CN115062658B - Modulation type recognition method for overlapping radar signals based on adaptive threshold network - Google Patents

Abstract

The invention discloses an overlapping radar signal modulation type identification method based on a self-adaptive threshold network, which comprises the following steps: s1, simulating an original radar signal and modulation parameters; s2, overlapping original radar signals; s3, extracting time-frequency domain features; s4, constructing a basic module; s5, constructing Inception modules; s6, constructing a self-adaptive threshold module; s7, constructing a probability module; s8, forming an SE-INCEPATNET network; s9, training an SE-INCEPATNET network; and S10, identifying the modulation type of the overlapped radar signals by adopting the SE-INCEPATNET network after training. According to the invention, the time-frequency analysis method is utilized to extract the characteristics of the overlapped radar signals, the depth convolution network Inception is based on the characteristics of different receptive fields, the SE module is used for reducing noise influence, the self-adaptive threshold module is constructed to solve the problem of difficult threshold setting in multi-classification tasks, and the recognition accuracy under the condition of low signal-to-noise ratio is improved while the recognition of the overlapped radar signals is realized.

Description

Modulation type recognition method for overlapping radar signals based on adaptive threshold network

Technical Field

The present invention belongs to the technical field of radar signal processing, and is particularly suitable for radar emitter signal recognition in electronic countermeasures, and more particularly relates to a method for identifying the modulation type of overlapping radar signals based on an adaptive threshold network.

Background technique

Modulation recognition is widely used in threat level assessment and combat strategy formulation.

As the pulse flow density increases, the radar receiver will receive multiple radar signals at the same time. These signals overlap not only in the time domain, but also in the frequency domain. Under low signal-to-noise ratio, overlap and noise will cause some features of the signal to become chaotic, thereby increasing the difficulty of modulation recognition. For the overlap problem, the instantaneous characteristics and statistics of radar signals are often used for modulation recognition, reflecting the inherent characteristics of the signal. The literature "Lu Mingquan, Xiao Xianci, and Li Lemin, Modeling-based features extraction of multiple signals for modulation recognition, in ICSP’98.1998Fourth International Conference on Signal Processing (Cat.No.98TH8344), 1998, vol.2, pp.1385–1388 vol.2" uses the instantaneous frequency and bandwidth of overlapping radar signals to achieve modulation recognition, but it is not suitable for signals with overlapping frequency domains. The literature "Kuang-dai Li, Li-liGuo, Rong Shi, and Dan Wu, Modulation recognition method based on high ordercyclic cumulants for time-frequency overlapped two-signal in the single-channel, in 2008 Congress on Image and Signal Processing, 2008, vol.5, pp.474–478." uses cyclic statistical features and minimum error criterion to identify overlapping signals. The recognition accuracy can only reach 90% when the signal-to-noise ratio is greater than 10dB.

For modulation recognition under low signal-to-noise ratio, neural networks can mine the robustness features of signals by learning from a large amount of data, so they are usually used for modulation recognition. The literature "Shunjun Wei, Qizhe Qu, Hao Su, Mou Wang, Jun Shi, and Xiaojun Hao, Intra-pulse modulation radar signal recognitionbased on cldn network, IET Radar, Sonar & Navigation, vol. 14, no. 6, pp. 803–810, 2020." combines CNN, Long Short-Term Memory (LSTM), and deep neural network (DNN) to build a recognition network with an accuracy of 90% at -6dB. However, the above literature only recognizes a single radar signal, and there is still a lack of relevant research on the recognition of overlapping radar signals under low signal-to-noise ratio. The paper "Yehan Ren, Weibo Huo, Jifang Pei, Yulin Huang, and Jianyu Yang, Automatic modulation recognition for overlapping radar signals based on multi-domain se-resnext, in 2021IEEE Radar Conference (Radar Conf21), 2021, pp. 1–6." proposed a network for single-label classification to identify overlapping signals, but the flexibility of the single-label classification network is poor. Therefore, the recognition of overlapping radar signals under low signal-to-noise ratio still needs further research.

Summary of the invention

The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for identifying the modulation type of overlapping radar signals based on an adaptive threshold network, which can improve the recognition accuracy under low signal-to-noise ratio conditions while realizing overlapping radar signal recognition.

The objective of the present invention is achieved through the following technical scheme: a method for identifying the modulation type of overlapping radar signals based on an adaptive threshold network, comprising the following steps:

S1, simulate original radar signal and modulation parameters;

S2, overlapping the simulated original radar signals to generate overlapping radar signals;

S3, extracting the time-frequency domain features of the overlapping radar signals, and representing them in the form of a time-frequency diagram, where the horizontal axis represents time and the vertical axis represents frequency;

S4, build basic modules;

S5. Construct Inception module: The Inception module contains multiple basic modules with different convolution kernel sizes, which are used to extract information of different dimensions in the time-frequency graph and combine them together;

S6, build an adaptive threshold module: including the sequential fully connected layer, ReLU activation function, fully connected layer and Sigmoid activation; take the output of the Inception module in step S5 as input to obtain the corresponding threshold;

S7, build probability module: use the structure of sequentially connected fully connected layers, ReLU activation function and Sigmoid activation function, take the output of Inception module in step S5 as input, and get the posterior probability;

S8. The submodules constructed in steps S5 to S7 are combined into a SE-IncepAtNet network, specifically: the probability module and the adaptive threshold module are connected in parallel, and then cascaded after the Inception module to form a SE-IncepAtNet network; the input is the time-frequency diagram processed by S3, and the output is the classification result;

S9, training the SE-IncepAtNet network, specifically: inputting the time-frequency diagram of the overlapping radar signal into the SE-IncepAtNet network for forward propagation, and calculating the cost function value; using the gradient descent-based backward propagation algorithm to update the parameters of the SE-IncepAtNet network; iterating the backward propagation process until the cost function converges, thereby obtaining the trained SE-IncepAtNet network;

S10, using the SE-IncepAtNet network trained in step S9 to identify the modulation type of the overlapping radar signal.

Furthermore, the specific method of extracting time-frequency features in step S3 is: using FSST algorithm to extract time-frequency features, and the FSST calculation formula is:

Where g(0) represents the value of the sliding window function g(t) at time 0, δ() is the Dirac impulse function; V _f (η,t) represents the signal after STFT of the overlapping radar signal; ω represents the frequency after FSST redistribution, η represents the frequency before FSST redistribution, and t represents time. Represents the entire real number domain; /> is the instantaneous frequency, defined as:

Furthermore, the basic module includes sequentially connected convolutional layers, batch normalization layers BN, ReLU activation layers, maximum pooling layers and squeeze excitation modules SE; wherein SE is used for denoising, including sequentially connected global average pooling layers GAP, fully connected layers FC, ReLU activation function layers, fully connected layers FC, and Sigmoid activation function layers.

Furthermore, the detailed processing process of step S5 is as follows:

S5.1. The time-frequency graph is simultaneously input into four different basic modules for processing to obtain features of different dimensions. The convolution kernel sizes of the four basic modules are 3×7, 7×3, 5×5 and 3×3 respectively.

S5.2, input the output of each basic module in step S5.1 into the basic module with a convolution kernel size of 3×3;

S5.3, merging the outputs of the four basic modules in step S5.2;

S5.4. Input the fused output of step S5.3 into four basic modules with a convolution kernel size of 3×3 in sequence to obtain the output result.

Furthermore, the step S9 specifically includes the following steps:

S9.1, forward propagation;

S9.2. Calculate the cost function value, using the binary cross entropy loss function as the cost function. The calculation method is:

Where BCE is the binary cross entropy loss function, prob∈R ^1×n is the output probability, lab∈R ^1×n is the true value of the overlapping signal in the form of one-hot encoding, thre∈R ^1×n is the threshold, and n is the total number of modulation types;

S9.3. Update the network parameters based on the gradient descent back-propagation algorithm.

The beneficial effects of the present invention are as follows: the present invention aims at the problem of overlapping radar signal recognition, which is less studied at present, and proposes a recognition method that adapts to complex electromagnetic environment and high-density pulse flow conditions. The overlapping radar signal features are extracted by time-frequency analysis method, the features of different receptive fields are extracted based on the deep convolution network Inception module, the SE module is used for noise reduction to reduce the influence of noise, and the adaptive threshold module is constructed to solve the problem of difficult threshold setting in multi-classification tasks. The effective recognition of multiple overlapping radar signals in low signal-to-noise ratio environment is realized; while realizing the recognition of overlapping radar signals, the recognition accuracy under low signal-to-noise ratio conditions is improved. It has the advantages of flexibility, accuracy, strong generalization ability and strong robustness.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a method for identifying modulation types of overlapping radar signals according to the present invention;

FIG2 is a structural diagram of a basic module of the present invention;

FIG3 is a structural diagram of the Inception module of the present invention;

FIG4 is a structural diagram of an adaptive threshold module of the present invention;

FIG. 5 is a diagram showing the overlapping radar signal modulation type recognition result provided in this example.

Detailed ways

The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

As shown in FIG1 , a method for identifying modulation types of overlapping radar signals based on an adaptive threshold network of the present invention comprises the following steps:

S1, simulate original radar signal and modulation parameters;

Table 1 shows the original radar signal parameters of this embodiment. This embodiment simulates five typical radar signal modulation types, including linear frequency modulation (LFM), nonlinear frequency modulation (NLFM), conventional wave (CW) signal, frequency agility (FA) signal and binary frequency shift keying (Costas). The simulated signals are all baseband signals, with a sampling frequency of 700MHz, a bandwidth range of 100-300MHz, a pulse width range of 3-5us, a signal-to-noise ratio range of -12dB-5dB, and a signal overlap rate set to 30%-100%.

Table 1

parameter scope f _s 700MHz B 10～300MHz P 3～5us SNR -12dB～5dB Od 30%～100%

S2. Overlap the simulated original radar signals to generate overlapping radar signals. Table 2 shows the overlapping of radar signals, including the overlapping of 2 modulation types and 3 modulation types, as well as the overlapping of the same modulation type and different modulation types.

Table 2

S3. Extract the time-frequency domain features of the overlapping radar signals and express them in the form of a time-frequency diagram, where the horizontal axis represents time and the vertical axis represents frequency. The specific method for extracting the time-frequency features is: use the FSST (Fourier Synchronous Squeezed Transform) algorithm to extract the time-frequency features. The FSST calculation formula is:

Where g(0) represents the value of the sliding window function g(t) at time 0, δ() is the Dirac impulse function; V _f (η,t) represents the signal after STFT (short-time Fourier transform) of the overlapping radar signal; ω represents the frequency after FSST redistribution, η represents the frequency before FSST redistribution (the frequency after STFT transformation), and t represents time. represents the entire real number domain, which is equivalent to is the instantaneous frequency, defined as:

The different radar signals generated in step S2 are intertwined with each other, not only overlapping in the time domain, but also in the frequency domain. Traditional Fourier transform is not applicable to this situation. However, the time-frequency characteristics of the signal can reflect the relationship between its frequency and time. Since the frequency of radar signals of different modulation types varies differently over time, the time-frequency characteristics can be used to distinguish the overlapping signals.

S4, construct a basic module; as shown in Figure 2, the basic module includes a sequential convolution layer, a batch normalization layer BN, a ReLU activation layer, a maximum pooling layer and a squeeze excitation module SE; wherein SE is used for denoising, including a sequential global average pooling layer GAP, a fully connected layer FC, a ReLU activation function layer, a fully connected layer FC, and a Sigmoid activation function layer; the weight coefficient of each channel is learned through the SE submodule, and then the noise threshold is obtained by multiplying the weight by the GAP output; finally, the output of the SE module is used as a threshold to denoise the output result of the maximum pooling layer: the value in each feature channel is compared with its noise threshold, and the value less than the threshold is considered to be caused by noise, which is not conducive to modulation recognition, and is set to zero.

S5. Construct Inception module: The Inception module contains multiple basic modules with different convolution kernel sizes, which are used to extract information of different dimensions in the time-frequency graph and combine them together;

When the frequency of the radar signal changes slowly over time, it appears as a flat curve on the time-frequency graph, so a basic module with a convolution kernel size of 3×7 can be used to capture the characteristics of its frequency changing over time. Similarly, a basic module with a convolution kernel size of 7×3 can be used to capture the characteristics of the signal frequency changing rapidly over time. When multiple signals overlap, multiple frequency components will appear in the same time dimension on the time-frequency graph, so the basic block 7×3 can be used to obtain the number of overlapping signals. In addition, basic modules with convolution kernel sizes of 3×3 and 5×5 can be used to capture more detailed features. The outputs of the above basic modules are fused and cascaded to the basic modules to extract higher-dimensional features. "Basic module m×n" means a basic module with a convolution kernel size of m×n. As shown in Figure 3, the detailed processing process of this step is:

S5.1. The time-frequency graph is simultaneously input into four different basic modules for processing to obtain features of different dimensions. The convolution kernel sizes of the four basic modules are 3×7, 7×3, 5×5 and 3×3 respectively.

S5.2, input the output of each basic module in step S5.1 into the basic module with a convolution kernel size of 3×3;

S5.3, merging the outputs of the four basic modules in step S5.2;

S5.4. Input the fused output of step S5.3 into four basic modules with a convolution kernel size of 3×3 in sequence to obtain the output result.

S6, construct an adaptive threshold module: including a fully connected layer, a ReLU activation function, a fully connected layer and a Sigmoid activation, as shown in Figure 4; the output of the Inception module in step S5 is used as input to obtain a corresponding threshold; in a multi-label classification task, the posterior probability output by the network is compared with the threshold to obtain a classification result, and different threshold settings will achieve different classification performances. In order to avoid the difficulty of selecting the optimal threshold, the present invention uses an adaptive threshold block to obtain an adaptive threshold for each modulation type to improve the classification performance and generalization ability of the network.

S7, build probability module: use the structure of sequentially connected fully connected layers, ReLU activation function and Sigmoid activation function, take the output of Inception module in step S5 as input, and get the posterior probability;

S8. The submodules constructed in steps S5 to S7 are combined into a SE-IncepAtNet network, specifically: the probability module and the adaptive threshold module are connected in parallel, and then cascaded after the Inception module to form a SE-IncepAtNet network; the input is the time-frequency diagram processed by S3, and the output is the classification result;

After the original radar signal time series is processed in step S3, a two-dimensional time-frequency diagram is obtained. After the time-frequency diagram is input into the SE-IncepAtNet network, the posterior probability corresponding to each category can be obtained through the probability submodule, and the threshold can be obtained through the adaptive threshold submodule. The category corresponding to the data with a value greater than the threshold in the posterior probability is the classification result of the network structure.

S9, training the SE-IncepAtNet network, specifically: inputting the time-frequency diagram of the overlapping radar signal into the SE-IncepAtNet network for forward propagation, and calculating the cost function value; using the gradient descent-based backward propagation algorithm to update the parameters of the SE-IncepAtNet network; iterating the backward propagation process until the cost function converges, thereby obtaining a trained SE-IncepAtNet network; specifically including the following steps:

S9.1, forward propagation;

S9.2. Calculate the cost function value, using the binary cross entropy loss function as the cost function. The calculation method is:

Where BCE is the binary cross entropy loss function, prob∈R ^1×n is the output probability, lab∈R ^1×n is the true value of the overlapping signal in the form of one-hot encoding, thre∈R ^1×n is the threshold, and n is the total number of modulation types;

S9.3. Update the network parameters based on the gradient descent back-propagation algorithm.

S10, using the SE-IncepAtNet network trained in step S9 to identify the modulation type of the overlapping radar signal.

In order to demonstrate the recognition performance of the SE-IncepAtNet network proposed in the present invention, the accuracy and recall of SE-IncepAtNet are compared with those of AlexNet, ResNet and GoogleNet. As shown in FIG5 , the modulation type recognition result of the present invention is shown. FIG5( a ) is an image of the accuracy of the recognition result with the change of SNR; FIG5( b ) is an image of the recall of the recognition result with the change of SNR.

As can be seen from Figure 5, the recall and precision of SE-IncepAtNet are better than other networks at low SNR, because the SE module reduces the impact of noise. At high SNR, the SE module has little effect on recognition, and the adaptive threshold block improves its recognition performance. In addition, due to the Inception module, the performance of SE-IncepAtNet is better than ResNet and AlexNet. Simulations show that this method can effectively realize modulation recognition of overlapping radar signals at low SNR. At -10dB, the recall rate of this method is 90.6% and the precision rate is 95.0%. According to statistics, this method has a good performance in the recognition of each modulation type. The accuracy of CW is above 95%, the accuracy of LFM and NLFM is above 90%, and the accuracy of Costas and FA is above 85%.

Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical inspiration disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims (5)

1. A method for identifying modulation types of overlapping radar signals based on an adaptive threshold network, characterized in that it comprises the following steps: S1, simulate original radar signal and modulation parameters; S2, overlapping the simulated original radar signals to generate overlapping radar signals; S3, extracting the time-frequency domain features of the overlapping radar signals, and representing them in the form of a time-frequency diagram, where the horizontal axis represents time and the vertical axis represents frequency; S4, build basic modules; S5. Construct Inception module: The Inception module contains multiple basic modules with different convolution kernel sizes, which are used to extract information of different dimensions in the time-frequency graph and combine them together; S6, build an adaptive threshold module: including the sequential fully connected layer, ReLU activation function, fully connected layer and Sigmoid activation; take the output of the Inception module in step S5 as input to obtain the corresponding threshold; S7, build probability module: use the structure of sequentially connected fully connected layers, ReLU activation function and Sigmoid activation function, take the output of Inception module in step S5 as input, and get the posterior probability; S8, the submodules constructed in steps S5 to S7 are combined into a SE-IncepAtNet network, specifically: the probability module and the adaptive threshold module are connected in parallel, and then cascaded after the Inception module to form a SE-IncepAtNet network; the input is the time-frequency diagram processed by S3, and the output is the classification result; S9, training the SE-IncepAtNet network, specifically: inputting the time-frequency diagram of the overlapping radar signal into the SE-IncepAtNet network for forward propagation, and calculating the cost function value; using the gradient descent-based backward propagation algorithm to update the parameters of the SE-IncepAtNet network; iterating the backward propagation process until the cost function converges, thereby obtaining the trained SE-IncepAtNet network; S10, using the SE-IncepAtNet network trained in step S9 to identify the modulation type of the overlapping radar signal. 2. The overlapping radar signal modulation type recognition method based on adaptive threshold network according to claim 1 is characterized in that the specific method of extracting time-frequency features in step S3 is: using FSST algorithm to extract time-frequency features, and the FSST calculation formula is: Where g(0) represents the value of the sliding window function g(t) at time 0, δ() is the Dirac impulse function; V _f (η,t) represents the signal after STFT of the overlapping radar signal; ω represents the frequency after FSST redistribution, η represents the frequency before FSST redistribution, and t represents time. Represents the entire real number domain; /> is the instantaneous frequency, defined as: 3. According to the method for identifying the modulation type of overlapping radar signals based on an adaptive threshold network in claim 1, it is characterized in that the basic module includes a sequential convolutional layer, a batch normalization layer BN, a ReLU activation layer, a maximum pooling layer and a squeeze excitation module SE; wherein SE is used for denoising, including a sequential global average pooling layer GAP, a fully connected layer FC, a ReLU activation function layer, a fully connected layer FC, and a Sigmoid activation function layer. 4. The method for identifying modulation types of overlapping radar signals based on an adaptive threshold network according to claim 1, wherein the detailed processing process of step S5 is as follows: S5.1. The time-frequency graph is simultaneously input into four different basic modules for processing to obtain features of different dimensions. The convolution kernel sizes of the four basic modules are 3×7, 7×3, 5×5 and 3×3 respectively. S5.2, input the output of each basic module in step S5.1 into the basic module with a convolution kernel size of 3×3; S5.3, merging the outputs of the four basic modules in step S5.2; S5.4. Input the fused output of step S5.3 into four basic modules with a convolution kernel size of 3×3 in sequence to obtain the output result. 5. The method for identifying modulation types of overlapping radar signals based on an adaptive threshold network according to claim 1, wherein step S9 specifically comprises the following steps: S9.1, forward propagation; S9.2. Calculate the cost function value, using the binary cross entropy loss function as the cost function. The calculation method is: Where BCE is the binary cross entropy loss function, prob∈R ^1×n is the output probability, lab∈R ^1×n is the true value of the overlapping signal in the form of one-hot encoding, thre∈R ^1×n is the threshold, and n is the total number of modulation types; S9.3. Update the network parameters based on the gradient descent back-propagation algorithm.