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/TVT.2019.2943997, IEEE
Transactions on Vehicular Technology
1

A Machine Learning-Enabled Spectrum Sensing

Method for OFDM Systems
Jinfeng Tian, Peng Cheng, Zhuo Chen, Mingqi Li, Honglin Hu, Yonghui Li, and Branka Vucetic

Abstract—This paper addresses the spectrum sensing problem rendering them infeasible to handle the hidden node problems
in an orthogonal frequency-division multiplexing (OFDM) system ubiquitous in many spectrum sensing applications.
based on machine learning. To adapt to signal-to-noise ratio Machine learning (ML) based approaches have recently
(SNR) variations, we first formulate the sensing problem into
a novel SNR-related multi-class classification problem. Then, been proposed to provide a new avenue to general spec-
we train a naive Bayes classifier (NBC), and propose a class- trum sensing [8], including that for OFDM configuration.
reduction assisted prediction method to reduce spectrum sensing In [9], supervised and semi-supervised learning algorithms
time. We derive the performance bounds by translating the were developed for a unchanged RF environment with fixed
Bayes error rate into spectrum sensing error rate. Compared SNR, where the eigenvalues of the received signal covariance
with the conventional methods, the proposed method is shown
by simulation to achieve higher spectrum sensing accuracy, in matrix are utilized as the features. Given received signal
particular at critical areas of low SNRs. It offers a potential energy and likelihood ratio test statistic with different SNRs,
solution to the hidden node problem. binary classification based artificial neural network (ANN) was
Index Terms—Machine learning, spectrum sensing, naive
adopted in [10], and shown to outperform the classical energy
Bayes classifier. detection method. In addition to these localized spectrum sens-
ing methods, ML based cooperative spectrum sensing (CSS)
methods were also studied in the literature [11], [12]. For
I. I NTRODUCTION example, [11] comprehensively investigated the unsupervised
and supervised learning techniques in CSS for a fixed received
The ever-increasing demand for ubiquitous broadband ser- SNR for each node. It is worth noting that the aforementioned
vice and massive connection has led to explosive growth in work designs either an individual binary classifier for each
the requirement for spectrum resource. Cognitive radio, an SNR level exhibited in the received samples, or a unified
intelligent wireless technology, opens a potential communica- binary one for multiple SNR levels. The former requires the
tion paradigm to achieve more efficient and flexible spectrum accurate estimate of the current SNR on a continuous basis for
use. Spectrum sensing is one of the key enabling function the prediction in spectrum availability, which is usually very
in cognitive radio. It detects spectrum opportunities so that challenging. The latter finds it difficult to capture the complex
the secondary user (SU) can determine its access strategy [1]. relationship between the input features and the output labels
On the other hand, as a promising multicarrier transmission with a simple classifier structure.
technique to combat multipath fading, orthogonal frequency Bearing these limitations in mind, in this paper, we take a
division multiplexing (OFDM) has seen its wide adoption in different approach on the spectrum sensing for OFDM systems
a number of standards [2]. Therefore, spectrum sensing for while tapping the benefits of ML. The novelty of our ML
OFDM systems has received considerable attention in recent approach lies in the adoption of both the received signal
years. power and CP-induced correlation as the features. Further-
Among many OFDM sensing methods, energy detection more, different from the conventional methods in [8]–[12],
(ED) is widely adopted due to its simplicity and efficiency. we formulate the spectrum sensing problem as an SNR-related
However, it is sensitive to noise uncertainty (NU), and there multi-class one. This eliminates the tedious process of contin-
exists a signal-to-noise ratio (SNR) wall at low SNRs [3]. uous SNR estimation, while enabling a simple training model
Some specific characteristics of OFDM signal could also be for accurate prediction. Specifically, the main approaches and
utilized to identify the primary user (PU) spectrum usage. contributions of this paper are summarized as follows.
For example, it was shown in [4] and [5] that the correlation • First, we formulate the spectrum sensing problem as
induced by cyclic prefix (CP) can be leveraged to overcome multi-class classification problem, adapting to unknown
the SNR wall phenomenon in ED. However, the CP-based SNR variations. This means that the learning in our
sensing methods generally generate poor correlation, and can approach is one-off, regardless of the ensuing SNR vari-
not achieve good detection performance [6]. Recently, an ations.
asymptotic simple hypothesis test (ASHT) based detector • Second, we train a machine learning model with a naive
was proposed in [7] for OFDM signal sensing, where both Bayes classifier (NBC), which takes full advantage of
the received signal strength and CP-induced correlation are the statistical information of the data features and has
exploited. This method enjoys the advantage of improved a simple modeling structure. Furthermore, we propose
detection probability over ED and CP. In commonality, all a class-reduction assisted prediction method to reduce
these detection methods do not perform well at low SNRs, spectrum sensing time.

0018-9545 (c) 2019 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/TVT.2019.2943997, IEEE
Transactions on Vehicular Technology
2

∗
• Third, we analyze the error probability introduced by quantifies the CP-induced correlation1 , where < {·} and (·)
class reduction and NBC, which results in the expression denote the real part and complex conjugate, respectively. Note
for the spectrum sensing error rate (SSER). Then its that these two statistics were used in [7] for an approximate
performance bounds are developed based on the bounds likelihood ratio test with the estimated noise power. In this
of Bayes error rate. paper, we will utilize F1 and F2 in a totally different way, and
• Finally, simulation results show that the proposed method develop a novel data-driven based spectrum sensing method.
can achieve more accurate spectrum sensing compared to
the conventional ones, in particular at low SNRs, offering III. T HE P ROPOSED M ACHINE L EARNING BASED
a potential solution to the hidden node problem. S PECTRUM S ENSING M ETHOD
The rest of the paper is organized as follows. Section II In this section, we elaborate on the proposed machine-
introduces the system model and input features. In Section learning based sensing method. We first formulate the spec-
III, a class-reduction assisted NBC method is proposed based trum sensing process into an SNR-related multi-class classi-
on the formulated SNR-related multiclass classification model. fication problem. Then we train a NBC model. After that,
Its performance bounds are derived in Section IV. Numerical we develop a novel class-reduction assisted NBC to reduce
results are provided in Section V to validate the benefits and the number of candidate classes and determine the spectrum
effectiveness of our methods. Finally, conclusions are provided occupancy.
in Section VI.

A. SNR-Related Multi-Class Classification Problem Formula-

II. P RELIMINARIES
tion
We consider an OFDM signal s(n) with NF F T subcarriers,
Based on the central limit theorem, the average received
whose length of CP is Ng . It is transmitted over a multipath
signal power F1 follows a Gaussian distribution which is
fading channel with the maximum tap order Lh (Lh < Ng ).
distributed as [14, Eq. (2)]
The complex gain and multipath delay of the l-th channel tap
are denoted by h(l) and τl . Without the loss 2σ 4
PLofh −1
generality, the 2
2
H0 : F1 ∼ N (σw , KNw )
channel is assumed to be normalized, i.e., l=0 |h(l)| = 1. 2 2 2 (5)
2 2(σz +σw )
H1 : F1 ∼ N (σz2 + σw , KN ).
This entails the convenience of SNR calculation at the receiver,
and it does not affect the resulting performance presentation. Similarly, for a large number of observation samples, the CP-
Then the received signal can be represented as induced correlation F2 follows a Gaussian distribution as [4],
h −1
LX [15, Eq. (23)]
r(n) = h(l)s(n − τl ) + w(n), 0 ≤ n ≤ KN − 1, (1) H0 : F2 ∼ N (0, 2KN
w
)
σ4
g
l=0
(σz2 +σw
2 2
) +2σz4
(6)
H1 : F2 ∼ N (σz2 , ).
where K is the number of OFDM symbols, N is the total 2KNg
length of an OFDM symbol with N = NF F T + Ng , and Given F1 and F2 as feature inputs, we aim to correctly
w(n) denotes a zero-mean additive white Gaussian noise recognize the channel occupancy by a data-driven classifier.
2
(AWGN) with variance σw . The transmitted signal s(n) is However, for either H0 or H1 , the distributions of F1 and F2
assumed Pto have a fixed power σs2 . It is clear from (1) that are not fixed but change with σw 2
. In this case, the conven-
Lh −1
z(n) = l=0 h(l)s(n − τl ) is the received signal without tional binary classification in machine learning cannot adapt
noise, whose power satisfies σz2 = σs2 . The received SNR is to the SNR variations. This motivates us to develop a new
denoted by σz2 /σw
2
. SNR-related multi-class classification method. Specifically, let
Given a sequence of observations r(n), the aim of spectrum (m) (m)
F(m) = [F1 , F2 ]T denote the feature vector of the m-th
sensing is to determine whether primary user is present or training sample (m = 0, · · · , M − 1), and
not, which can be formulated as a binary hypothesis testing
problem o(m) ∈ Ω = {C0 , C1 , C2 , C3 , · · · , C2I−2 , C2I−1 } (7)
H0 : r(n) = w(n)
| {z } | {z } | {z }
(2) SN R0 SN R1 SN RI−1
H1 : r(n) = z(n) + w(n),
denote the true class (label) corresponding to F(m) , where I
where H0 and H1 indicate the absence and presence of the is the number of discrete SNR levels, Ck represents the k-th
primary signal, respectively. Based on [7], we next introduce class, and Ω denotes the class set. Here, at each level SN Ri ,
two statistics F1 and F2 . In specific, there are two classes C2i and C2i+1 (e.g., C2 and C3 for
KN −1 SN R1 ), corresponding to H0 and H1 , respectively. Thus, we
1 X 2
F1 = |r(n)| (3) have a total number of 2I classes, and we attempt to train a
KN n=0
multi-class classifier, reflecting the relationship between input
represents the average received signal power, and F(m) and label o(m). It is worth noting that the 2I labels
for I different SNRs can be available by directly or indirectly
K−1 Ng −1
1 X X
F2 = < {r(l0 N + n)r∗ (l0 N + n + NF F T )} 1 The calculation of F requires time synchronization to locate the position
2
KNg 0 of the CP. In practice, F2 can be calculated by the sliding window based peak
l =0 n=0
(4) detection method without time synchronization [13].

0018-9545 (c) 2019 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/TVT.2019.2943997, IEEE
Transactions on Vehicular Technology
3

obtaining the PU occupancy within the training period. For our method features reduced spectrum sensing time, as well
example, checking the operational time table of the PU or as excellent mathematical tractability.
placing a measurement equipment of high sensitivity adjacent Specifically, our method takes advantage of F1 to reduce
to the PU. the search scope of candidate classes. Assume that the testing
sample F∗ belongs to class Ck∗ ∈ Ω. Our class-reduction
B. NBC Model Training criterion can be written as
The goal of training the NBC is to learn the unknown func- |F1∗ − µ̂1,k | ≤ b0 σ̂1,k , (13)
tion of posterior probability p(Ck |F) over all possible classes
of Ck . According to Bayes rule, p(Ck |F) ∝ p(Ck ) · p(F|Ck ), where b0 is a predefined parameter. Here, based on the rule
where p(Ck ) and p(F|Ck ) denote the prior probability and of standard deviation and coverage in Gaussian distribution2 ,
likelihood function, respectively. Here, F1 and F2 are not inde- we roughly identify all the possible Gaussian distributions
pendent. However, the NBC makes a fundamental assumption (with different mean and variance, i.e., all the possible
of conditionally independent features, i.e., classes) that F1∗ belongs to. In specific, we include the
selected class Ck satisfying (13) into a candidate class subset
p(F|Ck ) = p(F1 |Ck ) · p(F2 |Ck ). (8) Ωsub,k∗ ∈ Ω. Based on updating the prior probability p(Ck )
This can substantially reduce the computational complexity from o(m) ∈ Ωsub,k∗ , we then adopt (12) to determine the
and is found to work well in many real-world cases [16], [17]. predicted class CkN B within Ωsub,k∗ . As a result, the binary
Consequently we follow this assumption in our approach. hypothesis Hj can be determined by j = mod (kN B , 2), where
Under this assumption, the NBC rule which resorts to find mod denotes the modulus operator.
the maximum posterior probability can be described as
( 2
) D. Algorithm Summary
Y
arg max p(Ck ) p(Fq |Ck ) . (9) The procedure of the proposed machine learning based
Ck q=1 spectrum sensing method can be summarized as follows.
Therefore, the naive Bayes learning algorithm transforms 1) Based on the features of F1 and F2 , we formulate the
into estimating p(Ck ) and p(Fq |Ck ). In fact, p(Ck ) can be spectrum sensing problem as an SNR-related multi-class
estimated by the frequency of Ck from the class set Ω. For classification problem.
p(Fq |Ck ), it is clear from (5) and (6) that it follows a Gaussian 2) With the collected training samples and their labels, we
distribution. Then, in training the NBC, we need to estimate train an NBC to learn the function of p(Ck |F) over
2
the mean µq,k 2
variance σq,k of each p(Fq |Ck ) based on the all possible Ck , where p(Ck ), µ̂q,k and σ̂q,k can be
 and
(m) (m) obtained.
training set F , o . The maximum likelihood estimates
for the mean and variance of p(Fq |Ck ) can be written as 3) The class prediction can be implemented by (12). Fur-
thermore, to avoid the complexity of the multi-class
(m)
Fq δ(o(m) = Ck )
P
classifier increasing with the class number, we use (13)
µ̂q,k = m P (m) = C )
(10) to select all the possible Gaussian distributions (classes)
m δ(o k
that F1∗ belongs to, obtaining a reduced candidate class
(m)
(Fq − µ̂q,k )2 δ(o(m) = Ck )
P
2 subset Ωsub,k∗ . Then the trained NBC as (12) is applied
σ̂q,k = m P , (11)
( m δ(o(m) = Ck )) − 1 to predict CkN B within Ωsub,k∗ .
4) We finally determine the binary hypothesis Hj based on
where δ(o(m) = Ck ) equals 1 if o(m) = Ck and 0 otherwise. In j = mod (kN B , 2).
other words, δ(o(m) = Ck ) aims to select only those training
examples for which o(m) = Ck . IV. P ERFORMANCE A NALYSIS
Next, we analyze the SSER of the proposed method. Basi-
C. Class Prediction and Spectrum Sensing cally, the error is determined by both pCR and pN BC
, referring
e e
2 to the error probability induced by the class reduction (c.f.
With µ̂q,k and σ̂q,k obtained in (10) and (11), now we can
predict the class CkN B associated with a new sample F∗ = (13)) and the NBC, respectively.
[F1∗ , F2∗ ]T as
( 2 (F ∗ −µ̂ )2
) A. Class Reduction
Y 1 − q 2 q,k
2σ̂
CkN B = arg max p(Ck ) exp q,k . (12) We first quantify the cardinality of the candidate class subset
Ck σ̂
q=1 q,k after class reduction. Based on (5), the mean and standard
However, we find it time-consuming to calculate (12) for deviation of F1 in class Ck (Ck corresponds to SN Ri under
each Ck , entailing excess spectrum sensing time when I is Hj ) can be written in a compact form as
potentially large due to the wide SNR range. Furthermore, its SN Ri
 
µ1,k = 10− 10 + j σz2 , j = 0 or 1 (14)
SSER performance analysis becomes very complicated due to
multiple classes. In this case, we adopt a different approach 2 If X ∼ N (µ, σ 2 ), then we have P {|X − µ| ≤ b σ} = 0.997 for
0
and propose a novel class-reduction based prediction method b0 = 3. In other words, we have a probability of 0.003 to fail to select the
with a comparable classification accuracy. Most importantly, possible distribution X belongs to.

0018-9545 (c) 2019 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/TVT.2019.2943997, IEEE
Transactions on Vehicular Technology
4

and B. SSER Performance

r
2 The remaining task is to derive pNe
BC
based on Bayes error
σ1,k = µ1,k , (15)
KN rate, arriving at the final goal SSER. For a binary classification
problem (Nk∗ = 2), the Bayes error rate can be given by
respectively. Let us consider |F1∗ − µ1,k∗ | ≤ b0 σ1,k∗ . To 0
k,k
pick out all the possible classes that any F1∗ belongs to, an EBayes = p(Ck )pe (Ck0 |Ck ) + p(Ck0 )pe (Ck |Ck0 ), (23)
equivalent form of (13) can be written as R
where pe (Ck0 |Ck ) = R 0 p(F|Ck )dF and Rk0 is the decision
k
0
k,k
|µ1,k∗ − µ1,k | ≤ b0 (σ1,k∗ + σ1,k ), (16) region for Ck0 . Note that the direct computation of EBayes is
very difficult; instead, we can resort to its upper and lower
∗
SN Ri bounds [18, Eq. (8)].
where µ1,k∗ = (10− 10 + j ∗ )σz2 . Similarly, Ck∗ corre-
∗
sponds to SN Ri under Hj ∗ . In this way, all the Ck that Let us denote k = 2i. It is clear that H0 and H1 under
satisfy the criterion (16) form Ωsub,k∗ with the cardinality SN Ri correspond to Ck and Ck+1 , respectively.
 Assume that
|Ωsub,k∗ | = Nk∗ . Ωsub,k = {Ck , Ck0 } and Ωsub,k+1 = Ck+1 , C(k+1)0 . It is
∗ worth noting that k 0 may not be k + 1 and (k + 1)0 may not be
We define SN Ri − SN Ri = a∆, where ∆ > 0 is the
k due to the small SNR granularity ∆. For example, the two
discrete SNR granularity and a is a constant. Combining (14)-
candidate classes might be C0 and C3 (c.f. (7)). According
(16), we can obtain
to [19, Eq. (19)], the error detection probability of spectrum
D1i
∗
,j ∗
(j) ≤ a ≤ D2i
∗
,j ∗
(j) (17) sensing is defined by p(H0 )∗pe (H1 |H0 )+p(H1 )∗pe (H0 |H1 ).
Thus such error detection probability induced by the NBC can
with be written as
& ( ∗ ∗
! )' pN BC
= αp(Ck )pe (Ck0 |Ck ) + βp(Ck+1 )pe (C(k+1)0 |Ck+1 ),
∗ ∗ 10 g1i ,j ∗ ∗
e
D1i ,j
(j) = max log − g2i ,j , g3i ,max (24)
∆ b1 where α = δ(mod(k 0 , 2) = 1) and β = δ(mod((k + 1)0 , 2) =
(18) 0). As Ck and Ck0 have the similar statistics, we can approx-
and imately obtain pe (Ck0 |Ck ) = pe (Ck |Ck0 ). Similarly, we have
∗ ∗
 
10  ∗ ∗ ∗
 ∗  pe (C(k+1)0 |Ck+1 ) = pe (Ck+1 |C(k+1)0 ). Therefore, (24) can
D2i ,j (j) = min log b1 g1i ,j − g2i ,j , g3i ,min , be rewritten as
∆
(19) k,k 0
k+1,(k+1)0
√ 2 ∗ pN
e
BC
= αp(Ck )EBayes + βp(Ck+1 )EBayes . (25)
1+b0 i∗ ,j ∗ SN Ri i∗ ,j
√ 2 , g1 ∗
where b1 = KN
= 1 + 10 10 · j , g2 =
∗
1−b0 KN Finally, the SSER of the proposed method can be obtained as
SN Ri ∗ i∗ t
10 · j, and g3i ,t = SN R ∆−SN R , t ∈ {max, min}.
10
Ess = pCR + (1 − pCR N BC
max e e )pe . (26)
SN R and SN Rmin separately denote the maximum and 0
k,k
minimum SNR level. Therefore, Nk∗ can be calculated as As discussed before, EBayes in (25) is difficult to compute.
Based on [18, Eq. (8)], we can analytically derive the upper
1
X ∗
,j ∗ and lower bounds of Ess .
Nk∗ = Ni (j), (20)
j=0
V. N UMERICAL R ESULTS
where The effectiveness of the proposed method is demonstrated
∗ by simulation in the second generation terrestrial digital video
,j ∗
Ni (j) broadcasting (DVB-T2) system [20]. In specific, each OFDM
∗
,j ∗ ∗
,j ∗ ∗ ∗ ∗
,j ∗
D2i (j) − D1i if D2i ,j (j) ≥ D1i

(j) + 1, (j) symbol has 1024 subcarriers and unit power, and lasts for
= 126 µs, 1/8 of which is CP. The number of OFDM symbols
0, otherwise.
(21) used for spectrum sensing is 30. The 6-tap typical urban (TU6)
channel model is adopted [21]. Following the footnote 2, we
It is clear that the current number of classes Nk∗ is mainly choose b0 = 6. Three traditional threshold-based methods are
determined by SNR levels, channel occupancy, and the number included for comparison, namely, ED with 1 dB NU, CP-based
of OFDM samples. Our numerical results based on (20) show correlation detection [4], and ASHT-based detection [7]. For
that we can achieve Nk∗ = 2 with a large probability for a comparison with the ML-based method, the back propagation
wide SNR region. In addition, pCR e can be given by neural network (BPNN) is evaluated with 2 hidden layers, each
  of which has 5 neurons. To guarantee a favourable sensing
b0 performance, training samples need to be collected across a
pCR
e = 1 − erf √ , (22)
2 wide range of SNRs with small granularity. In the simulation,
Rx 2
a 1 dB SNR granularity is adopted for illustrative purpose.
where erf(x) = √2π 0 e(−t ) dt. According to (22), an optimal For each SNR under each binary hypothesis, the data set
b0 can be determined by the relationship between the standard consists of 5000 runs, among which Ntrain samples are used
deviation and coverage of the normal distribution. for training and the rest for testing.

0018-9545 (c) 2019 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/TVT.2019.2943997, IEEE
Transactions on Vehicular Technology
5

TABLE I 1
NBC with class reduction
ACCURACY OF ML ALGORITHMS USING DIFFERENT TRAINING SAMPLES 0.9 NBC without class reduction
BPNN
Number of training samples Ntrain Proposed BPNN 0.8 ASHT

Probability of correct detection

CP
10 0.898 0.7578 0.7 ED with 1dB NU
500 0.9421 0.8276
1000 0.9433 0.8367 0.6

1500 0.9441 0.8403 0.5

2000 0.9446 0.8433
0.4
2500 0.945 0.8441
0.3

0.2
Figs. 1 and 2 illustrate the probability of correct detection 0.1
and spectrum sensing error rate, respectively, where Ntrain =
0
1000 and the false alarm probability of the threshold-based −20 −15 −10 −5 0
SNR (dB)
methods is set to be 0.05. Fig. 1 shows that the class-reduction
assisted NBC method has the same detection performance as Fig. 1. Probability of correct detection versus SNR.
the NBC without class reduction. Furthermore, the proposed
NBC methods outperform the BPNN method and all the
0
10
threshold-based ones, significantly so at the lower end of SNR
scope. In addition, the ASHT-based method is superior to
CP-based method while the latter outperforms the ED-based

Spectrum sensing error rate

method with 1 dB NU. In Fig. 2, as the theoretical benchmark, −1
10
the upper and lower bounds of the sensing error rate are
also plotted. It can be found that performance curves of the
proposed NBC methods locate between the two bounds. The
NBC with class reduction
curves of all the thresh-based methods lie well above the upper 10
−2 NBC without class reduction
BPNN
bound. Note that for the ED-based method, its error detection ASHT
was not observed in the simulation for SNR ≥ −5 dB. CP
ED with 1dB NU
For generalization purpose, the receiver operating charac- Upper bound
Lower bound
teristics (ROC) for SNR = −18 dB is shown in Fig. 3. 10
−3

−20 −15 −10 −5 0

The optimal likelihood radio test (LRT) with the known noise SNR (dB)

variance [7] is plotted as the benchmark. It can be observed

Fig. 2. Spectrum sensing error rate versus SNR.
that the proposed method outperforms all the other methods.
As expected, more training samples result in better ROC
performance. When Ntrain = 1000, the proposed method 1

performs very close to the optimal LRT. 0.9

Table 1 illustrates the impact of the number of training 0.8
samples on the spectrum sensing accuracy of the proposed and
Probability of correct detection

0.7
BPNN methods, where SNR = −18 dB. It is clearly shown
0.6
that the proposed method achieves much higher classification
0.5
accuracy than the BPNN method. This supremacy originates Proposed, Ntrain=1000

from the fact that the proposed method utilizes the Gaussian 0.4 Proposed, Ntrain=10
BPNN, Ntrain=1000
statistical information of the data features. For both classifiers, 0.3
BPNN, Ntrain=10
the prediction accuracy enhances with the number of training 0.2 ASHT
samples. It can be observed that a quick convergence is 0.1
CP
ED with 1dB NU
achieved with a small number of training samples, and the Optimal LRT
0
accuracy improvement becomes slow when the number of 0 0.2 0.4 0.6
Probability of false alarm
0.8 1

training samples increases beyond 500.

Fig. 4 presents the computation complexity in terms of Fig. 3. ROC curves for different methods.
average CPU prediction time for the NBC method with and
without class reduction. It can be seen that, as the class number
increases, the average prediction time of NBC with class We formulated the sensing process as a novel SNR-related
reduction remains almost constant while the method without multi-class classification problem. Our approach removes the
class reduction grows approximately linearly. This verifies the inherent dependency on SNR in the conventional binary classi-
benefits of the class reduction. fication. A class reduction assisted NBC method was proposed
to train a machine learning model and reduce spectrum sensing
VI. C ONCLUSIONS time. We also derived the performance bounds of the spectrum
In this paper, we proposed an efficient spectrum sensing sensing error rate. Simulation results verified the effectiveness
method for OFDM systems from a data driven perspective. of the proposed method.

0018-9545 (c) 2019 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/TVT.2019.2943997, IEEE
Transactions on Vehicular Technology
6

0.7 [17] M. O. Mughal and S. Kim, “Signal classification and jamming detection
NBC with class reduction
NBC without class reduction
in wide-band radios using Naive bayes classifier,” IEEE Commun. Lett.,
0.6 vol. 22, no. 7, pp. 1398-1401, Jul. 2018.
[18] K. Tumer and J. Ghosh, “Bayes error rate estimation using classifier
ensembles,” Int’l J. Smart Eng. System Design, vol. 5, no. 2, pp. 95-110,
Average prediction time (ms)

0.5
2003.
[19] A. Taherpour, H. Mokhtarzadeh, and T. Khattab, “Optimized error
0.4
probability for weighted collaborative spectrum sensing in time- and
energy-limited cognitive radio networks,” IEEE Trans. Veh. Technol., vol.
0.3 66, no. 10, pp. 9035-9049, Oct. 2017.
[20] Digital Video Broadcasting (DVB); Frame structure, channel coding
0.2 and modulation for a second generation digital terrestrial television
broadcasting system (DVB-T2), ETSI Standard, EN 302 755, V1.3.1,
0.1
Apr. 2012.
[21] European Telecommunications Standards Institute, “Digital Video
Broadcasting (DVB): Measurement guidelines for DVB systems,” ETSI
0
10 20 30 40 TR 101 290 V1.2.1, May 2001.
Class number

Fig. 4. Average prediction time versus class number.

R EFERENCES

[1] A. Ali and W. Hamouda, “Advances on spectrum sensing for cognitive

radio networks: Theory and applications,” IEEE Commun. Surveys Tuts.,
vol. 19, no. 2, pp. 1277-1304, 2nd Quart. 2017.
[2] Y. Li and G. Stber, “Orthogonal frequency division multiplexing for
wireless communications,” Boston, MA, USA: Springer, 2006.
[3] R. Tandra and A. Sahai, “SNR walls for signal detection,” IEEE J. Sel.
Topics Signal Process., vol. 2, no. 1, pp. 4-17, Feb. 2008.
[4] Z. Lei and F. P. S. Chin, “Sensing OFDM systems under frequency
selective fading channels,” IEEE Trans. Veh. Technol., vol. 59, no. 4,
pp. 1960-1968, May 2010.
[5] S. Dikmese, Z. Ilyas, P. C. Sofotasios, M. Renfors, and M. Valkama,
“Sparse frequency domain spectrum sensing and sharing based on cyclic
prefix autocorrelation,” IEEE J. Sel. Areas Commun., vol. 35, no. 1, pp.
159-172, Jan. 2017.
[6] Y. Zeng, Y. C. Liang, and T. H. Pham, “Spectrum sensing for OFDM sig-
nals using pilot induced auto-correlations,” IEEE J. Sel. Areas Commun.,
vol. 31, no. 3, pp. 353-363, Mar. 2013.
[7] M. Jin, Q. Guo, J. Xi, Y. Li, and Y. Li, “On spectrum sensing of OFDM
signals at sow SNR: New detectors and asymptotic performance,” IEEE
Trans. Signal Process., vol. 65, no. 12, pp. 3218-3233, Jun. 2017.
[8] Y. Chen and H.-S. Oh, “A survey of measurement-based spectrum
occupancy modeling for cognitive radios,” IEEE Commun. Surveys Tuts.,
vol. 18, no. 1, pp. 848-859, 1st Quart., 2016.
[9] O. P. Awe, “Machine learning algorithms for cognitive radio wireless
networks,” Ph.D. thesis, Nov. 2015.
[10] M. R. Vyas, D. K. Patel, and M. Lopez-Benitez, “Artificial neural
network based hybrid spectrum sensing scheme for cognitive radio,” in
proc. IEEE Int. Symp. Personal, Indoor, and Mobile Radio Commun.
(PIMRC 2017), Montreal, Canada, Oct. 2017, pp. 1-7.
[11] K. M. Thilina, K. W. Choi, N. Saquib, and E. Hossain, “Machine
learning techniques for cooperative spectrum sensing in cognitive radio
networks,” IEEE J. Sel. Areas Commun., vol. 31, no. 11, pp. 2209-2221,
Nov. 2013.
[12] Y. Xu, P. Cheng, Z. Chen, Y. Li, and B. Vucetic, “Mobile collaborative
spectrum sensing for heterogeneous networks: A Bayesian machine
learning approach,” IEEE Trans. Signal Process., vol. 66, no. 21, pp.
5634-5647, Nov. 2018.
[13] K. Berbra, M. Barkat, F. Gini, M. Greco, and P. Stinco, “A fast spectrum
sensing for CP-OFDM cognitive radio based on adaptive thresholding,”
Signal Process., vol. 128, pp. 252-261, Nov. 2016.
[14] A. Vosoughi, J. R. Cavallaro, and A. Marshall, “A context-aware
trust framework for resilient distributed cooperative spectrum sensing in
dynamic settings,” IEEE Trans. Veh. Technol., vol. 66, no. 10, pp. 9177-
9191, Oct. 2017.
[15] S. Chaudhari, V. Koivunen, and H. V. Poor, “Autocorrelation-based
decentralized sequential detection of OFDM signals in cognitive radios,”
IEEE Trans. Signal Process., vol. 57, no. 7, pp. 2690-2700, Jul. 2009.
[16] P. Domingos and M. Pazzani, “On the optimality of the simple Bayesian
classifier under zero-one loss,” Machine Learning, vol. 29, pp. 103-130,
1997.

0018-9545 (c) 2019 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.