Professional Documents
Culture Documents
Khalfi 2017
Khalfi 2017
Fn
F1
F2
F3
Spectrum occupancy
measurements where m < n and k is the num-
ber of occupied bands, referred to as the spar-
sity level. The spectrum occupancy information
vector, xn1, is then recovered by minimizing the
l0-norm of xn1 subject to a constraint on the error
||ym1 – Fxn1||l22, where ym1 is the vector
representing the m measurements, F is a full-rank
sensing matrix, and is the discrete inverse Fou-
rier transform matrix. Due to its NP-hardness, Frequency bands
recovery heuristics (e.g., l1-norm minimization and
orthogonal matching pursuit [10]) have been pro-
posed in the literature for solving such problems. Blocks that are temporarily used. It should
From a practical viewpoint, the implementation of Discourage the search of occupied be highly favored during the search.
wideband spectrum sensing requires the use of bands in these blocks: sparser.
m amplifiers and then mixing the received ampli-
fied signals with pseudo-random waveforms at FIGURE 3. Sparsity-promoting wideband spectrum sensing.
Nyquist rates. After that, an integrator is applied
followed by an analog-to-digital converter that
takes samples at sub-Nyquist rate. This architec- 0.06
ture is known as analog-to-information converter Conventional
(AIC) sampler [10]. Linear SVR
An observation we make by investigating the Nonlinear SVR
Mis-detection
0.04
Batch gradient descent
existing compressive sensing-based approaches
is that they consider that the occupancy of wide- 0.02
band spectrum to be homogeneous, meaning
that the entire wideband spectrum is considered
as one single block with multiple bands, and the 0
300 400
0 100 200 500
sparsity level is estimated across all bands and Time index
considered to be the same for the entire wide-
band spectrum. However, in wideband spectrum FIGURE 4. Mis-detection performance evaluation of weighted compressive
assignment, applications of similar types (TV, sensing under different regression techniques compared to conventional
satellite, cellular, etc.) are often assigned bands approach [10].
within the same block, suggesting that wide-
band spectrum is heterogeneous. That is, band
occupancy patterns are not the same across the er than the blocks which are more likely to be
different blocks, since different application/user occupied. Since the number of occupied bands
types within each block can exhibit different traffic changes over time, the design of the weights w
behaviors; hence, wideband spectrum occupancy can be based on the average occupancy of every
– –
may vary significantly from one block to another, spectrum block i, ki, such that wi = 1/ ki [11]. Fur-
^
as illustrated in Fig. 2. This trend has indeed also thermore, if each block occupancy, k i, can be
been confirmed by recent measurement studies estimated through learning (e.g., using regression
[8]. techniques), better performance can be achieved
^
Incorporating this fine-grained sparsity struc- when setting w i = 1/ k i. In Fig. 4, we show the
ture into the formulation of wideband spectrum performance of the proposed weighted com-
occupancy recovery allows us to improve the pressive spectrum sensing approach with band
recovery performance and enhance the detection occupancy prediction (using different regression
accuracy of wideband spectrum sensing. Specifi- models) and compare it to a conventional wide-
cally, such a block-like sparsity structure allowed band spectrum sensing approach [10]. Note that
us to formulate the problem as a weighted l1-mini- in the non-cooperative case where IoTs perform
mization problem, thereby resulting in an algo- wideband spectrum sensing individually, there is
rithm that provides faster spectrum occupancy no signaling overhead (information exchange with
recovery with less sensing overhead [11]. The the other network entities to perform this task).
basic idea behind our algorithm is that the spec- However, in the cooperative case where multiple
trum blocks which are more likely to be occu- IoTs are involved in the sensing task, the signaling
pied are favored during the search. In addition, overhead becomes proportional to the number
blocks corresponding to critical applications or of cooperating IoTs. Table 2 shows the signaling
for which some occupancy information is known overhead associated with each of the approaches
are captured through careful design of block discussed in this article.
weights [11]. In essence, any additional knowl-
edge about spectrum utilization behavior can be Overcoming Network Edge Traffic Bottlenecks
incorporated and exploited so that faster recov- As discussed earlier, the massive IoT traffic that
ery of spectrum occupancy information can be 5G cells are required to support and enabling
achieved. Figure 3 illustrates some of these design edge cloud offloading will create severe conges-
elements, where the blocks that are more likely tion bottlenecks at the 5G network edges. One
to be unoccupied are encouraged to be spars- possible solution proposed in [12] to overcome
this challenge lies in leveraging D2D and com- cussed previously in the wideband spectrum
pressive sensing theory to reduce the number of sensing context. In addition, learning and predic-
connections established between the BSs and the tion approaches can also be used in conjunction
IoTs, and to reduce the amount of offloading traf- with the recovery approach to improve the per-
fic. D2D communication technology has been formance. This was considered in [13] where a
adopted in LTE-Advanced systems, but only for data gathering approach was proposed based on
public safety communications. When appropri- compressive sensing. The proposed scheme takes
ately exploited, D2D can offer great advantages. advantage of the correlation between data and
Higher throughput, low latency, better availability, introduces an autoregressive (AR) model in the
and new services, among other advantages, make recovery approach. IoTs can also be leveraged for
D2D an ideal candidate to help in the adoption performing wideband spectrum sensing but with
and support of IoTs by 5G. a focus on reducing the reported measurements’
With the use of compressive sensing, instead cost [14], which can be combined with the work of
of having all IoTs push their data to the BSs, the [13]. Under the assumption that the sensed signal
BSs can pull the data from only a subset of devices is sparse, a sparser basis can be found and can lead
and use compressive sensing to recover the data of to a more compressed signal than the frequency
all IoTs. Here the sparsity that allows the exploita- domain basis. The IoTs report the measurements to
tion of compressive sensing comes from the fact/ network nodes that perform simple addition of the
assumption that at a given time, only a few IoT measurements coming from the IoTs and report
devices experience changes in their memory, and them to the BS. This way, a constant communica-
hence only a few will need to upload their mem- tion cost is maintained (communication overhead
ory updates to the edge clouds. Specifically, con- is proportional to the number of network nodes).
sidering mIoTs with delay-tolerant requirements, At the BS, the different measurements are exploit-
every node multicasts to its neighbors a weighted ed to recover the wideband spectrum occupancy.
value of the updated data replica with a defined Table 2 summarizes the main proposed works
coefficient that corresponds to the coefficient of that exploit sparsity features to enable the support
the sensing matrix. When a node receives the of massive and heterogeneous IoTs.
weighted data replicas from other nodes, it adds
its corresponding update, if any, and multicasts it Open Research Problems and Directions
during its time slot. After exchanging the data rep- Despite the efforts made in exploiting the hidden
licas, the nodes turn to the sleep mode for energy sparsity structure for supporting IoTs through cel-
saving purposes. The BS pulls the measurements lular systems, key challenges remain that need to
from a few nodes, compared to the total number be overcome. We summarize here some of the
of mIoTs, where the number of these nodes should research directions we believe are worth investi-
satisfy a condition that depends on the total num- gating in the future.
ber of nodes and the number of nodes that have Wideband spectrum occupancy behaviors:
data updates. Since most of the mIoTs have no Although some research efforts aiming to exploit
update, the vector corresponding to the memory spectrum occupancy sparsity to reduce traffic over-
replicas is sparse. Compressive sensing theory can head have already been made, these approach-
accurately recover the data replicas for each IoT es either are generic (not specific to wideband
and support the corresponding ones through cloud spectrum access) or achieve limited performance
services. The proposed protocol shows that signal- improvements due to the assumptions made. For
ing overhead is considerably reduced (only m con- instance, the spectrum occupancy heterogeneity
nections are established with a BS), congestion is structure inherent to dynamic wideband spectrum
avoided, and latency is improved by placing cloud access is a feature that, when exploited properly,
services at the edge. The main shortcoming of this can allow for the design of more efficient compres-
approach is that it only works with homogeneous sive sampling approaches [11]. Also, a common
IoTs and assumes a fixed sparsity level. limitation of these existing approaches lies in the
Potential improvements can be achieved fact that the spectrum occupancy sparsity level is
through weighted compressive sensing as dis- considered constant and does not change over