Received: 4 December 2021 Revised: 17 February 2022 Accepted: 23 March 2022 IET Communications

DOI: 10.1049/cmu2.12393

ORIGINAL RESEARCH

Parameter estimation of multiple frequency-hopping signals based

on Bayesian compressive sensing and atomic norm soft
thresholding with missing observations

Hongbin Wang Bangning Zhang Heng Wang Binbin Wu Daoxing Guo

College of Communications and Engineering, Army Abstract

Engineering University, Nanjing 210007, People’s
Republic of China
In this study, the problem of estimating multiple frequency-hopping (FH) signal parameters
in situations in which random observations are missing is addressed. A uniform linear array
Correspondence (ULA) is utilized to receive FH signals, and the missing observations are considered to be
Daoxing Guo, College of Communications and equivalent to randomly missing elements in the space-time matrix obtained from the ULA.
Engineering, Army Engineering University, Nanjing
Bayesian compressed sensing (BCS) estimates the hopping time of received FH signals by
210007, People’s Republic of China.
Email: [email protected] the spatial information from the space-time matrix and restores the missing observations.
The estimated hopping time is implemented as the boundary to divide the signals such that
Funding information each segment contains a superposition of time-invariant multiple components. Then, the
Jiangsu Provincial Natural Science Foundation of instantaneous frequency (IF) of multiple components can be estimated precisely by atomic
China, Grant/Award Number: BK20191328
norm soft thresholding (AST) within each segment. After estimating the hopping time and
IF, the direction of arrival (DOA) of multiple signals is directly calculated using these two
parameters. The simulation results show that the proposed method is superior to existing
approaches. Provided that sufficient array elements are available, multi-FH signal parame-
ters can be estimated with satisfactory accuracy even when a large portion of observations
is missing.

1 INTRODUCTION (CS) [19] are the most representative methods. Methods suitable
for traditional time-frequency analysis are short-time Fourier
Frequency-hopping (FH) signals are generated by varying the transform (STFT) [20, 21] and Wigner-Ville distribution (WVD)
carrier frequency according to a pseudo-random pattern. Owing [22]. However, it is difficult to set the STFT window length to
to their inherent advantages, namely a low probability of being satisfy the high-precision representation of both the time and
intercepted, flexible networking capabilities, resistance to jam- frequency domains together. The bilinear transformation of
ming and multipath fading, FH signals have become an appro- WVD produces a cross-term, which causes false energy distri-
priate choice and are widely used in satellite communications butions, and which affects the normal analysis of the signals.
[1], wireless communications [2, 3], physical layer security [4, 5], Even if appropriate TF kernels were able to alleviate the influ-
smart grids [6, 7], underwater communication [8, 9], Internet of ence of the cross-term, it would still be a compromise between
Things (IoT) technology [10, 11], and unmanned aerial vehicles the suppression of the cross-term and the time-frequency
(UAV) [12, 13]. In the fields of interference analysis and commu- aggregation. Parameter estimation has been transformed into an
nication security, estimating the parameters of FH signals and underdetermined linear regression problem with a dual sparsity
tracking them is a highly challenging, but crucial, task when the penalty [18]. The penalty function can jointly control the intrin-
hopping patterns are unknown. sic sparsity and the smoothness of the estimation. However,
Among many FH spectrum representation and parame- this approach does not take missing observations into account.
ter estimation techniques, joint time-frequency (TF) analysis In practice, the measurements may experience sample failure as
[14–17], sparse linear regression [18] and compressive sensing a result of channel distortion/fading, line-of-sight obstruction,

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is
properly cited.
© 2022 The Authors. IET Communications published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology.

888 wileyonlinelibrary.com/iet-com IET Commun. 2022;16:888–901.

 17518636, 2022, 8, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cmu2.12393 by Cochrane France, Wiley Online Library on [12/07/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
WANG ET AL. 889

removal of samples contaminated by impulsive noise, and estimation. The spatial frequency is defined as the phase differ-
failures of collecting/storage equipment. In situations such as ence of the signals received by the adjacent array elements. BCS
these, it is meaningful to study parameter estimation of FH realizes the spatial frequency estimation in the case of missing
signals in the case of missing observations. The authors in observations, with a drastic change in the spatial frequency indi-
[23] investigated the FH parameter estimation in the case of cating the hopping time. The second is instantaneous frequency
missing observations, and they proposed a waveform-adaptive (IF) estimation of the received signals. The IF is interpreted as
TF kernel design by combining an automatically optimized the fixed frequency of the FH component within a dwell time.
pre-filtering window and the data-dependent adaptive optimal After BCS recovers the space-time matrix, the IF is estimated
kernel (AOK). This designed kernel could effectively sup- based on the AST in the case of complete data. The signal recov-
press the cross-terms and artefacts arising from interference ery is conducive to accurately acquiring the noise power, thereby
produced by the bilinear transformation. However, instead determining the regularization parameters in AST. It should be
of depending on the TF kernel to eliminate the interference pointed out that the estimation of the noise power is beyond
incurred by using bilinear transformation, a more appropriate the scope of this study. Finally, we utilize the estimated hopping
alternative is to directly estimate the signal parameters based on time and IF to directly calculate the incident angle of the signals
the sparse representation framework. Bilinear transformation to accomplish DOA estimation. The FH signals parameter esti-
is not necessary from the perspective of parameter estimation. mation method proposed in this paper can be applied to non-
The sparse representation more accurately reflects the inherent cooperative FH signal processing in military fields such as elec-
characteristics of the signals, which can avoid the shortcomings tronic countermeasures and anti-jamming analysis. It can also
of bilinear transformation and achieve high-precision estima- be used for cooperative blind FH signal processing. In the case
tion even in the case of missing observations. In recent years, of unsatisfactory channel quality, random missing observations
the continuous deepening of research on sparse reconstruction lead to the failure of traditional linear time-frequency analysis
technology [24], means it is no longer an obstacle to use fewer methods. The proposed method can estimate the parameters of
or even damaged samples for signal recovery. the FH signal in harsh channel conditions.
Atomic norm soft thresholding (AST), which is derived from The main contributions of this study are summarized as fol-
research on linear inverse problems [25], is based on sparse rep- lows:
resentation theory. This method uses atomic norm minimiza-
tion to transform frequency estimation into semi-definite plan- ∙ We propose a method to estimate the hopping time from
ning (SDP) in stochastic noise. The method is especially suit- spatial frequency, which is defined as the phase difference
able for the high-precision estimation of parameters. A param- between the measurements of adjacent array elements. The
eter estimation method based on AST was proposed for the hopping of the signal frequency causes a drastic change in the
analysis of FH signals [26]. Although this method can effec- spatial frequency, which can be utilized to sense the occur-
tively estimate the frequency of received signals, the hopping rence of hopping when observations are missing. In this way,
time estimation depends on the occurrence of redundant fre- the estimation of the hopping time is transformed into the
quency components when the frequency changes. This method estimation of the spatial frequency. For incomplete data, we
therefore makes it difficult to provide robust estimation results exploit BCS to estimate the spatial frequency and reconstruct
because of the sensitivity of the spectrum environment. the space-time matrix.
In addition, the regularization parameter in AST play a signif- ∙ We propose a method to estimate the IF of the received sig-
icant role. It not only controls the calculation results but also nals based on AST. After BCS restores the space-time matrix,
affects the convergence rate of the algorithm. This parame- AST estimates the IF of the signals received by multiple array
ter is related to the noise power. When observations are miss- elements. Compared with the bilinear time-frequency (TF)
ing, it is difficult to accurately obtain the noise power, which representation framework, the proposed method does not
causes the setting value to deviate from the optimal design of the produce cross-terms, and makes it unnecessary to design a
regularization parameter. Although this dilemma can be allevi- TF kernel to eliminate interference. In addition, the super-
ated when the missing observations are restored, an estimation resolution capability of AST also produces high-precision
method that does not require manual intervention or empiri- estimation results.
cal settings is demanded before signal reconstruction. In-depth ∙ In this study, the DOAs of the different FH components are
research of sparse Bayesian learning (SBL) led to the proposal directly calculated using the estimated spatial frequency and
of Bayesian compressed sensing (BCS) [27]. From a Bayesian IF. The algorithm is simple and effective and yields satisfac-
perspective, BCS provides an estimate for the posterior density tory results even in the case of missing observations.
function of sparse weights and additive noise when implement-
ing the compressive measurements. BCS hardly needs precise The remainder of this paper is organized as follows. Section 2
parameter settings to achieve signal reconstruction with incom- presents the preliminaries of the multiple FH signals model.
plete data (or missing observations). In this paper, a method In Section 3, a method for the estimation of multiple FH
for estimating the parameters of multiple FH signals based on signal parameters is proposed. The method, which is based on
BCS and AST by uniform linear array (ULA) reception is pro- BCS and AST, is designed from the perspective of space-time-
posed. We construct two frequency estimation problems from frequency analysis, and is capable of estimating the hopping
the space-time-frequency analysis. The first is spatial frequency time, IF, and DOAs in the case of missing observations.
 17518636, 2022, 8, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cmu2.12393 by Cochrane France, Wiley Online Library on [12/07/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
890 WANG ET AL.

Section 4 presents the simulation results. Finally, conclusions array manifold matrix is
are drawn in Section 5. [ ( ) ( ) ( )]
A (𝚿n ) = a 𝜓n,1 , a 𝜓n,2 , … , a 𝜓n,Kh , (4)

2 PRELIMINARIES AND PROBLEM where

STATEMENT
( ) [ ]T
a 𝜓n,k = 1, e− j 2𝜋d fn,k cos 𝜑k ∕c , … , e− j 2𝜋(M −1)d fn,k cos 𝜑k ∕c .
Let us consider the parameter estimation for the reception (5)
of multiple FH signals with unknown hopping patterns. The
signals consist of Kh frequency components, which can be We define the spatial frequency as the phase difference
expressed as [23]: between adjacent antenna elements. Therefore, the spatial fre-
quency of the kth component is
∑ ∑
Kh H
Δ Δ
s(t ) = Ak,h ΠTk (t − hTk ) e j 2𝜋 fk,h (t −hTk ) , (1) fspatialk = 𝜓k = fk d cos 𝜑k ∕c. (6)
k=1 h=1

√ We assume that the sampling period satisfies

where j = −1, ΠTk (t ) represents a normalized boxcar func-
T T
tion, which is equal to one for t ∈ (− k , k ] and 0 other- Ts >
Dk
, (7)
2 2
wise. The duration of each hop of the kth individual FH c
signal is represented by Tk , and Kh denotes the number of
where
FH emitters. In addition, the complex amplitude and fre-
quency of the hth tone in the kth system-wise dwell are
Dk = (M − 1)d cos 𝜑k , (8)
denoted as Ak,h and fk,h , respectively. H is the number of
tones.
Nyquist sampling is performed on the signal s(t ), and the which denotes the impingement of the maximum propagation
sampling frequency is fs , thus Δt = 1∕ fs denotes the sam- distance of the kth signal component on all ULA elements.
pling interval. The kth sampled discrete-time FH signal can be This assumption is reasonable in the multi-antenna recep-
derived as [23] tion of narrowband FH signals. Then the signals received by
the ULA can be represented as
∑
H
sk (n) = Ak,h e j 2𝜋 fk,h nΔt . (2) ⎛ x1,1 x2,1 … xn,1 … xN ,1 ⎞
h=1 ⎜x x2,2 … xn,2 … xN ,2 ⎟
X=⎜ ⎟,
1,2
(9)
We assume that Kh FH signals impinge onto an ULA with ⎜⋯ ⋯ … ⋯ … ⋯ ⎟
⎜ ⎟
antenna spacing d and M array elements, 1 ≤ k ≤ Kh . We ⎝x1,M x2,M … xn,M … xN ,M ⎠
suppose that d < 𝜆min ∕2, where 𝜆min represents the mini-
mum carrier wavelength corresponding to the maximum fre- where X is the space-time matrix, and xn,m (0 < n ≤ N , 0 < m ≤
quency fmax among the Kh components. There are Kh far- M ) denotes the received data for the mth array element of the
field and narrowband signals with the total bandwidth B = nth snapshot.
fmax − fmin , and fmax , fmin are the maximum and mini-
mum frequency of these signals, respectively. We assume that
the array elements are isotropic, ignoring the effect of mis- 3 ESTIMATION OF FH SIGNAL
match and mutual of the antenna. The incidence angle 𝜑k PARAMETERS
is assumed to remain almost constant within the processing
time. 3.1 Hopping time estimation based on BCS
The nth snapshot of the ULA outputs after Nyquist sampling with missing observations
can be expressed as
( ) The arbitrary length of the boxcar function in Equation (1) leads
xn = x(n) = A 𝝍n s(n) + V(n), (3) to spectral energy dispersion, making it difficult to estimate the
IF through direct linear TF analyses. However, when the hop-
where 0 < n ≤ N , and N is the number of snapshots of ping time is estimated correctly, we can divide the signal into
Nyquist sampling. V(n) = [V1 (n), V2 (n), … , VM (n)]T ∈ CM ×1 multiple segments according to the hopping time. The IF of
denotes the additive white Gaussian noise vectors. We assume each segment can be accurately estimated without energy dis-
that the additive noise is temporally and spatially white with persion caused by frequency change.
zero-mean and covariance 𝜎2 . s(n) = [s1 (n), s2 (n), … , sKh (n)]T ∈ We consider the time-frequency representation of multi-
CKh ×1 represents the source signal vector in Equation (2). The ple asynchronous FH signals, as shown in Figure 1. For FH
 17518636, 2022, 8, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cmu2.12393 by Cochrane France, Wiley Online Library on [12/07/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
WANG ET AL. 891

where 𝚽n is the partial canonical identity (PCI) matrix [29] of

size L × M (L < M ):
{
1 if i ∈ {1, 2, … , L} and j ∈ Ω
𝚽n i, j = , (12)
0 otherwise

where Ω ⊂ {1, 2, … , M } such that |Ω| = L and | ⋅ | denotes the

cardinality of the set Ω [29, 30]. 𝚽n consists of partially selected
and permuted rows of the identity matrix corresponding to the
nth column of X. 𝚯n ∈ CL×M , vn ′ is the zero-mean Gaussian
noise with variance 𝜎n2 . wn contains the K elements with the
largest magnitude, and the remaining M − K elements in wn are
set to zero.
BCS was originally designed to recover real-valued sparse
solutions. However, the application of complex values is
reflected in various practical scenarios, and BCS for complex
signal processing has attracted increasing attention. These com-
FIGURE 1 Time-frequency representation of multi-FH signals
plex values were decomposed into real and imaginary compo-
nents and the two components were processed independently
[31]. Motivated by this accomplishment, we constructed a com-
plex BCS framework. xn , yn , vn ′ , and wn in Equation (10) are
component1, hopping occurs at time B, and the frequency vectors of complex values, and F is the normalized inverse of
before and after this moment are f1 and f2, respectively. FH the Fourier matrix with the complex values. The sparse sig-
component2 hops from frequency f3 to f4 at time C. FH com- nal vector 𝝎n = [Re(wn ), Im(wn )]T ∈ ℝ2M ×1 satisfies the zero-
ponent3 has a frequency change at time D, and the frequency mean Gaussian distributions [27, 31]
hops from f5 to f6. If the hopping time B, C, and D can be
correctly detected, as shown by the dotted line in Figure 1, ∏
2M ( )
−1
p (𝝎n |𝜶 n ) =  wn,i |0, 𝛼n,i , (13)
we can re-divide the signals into four segments: AB, BC, CD,
i=1
and DE. The frequency in each segment is the superposition
of multiple components but no longer changes. For exam- where 𝜶 n is the hyperparameter vector to be determined
ple, the frequency components in section AB are f1, f3, and and controlling the sparseness of the signal vector xn .
f5, and the frequency components in section CD are f2, f4, −1
 (wn,i |0, 𝛼n,i ) is a zero-mean Gaussian density function with
and f5. precision (inverse-variance) 𝛼n,i . Based on a previous report [27,
We notice that the key to the signal segmentation method 31], we invoke a hierarchical prior on 𝜶 n
mentioned above lies in accurately estimating the hopping time.
Although the frequency of the received signals can be roughly
∏
2M
( )
estimated by the phase difference of the measured data [28], this p (𝜶 n |a, b ) = Γ 𝛼n,i |a, b , (14)
approach is challenging to apply to the situation where observa- i=1
tions are randomly missing. In this study, we determine the hop-
ping time by estimating the change in the spatial frequency. The the overall prior on 𝝎n is then evaluated as
problem of estimating the hopping time is transformed into the
problem of estimating the spatial frequency fspatialk . To solve ∏
2M ∞ ( ) ( )
−1
this problem for missing observations, we consider a method p (𝝎n |a, b ) =  wn,i |0, 𝛼n,i Γ 𝛼n,i |a, b d 𝛼n,i .
i=1
∫0
based on BCS.
(15)
The observations in xn = [xn,1 , xn,2, … xn,M ]T ∈ CM ×1 (xn is
the nth column of X) are randomly missing, which can be The density function Γ(𝛼n,i |a, b) is the gamma function. We
expressed as set a = b = 0 as a default choice which avoids a subjective
choice of a and b and leads to simplifications of computation.
yn = 𝚽n xn = 𝚽n Fwn + 𝚽n vn = 𝚯n wn + v′ n , (10) The integral term in Equation (15) corresponds to the Student-t
distribution [32]. With an appropriate choice of a, b, the Student-
where t distribution is strongly peaked at approximately wn,i = 0, and
therefore the prior in Equation (15) favours most wn,i being
zero. Similarly, a Gamma prior is performed on the inverse of
𝚯n = 𝚽n F, (11a)
the noise variance 𝛼n,0 = 1∕𝜎n2 , which is
( ) ( )
xn = Fwn + vn , (11b) p 𝛼n,0 = Γ 𝛼n,0 |c, d . (16)
 17518636, 2022, 8, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cmu2.12393 by Cochrane France, Wiley Online Library on [12/07/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
892 WANG ET AL.

We also let c = d = 0 as a default choice. In the BCS- where 𝜶 n,−i in Equation (21) is the same as 𝜶 n , except 𝛼n,i is
based approach, the posterior density function for 𝝎n can be removed. According to [27]
expressed analytically as a Gaussian distribution with mean and
covariance Δ
sn,i = 𝚿Tn,i C−1
n,−i 𝚿n,i , (23a)
𝝁n = 𝛼n,0 𝚺n 𝚿Tn yn , (17a)
Δ
qn,i = 𝚿Tn,i C−1
n,−i yn . (23b)
( )−1
𝚺n = 𝛼n,0 𝚿Tn 𝚿n + An , (17b) −1
The (𝜶 n , 𝛼n,0 ) has a unique maximum with respect to 𝛼n,i

where 2
sn,i
[ ] 𝛼n,i = , 2
if qn,i > sn,i , (24a)
Re (𝚯n ) −Im (𝚯n ) 2
qn,i − sn,i
𝚿n = , (18a)
Im (𝚯n ) Re (𝚯n )
𝛼n,i = ∞, 2
if qn,i ≤ sn,i , (24b)
( )
An = diag 𝛼n,1 , … , 𝛼n,2M , (18b) Equations (24a) and (24b) control the addition and deletion
of particular 𝚿n,i from the signal representation. The learning
[ ( ) ( )]T process is performed by operations sequentially for varying i.
yn = Re yn , Im yn . (18c)
Compared to the inverse operation of 𝚺n in Equation (17b),
The associated learning problem is the search for the param- of which the complexity is of the order of (2M )3 , the basis func-
−1
eters 𝜶 n and 𝛼n,0 . The empirical Bayesian estimate for 𝜶 n and tions 𝚿n,i are sequentially added into or deleted from the model
−1 to maximize the marginal likelihood in the algorithm mentioned
𝛼n,0 is determined by maximizing the marginal likelihood, or its
−1
above. This process is more efficient with a recursive implemen-
logarithm (𝜶 n , 𝛼n,0 ). tation, even in complex value evaluation.
{ } ( ) Once the algorithm converges, we can obtain the estimated
−1
𝜶 n , 𝛼n,0 −1
= arg max−1  𝜶 n , 𝛼n,0 value 𝝎ˆ n = [Re(ŵn ), Im(ŵn )]T of 𝝎n . Here, F in Equations (10),
𝛼n ,𝛼n,0 (11a) and (11b) represents the inverse of the Fourier orthogonal
( ( )) basis. Therefore, wn represents the frequency bins, and the posi-
−1
= arg max−1 log Pr 𝝎n |𝜶 n , 𝛼n,0 tion at which the energy of its estimated value is concentrated
𝜶 n ,𝛼n,0
ŵn can directly represent the estimated spatial frequency. Signal
1 ([ ]) reconstruction is achieved by
= − arg max 2L log(2𝜋) + log |C n | + yTn C −1
n yn ,
2 𝜶 n ,𝛼−1
𝕏̂ n = 𝔽𝜔
n,0
ˆn , (25)
(19)
where 𝕏̂ n = [Re(x̂n ), Im(x̂n )]T , 𝝎
ˆ n = [Re(ŵn ), Im(ŵn )]T , and
−1
where Cn = 𝛼n,0 I + 𝚿n An 𝚿Tn ∈ ℝ2L×2L , and Cn can be decom-
posed by [ ]
Re(F) −Im(F)
∑ 𝔽= . (26)
−1
Cn = 𝛼n,0 I+ −1
𝛼n,m −1
𝚿n,m 𝚿Tn,m + 𝛼n,i 𝚿n,i 𝚿Tn,i = Cn,−i Im(F) Re(F)
m≠i

−1
Suppose the length of the energy detection window is P, and
+ 𝛼n,i 𝚿n,i 𝚿Tn,i , (20) the energy detection of ŵn can be expressed as

where Cn,−i is C with the contribution of the basis function 𝚿n,i ∑ | ( )|2 | ( )|2
i+P
removed, i ∈ [1, … , 2M ]. En,k = |Re ŵn,k | + |Im ŵn,k | ,
−1 | | | | (27)
The (𝜶 n , 𝛼n,0 ) can be rewritten as [27] k=i
i = 1, … , M − P + 1.
( ) ( ) ( )
−1 −1 −1
 𝜶 n , 𝛼n,0 =  𝜶 n,−i , 𝛼n,0 + 𝓁 𝛼n,i , 𝛼n,0 , (21) When En,k > 𝜉, ̂fspatialk,n = ⌊i + P∕2⌋ is the kth estimated
spatial frequency, and 𝜉 is the spatial frequency detection
where threshold.
[ ] We assume that the incident angle of the signal is either basi-
( ) 1 ( )
2
qn,i cally invariant or undergoes small changes continuously within a
−1
𝓁 𝛼n,i , 𝛼n,0 = log 𝛼n,i − log 𝛼n,i + sn,i + ,
2 𝛼n,i + sn,i small range during the processing time. Compared with the sig-
nificant impact of frequency hopping on the spatial frequency,
(22) slight changes in the incident angle are insufficient to cause
 17518636, 2022, 8, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cmu2.12393 by Cochrane France, Wiley Online Library on [12/07/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
WANG ET AL. 893

ALGORITHM 1 Hopping time estimation based on BCS and ULA When m in Equation (29) is small, too many observations are
1: Input X, 𝜉, 𝜂, 𝜆min , d , M , N . missing, and the estimation of the spatial frequency is inaccurate
2: for n = 1 ∶ N do
or even impossible. To avoid this situation, we find the proba-
bility of m > th as
3: Extract xn from X;
4: Build a BCS framework and estimate 𝝎
ˆn ; ( )
∑
th
M
5: Estimate En,k and ̂fspatialk,n from 𝝎
ˆ n = [Re(ŵn ), Im(ŵn )]T ; P (m > th) = 1 − (1 − p)k pm , (31)
m
6: Arrange [ ̂fspatial1,n , ̂fspatial2,n , … , ̂fspatialK ,n ] in order according to m=0
h
different signal components;
7: Calculate Dk,n by Eq. (28);
where th is the threshold of the minimum retained observa-
tions in each column of X when the requirements for the spa-
8: if Dk,n+1 > 𝜂 then
tial frequency estimation are met. The increase in probabil-
9: Record n as the hopping time; ity P (m > th) is conducive to improving the parameter estima-
10: Add n to the set of hopping time NT = {n}; tion and signal reconstruction accuracy. Obviously, P (m > th)
11: endif increases monotonically with M . However, the increase in M
12: endfor means an increase in computational complexity and applica-
13: Output NT .
tion cost. Therefore, the value considered for M should be as
small as possible when the condition P (m > th) is met. Then
we can determine M by solving the following optimization
problem.

min M , (32a)
the spatial frequency to change beyond the detection thresh-
old, which can be guaranteed by a reasonable and controllable
subject to
threshold setting. We define the difference in the kth spatial fre-
quency as ( )
∑
th
M
| | 1− (1 − p)k pm > pth , (32b)
Dk,n = | ̂fspatialk,n − ̂fspatialk,n−1 | , (28) m=0 m
| |

when Dk,n > 𝜂, and 𝜂 is the hopping time detection threshold, ( )

N −1
which indicates that the kth signal component is hopping. In
summary, Algorithm 1 presents the proposed algorithm for esti- N −n
mating the hopping time. p=1− ( ), (32c)
N
N −n
3.2 Effect of missing observations on ULA
elements M , m, N , n ∈ Z + ,

We investigate a scenario in which the samples are randomly where Z + denotes a positive integer and pth is a feasible proba-
missing from observations X in Equation (9). For a certain col- bility threshold to meet the requirements of parameter estima-
umn X, the probability of k samples being missing and m sam- tion and signal reconstruction. When n, N , th, m, and pth are
ples being reserved is given, Equation (32a) can be easily solved by varying the value
of M .
( )
M
P= (1 − p)k pm , (29)
k
3.3 Frequency estimation of multiple FH
where m + k = M , and
signals
( ) In the approaches mentioned above, a method based on BCS
N −1 estimates the spatial frequency, which restores the missing
N −n observations of X in Equation (9). A change in the spatial
p=1− ( ), (30) frequency indicates that the received signals were frequency
N
hopping. As multiple FH signals are divided into several seg-
N −n ments by the hopping time, the signal in each segment is the
superposition of multiple frequency components and no longer
where N − n samples are randomly lost for each row of X in changes. In this section, we present the approach we followed
Equation (9) and n samples are retained. to estimate the IF in each segment based on AST. AST is an
 17518636, 2022, 8, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cmu2.12393 by Cochrane France, Wiley Online Library on [12/07/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
894 WANG ET AL.

where s = [s(1), s(2), … , s(n), … , s(N )] is the sampled discrete-

time FH signals matrix and s(n) denotes the nth snapshot of the
superimposed signal of all FH components, and V denotes the
noise matrix. Then the space-time matrix in Equation (3) can be
rewritten as

X = S + V, (36)

p
where xr,m = [xn,m , xn+1,m , … , xn+pL −1,m ] represents the pth
segment and mth row in X. The length of the pth segment
p
is pL . The pth segment and mth row in S denotes sr,m =
[sn,m , sn+1,m , … , sn+pL −1,m ], and the pth segment and mth row in
p p p p
V is vr,m = [vn,m , vn+1,m , … , vn+pL −1,m ], then xr,m = sr,m + vr,m .
p
FIGURE 2 Schematic diagram of instantaneous frequency estimation sr,m is the sparse combination of “atoms” a( f ). The frequency
p
of sr,m represents the IF of the pth segment and mth row in X.
p p
To estimate the frequency of sr,m from the observation of xr,m ,
extension of “off the grid” technology [33, 34] in noisy envi- a norm referred to as the atomic norm is defined as [38]
ronments and is based on atomic norm minimization [35–37].
{ }
AST is a method for accurate spectrum estimation and effec- ‖ p ‖ Δ p
‖sr,m ‖ = tr > 0 ∶ sr,m ∈ tr conv()
tively uses continuous atoms in an infinite dictionary. This ‖ ‖
super-resolution method demonstrates that a sparse sum of { }
complex sinusoids can be reconstructed exactly from time ∑ p ∑
= inf ci ∶ sr,m = ci ai , ci ≥ 0, ai ∈  ,
samples provided the frequencies are sufficiently far apart from i i
one another [33].
The signal received by each array element is independently (37)
used to estimate IF. Figure 2 shows a schematic diagram of the
method for IF estimation. where i denotes the ith atomic. The atomic norm ‖ ⋅ ‖ can be
As shown in Figure 2, the hopping time divides the received computed via semidefinite programming (SDP) [39]
signal into multiple segments. Within each segment (the middle
part of the red dotted line), each row in the space-time matrix is ‖ p ‖ 1( )
‖sr,m ‖ = min tr + u1,r , (38a)
extracted to estimate the IF. We express the estimated IF of the ‖ ‖ tr ,ur 2
pth segment as
[ p H
]
̃ tr sr,m
1 ∑
M
subjectto ≥ 0, (38b)
f̂IF p = f̂IF p,m , (33) p
̃
M sr,m T (ur )
m=1

where M ̃ is the number of effective estimation results of vec- where ur ∈ ℂ pL , and T (ur ) ∈ ℂ pL ×pL denotes a (Hermitian)
tor f̂IF p,m by different array elements. The effective estimation Toeplitz matrix with
results here are those for which the number of estimated IFs in
the vector f̂IF p,m is equal to Kh . ⎡ u1 u2 ⋯ u pL ⎤
⎢ H ⎥
We employ an AST-based method to estimate f̂IF p,m . Accord- ⎢ u2 u1 ⋯ u pL −1 ⎥
T (ur ) = ⎢ ,
⋮ ⎥⎥
(39)
⎢⋮ ⋮ ⋱
ing to the general method proposed in [33], we define the atoms
a( f , 𝜙) ∈ ℂ pL ( f ∈ [0, 1]and𝜙 ∈ [0, 2𝜋)) as ⎢uH
⎣ pL uH
p −1 ⋯ u1 ⎥⎦
L
[ ]T
a( f , 𝜙) = 1, e j (2𝜋 f +𝜙) , … , e j (2𝜋 ( pL −1 ) f +𝜙 ) = e j (2𝜋 fl +𝜙) , where ui denotes the ith entry of u, and uiH denotes the conju-
(34) gate transpose of ui .
where l ∈ {0, … , pL − 1}. Let  be a collection of atoms Accordingly, the dual norm of atomic norm ‖sr,m ‖ is
p
a( f , 𝜙) = a( f )𝜙, and its convex hull conv() satisfies the defined as [33]
requirement of being compact, centrally symmetric, and con-
tains the origin as an interior point. ‖ p ‖∗ ⟨ p p ⟩ ⟨ p i𝜙 ⟩
‖qr,m ‖ = sup qr,m , sr,m ℝ = sup qr,m , e a( f , 0) ℝ
We define S as the noise-free FH signal matrix, which can be ‖ ‖ 𝜙∈[0,2𝜋), f ∈[0,1]
expressed as
|⟨ p ⟩|
= sup | qr,m , a( f , 0) | . (40)
S = A(𝝍)s, (35) f ∈[0,1] | |
 17518636, 2022, 8, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cmu2.12393 by Cochrane France, Wiley Online Library on [12/07/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
WANG ET AL. 895

p ALGORITHM 2 IF estimation based on AST and ULA

We consider the minimization of the atomic norm ‖sr,m ‖ .
An AST problem can be established as [35, 36, 39] 1: Input X, 𝜏r , M , Kh , NT .
2: Divide the received signals into P segments according to the hopping
1‖ p p ‖2 ‖ p ‖ timeNT .
min ‖ x − s ‖ + 𝜏r ‖sr,m ‖ , (41)
p
sr,m 2 ‖ r,m r,m
‖ 2 ‖ ‖ 3: for p = 1 ∶ P do

4: ̃ = 0;
Initialize M
p
where 𝜏r is the regularization parameter. in (41) can be ‖sr,m ‖ 5: for m = 1 ∶ M do
p
computed by Equations (38a) and (38b), then the AST problem 6: Extract xr,m = [xn,m , xn+1,m , … , xn+pL −1,m ];
in (41) can be formulated as the following SDP by 7:
p
Estimate f̂IF p,m based on xr,m by solve Eq. (45a) to Eq. (45c);

𝜏r ( ) 1‖ p p ‖2
8: if the number of components in f̂IF p,m is Kh then
minp tr + u1,r + ‖xr,m − sr,m ‖ , (42a)
tr ,ur ,sr,m 2 2‖ ‖2 9: Record f̂IF p,m as the instantaneous frequency;
10: ̃ =M
M ̃ + 1;
[ p H
]
tr sr,m 11: endif
subjectto p
≥ 0. (42b)
12: endfor
sr,m T (ur )
13: Calculate f̂IF p by Eq. (33);
p ∗ 14: Incorporate f̂IF p into instantaneous frequency set FIF = {f̂IF p };
Given the optimal solution (tr∗ , u∗r , sr,m )
of Equations (42a)
15: endfor
and (42b) from the Vandermonde decomposition of T (ur ),
p 16: Output FIF .
then we can estimate the frequency of sr,m , so-called f̂IF p,m .
However, to directly localize the frequency, we consider the dual
problem of AST in Equations (42a) and (42b). The dual prob-
lems of AST can be established as follows [36]
pL −l
{
∑ 1, l =0
1 ‖ p ‖2 1 ‖ p p ‖2 Hk,k+l = (45c)
max ‖x ‖ − ‖xr,m − 𝜏r qr,m ‖ ,
p 2 ‖ r,m ‖2
(43a) 0, l = 1, 2, … , pL − 1,
qr,m 2‖ ‖2 k=1

|⟨ p ⟩| where H in Equation (45c) is a Hermitian matrix, qr,m ∗

p
subjectto sup || qr,m , a( f ) || ≤ 1, (43b) p
f | | denotes the conjugate transpose of qr,m . We apply the con-
vex optimization toolbox—the CVX toolbox in MATLAB—to
where the norm of qr,m is ‖qr,m ‖∗ , which represents the dual
p p solve the optimization problems in Equations (45a)–(45c). Esti-
mation of the exact frequency of f̂IF p,m can always be achieved
norm of the atomic norm sr,m . The calculation of ‖qr,m ‖∗ can
p p
when the frequency components are sufficiently separate.
be realized by Equation (40). Considering the constraints in Algorithm 2 describes the IF estimation based on AST and
Equation (43b), we can obtain ULA.
|∑ |
|⟨ p ⟩| p
‖ p ‖∗ | L p 2𝜋l f |
| | |
‖qr,m ‖ = sup | qr,m , a( f ) | = sup | qr,m,l e | ≤ 1.
‖ ‖ f ∈[0,1] | | 3.4 DOA estimation and network sorting
| f ∈[0,1] | l =1 |
| |
(44)
Therefore, ‖qr,m ‖∗ is the maximum absolute value of poly-
p The spatial frequency of each snapshot f̂spatialn is estimated
∑ pL p 2𝜋l f according to Algorithm 1, and the IF in each segment f̂IF p
nomial l =1 qr,m,l e attained on the unit circle. This polyno-
is estimated by Algorithm 2. The DOA estimation algorithm
mial is a so-called dual polynomial, which can be represented by
∑ pL p 2𝜋l f is implemented based on Equation (6). When both the spa-
Q( f ) = l =1 qr,m,l e . Thus, the constraints on dual norms tial frequency and the IF are estimated, the incidence angle
in Equation (44) are transformed into constraints on dual poly- ̂fspatial c ̂fspatial 𝜆̂ p,k
nomials |Q(̂ f )| ≤ 1. These constraints can be equivalent to the 𝜑̂ k = arccos( n,k
), that is 𝜑̂ k = arccos( n,k
), where
2𝜋 ̂fIF p,k d 2𝜋d
SDP form, and we rewrite the dual problem of AST as [33]
𝜆̂ p,k =
c
̂fIF
. However, the spatial frequency fspatialn,k and the IF
p,k
1 ‖ p ‖2 1 ‖ p p ‖2 fIF p,k are required to satisfy the pairing relationship, that is, they
max ‖x ‖ − ‖xr,m − 𝜏r qr,m ‖ ,
p 2 ‖ r,m ‖2
(45a)
qr,m 2 ‖ ‖2 belong to the same FH signal. We know that the incidence angle
of the signal remains basically invariant during the processing
subject to time. We calculate ̂fspatialn,k 𝜆̂ p,k for different k and perform clus-
tering until the number of clusters is Kh . In this way, the network
[ p ] sorting is completed according to the clustering results.
H −qr,m
≥ 0, (45b) The spatial incidence angle 𝜑k basically remains constant
p ∗
−qr,m 1 before and after frequency hopping. We assume that the
 17518636, 2022, 8, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cmu2.12393 by Cochrane France, Wiley Online Library on [12/07/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
896 WANG ET AL.

(p − 1)th segment and the pth segment are separated by nt . SDP problem, and the complexity is cubic to the dimension of
According to Equation (6), we have the Toeplitz matrix. Therefore, the complexity of the frequency
set estimation is O(P (2M )PL3 ), where M is the number of subar-
𝜆̂ IF p−1,k ̂fspatialnt −1,k 𝜆̂ IF p,k ̂fspatialnt ,k ray elements, P is the number of segments, and PL denotes the
cos 𝜑k = = , (46) length of P. Although the order of magnitude of the computa-
d d
tional complexity of the proposed method is equivalent to that
of the method in [23], the subsequent simulations clearly indi-
where 𝜆̂ IF p−1,k is the estimated wavelength of the kth signal
cate that the parameter estimation performance of the proposed
component in the (p − 1)th segment, and 𝜆̂ IF p−1,k = c∕ ̂fIF p−1,k . method is superior to that of the method proposed in [23].
̂fspatial is the estimated spatial frequency of the kth com-
nt −1,k
ponent according to the (nt − 1)th snapshot. Similarly, 𝜆̂ IF p,k is
the estimated wavelength of the kth signal component in the 4 SIMULATION AND ANALYSIS
pth segment, and ̂fspatialnt ,k is the estimated spatial frequency of
the kth component according to the nt th snapshot. Because it is In this section, numerical experiments are presented to evaluate
affected by estimation error and noise, Equation (46) is too strict the performance of the proposed algorithm in comparison with
to be realized. Therefore, we calculate 𝜓̂ n,k = 𝜆̂ IF p,k ̂fspatialn,k for those reported in the literature. In view of their effectiveness,
all n ∈ {1, , 2, … , N } in their respective segment and all k in the we employ the same simulation settings as [18, 23, 42]. In the
total Kh component. When the result can be divided into Kh following experiments, the SNR is defined as [18, 23, 42]
clusters with the known number of sources as Kh , the pairing ( )
is considered complete. Then, 𝜑k of the nth snapshot can be Δ
2
‖s‖2
calculated by SNR = 10log10 , (48)
Ns 𝜎2
( )
𝜓̂ n,k where s is the signal vector received by the ULA, Ns is the num-
𝜑̂ n,k = arccos . (47)
2𝜋d ber of time indices and 𝜎2 denotes the power of additive white
Gaussian noise.
In particular, two performance measures are defined to evalu-
3.5 Computational complexity ate the hopping time and the IF detection performance, respec-
tively. The ratio of the correct hopping time detection is defined
In this section, the complexity of the proposed method and four as [14, 23]
existing methods are analysed and compared, namely methods
1 ∑
based on STFT [40], sparse linear regression (SLR) [41], and Mc
Δ
sparse Bayesian learning (SBL) [42], in addition to the method Pt = D (i ), (49)
Mc i=1 t
of [23].
The computational complexity of STFT, O(N log2 L), where
where Mc is the number of Monte Carlo trails and Dt (i ) is the
L is the length of the short-time slide window, is the lowest
ratio of correct detections in the ith Monte Carlo trial. If the
among that of all the methods mentioned above [23]. The com-
estimated hopping instant is less than three observations away
plexity of the SLR-based method is O(N 2 L 2 ) for one itera-
from the associated true hopping instant, then the estimated
tion, which is implemented by using the alternating direction
hopping instant is declared the correct hopping time detection.
method of multiplier (ADMM), where L is the number of fre-
The ratio of incorrect IF detection is defined as [14, 23]
quency bins. In comparison, the computational complexity of
the SBL method mainly depends on the matrix inversion oper-
1 ∑
Mc
ation when calculating Equation (17b), which is O(P𝜍 3 + K g3 ) Δ
Ef = 1 − D (i ), (50)
[42], where K is the number of latent parameters and 𝜍 is chosen Mc i=1 f
to be a small value compared with P. In the previously proposed
approach [23], the computational complexity includes multiple where D f (i ) is the ratio of correct frequency detections in the
terms, and the order of the overall asymptotic computational ith Monte Carlo trial.
complexity is determined by the fastest-growing term, which is The DOA estimation error of FH signals is defined as follows
O(N 3 ), where N is the total length of signals. [43]
In our proposed scheme, the complexity of estimation of the
√
1 ∑ ( (i ) ̂ (i ) )2
hopping time is based on BCS. Because we utilize a fast algo-
rithm, the principled and efficient sequential addition and dele- RMS E𝜃 = 𝜃k − 𝜃k , (51)
tion of the candidate basis function to monotonically maximize I Kh k,i
the marginal likelihood replaces the calculation of the matrix
inverse operation of 𝚺. The calculation complexity is an order where I represents the total number of Monte Carlo trials and
of magnitude lower than that of SBL. The estimation of IF is an Kh is the number of signal components.
 17518636, 2022, 8, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cmu2.12393 by Cochrane France, Wiley Online Library on [12/07/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
WANG ET AL. 897

mated accuracy of the hopping time when as many as 30% of

the observations are missing is higher than that in [23], the esti-
mation performance at other missing proportions is lower than
that in [23]. The main reason is the inaccurate spatial frequency
estimation owing to the small number of array elements and low
SNR. When the number of array elements increases to 48 the
accuracy of the spatial frequency estimation also improves. Sim-
ulation in the SNR range 0–15 dB, with the same missing obser-
vations ratio, the estimation accuracy of the hopping time of the
proposed method is higher than that of the method in [23], as
shown in Figure 4b. In addition, regardless of whether M = 32
or M = 48, the improvement rate (curve slope) of the estima-
tion accuracy of the hopping time by the proposed method is
higher than that of the method in [23], which is mainly attributed
to the spatial signal processing gain resulting from multi-array
element reception.
The results in Figure 5 show that when M = 32 and SNR<3
FIGURE 3 Statistic comparisons among different methods for the dB, the IF detection performance of the proposed method is
correct hopping time detection ratio without missing observations equivalent to that of SBL and slightly worse than the method
in [23], but as the SNR increases, the detection error decreases
sharply. This phenomenon is mainly owing to the super-
The FH signals are generated as follows [23, 42]: The sam- resolution estimation of AST. When the SNR is sufficiently
pling frequency fs is 64 kHz. The first FH component hops at large, the infinite atom dictionary in the AST can achieve a
time index 16, the frequency before the hopping time is 13 kHz, nearly “off the grid” estimation effect, which can theoretically
and the frequency after the hopping time is 18 kHz. The second be infinitely close to the actual value. When M = 48, the estima-
FH component is active with the frequency of 28 kHz within tion performance of IF is superior to that of existing methods
the range of time index [0:31], and the frequency hops to 23 within the range of the simulated SNR. Two factors are mainly
kHz within the range of time index [32:63]. The third FH com- responsible for this phenomenon. The first is that the increase
ponent hops from 3 to 6 kHz within the range of time index in M improves the estimation accuracy of the hopping time,
[0:47] and [48:63], respectively. Figure 2 shows the TF trajec- thereby improving the signal segmentation accuracy, which is
tories of the FH signals and the signal with three components more conducive to the accurate estimation of IF. The second
have different DOAs 𝜃 = [40◦ , 58◦ , 78◦ ]. is that the increase in M weakens the influence of the estima-
To demonstrate the effectiveness of the proposed method tion error on the result. After outliers and average processing
with statistical results, 20,000 Monte Carlo trials were conducted are eliminated from multi-element estimation values, they are
with the input SNRs varying from 0 to 15 dB. We mainly con- closer to the actual value.
sider three critical parameters: the correct hopping time detec- Figure 6 compares the IF estimation performance of the pro-
tion ratio, the ratio of incorrect IF detection, and the DOA esti- posed method and the method in [23] with the same miss-
mation error. Motivated by [18, 23, 42], the different existing ing observations ratio and SNR, and the analysis results in
methods include STFT, SLR, SBL, and the method in [23] as Figure 6 are similar to those in Figure 5. It is worth noting that
the objects of comparison. although the AST-based method can directly estimate the IF
Figure 3 compares the correct hopping time detection ratio when observations are missing, it does not perform well in the
without missing observations, as calculated with STFT, SLR, case of incomplete data. Therefore, the proposed method uti-
SBL, and the method in [23]. The proposed approach outper- lizes BCS to recover the missing observations in the space-time
forms most of the existing methods. However, when SNR<2 matrix when estimating the hopping time. Estimation of the IF
dB and M = 32, the correct hopping time detection ratio of based on AST involves the complete data space-time matrix.
the proposed method is lower than the method in [23]. This When M = 48, the incorrect IF detection ratio of the proposed
phenomenon is mainly caused by inaccurate spatial frequency method is lower than that for the same proportion of missing
estimation owing to the small number of array elements at low observations in [23]. With the continuous increase in the SNR,
SNR. When the number of array elements increases to 48, the the advantages of the proposed method are becoming increas-
calculation of the correct hopping time detection ratio (without ingly prominent.
missing observations) is more accurate than that of the existing From the simulation above, it can be seen that when M =
method in the SNR range of 0–15dB. 48, the proposed method is more competitive than the existing
We additionally compared the proposed method with the techniques in the estimation accuracy of hopping time and IF,
method in [23] with respect to the same missing observations regardless of whether any data are missing.
ratio. Figure 4a presents a comparison of the estimation accu- To further illustrate the application of the proposed method
racy of the proposed method when M = 32 with the method in to different FH sequences, we consider the parameter estima-
[23]. These results show that, at low SNR, except that the esti- tion of wide-gap FH sequences and compare the estimation
 17518636, 2022, 8, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cmu2.12393 by Cochrane France, Wiley Online Library on [12/07/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
898 WANG ET AL.

FIGURE 4 Statistic comparisons with the method in [23] for the correct hopping time detection ratio with missing observations

performance with general FH sequences (sequences used in

[23, 42]).
As shown in Figures 7 and 8, in the case of 20% miss-
ing observations, whether M = 32 or M = 48, the estimation
performance of the hopping time and IF of the wide-gap FH
sequence is better than that of the general sequence. This is
mainly because the frequency distance between adjacent fre-
quency components in wide-gap FH sequences is large. Accord-
ing to the proposed method, the spatial frequency distance is
also increased, which is not easy to cause confusion between
spatial frequencies to obtain better estimation results of hop-
ping time. In the estimation of the IF, on the one hand, the
accurate estimation of the hopping time can bring about a more
precise signal division, which is beneficial to the estimation of
the IF. On the other hand, the more significant signal frequency
distance improves frequency estimation accuracy using AST.
Figure 9 shows the estimation bias of DOA for the different
FIGURE 5 Statistic comparisons among different methods for the SNR conditions of our work. As the SNR increases, the RMSE
incorrect IF detection ratio without missing observations calculated according to Equation (51) gradually decreases. In
the same proportion of missing observations, the RMSE for

FIGURE 6 Statistic comparisons with the method in [23] for the incorrect IF detection ratio with missing observations
 17518636, 2022, 8, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cmu2.12393 by Cochrane France, Wiley Online Library on [12/07/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
WANG ET AL. 899

FIGURE 9 Bias of DOA estimation in different SNRs

FIGURE 7 Statistic comparisons with the wide-gap FH sequences for the

correct hopping time detection ratio with 20% missing observations

FIGURE 10 Bias of DOA estimation in different SNRs for signals that

are close in space

late into the DOA estimation, resulting in increased errors. The

FIGURE 8 Statistic comparisons with the wide-gap FH sequences for the result is shown in Figure 10.
incorrect IF detection ratio with 20% missing observations

5 CONCLUSION
M = 48 is lower than that for M = 32. When SNR>7 dB and
there are no missing observations, the RMSE for M = 48 is less In this paper, we proposed a novel approach for estimating
than 1◦ . Even when as many as 30% of the observations are multi-FH signal parameters based on BCS and AST considering
missing, the RMSE for M = 48 is satisfactory. missing observations. A ULA was employed to receive multi-
We further explore DOA estimation results for signals that ple FH signals in different directions. Regardless as to whether
are close in space. We assume that the incidence angles of the observations were missing, BCS was used to estimate the spa-
3 FH signals are 54◦ , 58◦ and 61◦ , respectively. The pairing of tial frequency and restore the space-time matrix. The hopping
spatial frequencies and IF of different FH signal components time is determined as the point at which a sudden change in
depends on the clustering of incidence angles. When the angles the spatial frequency is detected. The signals are divided by the
are close, the clustering effect is poor, resulting in the deteri- hopping time. Each segment is a time-invariant signal whose IF
oration of DOA estimation performance. In addition, spatial can be estimated by an AST-based method. After completing
frequency estimation errors and IF estimation errors accumu- the spatial frequency and IF estimation, we calculated the DOA
 17518636, 2022, 8, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cmu2.12393 by Cochrane France, Wiley Online Library on [12/07/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
900 WANG ET AL.

directly by the two parameters. Unlike the FH parameter esti- the presence of an untrusted relay. IEEE Trans. Wireless Commun. 19(10),
mation method based on bilinear transformation, the proposed 6771–6785 (2020)
12. Kaplan, B., Kahraman, I., Gorcin, A., Cirpan, H.A., Ekti, A.R.:
method does not produce cross-terms, and no time-frequency
Measurement based FHSS?type drone controller detection at
kernel is required. In addition, the AST-based spectrum esti- 2.4GHz: an STFT approach. In: 2020 IEEE 91st Vehicular Tech-
mation method also has the advantage of super-resolution. nology Conference (VTC2020-Spring), pp. 1–6. IEEE, Piscataway
The simulation results show that when sufficient array ele- (2020)
ments are used, the proposed method is superior to exist- 13. Mototolea, D., Youssef, R., Radoi, E., Nicolaescu, I.: Non-cooperative low-
complexity detection approach for FHSS-GFSK drone control signals.
ing technology even if a large portion of data observations is
IEEE Open J. Commun. Soc. 1, 401–412 (2020)
missing. 14. Wei, S., Zhang, M., Wang, G., Sun, X., Zhang, L., Chen, D.: Robust
multi-frame joint frequency hopping radar waveform parameters esti-
ACKNOWLEDGEMENTS mation under low signal-noise-ratio. IEEE Access 7, 177198–177210
(2019)
This research was supported by the Jiangsu Provincial Natural
15. Fu, W., Li, X., Liu, N., Hei, Y., Wei, J.: Parameter blind estimation of
Science Foundation of China (No. BK20191328). frequency-hopping signal based on time-frequency diagram modification.
Wirel. Personal Commun. 97(3), 3979–3992 (2017)
CONFLICT OF INTEREST 16. Kanaa, A., ShaAmeri, A.Z.: A robust parameter estimation of FHSS sig-
nals using time-frequency analysis in a non-cooperative environment. Phys.
The authors declare no conflict of interest.
Commun. 26, 9–20 (2018)
17. Jin, Y., Liu, J.: Parameter estimation of frequency hopping signals based
DATA AVAILABILITY STATEMENT on the Robust S-transform algorithms in alpha stable noise environment.
The data that support the findings of this study are available AeuInt. J. Electron. and Commun. 70(5), 611–616 (2016)
from the corresponding author upon reasonable request. 18. Angelosante, D., Giannakis, G.B., Sidiropoulos, N.D.: Estimating multiple
frequency-hopping signal parameters via sparse linear regression. IEEE
Trans. Signal Process. 58(10), 5044–5056 (2010)
ORCID 19. Fu, W., Jiang, T.: A parameter estimation algorithm for multiple frequency-
Hongbin Wang https://orcid.org/0000-0002-0534-8749 hopping signals based on compressed sensing. Phys. Commun. 37, 100892
Binbin Wu https://orcid.org/0000-0001-6605-8270 (2019)
20. Fu, W., Hei, Y., Li, X.: UBSS and blind parameters estimation algorithms
for synchronous orthogonal FH signals. J. Syst. Eng. and Electronics 25(6),
REFERENCES 911–920 (2014)
1. Lee, S., Kim, S., Seo, M., Har, D.: Synchronization of frequency hopping by 21. Tong, L., Yinhui, T., Jun, L.: Parameter estimation of FH signals based on
LSTM network for satellite communication system. IEEE Commun. Lett. STFT and music algorithm. In: 2010 International Conference on Com-
23(11), 2054–2058 (2019) puter Application and System Modeling (ICCASM 2010), vol. 5, pp. 911–
2. Ma, J., Yang, Y.: A Generalized anti-interference low-ambiguity dual- 920. IEEE, Piscataway (2010)
frequency multiplexing modulation based on the frequency-hopping tech- 22. Qian, Y.: Estimation method of frequency hopping parameter of DS/FH
nique. IEEE Access 8, 95288–95300 (2020) signal based on improved SPWVD. Shipboard Electronic Countermeasure
3. Li, Y., Huo, K., Li, Q., Liu, X., Jiang, S., Liu, J., Ni, X.: A novel method (2015)
of wireless power transfer identification and resonance decoupling based 23. Liu, S., Zhang, Y.D., Shan, T., Tao, R.: Structure-aware bayesian
on frequency hopping communication. IEEE Access 7, 161201–161210 compressive sensing for frequency-hopping spectrum estimation with
(2019) missing observations. IEEE Trans. Signal Process. 66(8), 2153–2166
4. Yazicigil, R.T., Nadeau, P., Richman, D., Juvekar, C., Vaidya, K., (2018)
Chandrakasan, A.P.: Ultra-fast bit-level frequency-hopping transmitter for 24. Candes, E.J., Wakin, M.B.: An introduction to compressive sampling.
securing low-power wireless devices. In: 2018 IEEE Radio Frequency IEEE Signal Process Mag. 25(2), 21–30 (2008)
Integrated Circuits Symposium (RFIC), pp. 176–179. IEEE, Piscataway 25. Chandrasekaran, V., Recht, B., Parrilo, P.A., Willsky, A.S.: The convex
(2018) geometry of linear inverse problems. Foundations of Computational Math-
5. Sa Sousa, J., Vilela, J.P.: Uncoordinated frequency hopping for wireless ematics 12(6), 805–849 (2012)
secrecy against non-degraded eavesdroppers. IEEE Trans. Inf. Forensics 26. Runze, D., Ying, G., Hongguang, L., Ping, S., Yinsong, Y.: Parameter esti-
Secur. 13(1), 143–155 (2018) mation of multi frequency-hopping signals based on atomic norm. In: 2018
6. Zeng, Q., Li, H., Peng, D.: Frequency-hopping based communication net- 10th International Conference on Communication Software and Networks
work with multi-level qoss in smart grid: code design and performance (ICCSN), pp. 391–395. IEEE, Piscataway (2018)
analysis. IEEE Trans. Smart Grid 3(4), 1841–1852 (2012) 27. Ji, S., Xue, Y., Carin, L.: Bayesian compressive sensing. IEEE Trans. Signal
7. Malandra, F., Sanso, B.: A markov-modulated end-to-end delay analysis of Process. 56(6), 2346–2356 (2008)
large-scale RF mesh networks with time-slotted ALOHA and FHSS for 28. Zhang, K.-f., Guo, Y., Qi, Z.: A parameter estimation algorithm for mul-
smart grid applications. IEEE Trans. Wireless Commun. 17(11), 7116– tiple frequency-hopping signals based on sparse Bayesian method. Math.
7127 (2018) Prob. Eng. 2017, 6129120 (2017)
8. Schmidt, J.H.: Using fast frequency hopping technique to improve relia- 29. Ansari, N., Gupta, A.: Image reconstruction using matched wavelet esti-
bility of underwater communication system. Applied Sciences 10(3), 1–12 mated from data sensed compressively using partial canonical identity
(2020) matrix. IEEE Trans. Image Process. 26(8), 3680–3695 (2017)
9. Sun, D., Hong, X., Cui, H., Liu, L.: Iterative multichannel FH-MFSK 30. Jain, N., Gupta, A., Bohara, V.A.: PCI-MDR: missing data recovery in wire-
reception in mobile shallow underwater acoustic channels. IET Commun. less sensor networks using partial canonical identity matrix. IEEE Wireless
14(5), 838–845 (2020) Commun. Lett. 8(3), 673–676 (2019)
10. Jeon, W.S., Seo, S.B., Jeong, D.G.: Effective frequency hopping pattern for 31. Carlin, M., Rocca, P., Oliveri, G., Viani, F., Massa, A.: Directions-of-arrival
ToA estimation in NB-IoT random access. IEEE Trans. Veh. Technol. estimation through bayesian compressive sensing strategies. IEEE Trans.
67(10), 10150–10154 (2018) Antennas Propag. 61(7), 3828–3838 (2013)
11. Letafati, M., Kuhestani, A., Behroozi, H., Ng, D.W.K.: Jamming-resilient 32. Tipping, M.E: Sparse bayesian learning and the relevance vector machine.
frequency hopping-aided secure communication for internet-of-things in J. Mach. Learn. Res. 1(3), 211–244 (2001)
 17518636, 2022, 8, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cmu2.12393 by Cochrane France, Wiley Online Library on [12/07/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
WANG ET AL. 901

33. Tang, G., Bhaskar, B.N., Shah, P., Recht, B.: Compressed sensing off the 41. Angelosante, D., Giannakis, G.B., Sidiropoulos, N.D.: Estimating multiple
grid. IEEE Trans. Inf. Theory 59(11), 7465–7490 (2013) frequency-hopping signal parameters via sparse linear regression. IEEE
34. Lu, Z., Ying, R., Jiang, S., Liu, P., Yu, W.: Distributed compressed sensing Trans. Signal Process. 58(10), 5044–5056 (2010)
off the grid. IEEE Signal Process Lett. 22(1), 105–109 (2015) 42. Zhao, L., Wang, L., Bi, G., Zhang, L., Zhang, H.: Robust frequency-
35. Yang, Z., Xie, L.: On gridless sparse methods for line spectral estimation hopping spectrum estimation based on sparse bayesian method. IEEE
from complete and incomplete data. IEEE Trans. Signal Process. 63(12), Trans. Wireless Commun. 14(2), 781–793 (2015)
3139–3153 (2015) 43. Wan, J., Zhang, D., Xu, W., Guo, Q.: Parameter estimation of multi fre-
36. Tang, G., Bhaskar, B.N., Recht, B.: Near minimax line spectral estimation. quency hopping signals based on space-time-frequency distribution. Sym-
IEEE Trans. Inf. Theory 61(1), 499–512 (2015) metry 11(5), 648–793 (2019)
37. Li, S., Yang, D., Tang, G., Wakin, M.B.: Atomic norm minimization for
modal analysis from random and compressed samples. IEEE Trans. Signal
Process. 66(7), 1817–1831 (2018)
38. Chandrasekaran, V., Recht, B., Parrilo, P.A., Willsky, A.S.: The convex How to cite this article: Wang, H., Zhang, B., Wang,
geometry of linear inverse problems. Found. Computat. Math. 12(6), 805– H., Wu, B., Guo, D.: Parameter estimation of multiple
849 (2012) frequency-hopping signals based on Bayesian
39. Bhaskar, B.N., Tang, G., Recht, B.: Atomic norm denoising with applica- compressive sensing and atomic norm soft thresholding
tions to line spectral estimation. IEEE Trans. Signal Process. 61(23), 5987–
with missing observations. IET Commun. 16, 888–901
5999 (2013)
40. Cohen, L.: Time-frequency distributions: A review. Proc. IEEE 77(7), (2022). https://doi.org/10.1049/cmu2.12393
941–981 (1989)