IEEE - Aerospace and Electronic Systems - April 2023 - 21

Huang et al.
Figure 8.
CNN structure for the LPI waveform classification.
abovementioned choice. The details of these operations
are explained as follows:
Reshaping. Typically, there is a need to manipulate
the shape ofthe feature. For example, if the adaptive
features from VMD and EMD are 10 vectors with
the shape of 14096, the adaptive features can be
reshaped to a matrix with the shape of51280.
Squaring. In order to enhance the feature in the feature
fusion method, the squaring operation squares
the absolute value of data samples.
Downsampling. The downsampling operates on features
that do not show excellent classification performance.
For example, the CWD feature with a
shape of 20482048 can be downsampled to
51280, so that it can be unified with the exemplary
adaptive features previously.
Maxpooling. The maxpooling operation is a special
downsampling process, which uses the maximum
value ofthe subregions ofthe original data samples as
the downsampled value. In this project, this operation
is done to in part to help overfitting by providing an
abstracted form of the original representations, such
as the WVD feature. In addition, it reduces the
computational cost by reducing the number ofparameters
to learn in the CNNmodel.
For the CNN model, it is straightforward that the four
types of features can be combined by simply stacking the
matrix of the features, as the CNN model can easily handle
the stacking of the two-dimensional matrix. Figure 7
shows the procedure of the feature fusion with the four
types of features as images.
it has been shown with extensive studies that the CNN exhibits
excellent performance in classification problems for
images, such as handwriting recognition, face recognition,
object detection and classification [24]. Therefore, CNN is
selected to classify the fused feature from the feature fusion
approach of the LPI waveforms and thus achieve the LPI
waveform recognition.
BASIC STRUCTURE OF CNN
In this work, the basic structure ofCNN [24] is employed to
develop the proper CNN structure for the proposed LPI waveform
recognition system, as shown in Figure 8, which can be
described by a sequence of operations as Input-Conv-ReLUPooling-Flatten-FC-Output,
where Conv represents the convolution
layer, ReLU is the rectified linear unit, Pooling is the
maxpooling layer, and FC denotes the fully connected layer.
The hyperparameters of the CNN are designed based on the
basic CNN structure including the convolution filter size, the
number of convolution layers and FC layers, and the number
ofneurons in each layer. In addition, the regularization techniques,
such as dropout layers, are also used for determining the
optimal structure for the LPI waveform recognition problem
based on numerousMonte Carlo simulations.
CNN MODEL DESIGN
In the initial step of the CNN design, the number ofconvolution
layers and dense layers are determined with various simulations.
Three CNN structures are used in the simulations, as
CNN-BASED LPI WAVEFORM RECOGNITION SYSTEM
In recent years, numerous advances occur in various classification
tasks based on the improved neural network architectures,
algorithms, and optimization techniques collectively
known as deep learning (DL). For the two dimension images,
APRIL 2023
Table 2.
CNN Models with Different Structures
Index
Structure
1
2
3
IEEE A&E SYSTEMS MAGAZINE
1CNN+2DNN
3CNN+2DNN
5CNN+2DNN
Network depth
three layers
five layers
seven layers
21
IEEE - Aerospace and Electronic Systems - April 2023

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