Professional Documents
Culture Documents
Abstract
Considering a large-scale energy-harvesting wireless sensor network (EH-WSN) measuring compressible data, sparse
random projections are feasible for data well-approximation, and the sparsity of random projections impacts the
mean square error (MSE) as well as the system delay. In this paper, we propose an adaptive algorithm for sparse
random projections in order to achieve a better tradeoff between the MSE and the system delay. With the
energy-harvesting constraints, the sparsity is adapted to channel conditions via an optimal power allocation
algorithm, and the structure of the optimal power allocation solution is analyzed for some special case. The
performance is illustrated by numerical simulations.
Keywords: Sparse random projection; Energy harvesting; Wireless sensor network
© 2015 Ran and Oh; licensee Springer. This is an Open Access article distributed under the terms of the Creative Commons
Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction
in any medium, provided the original work is properly credited.
Ran and Oh EURASIP Journal on Wireless Communications and Networking (2015) 2015:113 Page 2 of 10
WSNs. Data collected from wireless sensors are typically problem which is converted into an optimal power
correlated and thus compressible in an appropriate trans- allocation problem for maximizing the system
form domain (e.g., the Fourier transform or wavelet)[8]. throughput. Unlike the conventional energy
Therefore, the main ideal of CS is that n data values can allocation problem, due to battery dynamics and
be well-approximated using only k << n transform coef- channel dynamics, the closed-form solution may not
ficients if the data are compressible [8-13]. In particular, be available. Therefore, we study a special case that
Wang et al. [13] propose a distributed compressive sens- the battery capacity is not bounded to find the
ing scheme for WSNs in order to reduce computational structure of the optimal solution. Specifically, in the
complexity and the communication cost. It considers an case the problem is converted into a convex
m × n sparse random matrix with entries that have a prob- optimization problem, then the closed-form solution
ability g of being nonzero, so that on average there are is obtained in terms of Lagrangian multipliers.
ng nonzeros per row. The resulted data-approximation
error rate is comparable to that of the optimal k-term The rest of paper is organized as follows. Section 2 gives
approximation if the energy of the signal is not concen- the system model and overview previously known results
trated in a few elements. Somehow, the sparsity factor on sparse random projections. Section 3 redefines the
g of random projections impacts the accuracy of signal sparsity and formulate the optimal sparsity problem for
reconstructions. Usually, the sparsity factor g is statisti- EH-WSNs. Section 4 considers a specific case and address
cally determined according to the amount of harvested the structure of the optimal solution. Section 5 provides
energy and is homogeneous for all sensors [14,15]. Rana the simulation results. Finally, Section 6 concludes the
et al. [14] only considered AWGN channels. Yang et al. paper.
[15] took into account fading channels and studied the
sufficient conditions guaranteeing a reliable and compu- 2 System model
tationally efficient data approximation for sparse random We consider a wireless sensor network of n sensor nodes,
projections. It is not surprising that the sparse random each of which measure a data value si ∈ C and is capable
projections based signal recovery is non-optimal since the of energy harvesting. We assume a Rayleigh-fading chan-
sparsity factor g is fixed during the entire transmission nel, and the channel coefficients, denoted as hij , where
slots and thus can not reflect the effect of channel condi- 1 ≤ i ≤ m denotes the slot index and 1 ≤ j ≤ n
tions. On the other hand, the system delay m, which is one denotes the sensor index, are independent and identically
of key quantities used to characterize the performance of distributed (i.i.d) and satisfy the complex Gaussian dis-
random projection-based CS schemes, is expected to be tribution with zero mean and unit variance. We further
as small as possible. Upon [14] and [15], we realize that assume the channel remains constant in each slot. Sensor
the lower bound of the system delay also being related to j first multiplies its data sj by some random projections
the sparsity of random projections and the larger g being φij ∈ R, then transmits in the ith time slot. At the receiver
the shorter delay may be achieved. Note that there is often
a tradeoff between the system delay and the data approxi-
n
mation [8,16]. Therefore, in this paper, we consider fading yi = hij φij sj + ei (1)
j=1
channels and the energy-harvesting constraints and study
the problem on adapting sparsity of random projections
according to full channel information, in order to improve where ei is a White Gaussian distributed noise with zero
the performance of signal recovery and reduce the sys- mean and variance σ 2 . After m time slots, the received
tem delay as well. To the best of our knowledge, very vector is given as
limited work such as [15] has touched upon this topic,
which however, only provides rough discussion. The main y = (H )s + e = Zs + e (2)
contributions are presented as follows:
where H = hij ∈ Cm×n , = φij ∈ Rm×n , Z = H ,
• Considering the wireless fading channels, we verify and the operation is the element-wise product of two
that the random projection matrix satisfies the matrices. The corresponding real-valued equation of (2) is
property that the inner product between two
projected vectors are preserved in expectation, and
Re {(H )} −Im {(H )} Re {s} Re {e}
then provide a lower bound of the system delay for ŷ = +
Im {(H )} Re {(H )} Im {s} Im {e]
achieving an acceptable data approximation error
probability. Re {s} Re {e}
= Ẑ +
• We give a new definition of the sparsity of random Im {s} Im {e}
projections and formulate the optimal sparsity (3)
Ran and Oh EURASIP Journal on Wireless Communications and Networking (2015) 2015:113 Page 3 of 10
where Re {A} and Im {A} denote the real part and the in addition, Let V be any set of n vectors {v1 , · · · , vn } ⊂ Rn .
imaginary part of matrix A, respectively, and Suppose a sparse random matrix ∈ Rm×n satisfies the
conditions as
Re {(H )} −Im {(H )}
Ẑ = (4) E[ φij ] = 0, E[ φij2 ] = 1, E[ φij4 ] = g
Im {(H )} Re {(H )} (9)
n with probability at least 1 − n−γ , the random projections
s = x = ψj xj (5) √1 u and √1 vi can produce an estimate âi for uT vi
m m
j=1
satisfying
where g is a factor which gives the probability of a mea- 3 Adaptive sparse random projections
surement and controls the degree of sparsity of random 3.1 Sparse random projections with channel fading
projections, e.g., if g = 1, the random matrix has no However, Wang et al. [13] only considered AWGN chan-
sparsity, and if g = log n/n, the expected number of nel. With the assumption of channel fading, we wonder
nonzeros in each row is log n. We can easily verify that the whether the inner products are still preserved by sparse
entries within each row are four-wise independent, while random projections. Redefine the sparse random projec-
the entries across different rows are fully independent, i.e., tion matrix as follows,
⎧
E[ φij ] = 0, E[ φij2 ] = 1, E[ φij4 ] = g +1 w.p. gij /2
1 ⎨
(7)
ψij = √ 0 w.p. 1 − gij (13)
g ij ⎩
Therefore, each random projection vector is pseudo- −1 w.p. gij /2
randomly generated and stored in a small space.
where gij gives the probability of a projection from sensor
Corollary 1. [13] Consider a data vector u ∈ Rn which node j at time slot i. The details of gij will be illustrated in
satisfies the condition the next section.
projection of two vectors x and y. Expectation and variance u − û2 ≤ (1 + )η u2 (19)
2 2
of the inner product of u and v are respectively
with probability at least 1 − n−γ , if the k largest transform
E uT v = xT y (14) coefficients in magnitude give an approximation with error
u − ûopt 2 ≤ η u2 .
2 2
⎛ ⎞
n
1 ⎝ T 2 2 3 Proof: Follow the approach of [13] and define m =
T
Var u v = x y + x22 y2 + 2 2⎠
− 3 xj yj m1 m2 . Partition the m × n matrix into m2 matrices
m gij
j=1
1 , 2 , · · · , m2 , each of size m1 ×n. Using the Chebyshev
(15) inequality, we have
n
O . For any pair of vectors u and vi , the
2 2
in EH-WSNs [14], where Ej denotes the harvested energy is the limited battery capacity. The battery overflow hap-
profile for node j. Usually, gj is predetermined and uniform pens when the reserved energy plus the harvested energy
regardless of nodes and time slots, i.e., gj = g. Obviously, exceeds the battery capacity, which, however, is not pre-
it is not a sophisticated definition because it does not con- ferred because the data rate can be increased if the energy
sider the different channel conditions of nodes and times is used in advance instead of overflowed. If we assume that
as well as the energy-harvesting constraints. Therefore, a there is an m which satisfies the condition (24), the opti-
more specific definition on sparsity is desired. We redefine mal problem minimizing the system delay is immediately
the sparsity of random projections as follow, converted into a throughput maximizing problem, which
can be formulated as follows:
p∗ij
gij = (21)
Eij
m 2
hij pij
maxpij log2 1 + n 2
where p∗ij is the allocated energy for node j during the ith i=1 l=1,l =j |hil | Eil + σ2
time slot. p∗ij is determined in term of full information s.t. (27)
consisting of past and present and future channel condi-
tions and amount of energy harvested. The case of full
i
i−1
pkj ≤ Ekj , k = 1, 2, · · · , m (28)
information may be justified if the environment is highly
k=1 k=1
predictable, e.g., the energy is harvested from the vibra-
tion of motors that turned on only during fixed operating
i
i
hours and line-of-sight is available for communications. Ekj − pkj ≤ Emax , k = 1, 2, · · · , m − 1. (29)
If the energy-harvesting profile Eij for each node is k=0 k=1
known in advance and kept constant during all transmis- pij ≥ 0, ∀i
sion time slots, the optimal sparsity problem is converted
into an optimal power allocation problem. But the rising
Note that the objective (28) is convex for all i since
question is which performance measurement will be used
it is a sum of log functions, and others are all affine
for power allocation. We know thta the performance of
constraints. Consequently, the optimization problem is a
random projection-based CS schemes is characterized by
convex optimization problem, and the optimal solution
two quantities, i.e., the data approximation error proba-
satisfies the Karush-Kuhn-Tucker (KKT) conditions [17].
bility (or the mean square error (MSE)) and the system
With the assumption that the initial battery energy E0j is
delay. Note that there is often a tradeoff between these two
always known by node j, define the Lagrangian function
quantities [16]. Under an allowable MSE η > 0, we thus
for any multipliers λi ≥ 0, μi ≥ 0, βi ≥ 0 as
define the achievable system delay D(η) as
i
i
Ekj − pkj ≤ Emax , k = 1, 2, · · · , m − 1. (26)
with additional complementary slackness conditions
k=0 k=1
pij ≥ 0, ∀i
i
i−1
m |hij |2 pij λi pkj − Ekj = 0, ∀i (31)
where i=1 log 2 1+ n 2 is the lower
l=1,l =j |hil | Eil +σ 2 k=1 k=1
bound of short-term throughput of node j and B is the
i i
required data information to transmit for each node. The μi Ekj − pkj ≤ Emax = 0, i < m (32)
constraint (26) is due to that the harvested energy can- k=0 k=1
not be consumed before its arrival, and the constraint (27) βi pij = 0∀i (33)
Ran and Oh EURASIP Journal on Wireless Communications and Networking (2015) 2015:113 Page 6 of 10
We apply the KKT optimality conditions to the 0, ∀k, we have αi ≤ αi+1 . If αi ≤ αi+1 , by defini-
−1
Lagrangian function (30). By setting ∂L/∂pij = 0, we tion αi = ln 2 m k=i λk , we get λi = 0 and λi >
obtain the unique optimal energy level p∗ij in term of 0. So the complementary slackness condition (32) only
Lagrange multipliers as i i−1
holds when k=1 pkj − k=1 Ekj = 0, which means
all stored energy should be used up before the current
1 ∗ transmission.
p∗ij = αi − (34)
γi
The case of Emax = ∞ represents an ideal energy
−1 buffer which refers to a device that can store any amount
where αi = ln 2 m
k=i (λk − μk )βi , μm = 0 and γi = of energy, does not have any inefficiency in charging,
n
|hij |
2
and does not leak any energy over time. As an example,
2 .
l=1,l =j |hil | Eil
consider a sensor node installed to monitor the health
of heavy duty industrial motors. Suppose the node oper-
3.3 Structural solution
ates using energy harvested from the machine’s vibra-
If the battery capacity is finite, the optimal water-level tions, the harvested energy is greater than the consumed
is not monotonic. Therefore, the structure of the opti- power and the health monitoring function is desired only
mal energy allocation cannot be described in a sim- when the motor is powered on. Proposition 3 presents
ple and clear way, and an online programming may be an analytically tractable structure of the optimal spar-
required. Since we are more interested in an offline sity. Intuitively, the harvested energy is reserved in the
power allocation structure, we study the following special battery for the use in the later transmission, in order
case. to reduce the effect of causality constraint and improve
the flexibility of harvested energy allocation. The opti-
mal water level can be obtained by the power allocation
Proposition 3. if Emax = ∞, the optimal water lev- policy and it is structured as follows: the water level
els are non-decreasing as αi ≤ αi+1 . In addition, the is non-decreasing and the harvested energy is used in
water level changes when all the energy harvested before a conservative way. Based on the structural properties,
the current transmission are used up. we can use the following reserve multi-stage waterfilling
Proof: Without the battery capacity constraint, the water algorithm modified based on [18], to achieve the
−1
level is given as αi = ln 2 m k=i λk . Since λk ≥ solution:
0
10
g=1
g=1/2
g=1/4
g=1/8
Mean Square Error
−1
10
−2
10
20 40 60 80 100 120 140
Transmission slots (m)
Figure 1 The MSEs comparison for sparse random projections with different degrees of sparsity.
Ran and Oh EURASIP Journal on Wireless Communications and Networking (2015) 2015:113 Page 7 of 10
Algorithm 1: Reserve multi-stage waterfilling of sparsity. The larger g is given, the smaller MSE is
algorithm with harvested energy [18] achieved. However, a larger g may bring great computa-
tional complexity. Therefore, the sparsity factor g should
1: Set t0 = 0, γ̂i = γi and Êij = Eij for i = 1, . . . , m
2: for all i = 1 to m do be carefully chosen in order to keep a balance between
3: for all k = m to ti−1 + 1 do the MSE and the complexity. Intuitively, when channel
4: Find αl so that kl=ti−1 +1 plj = k−1 Ê conditions are not good, a larger g should be selected
∗ l=ti−1 +1 lj
and plj = αl − γ̂ 1 for guaranteeing an acceptable MSE, whereas a smaller
l
−1 g should be selected for saving the computational com-
5: Update γ̂l = plj + γ̂1 plexity when channel conditions are good enough. This
k−1 l
6: If l=1 Êlj ≥ kl=1 plj , then ti = k motivates us to study adapting the sparsity of random pro-
7: end for k jections according to channel conditions for improving
8: If ti = m then exit the data-approximation performance as well as the system
9: endfor i
∗ delay.
10: ∗
pij = γ̂ − γ1i , i = 1, . . . , m
1
i Figures 2 and 3 compare the MSE performance obtained
by our proposed adaptive sparse random projection
(denoted as ‘Adaptive’ in the legend) with that obtained by
4 Simulation results the conventional sparse random projections (denoted as
We consider a EH-WSN containing n = 500 sensor nodes, ‘Fixed’ in the legend) with respect to the number of trans-
and a uniform energy-harvesting rate Eij = 2 dB for mission slots m for SNR = 15 dB and 30 dB, respectively.
all nodes. We evaluate the performance of the proposed The conventional sparse random projections with a fixed
adaptive sparse random projections. One of performance sparsity given as g = 1/4 is looked as a baseline since
measurements is the mean-square error (MSE) given as it achieves an acceptable MSE with a modest complex-
ity. We observe that the proposed adaptive sparse random
s − ŝ2 projections achieves better tradeoff between the MSE and
2
error = (35) the system delay than the conventional one does when k
s22
is either 10 or 5. However, the performance gap between
Figure 1 illustrates the data approximation performance the proposed scheme and the conventional one is getting
using sparse random projections for the different degrees smaller when SNR increases. That makes sense because
SNR=15dB
0
10
Fixed, k=10
Adaptive, k=10
Fixed, k=5
Adaptive, k=5
−1
10
Mean Square Error
−2
10
−3
10
20 40 60 80 100 120 140
Transmission slots (m)
SNR=30dB
0
10
Fixed, k=10
Adaptive, k=10
Fixed, k=5
Adaptive, k=5
−1
10
Mean Square Error
−2
10
−3
10
20 40 60 80 100 120 140
Transmission slots (m)
when the channel conditions is getting better, the bene- a fixed sparsity and the proposed one with respect to the
fits from the adaptive sparsity become limited. For both number of transmission (or the system delay) m for dif-
SNR = 30 dB and 15 dB, we notice that the case of k = 5 ferent SNRs. We still observe that the proposed scheme
provides better performance than the case of k = 10. outperforms the conventional one for both SNR = 20 dB
In Figure 4, we present the performance comparison and 30 dB resulting in a better tradeoff between the MSE
between the conventional sparse random projection with and the system delay. We also notice that, for both the
k=10
0
10
Fixed, SNR=20dB
Adaptive, SNR=20dB
Fixed, SNR=30dB
adaptive, SNR=30dB
Mean Square Error
−1
10
−2
10
20 40 60 80 100 120 140
Transmission slots (m)
SNR=30dB, k=5
130
Fixed
Adaptive
120
110
100
90
System delay (m)
80
70
60
50
40
30
0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Mean Square Error
Figure 5 Tradeoff between the MSE and the system delay for SNR = 30 dB and k = 5.
proposed scheme and the conventional scheme, there is better tradeoff between the MSE and the system delay.
not a performance difference between the case of SNR The optimal sparsity problem is turned into an opti-
= 20 dB and that of SNR = 30 dB when m < 80, but mal power allocation maximizing throughput with the
the MSE decreases as SNR increases when m is over 80. energy-harvesting constraints. An offline power alloca-
That is because m is also one of factors which control tion structure is available for a special case that the battery
the variance of the estimation illustrated in (15). If m is capacity is infinite. Simulation results have shown that
not sufficiently large, it is one of dominant factors which the proposed scheme achieves smaller MSEs than the
effect the MSE performance. Therefore, increasing SNR conventional scheme. Meanwhile, the proposed scheme
barely impacts the MSE performance. While m is large can also reduce the system delay given an accepted
enough, a very limited improvement of the MSE may be error rate. However, full channel information may not be
achieved by further increasing m, but SNR now becomes a always available. Therefore, for future work, we will study
dominant factor and increasing SNR may benefit the MSE adaptive sparse random projections with partial channel
performance. information.
Figure 5 shows tradeoffs between the system delay and Competing interests
the MSE for the proposed adaptive sparse random projec- The authors declare that they have no competing interests.
tions and the conventional ones when SNR = 30 dB and
k = 5. Consider the MSE 3×10−2 , the conventional sparse
Acknowledgements
This research was supported by Basic Science Research Program through the
random projection requires about m = 95 times trans- National Research Foundation of Korea (NRF) funded by the Ministry of
mission, while the proposed scheme only requires m = 78 Science, ICT & Future Planning (Grant. NRF-2012R1A1A1014392 and
NRF-2014R1A1A1003562).
times transmission. Consequently, the proposed scheme
achieves a better tradeoff compared to the conventional Author details
1 Department of Electrical and Computer Engineering, Ajou University, World
one.
cup-ro, Suwon, South Korea. 2 Department of Electrical and Information
Engineering, Soongsil University, Sangdo-ro, Seoul, South Korea.
5 Conclusions
In this paper, we proposed to adapt sparsity of random Received: 15 November 2014 Accepted: 10 March 2015
projections according to full channel information for EH-
WSNs. Compared to the conventional sparse random
References
projections which keep the sparsity constant for the 1. V Raghunathan, S Ganeriwal, M Srivastava, Emerging techniques for long
whole transmission slots, the proposed one achieves a lived wireless sensor networks. IEEE Comm. Mag. 44, 108–114 (2006)
Ran and Oh EURASIP Journal on Wireless Communications and Networking (2015) 2015:113 Page 10 of 10