A Novel Dual-Component Radar-Signal Modulation Recognition Method Based on CNN-ST

Abstract

Dual-component radar-signal modulation recognition is a challenging yet significant technique for electronic reconnaissance systems. To improve the lower recognition performance and the higher computational costs of the conventional methods, this paper presents a randomly overlapping dual-component radar-signal modulation recognition method based on a convolutional neural network–swin transformer (CNN-ST) under different signal-to-noise ratios (SNRs). To enhance the feature representation ability and decrease the loss of the detailed features of dual-component radar signals under different SNRs, the swin transformer is adopted and integrated into the designed CNN model. An inverted residual structure and lightweight depthwise convolutions are used to maintain the powerful representational ability. The results show that the dual-component radar-signal recognition accuracy of the proposed CNN-ST is up to 82.58% at −8 dB, which shows the better recognition performance of the CNN-ST over others. The dual-component radar-signal recognition accuracies under different SNRs are all more than 88%, which verified the fact that the CNN-ST achieves better recognition accuracy under different SNRs. This work offers essential guidance in enhancing dual-component radar signal recognition under different SNRs and in promoting actual applications.

1. Introduction

Efficient and high-accuracy radar-signal modulation recognition has emerged as a pivotal requirement in modern electronic warfare [1,2]. With the ever-increasing complicated electromagnetic environment, the interceptors are likely to be affected by the existing multiple potential emitters and result in an overlapping of the radar signals. As the first step for electronic reconnaissance, the precise detection of the signal modulation can provide promising evidence for the subsequent parameter estimation and jamming implementation [3]. Therefore, automatically recognizing the modulation of overlapping radar signals was of great significance, and has become an essential topic in signal processing.

Radar-signal modulation recognition techniques have been investigated in recent decades, and mainly contain radar signal feature extraction and modulation recognition [4,5]. Previous research mainly used the handcrafted feature to train a multi-class classifier of single-component radar signals [6]. However, the conventional extracted features in single-component radar signal recognition, such as high-order moments and cumulants [7] and power spectral density, as well as instantaneous frequency and phase, were inadequate in the circumstance of overlapping radar signals. Moreover, the potential hybrid category was rather random, and the computational features led to poor generalization performance. Because of the enormous advantages of deep learning, multi-class classification algorithms based on deep learning have been proposed for recognizing dual-component radar signals [8,9], which shed light on the capability of highly complicated feature representation [10]. Prior work has largely proved that the extracted representations through deep learning models were more robust and efficient than the handcrafted features of computer vision, which were also potentially applicable to the signal process.

The convolutional neural network (CNN) has been adopted for recognizing related tasks and achieved encouraging performance on radar-signal modulation recognition [11,12,13]. That is because it possessed a robust learning ability feature and higher classification performance compared with the traditional methods, especially for converting radar signals into time-frequency images and automatically extracting various feature details of images. In particular, ResNet [3,14], U-shaped network (U-Net) [15], Asymmetric Convolution Squeeze-and-Excitation [16], and their variants have witnessed remarkable success in radar-signal modulation recognition. ResNet improved the network performance by increasing depth and using residual connections, while it took a lot of computational resources and may overfit. U-Net reduced the size of the feature map by using encoders to superimpose convolution and pooling operations, while it reduced the efficiency and lost spatial information. All of the architecture benefited from a skip connection which incorporated semantic features with fine-grained features. However, the radar-signal modulation recognition method based on CNN makes it difficult to effectively mine the characteristics of dual-component radar signals.

Recently, the transformer has become a deep learning architecture specifically designed to work with sequential data. The main functions include sequence modeling and understanding, efficient parallel computing, and long-distance dependency modeling. Therefore, it has also been utilized in computer vision [17]. In [18], a dynamic transformer was proposed for configuring some tokens, while it would require higher computation requirements. Following this structure, a swin transformer was proposed [19], which computed self-attention within non-overlapping local windows and leveraged the shifted window partition to build connections among the windows of each preceding layer. Although the swin transformer model has made great progress in image classification [20], it has not been utilized in radar-signal modulation recognition.

As can be known from the above references, some methods of radar-signal modulation recognition at the same signal-to-noise ratio (SNR) have been proposed. However, the dual-component radar signal recognition method under different SNRs has not been reported to date. A majority of the proposed radar-signal modulation recognition methods cannot accurately classify dual-component radar signals under different SNRs. The recognition accuracy of radar-signal modulation recognition was relatively lower, particularly under a lower SNR. There is an urgent need to propose an efficient method for recognizing dual-component radar signals under different SNRs and to improve the recognition accuracy. Moreover, for promoting the actual applications in modern electronic warfare, the proposed method would also be used to recognize multi-component radar signals under different SNRs by training the network parameters and designing a multi-label classifier.

Therefore, to solve the above deficiencies, this paper presents a novel network model called CNN–swin transformer (CNN-ST) for recognizing randomly overlapping dual-component radar signals under different SNRs, which integrates the swin transformer into the designed CNN model and improves the recognition accuracy. The important contributions include the following: (1) a novel randomly overlapping dual-component radar signal recognition method under different SNRs based on CNN-ST is first presented; (2) it is inspired by the swin transformer with powerful global modeling capability, which integrates into the CNN; (3) for extracting detailed radar signal features, an inverted residual structure and lightweight depthwise convolutions are adopted; and (4) the proposed CNN-ST model achieves better recognition accuracy under different SNRs.

The related works of radar-signal modulation recognition are reviewed in Section 2; the dual-component radar signals and data preprocessing are explained in detail in Section 3; the novel network CNN-ST for classifying dual-component radar signals under different SNRs is fabricated in Section 4; The influences of dual-component radar signals under different SNRs on the recognition performance are fully investigated in Section 5; and some important conclusions are summarized in Section 6.

2. Related Work

CNN-based methods for radar-signal modulation recognition are first summarized, and then an overview of recent applications of transformers in computer vision is provided, especially image recognition.

2.1. Radar Signal Modulation-Recognition Methods Based on CNN

Owing to the rapid development of CNN in recent decades, the recognition methods of radar signals based on CNN have been proposed [21,22,23]. In [24], a CNN-based radar signal modulation-recognition technique was proposed, which offered significant improvement over the recent radar-signal modulation recognition technology. In [25], a generic training strategy improved the prediction performance of radar-signal modulation recognition. In [26], a radar emitter signal classification algorithm based on CNN was presented, and the simulation results indicated superior recognition accuracy. In [27], a feature fusion algorithm for radar signal automatic modulation recognition using CNN was proposed, and the simulation results revealed superior accuracy. In [28], a cost-efficient CNN for radar-signal modulation recognition was proposed, and the results demonstrated excellent generalization ability. In [29], a CNN for low-complexity and robust modulation recognition was proposed, which achieved better recognition performance. In [30], a CNN-LSTM to address automatic modulation recognition was proposed, and experimental results demonstrated superior recognition performance. In [31], a multi-class learning framework based on CNN under the same SNR was proposed, and the results demonstrated superior performance over others. In [32], an efficient deep CNN with feature fusion at low SNR was presented, and recognition performance was up to 84.38% at −12 dB. Although the CNN-based network has demonstrated better robustness in the single-component radar signal, it cannot be applied completely to dual-component radar signals. Because of the existing difference in time-frequency images and features, the CNN-based network may demonstrate poor recognition performance for dual-component radar signals.

2.2. Vision Transformer

The transformer block is composed of a multi-head self-attention (MSA), multiple-layer perceptron (MLP), and layer normalization (LN) module [33]. Recent research has been trying to investigate the benefits of transformers in computer vision. In [17], the transformer first applied to image classification for replacing the CNN model was proposed. Chen et al. [34] presented an ImageNet benchmark, and maximally excavated the capability of the transformer. Kolesnikov et al. [35] conducted image recognition with a transformer structure. Wu et al. [36] dynamically extracted visual tokens, and used visual transformers to perform the visual tokens. Jiang et al. [37] conducted the first pilot study in building a generative adversarial network model using only pure transformer-based architecture. Yuan et al. [38] designed a novel vision transformer, which reduced the parameter count. Touvron et al. [39] produced the transformer structure and recognition accuracy achieved 83.1% on the ImageNet dataset. Wang et al. [40] introduced the pyramid vision transformer. Liu et al. [19] designed a swin transformer based on the shifted-window strategy. Following that, Zheng et al. [20] proposed the swin transformer and multi-layer perceptron (ST-MLP) to classify the strawberry appearance, and the results demonstrated superior recognition performance.

The CNN-based methods have improved radar-signal modulation recognition performance. Inspired by these excellent works, we adopt the swin transformer blocks to obtain the global context information for the CNN model. In this paper, a new framework including CNN and the swin transformer (CNN-ST) is proposed. CNN-ST first applies the swin transformer to the radar-signal modulation recognition field, which increases the recognition performance, especially for dual-component radar-signal modulation recognition under different SNRs.

3. Signal Model and Time-Frequency Transformation

The dual-component radar signals under different SNRs are adopted, and the time-frequency transformation technology is implemented to obtain the detailed features and classify the radar signal types.

3.1. Dual-Component Radar Signals under Different SNRs

For recognizing intra-pulse modulation radar signals, even quadratic frequency modulation (EQFM), linear frequency modulation (LFM), normal signal (NS), binary phase-shift keying (BPSK), binary frequency-shift keying (2FSK), sinusoidal frequency modulation (SFM), FRANK, and four frequency-shift keying (4FSK) are adopted in this paper. The Gaussian white noise (GWN) is utilized to disturb those signals, for simulating the realistic application environment in the actual battlefield. The received dual-component radar signal is written as [31]

$$\mathrm{y}\left(t\right)={\sum }_{i=1}^k{A}_{i}rect\left(t/T_i\right)e^{j(2\pi f_{c_i}t+{\varnothing }_i\left(t\right)+{\varnothing }_{0_i})}+n\left(t\right) (k=1, 2)$$

(1)

where $n\left(t\right)$ is i-th noise component; k represents signal number; and $A_i$, $f_{c_i}$, ${\varnothing }_{0_i}$, and $T_i,$ denote amplitude, carrier frequency, initial phase, and pulse width of the signals, respectively.

This paper is mainly intended to classify dual-component signals under different SNRs, wherein dual-component signals are obtained by randomly overlapping two single-component signals, and thus k is set as 2. Given the received overlapping-signal sample dataset $D=\left\{(x_i,t_i)|1\le i\le N\right\}$ including $N$ samples with eight types of signals, the i-th sample is represented as $x_i$, and $t_{ij}=[t_{i1},t_{i2},\cdots ,t_{i8}]$ represents the true label vector $x_{i,}$ such that sample $x$ with a label vector $t=[0,1,0,0,0,0,1,0]$ indicates the presence of radar signals at the second and seventh positions, which are overlapped in $x$.

3.2. Time-Frequency Transformation

The common time-frequency transformation methods include the short-time Fourier Transform (STFT), the Wigner–Ville distribution (WVD), the Choi–Williams distribution (CWD), the pseudo Wigner–Ville distribution (PWVD), the smoothed pseudo Wigner–Ville distribution (SPWVD), and so on [41,42]. The SPWVD can express the radar signals in detail and effectively prevent cross-interference over other methods, and is adopted in this paper [43]. The SPWVD transformation is written as

$$S\left(t,f\right)=\iint x(t-\mu +\tau /2)·x^{*}\left(t-\nu -\tau /2\right)\bullet h\left(\tau \right)g\left(\mu \right)e^{-j2\pi f\tau }d\mu d\tau$$

(2)

where $\ast$ is the complex conjugate; $h\left(\tau \right)$ and $g\left(\mu \right)$ represent window functions; and $x\left(t\right)$ is the analytic signal $y\left(t\right)$.

Through SPWVD time-frequency transformation, 2184 types of dual-component radar signals under different SNRs are acquired. To save space, Figure 1 only illustrates four types of time-frequency images (TFIs) of dual-component radar signals under different SNRs. Therein, Figure 1a represents the overlap of EQFM and SFM at −2 dB; Figure 1b represents the overlap of EQFM at −2 dB and SFM at 12 dB; Figure 1c represents the overlap of EQFM at 12 dB and SFM at −2 dB; and Figure 1d represents the overlap of EQFM and SFM at 12 dB.

Figure 1 demonstrates that SNR exerts an essential effect on TFIs. The higher SNR results in obtaining clearer TFIs. It means that the more detailed features are extracted for the clearer TFIs, and better classification performance can be achieved in the following analyses.

4. Network Model CNN-ST

The network model CNN-ST is proposed in this paper, for a detailed extraction of the features of signals and for accurately recognizing the signal types. Firstly, the model framework of CNN-ST is described, and the motivation for choosing the CNN-ST model is explained. Next, the network model CNN-ST is designed. Finally, the swin transformer is utilized for enhancing the feature representation ability and reducing the loss of detailed information of dual-component radar signals under different SNRs.

4.1. Model Framework

To accurately classify randomly overlapping dual-component radar signals under different SNRs, this paper presents a novel network model CNN-ST, which possesses robust feature-extraction capability. To improve the focusing and information expression ability of dual-component radar signals under different SNRs, the network architecture based on CNN and the swin transformer is designed for extracting in detail the radar signal features. Figure 2 shows the network architecture of the CNN-ST model.

Figure 2 demonstrates that TFIs obtained by time-frequency transformation are first entered into the designed CNN-ST model, the radar signal features are then extracted in detail, and the signal types are finally classified. The CNN-ST model consists of the first convolutional layer, an average pooling layer, three bottleneck convolution blocks, four swin transformer blocks, a second convolutional layer, a nonlinearity layer Hard swish (H-swish), a third convolutional layer, an average pooling layer, an 8-dimension fully connected layer, and a multi-label classifier. Within this, softmax is the activation function of the multi-label classifier. The bottleneck convolution block aims to increase expressive ability. The expansion ratio changes the output channel number. The swin transformer block possesses powerful global modeling capability. For the deepening of the network, the cost of applying the nonlinear activation function H-swish could be reduced, and could largely decrease the number of parameters.

Table 1. CNN-ST parameters.

Operators	Layers	Size	Expansion Ratio	Stride	Output
Convolution	1	3 × 3	-	2	224 × 224 × 32
Average pooling	1	2 × 2	-	2	112 × 112 × 32
Bottleneck	1	-	1	1	112 × 112 × 16
Bottleneck	1	-	6	2	56 × 56 × 24
Bottleneck	1	-	1	2	28 × 28 × 32
Swin transformer	2	-	-	-	14 × 14 × 64
Swin transformer	2	-	-	-	14 × 14 × 96
Convolution	1	1 × 1	-	1	7 × 7 × 160
H-swish	1	-	-	-	7 × 7 × 160
Convolution	1	1 × 1	-	1	7 × 7 × 1280
Average pooling	1	7 × 7	-	1	1 × 1 × 1280
Fully connected	1	-	-	-	1 × 1 × 8

lists the CNN-ST parameters.

4.2. CNN Structure

To extract in detail the radar signal features, the CNN model is designed. Additionally, to maintain the representational ability, the designed CNN removes the nonlinearity in the narrow layers. The structure of CNN in CNN-ST is illustrated in detail in Figure 3.

Figure 3 demonstrates that the designed CNN includes the following: the initial fully convolutional layer; an average pooling layer; two bottleneck convolutional blocks, where the expansion ratio is 1, and the stride length is 1; and a bottleneck convolution block, where the expansion ratio is 6, and the stride length is 2. Within this, the bottleneck convolution block with a stride length of 1 contains the following: a convolutional layer, a batch normalization, an activation function H-swish, an average pooling layer in which a kernel size is 2 × 2 and a stride length is 2, a convolutional layer in which the kernel size is 1 × 1 and the stride length is 1, an activation function H-swish, a depthwise convolutional layer in which t the kernel size is 3 × 3 and the stride length is 2, an activation function ReLU6, a convolutional layer in which the kernel size is 1 × 1 and the stride length is 2, a batch normalization, and a dropout. Finally, a residual connection is utilized. The expansion ratio of the bottleneck convolution block is set as 1. The bottleneck convolution block with a stride length of 2 is the same as the one with a stride length of 1, except that there is no residual connection.

Depthwise separable convolutions are an essential component of many deep CNN architectures, and are adopted in this paper. A nonlinearity swish was introduced in [44,45], which was utilized as a replacement for ReLU and improved the classification performance of the dual-component radar signals under different SNRs. The nonlinearity is written as

$$\mathrm{s}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{h}x=x·\mathsf{\sigma }\left(x\right)$$

(3)

Nonlinearity brings in the non-zero cost and improves recognition accuracy. This is because it is much more expensive to calculate the sigmoid functions on mobile devices. The sigmoid function $\frac{ReLU6\left(x+3\right)}{6}$ is the same as in [46]. The small difference is that ReLU6 is used. Recently, a similar H-swish was proposed in [47], and the hard version of H-swish is defined as

$$H-\mathrm{s}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{h}\left[x\right]=x\frac{ReLU6\left(x+3\right)}{6}$$

(4)

4.3. Swin Transformer Block

To improve the feature representation performance and decrease the loss of detailed information of dual-component radar signals under different SNRs, this paper integrates the swin transformer into the designed CNN model, which possesses powerful global modeling capability [48]. Figure 4 illustrates the swin transformer structure in detail.

Figure 4 demonstrates that the swin transformer structure mainly includes LN, W-MSA, MLP, MSA, and SW-MSA, and a residual connection is used after each module, wherein the LN layer is added before each W-MSA and SW-MSA. Therefore, the output ${\mathrm{s}}^l$ of the l-th layer in the swin transformer is written as

$${\widehat{s}}^l=W-MSA\left(LN\left(s^{l-1}\right)\right)+s^{l-1}$$

(5)

$$s^l=MLP\left(LN\left({\widehat{s}}^l\right)\right)+{\widehat{s}}^l$$

(6)

where $s^l$ and ${\widehat{s}}^l$ represent the outputs of W-MSA and MLP. In W-MSA, the input is a patches sequence, $s^{l-1}\in R^{L\times D}$.

The outputs of SW-MSA and MLP modules are expressed as

$${\widehat{s}}^{l+1}=SW-MSA\left(LN\left(s^l\right)\right)+s^l$$

(7)

$$s^{l+1}=MLP\left(LN\left({\widehat{s}}^{l+1}\right)\right)+{\widehat{s}}^{l+1}$$

(8)

where ${\widehat{s}}^{l+1}$ and $s^{l+1}$ refer to the outputs of SW-MSA and MLP, respectively. The self-attention computation in W-MSA and SW-MSA is defined as

$$\mathrm{A}\mathrm{t}\mathrm{t}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\left(z_l\right)=SoftMax\left(Q{K}^T/\sqrt{d}+B\right)V$$

(9)

$$Q=s^l{W}_Q, K=s^l{W}_K, V=s^l{W}_V$$

(10)

where $W_Q$, $W_K$, and $W_V\in R^{D\times d}$ represent the three projection matrices’ parameters; Q, K, and V$\in R^{L\times d}$ represent the query, key, and value matrices, respectively; d denotes the dimension of query or key; and B$\in R^{L\times L}$ represents the relative position bias.

5. Experiment, Results, and Analysis

The recognized performance of CNN-ST for randomly overlapping dual-component radar signals under different SNR is evaluated. To demonstrate better classification performance over others, CNN-DQLN and CNN-Softmax in [49] are also utilized for conducting comparative analyses.

5.1. Datasets and Training Parameters

The typical eight types of signals that include 2FSK, BPSK, 4FSK, FRANK, NS, EQFM, LFM, and SFM are adopted in this paper.

Table 2. Parameters of various radar signals.

Types	Parameters	Ranges
2FSK	$\mathrm{Carrier} \mathrm{frequency} f_1$$, f_2$	0.01 to 0.46
	$\mathrm{Bandwidth} ∆f$	N/32 to N/8
BPSK	Barker codes	[5, 7, 9, 13]
	$\mathrm{Carrier} \mathrm{frequency} f_0$	0.1 to 0.4
	$T_s$	N/32 to N/16
4FSK	$\mathrm{Carrier} \mathrm{frequency} f_1$$\mathrm{to} f_4$	0.1 to 0.4
	$T_s$	N/32 to N/8
EQFM	$\mathrm{Carrier} \mathrm{frequency} f_1$$, f_2$	0.05 to 0.4
	$\mathrm{Bandwidth} ∆f$	0.05 to 0.3
FRANK	$\mathrm{Carrier} \mathrm{frequency} f_0$	0.1 to 0.4
	$T_s$	N/100 to N/50
	Phase number M	[4, 5, 6, 7]
LFM	$\mathrm{Initial} \mathrm{frequency} f_c$	0.01 to 0.45
	$\mathrm{Bandwidth} ∆f$	0.05 to 0.4
NS	$\mathrm{Carrier} \mathrm{frequency} f_0$	0.1 to 0.4
SFM	$\mathrm{Carrier} \mathrm{frequency} f_0$	0.05 to 0.15
	$\mathrm{Bandwidth} ∆f$	0.05 to 0.35

lists the detailed radar signal parameters. Note that the training, validation, and testing datasets contain randomly overlapping dual-component radar signals under different SNRs, which are preprocessed by using SPWVD transformation. To conduct the comparative analyses, the SNR ranges from −12 dB to 10 dB, and every 2 dB is selected. The number of each class of dual-component radar signals under different SNRs is set as 300, and a total of 1,419,600 samples are obtained in the training dataset. For the validation dataset, the number of each class of dual-component radar signals under different SNRs is set as 100, and a total of 473,200 samples were obtained. For the testing dataset, the number of each class of dual-component radar signals under different SNRs is set as 30, and a total of 5160 samples is obtained. The types and parameters of radar signals are the same as those in Reference [49].

The deep learning framework used in the analyses is Pytorch 1.11 and Python 3.9. Based on the computational complexity and recognition performance, the training parameters are set as follows: the batch size is 128; epoch is 100; epsilon is 0.001; momentum is 0.01; initial learning rate is set as 0.01; learning decay rate is 0.1 every 10 epochs; dropout is 0.8; weight decay is set as 1 × 10⁻⁵; and batch normalization with average decay is set as 0.999.

5.2. Results, Discussions, and Analysis

To explicitly explore the recognition accuracy of CNN-ST and conduct the comparative analyses, Figure 5 illustrates the overall recognition accuracies of CNN-ST, CNN-DQLN, and CNN-Softmax as a function of SNR. Therein, the testing dataset used in this subsection represents the dual-component radar signals under the same SNR, and the SNR used ranges from −10 dB to 10 dB.

Figure 5 demonstrates that an increase in SNR leads first to increasing overall recognition accuracies of CNN-ST, CNN-DQLN, and CNN-Softmax, and then leveling off. Therein, the increasing trend for CNN-ST at the lower SNR is faster than CNN-DQLN and CNN-Softmax, and achieves better recognition performance over others. The recognition performance of CNN-ST, CNN-DQLN, and CNN-Softmax are 50.88%, 37.37%, and 26.49%, at −10 dB, respectively. The recognition performance of CNN-ST is 13.51% higher than that of CNN-DQLN, and 24.39% higher than that of CNN-Softmax. The recognition accuracy of CNN-ST is 82.58% at −8 dB, while it is 74.74% and 57.02% for CNN-DQLN and CNN-Softmax, respectively. Moreover, the recognition accuracies can reach 100%, 98.77%, and 95.44% at 4 dB for CNN-ST, CNN-DQLN, and CNN-Softmax, respectively. Therefore, CNN-ST achieves superior recognition accuracy compared to CNN-DQLN and CNN-Softmax. The reason is that the adopted swin transformer in the CNN-ST can extract the more detailed features of the radar signals.

To clearly express the recognition performance of each type of radar signal, Figure 6 illustrates the variation in the recognition accuracy of eight types of signals with SNR.

Figure 6 demonstrates that the recognition accuracies of eight types of signals increase with increasing SNR, while the increasing trend levels off at the higher SNR. It means that the lower SNR exerts an important effect on recognition accuracy. Moreover, 2FSK and LFM demonstrate a superior recognition performance over the others. That is because the clearer the TFIs for 2FSK and LFM are, the easier the detailed features are extracted, and the better the recognition performance which can be obtained. The recognition accuracies of 2FSK, EQFM, 4FSK, BPSK, LFM, SFM, NS, and FRANK are 90.83%, 82.14%, 77.98%, 67.62%, 80.71%, 75.95%, 73.45%, and 71.91%, at SNR of −10 dB, respectively. Moreover, the recognition accuracies of all types of radar signals are basically up to 1 at larger than 0 dB. Therefore, the designed CNN-ST model demonstrates better recognition performance at the same SNR.

To investigate the influence of different SNRs on the recognition accuracy of dual-component signals, Figure 7 demonstrates the recognition accuracy as a function of different SNRs. Therein, the SNR is −12 dB to 4 dB; the X and Y axes represent the SNR under different radar signals; and the Z axis represents the classification accuracy. The dual-component radar signals include the following: overlapping 2FSK and 4FSK (2FSK-4FSK), overlapping 2FSK and BPSK (2FSK-BPSK), overlapping 2FSK and EQFM (2FSK-EQFM), overlapping 2FSK and FRANK (2FSK-FRANK), overlapping 2FSK and LFM (2FSK-LFM), overlapping 2FSK and NS (2FSK-NS), overlapping 2FSK and SFM (2FSK-SFM), overlapping LFM and 4FSK (LFM-4FSK), overlapping LFM and BPSK (LFM-BPSK), overlapping LFM and EQFM (LFM-EQFM), overlapping LFM and FRANK (LFM-FRANK), and overlapping LFM and SFM (LFM-SFM).

Figure 7 shows that the recognition accuracies of twelve types of randomly overlapping dual-component radar signals increase as different SNRs increase. When the different SNRs of the dual-component radar signal are higher than −2 dB, the increasing trend gradually levels off, and the recognition accuracies are up to 1. It means that the different SNRs of the dual-component radar signal exert less effect on the recognition accuracy, especially at the higher SNR. That is because the higher the different SNR is, the less the noise interferes with the TFIs is, and the easier it is for the effective features of the TFIs to be extracted. The recognition accuracies of 2FSK-4FSK, 2FSK-BPSK, 2FSK-EQFM, 2FSK-FRANK, 2FSK-LFM, 2FSK-NS, 2FSK-SFM, LFM-4FSK, LFM-BPSK, LFM-EQFM, LFM-FRANK, and LFM-SFM are up to 76.25%, 63.75%, 82.625%, 75.625%, 70.625%, 72%, 67.5%, 71.5%, 60.625%, 78.25%, 73.75%, and 70.625%, at different SNRs of −12 dB, respectively. When the SNR of 2FSK is −12 dB and of 4FSK it is 4 dB, the recognition accuracy of 2FSK-4FSK is up to 88.56%. For 2FSK at 4 dB and 4FSK at −6 dB, the recognition accuracy of 2FSK-4FSK is 1. When SNR of 2FSK is −4 dB and 4FSK is −2 dB, the recognition accuracy of 2FSK-4FSK is up to 1. For 2FSK at −6 dB and 4FSK at −8 dB, the recognition accuracy of 2FSK-4FSK is 93.125%. When SNR of 2FSK is −12 dB and BPSK is 4 dB, the recognition accuracy of 2FSK-BPSK is up to 81.875%. For 2FSK at 4 dB and BPSK at −6 dB, the recognition accuracy of 2FSK-BPSK is 98.75%. When SNR of 2FSK is −4 dB and BPSK is −2 dB, the recognition accuracy of 2FSK-BPSK is up to 96.125%. For 2FSK at −6 dB and BPSK at −8 dB, the recognition accuracy of 2FSK-BPSK is 85.25%. When SNR of 2FSK is −12 dB and EQFM is 4 dB, the recognition accuracy of 2FSK-EQFM is up to 92.625%. For 2FSK at 4 dB and EQFM at −6 dB, the recognition accuracy of 2FSK-EQFM is 1. When SNR of 2FSK is −4 dB and EQFM is −2 dB, the recognition accuracy of 2FSK-EQFM is up to 1. For 2FSK at −6 dB and EQFM at −8 dB, the recognition accuracy of 2FSK-EQFM is 96.5%. Note that the recognition accuracies of randomly overlapping dual-component signals are all more than 90% at −2 dB, which demonstrates better recognition performance. The reason is the powerful feature extraction and classification ability. Therefore, this work offers important experimental guidance in further enhancing recognition performance under different SNRs and promoting the actual applications.

The floating point of operations (FLOPs), parameters, and time are adopted to investigate computational complexity.

Table 3. Computational complexity of two methods.

Parameter	CNN-ST	CNN-DQLN
FLOPs (G)	3.7	5.2
Parameters (M)	20.6	34.9
Time (ms)	33	59

lists the computational complexity of CNN-ST and CNN-DQLN.

As can be seen from Table 3, the FLOPs, parameters, and the time of CNN-ST are all lower than CNN-DQLN. Therefore, the proposed CNN-ST demonstrates higher computational accuracy and lower computational complexity over others, and possesses enough computational novelty.

6. Conclusions

This paper presented a novel randomly overlapped dual-component radar signal recognition method under different SNRs based on a convolutional neural network–swin transformer (CNN-ST), for improving recognition performance. The overall model framework was first designed, and the swin transformer was subsequently adopted and integrated into the CNN model. An inverted residual structure and lightweight depthwise convolutions were used to maintain the powerful representational ability. The influences of the different SNRs on the recognition performance of dual-component radar signals were experimentally investigated. The results demonstrated that recognition accuracies of CNN-ST were up to 82.58% at −8 dB, which showed better recognition performance over others. The recognition accuracies of randomly overlapping dual-component radar signals under different SNRs were all more than 90% at −2 dB, which verified the superior recognition performance of the proposed CNN-ST model. The recognition accuracies of 2FSK-4FSK, 2FSK-BPSK, 2FSK-EQFM, 2FSK-FRANK, 2FSK-LFM, 2FSK-NS, 2FSK-SFM, LFM-4FSK, LFM-BPSK, LFM-EQFM, LFM-FRANK, and LFM-SFM are up to 76.25%, 63.75%, 82.625%, 75.625%, 70.625%, 72%, 67.5%, 71.5%, 60.625%, 78.25%, 73.75%, and 70.625% at SNR of −12 dB, respectively. This work provided essential guidance in enhancing recognition performance under different SNRs and promoting the actual applications.

For considering the more complex scenarios with randomly overlapping radar signals, future research will focus on classifying multi-component signals under different SNRs and promoting the actual applications in modern electronic warfare.