
In practice, the variables µs , σs2 and τi,s in (2) are need to i=1 s=1
2σs
be estimated in the localization process. In the ECM criterion Clearly, the objective in (6) is non-convex since its domain
(η) 2,(η) ϕ α
[1], both µs and σs on the iteration η are estimated for {ϕ ϕj 6= α i } is not continuous, and log10 d(ϕdj0,α i ) is not
ϕj |ϕ
(η) (η)
ϕ
updating the position j without τi,s . In this letter, both convex on the convex domain. Hence, it is difficult to find the
(η) (η)
the position ϕ j and τi,s would be jointly optimized. For the globally optimal solution to the problem in (6). To achieve a
sake of convenience, the index of the iteration η is not shown. convex estimator in (6), we introduce the following lemma.
To this end, the objective of designing the localization Lemma 1: If x ∈ √ Rn , we have
algorithm via GM-SDP is to obtain the ML estimate ϕ ∗j by 1) kx
xk∞ ≤ kx xk2 ≤ nkx xk∞ , 2) kxxk∞ ≤ kxxk1 ≤ nkx xk∞ .

1089-7798 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2017.2666157, IEEE
Communications Letters

By replacing the `2 norm in Φ2 with the `∞ norm, i.e., the Now (11) is a convex optimization problem, which can
Chebyshev norm, we have be solved by the existing numerical tools [10] to obtain
  the globally optimal solution ϕ ∗j of the original optimization
(ni,j − µs )2 (a) 1 d2 (ϕϕj , α i )  problem in (3).
⇔ max  log10

2σs2  σs 2
γi,s


" # (8)
(b) d2 (ϕ
ϕj , α i ) 2
γi,s IV. C RAMER -R AO L OWER B OUND ON E STIMATOR
⇐⇒ max 2 , , ξi,s , VARIANCE
σs γi,s σs d2 (ϕ
ϕj , α i )
It is easily found that our GM-SDP estimator ϕ P ) is
b (P
P0 +µs −Pi,j
where = 2
γi,s d20 10
. In (8), (a) results due to 1)
5β P )] = ϕ (P
ϕ (P
unbiased, i.e., E[b P ). The accuracy of an estimator
in Lemma, and (b) results based on that log10 x is a strictly may be illustrated by some costs associated with the estimate.
monotonic function when it is increasing in its entire domain The most commonly used cost is the mean squared error
(0, +∞). Then the primal optimization problem in (6) can be (MSE), or the estimator variance for an unbiased estimate.
expressed as It is well known that the Cramer-Rao lower bound (CRLB),
which is the inverse of the Fisher information matrix (FIM),
N
X represents a lower bound on the variance of any unbiased
Tr(ττ iη T ) + Tr(ττ iξ Ti )
 
min
τ,ξ
ϕ j ,τ,ξ estimators. Specifically, the variance of our GM-SDP estimator
i=1
will meet
s.t. C1, C2, C3 (9) n o
ϕj − α i k22 ≤ γi,s
C4 : kϕ 2
σs ξi,s E [b P ) − ϕ ] [b
ϕ (P P ) − ϕ ]T  J −1 ,
ϕ (P (12)
2 −1 −1
ϕj − α i k22 ≥ γi,s
C5 : kϕ σs ξi,s , ∀i, s , where the element [J ]v,r of FIM J is defined by
where
 
P j |ϕ
∂ ln p(P P j |ϕ
ϕj ) ∂ ln p(P ϕj )
τ i√= [τi,1 , . . . , τi,S T [J ]v,r = E · , v, r ∈ V. (13)
√] , T ∂ϕj,v ∂ϕj,r
η = [ln 2πσ1 , . . . , ln 2πσS ] ,
ξ i = [ξi,1 , . . . , ξi,S ]T . In(13), V is denoted as the set of dimensions in the coor-
Noting that kϕϕj − α i k22 = Tr(Ψ Ψ) − 2ϕ ϕTj α i + kα
αi k22 . Hence dinate axis. For a specific target node in our two-dimensional
we rewrite the reformulated problem in norm form as scenario (|V| = 2), we have
 2 X N
min kx
xk1 + kyy k1 10β (ϕj,v − αi,v )(ϕj,r − αi,r )
τ,ξ
ϕ j ,τ,ξ [J ]v,r = In , v, r ∈ V,
ln 10 φj − αi k42
kφ
s.t. C1, C2, C3 i=1
(14)
C4 : Tr(Ψ ϕT
Ψ) − 2ϕ αi k22 ≤ γi,s
j α i + kα
2
σs ξi,s where
2 −1 −1
C5 : Tr(Ψ ϕT
Ψ) − 2ϕ αi k22 ≥ γi,s
j α i + kα σs ξi,s , ∀i, s (10) ( 2 )
∇n p(n) [∇n p(n)]2
Z
C6 : Ψ = ϕ jϕ T
j ,Ψ ∈ S
2 In = E = dn. (15)
p(n) p(n)
C7 : xi = Tr(ττ iη T )
PS
C8 : yi = Tr(ττ iξ T In (15), p(n) ∼ s=1 τs N (µs , σs2 ), and In can be evalu-
i ), ∀i .
ated numerically by Monte Carlo integration [1]. Further, by
However, the problem in (10) is still non-convex since the defining the location estimation error as e = kbϕ −ϕ ϕk2 , its root
equality constraints C6 and C8 are not affine. To obtain a mean square error (RMSE) is lower bounded by [7]
convex optimization problem in (10), we relax C6 into an p p
inequality constraint Ψ  ϕ j ϕ Tj (semidefinite relaxation), E(e2 ) ≥ Tr[J −1 ] , CRLBCRLB(ϕ ϕ). (16)
PS PS
and relax C8 into yi = s=1 τi,s ξi,s ≥ s=1 ξi,s (Jensen’s
inequality). Further, using Schur complement, the constraints V. R ESULTS A NALYSIS
C5 and C6 in (10) can be expressed in a linear matrix A. Simulation Results: Influence of the number of Anchors
inequality (LMI) [9], shown in the rewritten optimization
Numerical simulations are conducted to evaluate the per-
problem as below.
formance of our proposed GM-SDP algorithm. We consider a
min kx
xk1 + kyy k1 two-dimensional square region with the size of 15 × 15 m2
τ,ξ
ϕ j ,τ,ξ
for the simulations. In the network, 120 sensors (including
s.t. C1, C2, C3, C4 20 anchors) are uniformly deployed. Due to space limit, the
√ 
ϕTj α i + kα
αi k22

Ψ) − 2ϕ
Tr(Ψ γi,s / σs detailed strategy for anchor selection is omitted. The noise is
C5 : √ ≥ 0, ∀i, s
γi,s / σs ξi,s assumed to follow a two-mode Gaussian mixture distribution

Ψ ϕj
 with τ1 = 0.37, µ1 = −4.36 dBm, σ1 = 5.22 dBm, τ2 = 0.63,
C6 : T ≥ 0, Ψ ∈ S2 µ2 = 1.73 dBm, σ2 = 4.09 dBm. Here, the setting of the mean
ϕj 1
S
and variance of each mixture component is an example of the
C7 : xi ≥ Tr(ττ iη T ) C8 : yi ≥
X
ξi,s , ∀i, s . realistic indoor noise measurements as mentioned in Section
s=1
I. The performance is evaluated in terms of RMSE using 100
(11) Monte Carlo runs with the number of anchors varying from 4

1089-7798 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2017.2666157, IEEE
Communications Letters

6 TABLE I
L OCALIZATION ACCURACY OF D IFFERENT A LGORITHMS
5
4
RMSE(m)

Methods RMSE(m) Maximum Error(m) p(|e| ≤ 2m)

3 WLS 3.43 15.58 31.79%
WLS
2 BFGS-3
SDP-3 2.23 8.71 59.28%
1
SDP-3
BFGS-3 2.96 13.90 42.05%
GM-SDP-2

BFGS-4 1.83 7.63 55.38%

0
CRLB

4 6 8 10 12 14 16 18 20 GM-SDP-2 2.06 9.86 53.84%

N
GM-SDP-3 1.67 6.56 63.08%

(a) RMSE versus the number of anchors.

1.0

0.9
The numerical results in Table I illustrate clearly the
0.8 performance comparison of various estimators on the
0.7

0.6 experiments. From Table I, we observe that GM-SDP-3

CDF

0.5

0.4
WLS

SDP-3
performs the best with the least value of RMSEs and the
0.3
BFGS-3

BFGS-4
maximum localization estimation errors among all methods.
0.2

0.1
GM-SDP-2

GM-SDP-3
In addition, the GM-SDP estimator with 2-component
0.0

0 1 2
error(m) 3 4 5
performs even better than BFGS with 3-component GMM,
which has shown the robustness of the ECM estimate of the
(b) CDF of localization errors for the different algorithms.
noise parameters and also confirmed the observations in [1].
Fig. 2. Localisation performance of the different algorithms.

VI. C ONCLUSION
to 20. In this letter, the additive noise in RSS measurement is
Fig. 2 (a) illustrates the RMSE vs. the number of anchors modeled by the Gaussian mixture model, and an improved
for various estimators including BFGS [1], WLS [6], SDP [7], node localization algorithm employing relaxations is proposed
and our proposed GM-SDP. CRLB value is also illustrated to achieve ML estimate of the node locations in wireless
in Fig. 2 (a) for comparison. It shows that increasing the source localization systems, in which the primal nonlinear and
number of anchors generally improves the estimate accuracy nonconvex optimization problem is transformed into a SDP
of all algorithms, revealing that more anchors will increase the problem to obtain the globally optimal solution. Simulation
dimension as well as facilitate accurately determination of the and experimental results are presented to confirm the perfor-
RSS measurements. Clearly, our proposed GM-SDP with only mance improvement of the proposed GM-SDP algorithm.
fewer mixture components needed to be measured for GMM
performs much better than other compared algorithms, which R EFERENCES
is indicated from the closest RMSE of GM-SDP to CRLB. [1] F. Yin, C. Fritsche, D. Jin, F. Gustafsson, and A. M. Zoubir, “Cooperative
Localization in WSNs Using Gaussian Mixture Modeling: Distributed
ECM Algorithms, IEEE Trans. Signal Process., vol. 63, no. 6, pp. 1448-
B. Experimental Results 1463, Mar. 2015.
[2] Y. Liu, F. Guo, L. Yang, and W. Jiang, “An improved algebraic solution
The performance of proposed GM-SDP algorithm is for TDOA localization with sensor position errors,” IEEE Commun. Lett.,
further evaluated by the experiments in a real wireless sensor vol. 19, no. 12, pp. 2218-2221, Dec. 2015.
network with RSS measurements. Without loss of generality, [3] S. F. Chuang, W. R. Wu, and Y. T. Liu, “High-Resolution AoA Estimation
for Hybrid Antenna Arrays,” IEEE Trans. Antennas Propag., vol. 7, no.
five Zigbee nodes (TI-CC2530), among which one is the 63, pp. 2955-2968, Jul. 2015.
target node and the other four are anchors, are used in the [4] S. Tomic, M. Beko, and R. Dinis, “RSS-Based Localization in Wireless
experiments. The network environment is a typical meeting Sensor Networks Using Convex Relaxation: Noncooperative and Cooper-
ative Schemes,” IEEE Trans. Veh. Technol., vol. 64, no. 5, pp. 2037-2050,
room where the anchors are fixed at the corners of the room May. 2015.
with the coordinates (0, 0) m, (6, 0) m, (6, 7) m, and (0, [5] W. Y. Hu, J. L. Lu, S. Jiang, W. Shu, and M. Y. Wu, “WiBEST: A
7) m, respectively. The RSS measurements are randomly hybrid personal indoor positioning system,” in Proc. IEEE WCNC, pp.
2149-2154, 2013.
collected in 195 different locations of this area. Fig. 2 (b) [6] G. Wang, H. Chen, Y. Li, and M. Jin, “On Received-Signal-Strength
illustrates the CDF of the localization estimation errors of Based Localization with Unknown Transmit Power and Path Loss Ex-
various algorithms including GM-SDP. It shows that the ponent,” IEEE Wireless Commun. Lett., vol. 1, no. 5, pp. 536-539, Oct.
2012.
performance of SDP-3 estimator is close to GM-SDP-2 and [7] R.W. Ouyang, A. K. Wong, and C. T. Lea, “Received Signal Strength-
BFGS-4, and BFGS-4 performs better than BFGS-3 due to its Based Wireless Localization via Semidefinite Programming: Noncoopera
approximation of the probability density of the measurement tive and Cooperative Schemes,” IEEE Trans. Veh. Technol., vol 59, no.
3, pp. 1307-1318, Mar. 2010.
noise with more Gaussian components. For the same reason [8] D. Wu, Y. Zhang, L. Bao, and A. C. Regan, “Location-Based Crowd-
as BFGS-4 vs. BFGS-3, our proposed GM-SDP-3 provides sourcing for Vehicular Communication in Hybrid Networks,” IEEE Trans.
the better performance than GM-SDP-2. More importantly, Intell. Transp. Syst., vol. 14, no. 2, pp. 837-846, June 2013.
[9] S. P. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
GM-SDP-3 performs the best among all algorithms and even Cambridge University Press, 2004.
more than 63.08% of its localization estimation errors are [10] M. C. Grant and S. P. Boyd, The CVX Users’ Guide Release 2.1.
within a very short range of 2 m. http://cvxr.com/cvx/doc/CVX.pdf, Oct. 2016.

1089-7798 (c) 2016 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.