978-1-5090-0690-8/16/$31.00 ©2016 IEEE
Improved Distributed Compressed Sensing for
Smooth Signals in Wireless Sensor Networks
Boyu Li, Fei Gao*, Xiaoyu Liu, Xia Wang
Key Lab of Wireless Sensor Networks, Yunnan Minzu University
Kunming, China, 650500
{liboyu_wsn, ynkmgaofei, xiaoyu1846916, wangxiacsu}@163.com
Abstract— the technology of the distributed compressed
sensing is thought as an extension of compressed sensing and it
makes applying multiple signals into compressed sensing possible.
A vital issue in distributed compressed sensing is to minimize the
difference between the original signal and the recovery signal. In
this paper, we improve the distributed compressed sensing for
smooth signals in wireless sensor networks. Firstly, we put
forward a new weighted method to obtain the common
component of all signals, and then one method of lossy coding for
shortening the length of common component is proposed. Most
importantly, we improve the calculation formula of the
distributed compressed sensing to ensure that the common
component can be received losslessly. The numerical results show
that,comparing with the distributed compressed sensing, the
improved distributed compressed sensing not only can use much
fewer measurements to recover the original signal, but also
enable the effect of signal recovery to be better than that of
traditional distributed compressed sensing.
Keywords— Distributed Compressed Sensing, JSM-1, Smooth
Signals, Wireless Sensors Networks
I. I
NTRODUCTION
The theory of compressed sensing (CS) was proposed by
Candès [1] and Donoho er.al [2] in 2006. The key idea of
compressed sensing is to recover a sparse signal from very few
non-adaptive, linear measurements by convex optimization
method. Comparing with traditional sampling method, CS
provides an information capturing paradigm with both
sampling and compression and still can recover the original
signal even under the condition that the sampled rate in CS can
is less than the conventional Nyquist rate.
Mathematically, the procedure of CS can be expressed as a
linear projection: Let
[]
12
,,,
T
N
xx=x be an N-dimensional
original signal and can be represented in sparse basis
NN
R
×
∈Ψ
as
=θΨx
(1)
where
[]
12
,,,
T
N
θθ θ
=
θ
is a sparse vector. The signal
x
is K-
sparse in basis
Ψ
, and this means only
N
entries are
nonzero in
θ
. Suppose y is an
M
-dimensional measurement
signal, which is formed from
M
linearly projected
measurements of
x
,
==y Φx ΦΨθ (2)
where Φ is an
N
×
sensing matrix and
N
.
From above introduction, it is easily to be seen that only
one signal can be applied into CS, while distributed
compressed sensing (DCS) makes multiple signals into CS
possible. Baron er.al [3] studied joint signal sparse models and
joint signal recovery algorithm in CS based on intra(spatial)-
correlation and inter(temporal)-correlation of signals, before
they proposed the DCS. So the technology of DCS can be
treated as an extension of CS.
In wireless sensor networks (WSNs), a large number of
sensor nodes are distributed over a region of interest. Also, in
continuous monitoring applications, sensors report their
readings in short intervals. Therefore, the signal they acquired
possess both inter-(spatial) and intra-(temporal) signal
correlations and are individually sparse in some basis. That’s
the reason why DCS can be applied into WSNs. And in [4],
Wakin et al defined three joint sparsity models(JSM) for
WSNs, including sparse common component + innovations
model (JSM-1), common sparse supports model (JSM-2) and
non-sparse common component + sparse innovations model
(JSM-3). In this paper, we focus on the sparse common
component + innovations (JSM-1).
A critical issue in DCS is to minimize the difference
between the original signal
x and the recovery signal
*
x ,
which is recovered from a significantly less number of
measurements contained in measurements signal
y . Right
now, there are two main approaches proposed for this problem.
One is designing sensing matrix and the other one is designing
recovery algorithm.
Generally, sensing matrices should satisfy sufficient
conditions in terms of mutual coherence and restricted isometry
property (RIP). Then, the sensing matrix can be applied in DCS
to recover the original signal. The mutual coherence of a
matrix, initially introduced in [5], measures the smallest angle
between each pair of its columns. RIP was initially introduced
in [6]. It measures the degree to which each subset of
k
column vectors of sensing matrix
Φ is close to an isometry.
Currently, sensing matrices include two categories, random
sensing matrices and deterministic sensing matrices. Random
sensing matrices contain Gauss sensing matrix [7], Bernoulli