IEEE - Aerospace and Electronic Systems - April 2023 - 25

Huang et al.
with different features. The optimal hyperparameters are
selected for each model, and the average of five trials of
model training and testing is used to compare the performance
of the LPI waveform recognition. For the hyperparameters
tuning, the number of filters of the
convolutional layer ranges in 24- 28 and the kernel size is
from 22to 55. The number of nodes in dense layers
ranges in 24-28. With the input feature dimension of
512 80 4, the batch size of 100, and 150 epochs of
training, the parameters of the CNN model are determined
coupled with the batch normalization layer and the dropout
layer. The final CNN structure for the fusion feature is
shown in Table 4. Figure 12 shows the probability of correct
classification (PCC) of the LPI waveform recognition
models, and Table 3 is the PCC of each method with different
SNR levels. As shown in Figure 12 and Table 4, the
fusion feature of the proposed approach achieves higher
PCC than other features over all the SNR range and therefore
outperforms (up to 41% improvement of PCC) the
traditional features including wavelet feature, CWD feature,
and WVD feature in the waveform recognition of the
simulated LPI signals. For the recognition of specific LPI
waveform type, Figure 13 shows the confusion matrices
of the fusion feature under several SNR levels, which
present the classification details of the different LPI waveforms.
In order to examine the recognition performance,
SNR levels of -10 dB, -15 dB, -18 dB, and -20 dB are
selected to present the recognition results with 100 signal
samples in each type of the LPI signals. In the confusion
matrix, each row of the matrix represents the actual waveform
type while each column represents the predicted signal
type. Therefore, the diagonal entries represent the
number of correctly classified signal type, while other
entries show the number of signal classified incorrectly.
According to the figure, with the SNR levels going down
from (a) to (d), the numbers in the diagonal entries
decrease, which means the number of correct recognition
for different types of waveforms decreases. For each SNR
level in the figure, the number of incorrect P1 and P4 signals
classification is more than other signal types, which is
mainly due to the similarity of the phase code of the two
types of signals that makes them difficult to classify. It is
also noticed that the noise radar has a high recognition
rate. This is because the noise radar signal used in this
study is generated by adding noise to the LFM signals.
After adding the channel-based AWGN, the noise radar
signals are exactly the LFM signals with more noise-contaminated
compared with other phase-coded LPI signals.
Therefore, the trained classification model is prone to classify
noisy signals to noise radar signal. Overall, the LPI
waveform recognition model with fusion feature achieves
PCC greater than 70% even when the SNR is down to
-20 dB, which demonstrates the superior performance of
the waveform recognition when noise exists.
APRIL 2023
CONCLUSION
In this article, we developed an AFC framework based on
adaptive features and predefined analytical features for the
LPI radar waveform recognition. A feature fusion process
consists ofempirical VMD and EMD features, and analytical
features fromWVD,CWD, and wavelet analysis. Fusion
features composed ofadaptive features and other traditional
analytical features are further optimized for the LPI waveform
recognition. A CNN-based LPI waveform recognition
system processes the fusion feature. The proposed FALPINE
approach is evaluated with simulated LPI waveforms
introduced in the literature and the performance ofthe developed
LPI waveform recognition system is compared with
the state-of-the-art works. Experimental results show that
the proposed approach achieves higher PCC than other
works, which demonstrates the impressive performance of
the FALPINE approach. In the future, more investigations
will be performed concerning the adaptive features to determine
ifmore inherent features of the signals can be applied
in signal recognition. Although the neural networks, such as
CNN, can help extract features itself to some extent, the
well-designed features that can present the signals still play
an important role in the efficient recognition of LPI waveforms.
Moreover, the FALPINE approach is an open framework
that can also be extended to other types of signals in
other areas, such as communication signals.
ACKNOWLEDGMENT
The views and conclusions contained herein are those of
the authors and should not be interpreted as necessarily
representing the official policies or endorsements, either
expressed or implied, of the Air Force Research Laboratory
or the U.S. government.
REFERENCES
[1] P. E. Pace, Detecting and Classifying Low Probability of
Intercept Radar. Norwood, MA, USA: Artech House, 2009.
[2] Z. Tian, E. Blasch, W. Li, G. Chen, and X. Li,
" Performance evaluation of distributed compressed wideband
sensing for cognitive radio networks, " in Proc. 11th
Int. Conf. Inf. Fusion, 2008, pp. 1-8.
[3] Z. Wang, E. Blasch, G. Chen, D. Shen, X. Lin, and
K. Pham, " A low-cost, near-real-time two-UAS-based
UWB emitter monitoring system, " IEEE Aerosp. Electron.
Syst. Mag., vol. 30, no. 11, pp. 4-11, Nov. 2015,
doi: 10.1109/MAES.2015.140219.
[4] W.-Q. Wang, " Potential transmit beamforming schemes
for active LPI radars, " IEEE Aerosp. Electron. Syst.
Mag., vol. 32, no. 5, pp. 46-52, May 2017, doi: 10.1109/
MAES.2017.150259.
IEEE A&E SYSTEMS MAGAZINE
25
http://dx.doi.org/10.1109/MAES.2015.140219 http://dx.doi.org/10.1109/MAES.2017.150259 http://dx.doi.org/10.1109/MAES.2017.150259
IEEE - Aerospace and Electronic Systems - April 2023

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - April 2023
