ArticlePDF Available

Radar Active Jamming Recognition under Open World Setting

MDPI
Remote Sensing
Authors:

Abstract and Figures

To address the issue that conventional methods cannot recognize unknown patterns of radar jamming, this study adopts the idea of zero-shot learning (ZSL) and proposes an open world recognition method, RCAE-OWR, based on residual convolutional autoencoders, which can implement the classification of known and unknown patterns. In the supervised training phase, a residual convolutional autoencoder network structure is first constructed to extract the semantic information from a training set consisting solely of known jamming patterns. By incorporating center loss and reconstruction loss into the softmax loss function, a joint loss function is constructed to minimize the intra-class distance and maximize the inter-class distance in the jamming features. Moving to the unsupervised classification phase, a test set containing both known and unknown patterns is fed into the trained encoder, and a distance-based recognition method is utilized to classify the jamming signals. The results demonstrate that the proposed model not only achieves sufficient learning and representation of known jamming patterns but also effectively identifies and classifies unknown jamming signals. When the jamming-to-noise ratio (JNR) exceeds 10 dB, the recognition rate for seven known jamming patterns and two unknown jamming patterns is more than 92%.
This content is subject to copyright.
Citation: Zhang, Y.; Zhao, Z.; Bu, Y.
Radar Active Jamming Recognition
under Open World Setting. Remote
Sens. 2023,15, 4107. https://doi.org/
10.3390/rs15164107
Academic Editors: Qian Du, Yanni
Dong and Xiaochen Yang
Received: 30 June 2023
Revised: 17 August 2023
Accepted: 18 August 2023
Published: 21 August 2023
Copyright: © 2023 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
remote sensing
Article
Radar Active Jamming Recognition under Open World Setting
Yupei Zhang 1, Zhijin Zhao 2,* and Yi Bu 3
1School of Electronic and Information, Hangzhou Dianzi University, Hangzhou 310018, China;
2143040036@hdu.edu.cn
2School of Communication Engineering, Hangzhou Dianzi University, Hangzhou 310018, China
3
School of Electrical and Computer Engineering, Royal Melbourne Institute of Technology (RMIT University),
Melbourne 3000, Australia; prototyperimac@gmail.com
*Correspondence: zhaozj03@hdu.edu.cn
Abstract:
To address the issue that conventional methods cannot recognize unknown patterns
of radar jamming, this study adopts the idea of zero-shot learning (ZSL) and proposes an open
world recognition method, RCAE-OWR, based on residual convolutional autoencoders, which can
implement the classification of known and unknown patterns. In the supervised training phase, a
residual convolutional autoencoder network structure is first constructed to extract the semantic
information from a training set consisting solely of known jamming patterns. By incorporating center
loss and reconstruction loss into the softmax loss function, a joint loss function is constructed to
minimize the intra-class distance and maximize the inter-class distance in the jamming features.
Moving to the unsupervised classification phase, a test set containing both known and unknown
patterns is fed into the trained encoder, and a distance-based recognition method is utilized to classify
the jamming signals. The results demonstrate that the proposed model not only achieves sufficient
learning and representation of known jamming patterns but also effectively identifies and classifies
unknown jamming signals. When the jamming-to-noise ratio (JNR) exceeds 10 dB, the recognition
rate for seven known jamming patterns and two unknown jamming patterns is more than 92%.
Keywords:
radar signals; jamming recognition; deep learning; zero-shot learning; convolutional
neural network
1. Introduction
Radar faces increasingly complex electronic countermeasures, with various new types
of radar jamming patterns continuously emerging as challenges [
1
,
2
]. The accurate recogni-
tion of radar jamming is a precondition and key to implementing anti-jamming measures,
and automatic recognition of radar jamming patterns can effectively improve the target
detection and tracking performance of radar. Therefore, jamming pattern recognition has
always been a research hotspot in anti-jamming technology [
3
,
4
]. Suppression jamming
by emitting high-power noise signals is not only effective for linear frequency modulated
(LFM) radar systems but also for other modulated radar systems, making it the most widely
used [
5
]. Therefore, this paper focuses on the recognition of suppression jamming. Its main
patterns include amplitude modulation jamming (AMJ), frequency modulation jamming
(FMJ), comb spectrum jamming (CJ), phase modulation jamming (PMJ), swept jamming
(SJ), etc. Conventional jamming pattern detection and classification methods are generally
based on feature engineering. Firstly, multi-dimensional features of signal in time domain,
frequency domain, and transform domain are extracted, including features such as moment
kurtosis, moment skewness, envelope fluctuation, noise factor [
6
9
], singular spectrum
features [
10
], bispectrum features [
11
] and other signal features. Then, with the help of
machine learning-based classifiers, such as support vector machine (SVM ) [
12
], decision
trees [
13
], and back propagation (BP) neural networks [
12
], classifications are accomplished.
However, feature engineering-based classification methods are time-consuming and require
Remote Sens. 2023,15, 4107. https://doi.org/10.3390/rs15164107 https://www.mdpi.com/journal/remotesensing
Remote Sens. 2023,15, 4107 2 of 24
expert experience, especially when the transform domain features are large, the classifica-
tion performance has limited room for improvement, and the recognition rate is low in
strong noise environment.
With the development of deep learning (DL) technology, the feature extraction capa-
bility of neural networks has been improving, and classification methods based on DL are
emerging. In [
14
], A novel hybrid framework of optimized deep learning models com-
bined with multi-sensor fusion is developed for condition diagnosis of concrete arch beam,
and the results demonstrate that the method can achieve the classification of structural
damage with limited sensors and high levels of uncertainties.
Convolutional neural networks (CNNs), due to their network architecture which
incorporates weight sharing and small local receptive fields, have significantly reduced the
number of node connections compared to conventional neural networks. This simplification
of network connections has made CNNs widely applied in deep learning models [
15
17
].
CNNs can train their parameters using jamming signals, eliminating the need for manual
feature extraction and the design of decision trees for classification criteria. As a result,
CNNs have been extensively used in the research of classifying and recognizing radar
jamming signals [18].
In [
19
], a jamming recognition algorithm based on improved LeNet
·
CNN network
was designed which extracted one-dimensional radar received signals and adjusted the
network structure parameters to achieve optimal performance for the recognition of jam-
ming signals. Ref. [
20
] obtained the time-frequency spectrogram of jamming signals by
short-time Fourier transform, combined with the improved VGGNet
·
16 network model for
feature learning and training, and the simulation verified that the algorithm is still effective
for the identification of six kinds of mixed jamming. Ref. [
21
] adopted an adaptive cropping
algorithm to crop most of the redundant information of the time-frequency image and kept
the complete information of the jamming in the CNN for training, and finally achieved the
recognition of nine kinds of jamming signals with high accuracy and fast iteration. In [
22
],
a 1D CNN-based radar jamming signal classification model was proposed to achieve the
classification of 12 typical jamming signals by putting the real and imaginary parts of
jamming signals into the parallel network for training. In [
23
], a CNN was constructed
using the real and imaginary parts of the signal as inputs. With sufficient training samples,
this method demonstrated excellent recognition capabilities. The mentioned papers pri-
marily address the issue of recognizing jamming when there are sufficient labeled samples.
However, [
6
,
24
,
25
] considered the case of insufficient labeled samples. Ref. [
6
] proposed
a method based on a time-frequency self-attentive mechanism. The recognition rate for
most of the patterns of jamming reaches 90% when the samples with labels account for 3%.
In [
24
], Tian et al. inputed features obtained through empirical mode decomposition and
wavelet decomposition into the network. Simulations conducted on a dataset consisting
of only 8400 samples showed that the recognition rates for four types of jamming were
all above 90% when the JNR exceeded 6 dB. In [
25
], a large number of unlabeled samples
were first used to train an jamming recognition network to extract valuable features. Then,
a small number of labeled samples were used to improve the classification accuracy.
Although the application of deep learning technology in radar jamming recognition
is rapidly developing, the current methods still suffer from the closed-set assumption,
i.e., the existing methods assume that the jamming patterns are included in the training
set. However, in the actual battlefield environment, the enemy may invent new jamming
patterns, making it challenging to collect data for all patterns in the training set, so the
actual radar jamming environment is an open set scenario, i.e., the test environment is likely
to have jamming patterns that do not exist in the training jamming library. In the actual
open set jamming scenario, when a jamming pattern that does not exist in the training
jamming library appears in the test environment, the existing radar jamming identification
methods will incorrectly identify this unknown jamming as one of the known jamming
patterns. In [
26
], Zhou et al. first investigated the open set recognition problem for radar
jamming; however, the method can only detect or reject the unknown patterns, but cannot
Remote Sens. 2023,15, 4107 3 of 24
effectively identify the unknown patterns. How to further classify these unknown pattern
signals remains a challenging task, and it falls under the research domain of open world
recognition (OWR).
Currently, OWR techniques have been applied in target recognition of synthetic aper-
ture radar (SAR) images. In [
27
], a hierarchical embedding and incremental evolutionary
network (HEIEN) was designed for when there are fewer unknown target training sets
in open scenarios, which requires only a small number of unknown target samples for
effective model training. A more stable feature space was built in [
28
], which has better
interpretability. In testing, experiments on a dataset containing seven known targets and
one unknown target show that the method improves the reliability of recognizing unknown
targets. In [
29
], Song et al. used physical EM simulation images of targets at different
azimuths as training data in order to learn the features of unknown targets. An accuracy of
91.93% can be achieved in a recognition task with a dataset containing nine known targets
and one unknown target.
Nevertheless, OWR is just starting in radar jamming pattern recognition. Zero-shot
learning (ZSL) [
30
,
31
] is an effective approach to address the challenge of open world
recognition. The most typical implementation of ZSL is based on feature mapping. The goal
of this approach is to learn the mapping relationship between the original signal space and
the semantic feature space using the known jamming patterns. This learned relationship is
then generalized to the unknown pattern dataset, enabling recognition and classification
of unknown patterns using the semantic features. ZSL can be classified into two types:
traditional zero-shot learning (TZSL) and generalized zero-shot learning (GZSL) [
32
]. TZSL
assumes that the training patterns are known, while the testing patterns are unknown,
and there is mutual exclusion between the training and testing patterns. GZSL assumes
that the training patterns include both known and unknown patterns in the testing phase.
GZSL has a more relaxed experimental setting, which better reflects real-world scenarios.
Therefore, in this paper, we consider the GZSL scenario.
To address the challenge of existing methods being unable to classify unknown jam-
ming patterns specifically, this paper adopts the idea of ZSL and conducts research on
jamming patterns recognition in an open world scenario. We propose a residual convolu-
tional autoencoder-based radar jamming open world recognition algorithm, abbreviated
as RCAE-OWR.
In summary, the main contributions of this paper are as follows:
In order to address the limitations of existing radar jamming pattern recognition meth-
ods, which are mostly closed-set recognition or simply rejecting unknown patterns, we
propose a zero-shot learning approach based on residual conventional autoencoders.
This method does not require prior information about the patterns of jamming and
can classify both known and unknown patterns using distance-based recognition
methods.
A hybrid loss function consisting of cross-entropy loss, center loss and reconstruction
loss is introduced to recognize different patterns of jamming signals. Where the cross-
entropy loss makes the features obtained from the mapping network divisible, and to
some extent widens the distance among different patterns, the center loss makes it
easier to delineate the boundaries of the various patterns, and the reconstruction loss
ensures that the most essential characterization of the pattern features is learned from
the known patterns dataset.
Through extensive experimental simulations, we evaluate the open world recognition
performance of the proposed algorithm and investigate the influence of JNR and the
number of unknown patterns on the algorithm’s performance. The simulation results
demonstrate that the proposed algorithm achieves effective recognition of both known
and unknown patterns, especially in high-JNR environments.
Remote Sens. 2023,15, 4107 4 of 24
2. Jamming Signal Modeling
In radar jamming and anti-jamming training, the signals received by the radar receiver
include radar echo signals, jamming signals and noise signals, which can be expressed as
follows:
S(t) = so(t) + J(t) + v(t), (1)
where
S(t)
denotes the total signal received by the radar,
so(t)
represents the echo signal,
J(t)means the jamming signal, and v(t)is the noise signal.
2.1. Echo Signal
Typical modulation types of radar echo signals include continuous wave (CW), linear
frequency modulation (LFM), phase shift keying (PSK), frequency shift keying (FSK), etc.
LFM, with its frequency linearly changing over time, is the most widely used. Therefore,
in this paper, we focus on studying the typical echo signal of LFM, which can be expressed
as follows [33]:
so(t) =
N
i=1
recttiTr
Tpej2π(foB
2)(tiTr)+KLF
2(tiTr)2 (2)
where
i
denotes the pulse sequence; rect
tiTr
Tp
represents the
i
-th rectangular pulse with
a width of Tp;Tris the pulse repetition period; Tpis the pulse width; fois the radar center
frequency; Bmeans the bandwidth; and KLF =B
Tpis the LFM coefficient.
2.2. Jamming Signal
In this paper, we select nine typical radar active jamming patterns as the research ob-
jects, including: radio noise jamming (RNJ) [
26
], amplitude modulation jamming (AMJ) [
34
],
frequency modulation jamming (FMJ) [
34
], comb spectrum jamming (CJ) [
13
], phase modu-
lation jamming (PMJ) [
34
], linear sweep frequency jamming (LSFJ) [
34
], non-linear sweep
frequency jamming (NLSFJ) [
35
], hopping frequency jamming (HFJ) [
35
] and periodic
Gaussian pulse jamming (PGJ) [36].
2.2.1. Radio Noise Jamming
RNJ is a narrowband Gaussian process generated by filtering and then amplifying the
noise signal. Its mathematical model is expressed as:
J(t) = Ur(t)ej(ωt+ϕ)(3)
where
Ur(t)
obeys a Gaussian distribution,
ω
is the carrier frequency, and
ϕ
is the initial
phase, following a uniform distribution on [0, 2π].
2.2.2. Amplitude Modulation Jamming
The model of AMJ is represented as:
J(t) = [U0+KAM Un(t)]ej(ωt+ϕ)(4)
where
Un(t)
is zero-mean Gaussian white noise,
U0
is the carrier voltage, and
KAM
is the
amplitude modulation index.
2.2.3. Frequency Modulation Jamming
FMJ is a type of barrage jamming, and its model is expressed as:
J(t) = U0·ejωt+2πKFM Rt
0u(t0)dt0+ϕ. (5)
Remote Sens. 2023,15, 4107 5 of 24
where
KFM
is the frequency modulation index and
u(t)
is a zero-mean stationary ran-
dom process.
2.2.4. Comb Spectrum Jamming
CJ
consists of multiple narrowband noise frequency modulation signals, and ex-
presses as:
J(t) =
m
i=1
Uiejh2πfit+2πKFM Rt
0u(t)dt+ϕi. (6)
where
Ui
is the amplitude,
fi
represents the frequency where the comb teeth appear, and
m
is the number of frequencies.
2.2.5. Phase Modulation Jamming
The model of PMJ is:
J(t) = U0·ej(ωt+KPMU u(t)+ ϕ)(7)
where KPM is the phase modulation index.
2.2.6. Linear Sweep Frequency Jamming
LSFJ varies linearly with time in a frequency band, and expresses as:
J(t) = U0ej(2πf0t+2πKFM t2+ϕ)(8)
where f0is the initial frequency.
2.2.7. Non-Linear Sweep Frequency Jamming
NLSFJ is similar to LSFJ, except that the instantaneous frequency magnitude of the
jamming signal varies continuously with the square of time, and expresses as:
J(t) = U0ej(2πf0t+2πKFM t3+ϕ)(9)
2.2.8. Hopping Frequency Jamming
HFJ is a wideband non-stationary signal in which the frequency changes over time.
The model of HFJ can be written as:
J(t) =
N
n=1
Anej(2πfnt+ϕn)·p(tnTH). (10)
where
{An}
represents the amplitude sequence,
{fn}
is the pseudo-random frequency
sequence,
ϕn
is the random phase sequence,
TH
is the hop duration, and
p(t)
is the base
pulse signal with a pulse width of TH.
2.2.9. Periodic Gaussian Pulse Jamming
PGP is a widely used active suppression jamming. The PGP is expressed as:
J(t) = Φ(t), 0 <t<τ
0, τ<t<T(11)
where
Φ(t)
represents a Gaussian function with a mean of 0 and a variance of 1,
τ
denotes
the pulse duration, and Tis the pulse period of the jamming.
The time-domain waveforms of the nine patterns are shown in Figure 1. As can be seen
from the figure, there is little difference in the time domain between the various patterns
of suppressed jamming signals, which are more difficult to distinguish manually and are
easily disturbed by noise.
Remote Sens. 2023,15, 4107 6 of 24
012
t/s
(a)
10-6
-1
0
1
Magnitude real part
012
t/s
(b)
10-6
-1
0
1
Magnitude real part
012
t/s
(c)
10-6
-1
0
1
Magnitude real part
012
t/s
(d)
10-6
-1
0
1
Magnitude real part
012
t/s
(e)
10-6
-1
0
1
Magnitude real part
012
t/s
(f)
10-6
-1
0
1
Magnitude real part
012
t/s
(g)
10-6
-1
0
1
Magnitude real part
012
t/s
(h)
10-6
-1
0
1
Magnitude real part
012
t/s
(i)
10-6
-0.5
0
0.5
Magnitude real part
Figure 1.
Time-domain waveforms of nine patterns of radar jamming signals. From (
a
) to (
i
): (
a
) RNJ,
(b) AMJ, (c) FMJ, (d) CJ, (e) PMJ, (f) LSFJ, (g) NLSFJ, (h) HFJ, (i) PGJ.
3. Proposed Method
Assuming that the dataset of jamming signals received by the radar is
T
, it consists of
Nc+Nu
kinds of patterns of jamming. The subset
Tc
is composed of samples of
Nc
kinds of
known patterns, while the subset
Tu
is composed of samples of
Nu
kinds of unknown patterns.
The two subsets are complementary, meaning
TcTu=T
and
TcTu=
. For the known
patterns set
Tc=nxc
i,yc
i,zyc
i|xc
i Xns×Nd
c,yc
i Yc={1, 2, ·· · ,Nc},zyc
iZNc×Nk
co
con-
taining
nc
samples of
Nc
kinds of patterns, where
xc
i
represents a
Nd
-dimensional vector of
the
i
-th sample,
yc
i
is the label of the sample, and
zyi
denotes the
Nk
-dimensional semantic
vector describing the features of the sample’s corresponding pattern to be obtained by
supervised learning. Similarly, for the unknown patterns dataset consisting of
nu
samples
of
Nc
kinds of patterns,
Tu=nxu
i,yu
i,zu
yu
io
, where
xu
i Xnu×Nd
u
is the
Nd
-dimensional
feature vector of the
i
-th sample,
yu
i Yu={Nc+1, Nc+2, . . . , Nc+Nu}
represents the
label of the sample, and
zyu
iZNu×Nk
u
denotes the
Nk
-dimensional semantic information of
the sample’s corresponding pattern.
For the generalized zero-shot learning classification task considered in this paper,
the supervised training phase only allows the utilization of dataset of known patterns
Tc
.
However, the objective is to ensure that the model trained in the supervised training phase
can accurately classify
Xc Xu
into the
Nc+Nu
-dimensional space
Yc Yu
during the
unsupervised classification phase.
Figure 2illustrates the network framework of the RCAE-OWR algorithm, which
consists of a supervised training phase and an unsupervised classification phase. In the
supervised training phase, the residual convolutional autoencoder (RCAE) network is
used to extract semantic features of known jamming patterns. Meanwhile, the network
is trained using center loss, cross-entropy loss, and reconstruction loss. In [
22
,
37
], time-
domain signals are directly used as inputs to the network to extract the deep features
of different signals, and the simulation results prove that the direct time-domain signal-
based recognition methods obtain good recognition performance in terms of accuracy and
speed, demonstrating a huge potential for radar signal processing. Motivated by [
22
,
37
],
in this paper, we also directly feed the IQ data of the jamming signals into the model.
Then, after the supervised training, the unsupervised classification phase was entered.
At this phase, the parameters of encoder of the RCAE are kept fixed, and both known and
unknown jamming samples are input to the encoder to obtain their semantic features. Then,
Remote Sens. 2023,15, 4107 7 of 24
a distance-based discriminative method is employed to achieve open world recognition of
radar jamming signals.
Figure 2. Radar jamming recognition framework of RCAE-OWR.
3.1. Supervised Training
The supervised training phase mainly consists of an autoencoder network (AE) and a
supervised classification network classifier. The autoencoder is divided into two parts: the
encoder and the decoder. In the supervised training phase, the focus is on constructing the
mapping relationship between the time-domain signal and the semantic features.
3.1.1. Residual Convolutional Autoencoder
Due to the simplicity of the traditional AE structure, this study primarily considers
the Residual Convolutional Autoencoder (RCAE). It replaces the fully connected layers
in AE with convolutional layers and pooling layers, inheriting the characteristics of an
autoencoder. This enables better feature learning and improves the efficiency of feature
learning in AE. To prevent degradation in recognition performance, a residual network
structure is employed. The input signals are IQ dual-channel data with a length of 512,
resulting in a dimension of 2 ×512. Additionally, to maintain the vector dimensions after
convolution, we have set the convolution kernel size to 3
×
3, padding = 1 and stride = 1.
RCAE is based on the semantic autoencoder (SAE) architecture [
31
], and SAE enables
mapping functions learned from known patterns to be better generalized to unknown
patterns, which can effectively resist the domain shift problem [
38
]. In the encoding
process, convolutional operations are used to extract features from input samples and
obtain semantic vectors. Then, the decoder utilizes transpose convolution to reconstruct
the semantic vectors and restore them to the original inputs.
The basic components of RCAE include the input layer, convolutional layers, semantic
layer, deconvolutional layers, and output layer.
The RCAE designed in this paper is illustrated in Figure 3. It replaces the fully
connected layers (FC) in the AE with convolutional layers and pooling layers, inheriting the
characteristics of the AE for better feature learning. In addition, to prevent degradation of
recognition performance, a residual structure is employed. The encoder uses convolutional
operations to extract features from the input samples to obtain the semantic vector; the
decoder utilizes transposed convolution to reconstruct the semantic vector and reduce it to
the original signal.
The basic components of RCAE include: input layer, convolutional layer, semantic
layer, deconvolutional layer and output layer.
Remote Sens. 2023,15, 4107 8 of 24
3×3
Conv
64
padding=1
3×3
Conv
128
padding=1
3×3
Conv
256
padding=1
3×3
Conv
512
padding=1
Flatten
FC
1024
FC
512
FC
256
FC
C
3×3
Conv
64
padding=1
3×3
Conv
128
padding=1
3×3
Conv
256
padding=1
3×3
Conv
512
padding=1
Flatten
FC
1024
FC
512
FC
256
Encoder, E(·)
Decoder, D(·)
Classifier
Input
2×512
Output
2×512
z
Figure 3. Residual Convolutional Autoencoder Network Structure.
In the encoding part, the input layer receives the input data
xc
i
and passes it to the
encoder. The encoder gradually extracts the semantic features of the input data through
multiple convolutional layers and their residual structures, denoted as
z=Exc
i
, where
E(·)denotes the mapping function of the encoder.
The convolutional layers extract features from the input data
xc
i
using convolutional
operations. The data processing can be described as follows:
gi=f[conv(x,kconv
i)+bconv
i]. (12)
where
kconv
i
is the
i
-th convolutional kernel matrix,
conv(·)
denotes the convolution opera-
tion,
bconv
i
represents the
i
-th bias term,
f(·)
represents the activation function. The feature
layer integrates the diverse features extracted by the convolutional layers and outputs the
semantic vector z.
In the decoding part, the semantic feature
z
is up-sampled through transpose con-
volutional operations, aiming to reconstruct the original signal
˜
xc
i
based on the semantic
features. This process is denoted as
˜
xc
i=D(z)
, where
D(·)
denotes the mapping function
of the decoder. Finally, the reconstructed results are outputted through the output layer.
The data processing can be described as follows:
qi=fhdeconvz,kdec
i+bdec
ii(13)
where
kdec
i
represents the i-th convolutional kernel matrix, deconv
(·)
denotes the transpose
convolution operation, bdec
irepresents the i-th bias term.
The training process of
RCAE
aims to minimize the reconstruction error, ensuring the
effectiveness of feature extraction. It can be expressed as:
Lre =1
2M
M
i=1k˜
xc
ixc
ik2
2(14)
where
M
represents the number of samples in a batch,
˜
xc
i
is the signal reconstructed by
xc
i
through the RCAE network.
To encourage the feature vectors of the same patterns to be close to their corresponding
pattern centers and far from centers of different patterns, a center loss is introduced.
During model training, the center loss assigns a center for each jamming pattern. Assuming
the input sample is
xc
i
with label
yc
i
, and the center for pattern
yi
is denoted as
zc
yc
i
. The center
loss can be defined as:
Lcl =1
M
M
i=1
E(xc
i)zc
yc
i
2
2(15)
Remote Sens. 2023,15, 4107 9 of 24
During the model iteration process, the selection of pattern center
zc
yc
i
is an important
issue. Theoretically, the optimal center for pattern
yc
i
would be the mean of all the feature
vectors in that pattern. However, calculating the mean for all samples in each iteration
would impose additional computational cost and reduce the efficiency of the model. To ad-
dress this, the pattern center are initialized randomly, and then updated separately for each
batch. The update process is as follows:
zc
yc
i=
M
m=1δyc
i=yc
m·zc
yc
iE(xc
m)
0.1 +M
m=1δyc
i=yc
m(16)
Zt+1
yc
i=Zt
yc
iα·Zc
yc
i(17)
where
δ(yi=k)=
1 when
yi=k
; otherwise, it is 0.
zt
yc
i
is the semantic center of
yc
i
at the
t-th epoch, αis the learning rate.
3.1.2. Supervised Classifier
The encoder, acting as a feature extractor, takes the encoded features and feeds them
into a fully connected layer followed by a softmax classifier, which outputs the predicted
label. The loss function for this process utilizes cross-entropy between the predicted label
and the true label.
Lce =1
M
M
i=1
yilog(ˆyi)(18)
where
yi
represents the true label
xc
i
in one-hot format, and
ˆyi
represents the predicted
probability vector.
The reconstruction loss
Lre
ensures that the reduced-dimensional semantic features
are representative, the central loss
Lcl
promotes cohesion among feature vectors of the
same jamming pattern, and the cross-entropy loss
Lce
enhances the discriminative ability of
feature vectors among different jamming patterns. To achieve both increased inter-class
distance and reduced intra-class distance, a joint loss function is proposed:
L=Lce +λcl Lcl +λre Lre (19)
where
λcl
and
λre
are constants used to balance the weights of the three loss functions.
The detailed gradient θLof Lis shown in Appendix A.
The network parameters during the supervised training phase are updated as shown
in Algorithm 1.
Algorithm 1 Pseudocode for supervised training of RCAE-OWR
Input: known jamming patterns dataset xc
1,yc
1,··· ,(xc
n,yc
n)and hyperparameter.
Output: Optimal parameters θ.
1: while epoch =1, . . . , Ndo
2: for each batch with size Mdo
3: Feed in a batch xc
1,yc
1,··· ,xc
M,yc
M.
4: Calculate Zc
yc
ivia Equation (16).
5: Calculate Zt
yc
ivia Equation (17).
6: Calculate Lce via Equation (18).
7: Calculate Lcl via Equation (15).
8: Calculate Lre via Equation (14).
9: Calculate Lvia Equation (19).
10: Update θ:θθηθL.
11: end for
12: end while
Remote Sens. 2023,15, 4107 10 of 24
3.2. Unsupervised Classification
After learning the mapping relationship between the original time-domain signals and
the semantic feature space in the supervised training phase, this learned relationship can be
generalized to the unknown jamming patterns dataset. Finally, the unknown patterns can
be recognized and classified using the semantic features. Inspired by [
39
], a distance-based
classification rule is proposed.
3.2.1. Known Jamming Patterns Classification
Sequentially, the
nc
known jamming samples from the training set are inputted into
the trained RCAE-OWR model to extract semantic features. The semantic center vector
zc
k
corresponding to the k-th known jamming pattern is defined as:
zc
k=nc
i=1δ(yi=k)zyc
i
nc
i=1δ(yi=k),k=1, 2, . . . , Nc(20)
where
zyc
i
denotes the semantic feature extracted from the
i
-th input sample
xc
i
. When
a test sample
t
is input, its semantic features
E(t)
are extracted by the encoder, and its
Mahalanobis distance to the center vectors of each known jamming pattern is calculated as:
d(E(t),zc
k)=qE(t)zc
kT1
kE(t)zc
k(21)
where
k
is the diagonal covariance matrix corresponding to
zk
, and
1
k
is its inverse matrix.
Let
d1=minkdE(t),zc
k
. If
d1Θc
, then
t
belongs to the known jamming patterns.
Here,
Θc=λc×
3
N
is a given threshold determined by the 3
σ
criterion [
34
], where
λc
is
a coefficient and
N
is the dimension of
k
. In this case, the label
yt
of
t
can be determined
as follows:
yt=arg min
k
(d(E(t),zc
k)) (22)
when
d1>Θc
,
t
belongs to the unknown jamming patterns, and its label
yt
in this case is
described in Section 3.2.2.
3.2.2. Unknown Jamming Patterns Classification
Let
TU
represent the dataset of unknown jamming patterns. If
TU=
, then
t
belongs
to the first sample of a new unknown jamming pattern, and its label is
yt=U1
. The sample
t
is then recorded in the
Tu
. If
TU6=
, the following rules are applied to determine whether
t
belongs to a previously recorded unknown jamming pattern or a new unknown jamming
pattern. First, let zu
irepresent the semantic center vector of pattern Uiin the TU:
zu
i=k∈Tuδ(yk=Ui)zyu
k
k∈Tuδ(yk=Ui),i=1, 2, · · · ,N(23)
where
N
is the number of patterns that already exist in
TU
.
zyu
j
denotes the semantics
extracted from the
j
-th unknown pattern sample
xu
j
. Let
d2=minzidE(t),zu
i
. If
d2Θu
,
then
t
belongs to an existing unknown jamming pattern. Here,
Θu=λumax(dc,du)
is a
given threshold, where λuis a coefficient, dcand duare defined as follows:
dc=max
¯zc
k
d(E(t),¯zc
k)(24)
du=max
¯zu
i
d(E(t),¯zu
i)(25)
In this case, the label ytof tcan be determined as follows:
yt=arg min
ui
(d(E(t),zu
i)) (26)
Remote Sens. 2023,15, 4107 11 of 24
Otherwise,
t
belongs to a new unknown jamming pattern, and its label is
yt=UN+1
.
The test sample tis then recorded in the Tu.
In summary, the recognition rules for the unsupervised classification phase are pre-
sented in Algorithm 2.
Algorithm 2 Pseudocode for unsupervised classification of RCAE-OWR
Input: test sample tiand well-trained weight of Encoder, θc.
Output: Jamming pattern yi.
1: for i=1, . . . , Ntest do
2: Calculate semantic vector ziof ti
3: d1=minkdE(t),zc
k
4: d2=minidE(t),zu
i
5: dc=max
¯zc
k
dE(t),¯zc
k
6: du=max
¯zu
i
dE(t),¯zu
i
7: Θu=λumax(dc,du)
8: if d1Θcthen
9: yi=arg min
kdE(t),¯zc
k
10: else if d1>Θcand TU=then
11: add tto Tu
12: yi=U1
13: else if d1>Θc,TU6=and d2Θuthen
14: yi=arg min
UidE(t),¯zu
i
15: else
16: yi=UN+1
17: add t to Tu.
18: end if
19: end for
4. Performance Evaluation
4.1. Simulation Parameter Settings
The simulations are performed on a PC with an Intel(R) Core(TM) i7-9700 CPU and a
GeForce RTX2070s GPU. The deep learning framework PyTorch is used for training and
testing the neural network. The RCAE-OWR network is initialized with random weights,
and the learning rate is set to 0.001 . The batch size is set to 256, and the number of epochs
is set to 250. The values of
λcl
,
λre
,
λc
, and
λu
are set to 0.02, 10, 0.5, and 1.5, respectively.
Furthermore, grid search is applied to ascertain the optimal hyperparameters. Specifically,
we train the model for each possible combination among all the candidate parameters by
loop traversal and pick the parameter combination that minimizes the validation set error
as the optimal hyperparameters.
4.2. Original Dataset
The jamming signals are generated using MATLAB. The dataset includes the nine
kinds of radar active jamming signals mentioned in Section 2.2, with detailed parameters
shown in Table 1. The data description is shown in Table 2. Gaussian white noise is added
during the simulation, and the JNR ranges from
4
dB
to 18
dB
with a step size of 2
dB
.
For each JNR, 1000 samples are generated for each type of radar active jamming signal.
In the following experiments,
Nc
kinds of the jamming patterns will be treated as known
jamming, while the remaining 9
Nc
kinds of patterns will be treated as unknown jamming.
Remote Sens. 2023,15, 4107 12 of 24
Table 1. Nine kinds of Radar Active Jamming Parameter Settings.
Jamming Pattern Parameter Description
RNJ Carrier frequency: 60110 MHz.
AMJ Carrier frequency: 60110 MHz, bandwidth: 510 MHz, KAM: 0.10.9
FMJ Carrier frequency: 60110 MHz, bandwidth: 510 MHz, KFM: 0.10.9
CJ Carrier frequency: 60110 MHz, bandwidth: 510 MHz,
number of bands: 24
PMJ Carrier frequency: 60110 MHz, bandwidth: 510 MHz, KPM : 0.10.9
LSFJ Starting frequency: 110 MHz, ending frequency: 50100 MHz.
NLSFJ Starting frequency: 110 MHz, ending frequency: 50100 MHz.
HFJ N= 20, {fc}: [10, 100] MHz, TH: 3264 µs
PGJ Pulse period T: [2.5, 10] µs . Duty cycle: [1/8, 1/2]
Table 2. The original dataset.
Total Samples Samples of Each Pattern Samples Each JNR Feature Dimension Patterns
108,000 12,000 1000 2 ×512 9
patterns
RNJ, AMJ, FMJ, CJ, PMJ, LSFJ, NLSFJ, HFJ, PGJ
number of JNR values JNR values
12 4, 2, 0, 2, 4, 6, 8, 10, 12, 14, 16, 18
4.3. Training Process
Firstly,
Nc
patterns are selected from the original dataset as known patterns, and the
remaining 9
Nc
patterns are considered unknown patterns. The samples from known pat-
terns are used not only for training the network but also for testing. However, the samples
from unknown patterns are exclusively used for testing purposes. To reduce the impact of
noise on the training results, during the training phase, only the known pattern samples
with a JNR higher than 12dB are used to train the RCAE-OWR network. For each JNR,
700 samples are randomly selected from the known pattern samples to form the training
set, and another 100 samples are randomly chosen to construct the validation set. In the
testing phase, a wider range of JNR values is used, namely, JNR =
4–18 dB. For each of
the nine jamming patterns, 200 samples are randomly selected at each JNR value to form
the test set. It is important to note that the training set, validation set and test set for the
known patterns samples are mutually exclusive and do not overlap with each other.
Figures 46show the curves of cross-entropy loss
Lce
, central loss
Lcl
and reconstruc-
tion loss
Lre
of the proposed RACE-OWR algorithm with the number of iterations. It can
be seen from the figures that the three losses are gradually decreasing and converging
after the epoch exceeds 50, which illustrates the effectiveness and the fast convergence
and low complexity of the algorithm. In the convergence stage, the reconstruction loss
approaches approximately 0.274, which indicates that the features extracted by the network
designed in this paper can effectively represent the original signal and further provide a
guarantee for unsupervised classification. The cross-entropy loss converges to 0.82, which
contributes to the separation of different patterns, while the central loss converges to 0.12,
which contributes to the aggregation of samples from the same pattern.
Remote Sens. 2023,15, 4107 13 of 24
0 50 100 150 200 250
epochs
0
5
10
15
20
25
Lce
Figure 4. Cross-entropy loss.
0 50 100 150 200 250
epochs
0
50
100
150
200
250
300
350
400
450
500
Lcl
Figure 5. Center loss.
Remote Sens. 2023,15, 4107 14 of 24
0 50 100 150 200 250
epochs
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
Lre
Figure 6. Reconstruction loss.
4.4. Impact of Different Combinations of Unknown Patterns on Classification Performance
In order to investigate the influence of different combinations of unknown patterns on
the algorithm’s classification performance, we conducted experiments by randomly select-
ing two jamming patterns from the dataset as unknown patterns. We define two metrics
for evaluation: true known rate (TKR) and true unknown rate (TUR). TKR is calculated
as
TK/K
, and
TUR
is calculated as
TU/U
, where
TK
and
TU
represent the number of
correctly identified samples from known and unknown patterns, respectively, while
K
and
U
represent the total number of samples from known and unknown patterns. For ease of
comparison, we ensure that the test set’s JNR is consistent with the training set, ranging
from 12 to 18
dB
. The experimental results are summarized in Table 3. From the table, it
can be observed that at higher JNR values, the algorithm effectively distinguishes between
known and unknown patterns, validating the efficacy of our proposed approach. However,
under the three combinations tested, neither TKR nor TUR could reach 1. This indicates
that some samples from the known patterns are misclassified as unknown, and vice versa.
Table 3. Performance with different unknown pattern combinations.
Indicator Unknown Patterns
RNJ, HFJ AMJ, PGJ RNJ, NLSFJ
Average accuracy 0.957 0.965 0.978
TKR 0.966 0.982 0.994
TUR 0.999 0.999 0.995
4.5. Impact of λcon Algorithm Performance
The value of
Θc
, which determines the proportion of jamming signal samples classified
as known or unknown patterns, is influenced by the value of
λc
. In Section 3.2.1,
Θc
is
defined as
Θc=λc×
3
N
, where
N
is a constant representing the dimension of the
matrix. Therefore, this section investigates the impact of
λc
on the algorithm’s performance.
Since
TKR
and
TUR
cannot both increase with the increase in
λc
, a weighted true rate
(WTR) [
39
] is defined to balance these two metrics, i.e.,
WTR =ξ×TKR + (
1
ξ)×TUR
,
where
ξ
is a balancing factor set to 0.5 , indicating equal importance given to
TKR
and
Remote Sens. 2023,15, 4107 15 of 24
TUR
. Figure 7shows the curves of
TKR
,
TUR
, and
WTR
as
λc
varies. It can be observed
that as
λc
increases,
TKR
increases while
TUR
decreases. This is because as
λc
increases,
the value of
Θc
increases, so more samples are discriminated as known jamming patterns
and fewer samples are discriminated as unknown jamming patterns. The value of
WTR
initially increases and then decreases with the increase in
λc
, reaching its maximum at a
value of approximately 0.50.7.
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
c
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
true rate/%
TKR
TUR
WTR
Figure 7. The impact of λcon recognition rate.
4.6. Ablation Study
In this research, we conduct an ablation study to comprehend the individual contri-
butions of component loss functions in the context of the RCAE-OWR learning model.
By systematically eliminating specific component loss functions, we investigate their im-
pact on the overall performance. The experimental findings, as summarized in Table 4,
reveal crucial insights into the significance of each loss function. Notably, the removal of
Lce
(Cross-Entropy Loss) leads to the most substantial performance degradation, resulting
in an accuracy of 95.1%. These results underscore the paramount role played by the cross-
entropy loss in the learning process, surpassing the influence of the other two component
loss functions, namely, Lre and Lct. Furthermore, we observe that the incorporation of Lre
enhances the accuracy by 3.3%. This improvement can be attributed to the ability of
Lre
to
preserve essential semantic features, thereby facilitating the differentiation between known
and unknown patterns. Additionally, the utilization of the center loss function,
Lct
yields a
minor enhancement in our learning model’s performance.
Table 4.
Ablation study on RCAE-OWR.
Lce
: cross entropy loss.
Lcl
: center loss.
Lre
: reconstruc-
tion loss.
Variant True Rate
RCAE-OWR(Lce ,Lcl and Lre ) 0.995
without Lce 0.951
without Lcl 0.982
without Lre 0.962
4.7. Open set Recognition Performance
When
Nc
is set to 7 as well as PGJ and NLSFJ as unknown patterns, the accuracy of the
RCAE-OWR algorithm for the seven known jamming patterns and the detection rate for
Remote Sens. 2023,15, 4107 16 of 24
the two unknown jamming patterns are shown in Figure 8. It can be observed that as the
JNR increases, the recognition rates for various patterns of jamming also increase.This is
because with a higher JNR, the signal becomes cleaner, and the extracted features become
more discriminative. When the JNR is below
2 dB, the various jamming signals are
submerged in the noise, and the extracted semantic features are not sufficient to separate
them. As a result, samples of known jamming patterns are often recognized as samples
of unknown jamming patterns, leading to a higher detection rate for unknown jamming
patterns compared to known jamming patterns.
-4 -2 0 2 4 6 8 10 12 14
JNR/dB
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Accuracy/%
RNJ
AMJ
FMJ
CJ
PMJ
LSFJ
HFJ
unknown
Figure 8. Open set recognition of RCAE-OWR.
Table 5presents the accuracy of the RCAE-OWR algorithm and the supervised-based
algorithm based on seven known jamming patterns when the JNR is set to 10dB. For a fair
comparison, the network structure of the supervised-based algorithm is consistent with
the encoder of the RCAE-OWR network in this paper. The difference lies in the fact that
the supervised-based algorithm directly outputs the probabilities of the seven jamming
patterns as the recognition results, while RCAE-OWR does not have prior knowledge
of the specific number of patterns. Therefore, Algorithm 2is used for jamming pattern
recognition in RCAE-OWR. As the supervised-based algorithm is designed for closed-
set recognition, it achieves higher detection rates. On the other hand, the RCAE-OWR
algorithm is designed for open-set recognition, which means that known jamming patterns
can be wrongly recognized as unknown jamming patterns, and unknown jamming patterns
can be misclassified as known jamming patterns. As a result, the recognition rate of
RCAE-OWR is lower than that of the supervised-based algorithm.
Table 5. Comparison of detection performance between two algorithms.
Method RNJ AMJ FMJ CJ PMJ LSFJ HFJ
supervised 98% 100% 99% 95% 92% 98% 95%
RCAE-OWR 96% 100% 98.5% 93.5% 89% 97.5% 90.5%
The confusion matrix of RCAE-OWR for known jamming patterns and unknown
jamming patterns is shown in Figure 9when JNR is set to 18 dB. It can be observed from
the figure that almost all samples are correctly classified into their respective patterns. This
is because with higher levels of jamming, the extracted signal features better represent
the original signal, resulting in increased intra-class cohesion and expanded inter-class
separation. However, it is also evident that there are cases where known jamming patterns
Remote Sens. 2023,15, 4107 17 of 24
are recognized as unknown jamming patterns, and vice versa. Therefore, our algorithm
needs to ensure a high detection rate for unknown jamming patterns while minimizing the
impact on the recognition of known jamming patterns.
Figure 9. RCAE-OWR open set recognition results when JNR = 18 dB.
4.8. Analysis of Unknown Jamming Patterns Recognition Performance in RCAE-OWR
To evaluate the recognition performance of RCAE-OWR for unknown jamming pat-
terns,
Precision
,
Recall
, and
F
1 are introduced as performance metrics [
6
].
Precision
repre-
sents the ratio of the number of correctly classified samples to the sum of correctly classified
samples and incorrectly classified samples.
Recall
represents the ratio of the number of cor-
rectly classified samples to the total number of samples.
F
1 is defined as the harmonic mean
of Precision and Recall, given by F1=2×Precision ×Recal l/(Precision +Recal l)[40].
4.8.1. Performance Comparison of Different Methods
In order to validate the effectiveness of our proposed RCAE-OWR, two other OWR-
based methods, IOmSVM [
41
] and SNN-OWR [
35
] are implemented for comparison.
The features extracted in the IOmSVM method are referenced from [
12
]. We utilize Open-
ness [
42
] to measure the complexity of the open world task to describe the weight of the
number of unknown patterns. Openness is defined as :
Openness =1sNtrain
Ntest (27)
where
Ntrain
denotes the number of patterns contained in the training set and
Ntest
denotes
the number of patterns contained in the test set. When the Openness is 0 , the task will
move to closed-set recognition. In this paper, we set
Ntest
to 9 and keep changing the value
of
Ntrain
for experiments;
F
1 score is used to measure the performance of the algorithm
and the results are shown in Figure 10. From the figure, it can be seen that as the Openness
increases, the classification accuracy of the three methods decreases, which is attributed
to the fact that more information about unknown patterns is fed into the model, posing a
serious challenge to the classification of the model. Meanwhile, RCAE-OWR has the highest
recognition accuracy at the same Openness, e.g., when the Openness is 0.25, the average
F
1
score of RCAE-OWR is 4.29% and 10.26% higher than that of SNN-OWR and IOmSVM,
respectively.
Remote Sens. 2023,15, 4107 18 of 24
0.15 0.2 0.25 0.3 0.35 0.4
openness
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Average F1/%
IOmSVM
SNN-OWR
RCAE-OWR
Figure 10. F1 scores against varying Openness under different methods.
Figure 11 illustrates the performance of the three methods as the JNR varies, with
Nu
set to 3. Notably, the detection performance of all three methods exhibits improvement as
JNR increases, and the differences in
F
1 scores diminish with higher JNR values. Among the
three OWR-based methods, RCAE-OWR stands out with the highest classification accuracy,
particularly at low JNR. For instance, when the JNR is 8 dB, the average
F
1 score of
RCAE-OWR surpasses that of SNN-OWR and IOmSVM by 6.85% and 12.23%, respectively.
4 5 6 7 8 9 10 11 12
JNR/dB
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Average F1/%
IOmSVM
SNN-OWR
RCAE-OWR
Figure 11. F1 scores against JNR under different methods.
Table 6presents the offline training time and online recognition time of the three
methods. It can be observed from the table that the IOmSVM method has the shortest
training time, and likewise the recognition accuracy is the worst within the three methods.
Remote Sens. 2023,15, 4107 19 of 24
Both SNN-OWR and RCAE-OWR require longer training times; however, once the models
are trained effectively, they still achieve real-time jamming recognition.
Table 6. Runtime (in seconds) of different methods.
Method Offline Training Online Recognition
IOmSVM 753 0.0075
SNN-OWR 1506 0.0071
RCAE-OWR 1427 0.0062
4.8.2. Impact of Unknown Patterns on RCAE-OWR Performance
To investigate the influence of the number of unknown jamming patterns on the
algorithm’s recognition performance, we randomly selected two, three, four, five and six
kinds of patterns out of the nine kinds of jamming patterns as unknown jamming patterns,
while the remaining patterns were considered known jamming patterns. The experimental
results are presented in Table 7, with JNR set to 12
dB
. From the table, it can be observed
that as the number of unknown jamming patterns decreases, the algorithm’s detection
performance improves, as indicated by an increase in the
F
1 score. For example, when the
number of unknown jamming patterns is reduced from four to two, the average
F
1 score
increases by 0.070. However, when the number of unknown patterns surpasses the number
of known patterns, it poses a challenge to the algorithm’s detection performance. For ex-
ample, when increasing the number of unknown patterns from four to five, the average
F1 score decreases by 0.174.
Table 7. Effect of Unknown jamming patterns on RCAE-OWR Performance.
Number of Patterns Pattern Prec ision R ecall F1
2Pattern 1 0.941 0.964 0.952
Pattern 2 0.979 0.940 0.959
3
Pattern 1 0.916 0.940 0.927
Pattern 2 0.929 0.928 0.928
Pattern 3 0.922 0.900 0.910
4
Pattern 1 0.861 0.930 0.894
Pattern 2 0.907 0.920 0.913
Pattern 3 0.912 0.820 0.863
Pattern 4 0.909 0.841 0.873
5
Pattern 1 0.841 0.724 0.778
Pattern 2 0.781 0.696 0736
Pattern 3 0.701 0.680 0.690
Pattern 4 0.623 0.713 0.663
Pattern 5 0.542 0.692 0.692
6
Pattern 1 0.779 0.598 0.677
Pattern 2 0.712 0.545 0.613
Pattern 3 0.603 0.536 0.567
Pattern 4 0.481 0.533 0.505
Pattern 5 0.365 0.452 0.402
Pattern 6 0.243 0.405 0.301
4.8.3. Recognition Performance of RCAE-OWR for Unknown Jamming Patterns
The recognition rates of RCAE-OWR for unknown jamming patterns with JNR of
4
14 dB are shown in Figure 12. It can be observed from the graph that when the number
of unknown jamming patterns is fixed, the average
F
1 score increases with the increase in
JNR. This is because at lower JNR levels, a large number of known jamming patterns are
incorrectly classified as unknown, resulting in poor classification performance. As the JNR
Remote Sens. 2023,15, 4107 20 of 24
increases, the characteristics of different jamming signals become more distinct, and the
samples of known patterns wrongly classified as unknown patterns decrease, leading
to improved performance in classifying unknown patterns. Additionally, at lower JNR
levels, there is a significant difference in the average
F
1 score among the three different
jamming patterns. However, as the JNR increases, this difference decreases. For example, at
JNR = 4 dB, the average
F
1 score for two unknown patterns is 0.3 and 0.65 higher than that
for three and four unknown patterns, respectively. At JNR = 10 dB, the average
F
1 score
for two unknown patterns is 0.08 and 0.15 higher than that for three and four unknown
patterns, respectively. This also highlights the importance of improving the JNR of the
dataset as a prerequisite and guarantee for open world recognition of radar jamming signals.
4 5 6 7 8 9 10 11 12 13 14
JNR/dB
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Average F1/%
Nu=2
Nu=3
Nu=4
Figure 12. Recognition performance of RCAE-OWR for different unknown jamming patterns.
4.8.4. Performance of RCAE-OWR for Unknown Patterns under Different Combinations
To investigate the influence of JNR on the algorithm’s performance with varying
combinations of unknown patterns, we randomly extract three or four kinds of signals from
the dataset to form the unknown patterns set, while the remaining are considered as known
patterns for experimental evaluation. The results of these experiments are presented in
Figure 13. From the figure, it can be seen that as the JNR increases, the overall classification
accuracy also improves; this is because as the JNR increases, the signal is less affected by
noise and is more discriminative. Notably, when the number of unknown patterns is fixed,
the classification accuracy consistently increases with higher JNR values. Additionally,
regardless of the specific combination of unknown patterns data sets, the classification accu-
racy exhibits minimal variation under the same JNR condition, highlighting the robustness
of the RCAE-OWR algorithm.
Remote Sens. 2023,15, 4107 21 of 24
4 5 6 7 8 9 10 11 12
JNR/dB
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Average F1/%
unknown patterns =RNJ,FMJ,LSFJ
unknown patterns =FMJ,NLSFJ,HFJ
unknown patterns =AMJ,HFJ,PGJ
unknown patterns = FMJ,CJ,PMJ,NLSFJ
unknown patterns =RNJ,AMJ,FMJ,CJ
unknown patterns =RNJ,CJ,LSFJ,PGJ
Figure 13. Performance under different unknown patterns.
5. Conclusions
In this study, we addressed the problem of open world recognition for radar jamming
signals. Leveraging the idea of zero-shot learning, we proposed an open world recognition
method called RCAE-OWR based on residual autoencoders. This method utilizes the
features extracted by the encoder network to form the semantic centers of known jamming
patterns. Then, a distance-based approach is proposed to classify known and unknown
jamming patterns. Experimental results demonstrate that the RCAE-OWR algorithm can
effectively recognize unknown jamming patterns, especially under high JNR. Although the
proposed RCAE-OWR algorithm has demonstrated excellent classification performance in
suppression jamming signals, there are still some limitations and drawbacks to address.
For instance, this paper focused only on common jamming signals, while there are other
jamming patterns, such as agile noise jamming and deceptive jamming. In the future, these
patterns should be incorporated into our method. Additionally, deep learning has been
widely applied to intelligent radar jamming signal recognition with a balanced training set.
However, in most cases, an imbalanced training set is inevitable. Therefore, considering
open world recognition of radar jamming under class-imbalanced conditions will be a
challenging yet valuable task in future research.
Author Contributions:
Conceptualization Z.Z. and Y.Z.; methodology, Y.Z. and Z.Z.; software, Y.B.
and Y.Z.; validation, Z.Z.; formal analysis, Z.Z.; investigation, Z.Z. and Y.Z.; resources, Z.Z.; data
curation, Y.B.; writing—original draft preparation, Y.Z.; writing—review and editing, Y.B. and Y.Z.;
visualization, Z.Z.; supervision, Z.Z.; project administration, Y.Z.; funding acquisition, Z.Z. All
authors have read and agreed to the published version of the manuscript.
Funding:
This work was supported by the State Key Program of National Natural Science of China
(Grant No. U19B2016) and Zhejiang Provincial Key Lab of Data Storage and Transmission Technology,
Hangzhou Dianzi University.
Acknowledgments:
The authors would like to thank an Editor who handled the manuscript and the
anonymous reviewers for their valuable comments and suggestions.
Conflicts of Interest: The authors declare no conflict of interest.
Remote Sens. 2023,15, 4107 22 of 24
Appendix A
Appendix A.1
Lemma A1.
Assuming the
θLce
,
θLcl
and
θLle
are gradients of
Lce
,
Lcl
and
Leε
, respectively,
then there is:
θL=θLce +λcl θLcl +λreθLre
=1
M
i
(S(E(xc
i)) yc
i)+
i
E(xc
i)zc
yc
i+
i
˜
xc
ixc
i!(A1)
Proof. because L=Lce +λcl Lcl +λre Lre,eLcan be notated as:
θLt=θLce +λrl θLkl +λctθLct (A2)
Notice that
Lce =1
M
M
i=1Lce(yi
,
ˆyi)
, where
Lce(yi
,
ˆyi) = Nc
j=1
yc
i,jlog(ˆ
yc
i,j)
, Let
u=[u1,u2, . . . , uNc]
denote the output vector of the last layer of the network. After softmax
operation,
u
yields
ˆyi
. That is,
ˆyj
i=SExc
ij=euj
keuk
. where
S(·)
represents the soft-
max operation, and
S(E(xi))j
and
uj
are the
j
-th elements of
SExc
i
and
u
, respectively.
According to the chain derivative rule:
Lce
u=Lce
p
p
u(A3)
We have:
Lce
uj
=
m
Lcem
SExc
im
SExc
im
uj
=
m
yc
i,mlog SExc
im
Exc
im
SExc
im
uj
=
myc
i,m
SExc
im
SExc
im
uj
Simplify to get:
SExc
im
uj
=(SExc
ij1SExc
ij,m=j
SExc
ijSExc
im,m6=j(A4)
Hence, Lce
ucan be rewritten as:
Lce
uj
=
m
Lcem
SExc
im
SExc
im
uj
=
m
yc
i,mlog SExc
im
Exc
im
SExc
im
uj
=S(E(xc
i))j1S(E(xc
i))j+
m6=jyc
i,m
SExc
imS(E(xc
i))jS(E(xc
i))m
=S(E(xc
i))jyc
i,j
Remote Sens. 2023,15, 4107 23 of 24
therefore, Lce
u=SExc
iyc
i, according to Lce =1
MM
i=1Lce(yi, ˆyi), we have:
Lce
u=1
M
i
S(E(xc
i)) yc
i(A5)
Similarly, Lcl
u=Lcl
E(xc
i), and Lcl =1
MM
i=1
Exc
izc
vc
i
2
2, we have:
Lcl
u=Lcl
E(xi)=1
M
i
E(xc
i)zc
yc
i(A6)
Similarly,
Lre
u== Lre
Exc
i=1
M
i
˜
xc
ixc
i(A7)
In summary, it can be concluded that:
θL=θLce +λcl θLcl +λreθLre
=1
M
i
(S(E(xc
i)) yc
i)+
i
E(xc
i)zc
yc
i+
i
˜
xc
ixc
i!(A8)
References
1.
Du, C.; Cong, Y.; Zhang, L.; Guo, D.; Wei, S. A Practical Deceptive Jamming Method Based on Vulnerable Location Awareness
Adversarial Attack for Radar HRRP Target Recognition. IEEE Trans. Inf. Forensics Secur. 2022,17, 2410–2424. [CrossRef]
2.
Xu, H.; Quan, Y.; Zhou, X.; Chen, H.; Cui, T.J. A Novel Approach for Radar Passive Jamming Based on Multiphase Coding Rapid
Modulation. IEEE Trans. Geosci. Remote Sens. 2023,61, 1–14. [CrossRef]
3.
Liu, Z.; Zhao, S. Unsupervised Clustering Method to Discriminate Dense Deception Jamming for Surveillance Radar. IEEE Sensors
Lett. 2021,5, 1–4. [CrossRef]
4.
Lv, Q.; Quan, Y.; Feng, W.; Sha, M.; Dong, S.; Xing, M. Radar Deception Jamming Recognition Based on Weighted Ensemble CNN
With Transfer Learning. IEEE Trans. Geosci. Remote Sens. 2022,17, 1–11. [CrossRef]
5.
Hua, Q.; Zhang, Y.; Wei, C.; Ji, Z.; Jiang, Y.; Wang, Y.; Xu, D. A Self-Supervised Method Based on CV-MUNet++ for Active
Jamming Suppression in SAR Images. IEEE Trans. Geosci. Remote Sens. 2023,61, 1–16. [CrossRef]
6.
Luo, Z.; Cao, Y.; Yeo, T.S.; Wang, Y.; Wang, F. Few-Shot Radar Jamming Recognition Network via Time-Frequency Self-Attention
and Global Knowledge Distillation. IEEE Trans. Geosci. Remote Sens. 2023,61, 1–12. [CrossRef]
7.
Huang, S.; Feng, Z.; Zhang, Y.; Zhang, K.; Li, W. Feature based modulation classification using multiple cumulants and antenna
array. In Proceedings of the 2016 IEEE Wireless Communications and Networking Conference, Doha, Qatar, 3–6 April 2016; pp. 1–5.
8.
Chen, T.; Gao, S.; Zheng, S.; Yu, S.; Xuan, Q.; Lou, C.; Yang, X. EMD and VMD Empowered Deep Learning for Radio Modulation
Recognition. IEEE Trans. Cogn. Commun. Netw. 2023,9, 43–57. [CrossRef]
9.
Zhu, F.; Jiang, Q.Q.; Lin, C.; Xiao, Y. Typical wide band EMI identification based on support vector machine. Syst. Eng. Electron.
2021,43, 2400–2406.
10.
Lu, Y.; Li, S. CFAR detection of DRFM deception jamming based on singular spectrum analysis. In Proceedings of the 2017 IEEE
International Conference on Signal Processing, Communications and Computing (ICSPCC), Xiamen, China , 22–25 October 2018;
pp. 1–6.
11.
Yang, S.; Tian, B.; Zhou, R. A jamming identification method against radar deception based on bispectrum analysis and fractal
dimension. J. Xian Jiaotong Univ. 2016,50, 128–135.
12.
Xu, C.; Yu, L.; Wei, Y.; Tong, P. Research on Active Jamming Recognition in Complex Electromagnetic Environment. In
Proceedings of the 2019 IEEE International Conference on Signal, Information and Data Processing (ICSIDP), Chongqing, China,
11–13 December 2019; pp. 1–5.
13.
Wei, Y.; Li, Y.; Zhang, J.; Tong, P. Radar Jamming Recognition Method Based on Fuzzy Clustering Decision Tree. In Proceedings of
the 2019 IEEE International Conference on Signal, Information and Data Processing (ICSIDP), Chongqing, China, 11–13 December
2019; pp. 1–5.
14.
Yu, Y.; Li, J.; Li, J; Xia, Y.; Ding, Z.; Samali, B. Automated damage diagnosis of concrete jack arch beam using optimized deep
stacked autoencoders and multi-sensor fusion. Dev. Built Environ. 2023,14, 100128. [CrossRef]
15.
Zhang, Y.; Zhao, Z. Limited Data Spectrum Sensing Based on Semi-Supervised Deep Neural Network. IEEE Access
2021
,9,
1813–1817. [CrossRef]
16.
Zhang, Y.P.; Zhao, Z.J. Semi-supervised deep learning using pseudo labels for spectrum sensing. J. Nonlinear Convex Anal.
2022
,
23, 1913–1929.
Remote Sens. 2023,15, 4107 24 of 24
17.
Yu, Y.; Hoshyar, A.N.; Samali, B.; Zhang, G.; Rashidi, M.; Mohammadi, M. Corrosion and coating defect assessment of coal
handling and preparation plants (CHPP) using an ensemble of deep convolutional neural networks and decision-level data
fusion. Neural Comput. Appl. 2023,35, 18697–18718. [CrossRef]
18.
Lv, Q.; Quan, Y.; Sha, M.; Feng, W.; Xing, M. Deep Neural Network-Based Interrupted Sampling Deceptive Jamming Countermea-
sure Method. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2022,15, 9073–9085. [CrossRef]
19.
Zhao, Q.; Liu, Y.; Cai, L.; Lu, Y. Research on electronic jamming identification based on CNN. In Proceedings of the 2019 IEEE
International Conference on Signal, Information and Data Processing (ICSIDP), Chongqing, China, 11–13 December 2019.
20.
Li, M.; Ren, Q.; Wu, J. Interference classification and identification of TDCS based on improved convolutional neural network. J.
Phys. Conf. Ser. 2020,16, 121–155. [CrossRef]
21.
Liu, Q.; Zhang, W. Deep learning and recognition of radar jamming based on CNN. In Proceedings of the 2019 12th International
Symposium on Computational Intelligence and Design (ISCID), Hangzhou, China, 14–15 December 2019; pp. 208–212.
22.
Shao, G.; Chen, Y.; Wei, Y. Convolutional neural network-based radar jamming signal classification with sufficient and limited
samples. IEEE Access 2020,8, 80588–80598. [CrossRef]
23.
Shao, G.; Chen, Y.; Wei, Y. Deep fusion for radar jamming signal classification based on CNN. IEEE Access
2020
,8, 117236–117244.
[CrossRef]
24.
Tian, X.; Chen, B.; Zhang, Z. Multiresolution Jamming Recognition with Few-shot Learning. In Proceedings of the 2021 CIE
International Conference on Radar (Radar), Haikou, Hainan, China, 15–19 December 2021; pp. 2267–2271.
25.
Luo, H.; Liu, J.; Wu, S.; Nie, Z. A Semi-Supervised Deception Jamming Discrimination Method. In Proceedings of the 2021 IEEE
7th International Conference on Cloud Computing and Intelligent Systems (CCIS), Xi’an, China, 7–8 November 2021; pp. 428–432.
26.
Zhou, Y.; Shang, S.; Song, X.; Zhang, S.; You, T.; Zhang, L. Intelligent Radar Jamming Recognition in Open Set Environment Based
on Deep Learning Networks. Remote Sens. 2022,14, 2410–2424. [CrossRef]
27.
Wang, L.; Yang, X.; Tan, H.; Bai, X.; Zhou, F. Few-Shot Class-Incremental SAR Target Recognition Based on Hierarchical
Embedding and Incremental Evolutionary Network. IEEE Trans. Geosci. Remote Sens. 2023,61, 1–11. [CrossRef]
28.
Wei, Q. R.; He, H.; Zhao, Y.; Li, J.A. Learn to Recognize Unknown SAR Targets From Reflection Similarity. IEEE Geosci. Remote
Sens. Lett. 2022,19, 1–5. [CrossRef]
29.
Song, Q.; Chen, H.; Xu, F.; Cui, T.J. EM Simulation-Aided Zero-Shot Learning for SAR Automatic Target Recognition. IEEE Geosci.
Remote Sens. Lett. 2020,17, 1092–1096. [CrossRef]
30.
Bernardino, R.P.; Torr, P.H.S. An embarrassingly simple approach to zero-shot learning. In Proceedings of the 32nd international
conference on Machine learning (ICML’15), Lille, France, 6–11 July 2015.
31.
Kodirov, E.; Xiang, T.; Gong, S. Semantic Autoencoder for Zero-Shot Learning. In Proceeding of the 2017 IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA, 21–26 July 2017; pp. 4447–4456.
32.
Zhang, L.; Xiang, T.; Gong, S. Learning a Deep Embedding Model for Zero-Shot Learning. In Proceeding of the 2017 IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA, 21–26 July 2017; pp. 3010–3019.
33.
Ma, B.J.; Qi, J.; Su, H.Y.; Zhang, X.F. One-dimensional Radar Active Jamming Signal Recognition Method Based on Bayesian Deep
Learning. Signal Process. 2023,39, 235–243.
34.
Qu, Q.; Wei, S.; Liu, S.; Liang, J.; Shi, J. JRNet: Jamming Recognition Networks for Radar Compound Suppression Jamming
Signals. IEEE Trans. Veh. Technol. 2020,69, 15035–15045. [CrossRef]
35.
Tang, Y.; Zhao, Z.; Chen, J.; Zheng, S.; Ye, X.; Lou, C.; Yang, X. Open world recognition of communication jamming signals. China
Commun. 2023,20, 199–214. [CrossRef]
36.
Wang, J.Q. Study and Implementation of Radar Active Jamming Type Discrimination. Master’s Thesis, Xidian University, Xi’an,
China, 2014.
37.
Zhang, H.; Yu, L.; Chen, Y.; Wei, Y. Fast Complex-Valued CNN for Radar Jamming Signal Recognition. Remote Sens.
2021
,13,
2867. [CrossRef]
38.
Fu, Y.; Hospedales, T.M.; Xiang, T.; Gong, S. Trans-ductive multi-view zero-shot learning. IEEE Tran. PAMI
2015
,37, 2332–2345.
[CrossRef]
39.
Dong, Y.; Jiang, X.; Zhou, H.; Lin, Y.; Shi, Q. SR2CNN: Zero-Shot Learning for Signal Recognition. IEEE Trans. Signal Process.
2021
,
69, 2316–2329. [CrossRef]
40.
Yang, L.; Li, Q.; Shao, H.Z. An Open Set Recognition Algorithm of Electromagnetic Target Based on Metric Learning and Feature
Subspace Projection. Acta Electron. Sin. 2022,17, 1310–1318.
41. Jleed, H.; Martin, B. Incremental multiclass open-set audio recognition. Int. J. Adv. Intell. Inform. 2022,8, 251–270. [CrossRef]
42.
Geng, X.; Dong, G.; Xia, Z.; Liu, H. SAR Target Recognition via Random Sampling Combination in Open-World Environments.
IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2023,16, 331–343. [CrossRef]
Disclaimer/Publisher’s Note:
The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.
... With JNR ranging from 10 dB to 15 dB, the average recognition rate for eight jamming types reached 88%, outperforming supervised neural networks. Later, Zhang et al. [167] applied the concept of ZSL to propose an open-world recognition method called the RCAE-OWR algorithm. During supervised training, they built a residual convolutional autoencoder network structure to 23 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING extract semantic information solely from the known training set. ...
Article
Full-text available
In the modern warfare with complex and changeable electromagnetic environment, radar jamming is getting more complex and realistic, which poses a serious threat to radar, jamming recognition has become a hot topic in the field of electronic countermeasures. To make effective anti-jamming measures, numerous jamming recognition methods have been proposed. This paper presents a systematic review of jamming recognition for this topic. Specifically, firstly building a system framework for jamming models, including deception jamming, suppression jamming, and smart jamming, thoroughly explaining the operational mechanisms. Then, recognition methods based on traditional machine learning are summarized, and delves into the advantages and disadvantages of feature extraction methods and classifiers. Furthermore, the focus shifts to neural network-based methods, such as shallow neural network methods and deep neural network methods. In particular, restricted sample strategies are also discussed as potential future directions. Finally, concluding on the current status of jamming recognition methods and the prospects for future work are made. This paper provides a reference for the research of radar jamming recognition.
... According to the format of the input radar jamming signal, these CNN-based methods can be divided into onedimensional methods [16,17] and two-dimensional methods [18,19]. For one-dimensional methods, radar jamming echo sequences are directly fed into deep-learning models [20,21]. Unlike one-dimensional methods, for two-dimensional methods, echo sequences are usually transformed into radar jamming images using a short-time Fourier transform. ...
Article
Full-text available
Radar jamming recognition is a key step in electronic countermeasures, and accurate and sufficient labeled samples are essential for supervised learning-based recognition methods. However, in real practice, collected radar jamming samples often have weak labels (i.e., noisy-labeled or unlabeled ones), which degrade recognition performance. Additionally, recognition performance is hindered by limitations in capturing the global features of radar jamming. The Transformer (TR) has advantages in modeling long-range relationships. Therefore, a weakly supervised Transformer is proposed to address the issues of performance degradation under weak supervision. Specifically, complementary label (CL) TR, called RadarCL-TR, is proposed to improve radar jamming recognition accuracy with noisy samples. CL learning and a cleansing module are successively utilized to detect and remove potentially noisy samples. Thus, the adverse influence of noisy samples is mitigated. Additionally, semi-supervised learning (SSL) TR, called RadarSSL-PL-TR, is proposed to boost recognition performance under unlabeled samples via pseudo labels (PLs). Network generalization is improved by training with pseudo-labeling unlabeled samples. Moreover, the RadarSSL-PL-S-TR is proposed to further promote recognition performance, where a selection module identifies reliable pseudo-labeling samples. The experimental results show that the proposed RadarCL-TR and RadarSSL-PL-S-TR outperform comparison methods in recognition accuracy by at least 7.07% and 6.17% with noisy and unlabeled samples, respectively.
Article
Full-text available
In view of the problems of ineffective feature extraction and low detection accuracy in existing detection system, this study presents a novel machine vision-based approach composed of an ensemble of deep convolutional neural networks (CNNs) and improved Dempster-Shafer (D-S) theory-based data fusion to evaluate corrosion and coating defect of coal handling and preparation plants. To start with, the structural surface image is sent to each transferred CNN for initial defect identification. Then, an improved D-S fusion algorithm is proposed to combine the identification results from different CNNs, which are vectors consisting of statistical indicators of all the potential damage severity categories. The decision-level fusion of different CNNs can effectively improve image classification. To validate the performance of the proposed method, a dataset made of 3593 surface images with different defect severities captured from mining infrastructure in field is established together with data augmentation. The validation result demonstrates that the proposed method is able to effectively improve the recognition accuracy of defect severity and reduce the wrong recognition rate. Finally, the robustness of the proposed approach is also appraised by polluting the images with different types and intensities of noise, with satisfactory results.
Article
Full-text available
Jamming recognition is an essential step in radar detection and anti-jamming in the complex electromagnetic environment. When radars detect an unknown type of jamming that does not occur in the training set, the existing radar jamming recognition algorithms fail to correctly recognize it. However, these algorithms can only recognize this type of jamming as one that already exists in our jamming library. To address this issue, we present two models for radar jamming open set recognition (OSR) that can accurately classify known jamming and distinguish unknown jamming in the case of small samples. The OSR model based on the confidence score can distinguish known jamming from unknown jamming by assessing the reliability of the sample output probability distribution and setting thresholds. Meanwhile, the OSR model based on OpenMax can output the probability of jamming belonging to not only all known classes but also unknown classes. Experimental results show that the two OSR models exhibit high recognition accuracy for known and unknown jamming and play a vital role in sensing complex jamming environments.
Article
Synthetic aperture radar (SAR) system is susceptible to electromagnetic jamming during imaging, which seriously affects the subsequent interpretation of SAR images. Aiming at the problem of active suppressive jamming, this paper proposes a suppression method of SAR suppressive jamming based on self-supervised complex-valued deep learning, which consists of a novel complex-valued jamming suppression network CV-MUNet++ and a self-supervised training strategy. CV-MUNet++ could fully use the amplitude and phase information of complex-valued SAR images. The network’s weights, activation functions, and convolution operations are designed for complex domain processing. The different information representations of target and jamming in amplitude and phase in SAR images are mined to achieve jamming suppression. The self-supervised training strategy is proposed to solve the problem of relying heavily on manually labeled samples in the traditional network training process and is suitable for application scenarios where ground truth is difficult to obtain under complex jamming. The experimental results show that the proposed method could effectively suppress the active jamming of complex backgrounds and has the ability to self-supervised intelligent jamming suppression.
Article
Radar jamming recognition aims to accurately recognize the type of jamming to provide guidance for radar countermeasures. Although previous deep learning-based methods have made promising performance, they mainly rely on convolutional neural network (CNN) based on local processing, which ignore the global information in the time-frequency domain data of the jamming signal and also require much inference time to obtain the final recognition results. In this paper, a novel few-shot jamming recognition network via time-frequency self-attention and global knowledge distillation (JR-TFSAD) is proposed by jointly considering the global information in the time-frequency spectrum of the jamming signal and the real-time performance of the recognition network. A time-frequency self-attention model (TFSA) is proposed to extract the global deep features of radar jamming signals by learning the correlation between two arbitrary points in the time-frequency spectrum of the jamming signal, thus improving the recognition accuracy. Moreover, to effectively reduce the inference time of the method while preserving the recognition accuracy, a global knowledge distillation model (GKD) is further constructed to perform jamming recognition by distilling the global knowledge from the TFSA model. The experimental results on the simulated and measured mixed dataset verify that the proposed method has higher recognition accuracy and shorter inference time compared to the state-of-the-art methods.
Article
To improve the recognition ability of communication jamming signals, Siamese Neural Network-based Open World Recognition (SNN-OWR) is proposed. The algorithm can recognize known jamming classes, detect new (unknown) jamming classes, and unsupervised cluseter new classes. The network of SNN-OWR is trained supervised with paired input data consisting of two samples from a known dataset. On the one hand, the network is required to have the ability to distinguish whether two samples are from the same class. On the other hand, the latent distribution of known class is forced to approach their own unique Gaussian distribution, which is prepared for the subsequent open set testing. During the test, the unknown class detection process based on Gaussian probability density function threshold is designed, and an unsupervised clustering algorithm of the unknown jamming is realized by using the prior knowledge of known classes. The simulation results show that when the jamming-to-noise ratio is more than 0dB, the accuracy of SNN-OWR algorithm for known jamming classes recognition, unknown jamming detection and unsupervised clustering of unknown jamming is about 95%. This indicates that the SNN-OWR algorithm can make the effect of the recognition of unknown jamming be almost the same as that of known jamming.
Article
It is difficult to realize effective synthetic aperture radar (SAR) automatic target recognition (ATR) in open scenarios because the ATR model cannot continuously learn from new classes with limited training samples. When adding new classes to the previously trained model, the capability of recognizing old classes may lose due to severe overfitting. To tackle this problem, a few-shot class-incremental SAR ATR method, namely hierarchical embedding and incremental evolutionary network (HEIEN), is proposed in this paper. Firstly, a hierarchical embedding network and a hybrid distance-based classifier is constructed for basic feature extraction and classification. Then, in order to obtain more accurate decision boundaries, an adaptive class-incremental learning module is designed to adjust the weights of classifiers in all tasks by collecting context information from the past to the present. Finally, a pseudo incremental training strategy is designed to enable effective model training with only a few samples. Experimental results on the MSTAR benchmark data set have illustrated that HEIEN performs well with remarkable advantages in few-shot class-incremental SAR ATR tasks.
Article
Electromagnetic controlled surface (ECS) which can regulate the amplitude and phase of electromagnetic wave reflection characteristics has attracted extensive attention in recent years because of its radar stealth effect by reducing the radar cross section (RCS). However, radar detection cannot be avoided only by stealth in the energy field gradually. It is a novel pulse Doppler radar jamming mode proposed in this paper to utilize the self-correlation of radar-transmitted waveforms to result in signal processing and radar detection failure through the rapid time-domain variation of phase coding ECS. In principle, phase modulation alters the intra-pulse characteristics of the original signal and fundamentally interferes with the processing of the radar signal. In this paper, random and periodic phase coding sequences are proposed according to different forms of radar jamming. And through the derivation formula and simulation experiment, the jamming effect under multi-phase modulation is verified. Among them, the target energy is dispersed to the surroundings to form a wide envelope through random coding modulation, which leads to noise barrage jamming. While we use periodic sequence coding modulation, the target is shifted in the range-Doppler domain, causing misplaced coherence on the radar echo and deceptive jamming to the radar. Moreover, the technology can also be widely used in synthetic aperture radar (SAR) imaging, microwave measurement and other remote sensing fields. To further confirm the accuracy and efficiency of the proposed method, multi-parameter contrast experiments and parameter sensitivity analyses are conducted.
Article
A novel hybrid framework of optimized deep learning models combined with multi-sensor fusion is developed for condition diagnosis of concrete arch beam. The vibration responses of structure are first processed by principal component analysis for dimensionality reduction and noise elimination. Then, the deep network based on stacked autoencoders (SAE) is established at each sensor for initial condition diagnosis, where extracted principal components and corresponding condition categories are inputs and output, respectively. To enhance diagnostic accuracy of proposed deep SAE, an enhanced whale optimization algorithm is proposed to optimize network meta-parameters. Eventually, Dempster-Shafer fusion algorithm is employed to combine initial diagnosis results from each sensor to make a final diagnosis. A miniature structural component of Sydney Harbour Bridge with artificial multiple progressive damages is tested in laboratory. The results demonstrate that the proposed method can detect structural damage accurately, even under the condition of limited sensors and high levels of uncertainties.