Frequency spectrum neural network model, recognition system composed of same and recognition method
By using the spectrum neural network model in individual recognition of radar radiation sources, combined with FFT pooling and wavelet pooling technology, the traditional method solves the problems of unstable performance and feature dimensionality reduction loss information, and achieves higher recognition accuracy and robustness.
Patent Information
- Application Number
- CN202411974014.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-30
- Publication Date
- 2025-04-25
AI Technical Summary
Existing deep learning algorithms are difficult to fully tap potential data information in individual recognition of radar radiation sources, resulting in unstable recognition performance. Traditional convolutional neural networks will lose key information when feature dimensionality reduction, affecting recognition performance.
A spectrum neural network model is proposed, including neural network, pooling layer and fully connected layer. By using FFT pooling layer or wavelet pooling layer to pool the signals in frequency domain, retain more subtle information, and use low-frequency frequency component criterion or highest recognition accuracy criterion for identification.
Experimental results on the real radar radiation source data set show that compared with GAP pooling, the proposed method improves the recognition performance by nearly 5%, and performs better under the conditions of sample imbalance and small samples, and is robust.
Smart Images
Figure CN119884826A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a spectrum neural network model and an identification system and an identification method composed of the same, in particular to a spectrum neural network model for individual identification of radar radiation sources and an identification system and an identification method composed of the same, belonging to the technical field of radar radiation source identification. Background Art
[0002] Radar emitter individual identification technology (Specific Emitter Identification, SEI) has important application value in the fields of electronic warfare, radio monitoring, spectrum sensing and electronic intelligence. SEI technology refers to obtaining individual characteristics by measuring the external characteristics of intercepted electromagnetic signals, and then uniquely identifying the individual radio emitters of interest based on the intelligence database. Traditional radar emitter individual identification methods based on manually designed features usually rely on manually designed features or shallow learning models. These methods often find it difficult to fully mine the potential information in the data. At the same time, the subtlety of the individual characteristics of radar emitters, the complexity of the electromagnetic environment, and the differences in receiver performance also increase the difficulty of the emitter identification task, making it difficult for traditional methods to achieve ideal results in actual application scenarios.
[0003] In recent years, researchers have begun to introduce deep learning technology into the field of radar emitter identification, trying to learn the abstract features of emitters from raw data through deep neural networks to improve recognition performance. Compared with traditional hand-designed feature extraction methods, neural networks have better adaptability and generalization capabilities, and can automatically learn complex features in spatiotemporal data.
[0004] Ding proposed a convolutional neural network recognition method based on bispectral feature extraction, which achieved high recognition accuracy in simulation and software radio experiments.
[0005] He proposed a SEI identification algorithm based on the LTSM model, which adopted a combination of features such as empirical mode decomposition (EMD), intrinsic time-scale decomposition (ITD) and variational mode decomposition (VMD) for feature extraction.
[0006] Wu proposed a SEI identification algorithm based on contrastive learning to solve the problem of insufficient labeled data.
[0007] Gong An unsupervised SEI framework based on information maximization generative adversarial networks (InfoGANs) and radio frequency fingerprint embedding (RFFE) is proposed.
[0008] Qian et al. proposed a sparse representation based multi-class recognition (MSRI) algorithm that combines neural networks with sparse representation based classification (SRC).
[0009] The above-mentioned deep learning method provides a new idea for radar emitter individual recognition technology. However, the existing deep learning algorithm still has some problems when processing radar emitter data. First, due to the subtlety of the individual characteristics of radar emitters, traditional deep neural networks often find it difficult to fully mine the potential information of the data, resulting in unstable recognition performance. Secondly, traditional convolutional neural networks usually use methods such as GAP to reduce feature dimensionality, which will cause a large amount of key information to be lost, affecting the further improvement of recognition performance. Finally, when processing large-scale data, sample imbalance and small samples, the performance of deep learning networks is even more limited. Summary of the invention
[0010] The purpose of the present invention is to solve the problems existing in the prior art and to provide a spectral neural network model and a recognition system and a recognition method composed of the spectral neural network model.
[0011] In order to solve the above problems, the present application is implemented through the following technical solutions: A spectrum neural network model, which is special in that: it is oriented to the individual identification of radar radiation sources, and comprises a neural network, a pooling layer, and a fully connected layer, wherein the pooling layer is connected to the back of the neural network, and the fully connected layer is connected to the back of the pooling layer; The neural network is used to extract data features; the pooling layer pools the data output by the neural network; the fully connected layer is used to realize individual identification of radar radiation sources; Furthermore, the neural network is a neural network of different depths such as resnet34, resnet50, and resnet101; Furthermore, the pooling layer is a GAP pooling layer, an FFT pooling layer or a wavelet pooling layer, and the pooling layer is selected according to different needs; Furthermore, the GAP pooling layer performs maximum average pooling processing, and the feature dimension before pooling is , for each channel feature Perform global average pooling. After pooling, each channel feature is averaged to a number, and the pooling dimension is ; Furthermore, the FFT pooling layer performs FFT transformation, and uses FFT transformation to decompose the features into zero-frequency frequency components, low-frequency frequency components and high-frequency frequency components. The FFT transformation does not change the feature dimension, but only changes the feature from the time domain dimension to the frequency domain dimension; Furthermore, the wavelet pooling layer performs wavelet t transform, and uses wavelet transform to decompose the features into approximation coefficients, horizontal detail coefficients, vertical detail coefficients and diagonal detail coefficients. The wavelet transform will change the feature dimension, and the transformation dimension is changed from becomes ; Furthermore, the fully connected layer uses a low-frequency component criterion or a maximum recognition accuracy criterion to realize individual recognition of radar emitters; The radar radiation source individual identification system composed of the above neural network model is special in that it includes a signal acquisition module for acquiring radar radiation source signals; a signal receiving module for receiving the signals acquired by the signal acquisition module; a signal preprocessing module for receiving the signals of the signal receiving module and performing Fourier transform processing; The spectral neural network model is used to perform pooling and recognition on the Fourier transformed signal.
[0012] Furthermore, the signal acquisition module is a radar, and 8 radars of the same model are set to be turned on to collect signals at the same time; Furthermore, the signal receiving module is a receiver, which adopts a portable ultra-wideband real-time spectrum analyzer, and the receiver's instantaneous receiving bandwidth is 80 MHz.
[0013] A method for identifying individual radar radiation sources, which is special in that it includes the following steps: Step 1: Perform FFT preprocessing on the received data, i.e., the I / Q data stream, to obtain a data set D; Step 2: Build a neural network, provide neural networks of different depths such as resnet34, resnet50, and resnet101, and select different network structures to extract data features according to the characteristics of data set D; Step 3: Pool the output of the network structure in step 2; Step 4: Select frequency components for the data after pooling in step 3, classify the data, and realize individual identification of radar emitters.
[0014] Furthermore, in step 1: assuming that the received radar radiation source signal has been sorted into individual radiation source signals, considering the nonlinear system response of the power amplifier, the high-frequency signal is input It is expressed as: , Where: represents the radar baseband modulated signal, Indicates the signal carrier frequency, Indicates the signal sampling frequency; Taylor polynomial amplifier output Describe it as: , Where: are Taylor polynomial coefficients, is the order of the polynomial; The sampled signal R from the i-th radiation source intercepted by the receiver (i) (nT) is expressed as: , Where: is the channel fading coefficient from the transmitter to the receiver, is the channel additive noise.
[0015] Furthermore, the Fourier transform in step 1 is specifically as follows: Considering the real-time nature of the algorithm, the received signal is not subjected to complex feature extraction. Instead, it is subjected to FFT transformation to ensure the consistency of the subsequent network input dimensions. The details are as follows: , , Where: Indicates The radiation source at time The received signal at represents the signal length of the sampling point, represents an imaginary unit; right Perform absolute value processing and normalization, that is, , and the length is The input features are sample features; Furthermore, the data set in step 1 ,in represents the sample characteristics, Indicates the label to which the sample belongs. Furthermore, FFT pooling is used in step 3, and the specific process is as follows: Assume that the input feature before pooling is ,right The 2D Discrete Fourier Transform (DTF) is: , Where: is the spectrum component after DFT transformation, and They are the input features height and width, and , (Why can't the next line come up? If it is a specific format, ignore this opinion and delete the content in brackets) , ; The input feature before FFT pooling is , the dimension is , after FFT transformation, the characteristic is , the size remains unchanged, only the original time domain representation is converted into frequency domain representation; Furthermore, after the FFT pooling in step 3, if and only if and hour, , It is essentially the expression of global average pooling GAP, that is, all input features are The mean is obtained by summing and averaging; Furthermore, wavelet pooling is used in step 3, and the specific process is as follows: Assume that the input feature before pooling is , using Haar wavelet function Perform a 2D wavelet transform and obtain its approximate coefficients, horizontal detail coefficients, vertical detail coefficients, and diagonal detail coefficients: , , , , Where: represents the number and scale of decomposition, and is the position coordinate of the transformed feature, and denote the scale function and detail function respectively; when and hour, It is expressed as: , The input feature before wavelet pooling is , the dimension is , the characteristics after wavelet transformation are , the size becomes , the original two-dimensional signal is divided into wavelet coefficients of different scales and directions.
[0016] Furthermore, the step 4 is specifically as follows: The frequency component adopts the low-frequency frequency component criterion or the highest recognition accuracy criterion. After FFT transformation and wavelet transformation, the main information of the feature map is concentrated in the low-frequency area, and the low-frequency frequency component criterion is adopted: Assume that after transformation It means that the frequency component is selected by using the low-frequency component criterion. , For FFT transform or wavelet change dimension: ; right frequency components for recognition accuracy test, select the first K corresponding frequency components with the highest recognition accuracy, and use the highest recognition accuracy criterion: P = [ p 1 , p 2 , ⋯ p K ] = max K ( p 1 , p 2 ⋯ p MN ) , Where: The representative is from Select the K frequency components with the highest recognition accuracy from the largest recognition accuracy, and record them as [ p 1 , p 2 , ⋯ p K ] , discard the remaining frequency components; Directly [ p 1 , p 2 , ⋯ p K ] Identify and obtain the radar radiation source identification result.
[0017] This application combines the characteristics of individual radar emitter signals and proposes a new frequency domain pooling method - FFT pooling and wavelet pooling. Experimental results on a real radar emitter data set show that compared with GAP, the recognition performance of this application's method is improved by nearly 5% under the same conditions, and under conditions such as sample imbalance and small samples, the recognition performance of this application's method is better than that of GAP and other methods, and it is more robust.
[0018] The present invention proposes a method of pooling feature maps from the frequency domain perspective. By pooling signals in the frequency domain, the characteristic information of radar emitters can be better extracted. Based on the characteristics of individual signals of radar emitters, FFT pooling and wavelet transform pooling methods are proposed: FFT pooling: Use fast Fourier transform (FFT) to perform spectrum analysis on the signal and retain the corresponding frequency components in the frequency domain. Compared with global average pooling, it can retain more subtle information; Wavelet transform pooling: Wavelet transform is used to perform multi-scale analysis on the signal and retain the corresponding frequency components in the frequency domain. Wavelet transform can capture the time domain and frequency domain characteristics of the signal at the same time, which helps to improve the recognition accuracy; In terms of frequency component retention, low-frequency component criteria and maximum recognition accuracy criteria are proposed. These methods can selectively retain or discard certain components in the spectrum according to the characteristics of the signal, thereby improving the neural network's ability to extract subtle features. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 It is a structural schematic diagram of the spectrum neural network model of the present invention; Figure 2 It is a structural schematic diagram of the identification system of the present invention; Figure 3 is a flow chart of the identification method of the present invention; Figure 4 It is a framework diagram of the identification method of the present invention; Figure 5 This is the pooling principle diagram; Figure 6 This is a schematic diagram of the transformation before and after the highest recognition accuracy criterion; Figure 7 This is a schematic diagram of FFT pooling recognition accuracy; Figure 8 After FFT pooling Loss cost function change curve; Fig. 9 This is a schematic diagram of wavelet transform pooling recognition accuracy; Fig.10 After wavelet transform pooling Loss cost function change curve. DETAILED DESCRIPTION
[0020] The following is a specific implementation of the present invention with reference to the accompanying drawings to further illustrate the composition of the present invention.
[0021] In order to facilitate understanding of the embodiments of the present invention, abbreviations and key terms that may be involved in the embodiments of the present invention are first explained or defined.
[0022] Embodiment 1. A spectrum neural network model, such as Figure 1 As shown, for radar radiation source individual identification, it includes a neural network, a pooling layer, and a fully connected layer. The pooling layer is connected to the back of the neural network, and the fully connected layer is connected to the back of the pooling layer. The neural network is used to extract data features; the pooling layer pools the data output by the neural network; the fully connected layer is used to realize individual identification of radar radiation sources; Furthermore, the neural network is a neural network of different depths such as resnet34, resnet50, and resnet101; Furthermore, the pooling layer is a GAP pooling layer, an FFT pooling layer or a wavelet pooling layer. The pooling layer is selected according to different needs. Figure 5 As shown; Furthermore, the GAP pooling layer performs maximum average pooling processing, and the feature dimension before pooling is , for each channel feature Perform global average pooling. After pooling, each channel feature is averaged to a number, and the pooling dimension is ; Furthermore, the FFT pooling layer performs FFT transformation, and uses FFT transformation to decompose the features into zero-frequency frequency components, low-frequency frequency components and high-frequency frequency components. The FFT transformation does not change the feature dimension, but only changes the feature from the time domain dimension to the frequency domain dimension; Furthermore, the wavelet pooling layer performs wavelet t transform, and uses wavelet transform to decompose the features into approximation coefficients, horizontal detail coefficients, vertical detail coefficients and diagonal detail coefficients. The wavelet transform changes the feature dimension, and the transformation dimension is changed from becomes ; Furthermore, the fully connected layer adopts a low-frequency component criterion or a maximum recognition accuracy criterion to realize individual recognition of radar emitters.
[0023] Embodiment 2. A radar radiation source individual recognition system composed of the neural network model of Embodiment 1, such as Figure 2 As shown, it includes a signal acquisition module for acquiring radar radiation source signals; a signal receiving module for receiving the signals acquired by the signal acquisition module; The signal preprocessing module is used to receive the signal from the signal receiving module and perform Fourier transform processing; the spectrum neural network model is used to perform pooling processing and recognition on the signal after Fourier transform.
[0024] Furthermore, the signal acquisition module is a radar, and 8 radars of the same model are set to be turned on to collect signals at the same time; Furthermore, the signal receiving module is a receiver, which adopts a portable ultra-wideband real-time spectrum analyzer, and the receiver's instantaneous receiving bandwidth is 80 MHz.
[0025] Embodiment 3. A method for identifying individual radar radiation sources, the process of which is as follows: Figure 3 As shown, the framework diagram is as follows Figure 4 As shown, the following steps are included: Step 1: Perform FFT preprocessing on the received data, i.e., the I / Q data stream, to obtain a data set D; Step 2: Build a neural network, provide neural networks of different depths such as resnet34, resnet50, and resnet101, and select different network structures to extract data features according to the characteristics of data set D; Step 3: Pool the output of the network structure in step 2; Step 4: Select frequency components for the data after pooling in step 3, classify the data, and realize individual identification of radar emitters.
[0026] Furthermore, in step 1: assuming that the received radar radiation source signal has been sorted into individual radiation source signals, considering the nonlinear system response of the power amplifier, the high-frequency signal is input It is expressed as: , Where: represents the radar baseband modulated signal, Indicates the signal carrier frequency, Indicates the signal sampling frequency; Taylor polynomial amplifier output Describe it as: , Where: are Taylor polynomial coefficients, is the order of the polynomial; The sampled signal R from the i-th radiation source intercepted by the receiver (i) (nT) is expressed as: , Where: is the channel fading coefficient from the transmitter to the receiver, is the channel additive noise.
[0027] Furthermore, the Fourier transform in step 1 is specifically as follows: Considering the real-time nature of the algorithm, the received signal is not subjected to complex feature extraction. Instead, it is subjected to FFT transformation to ensure the consistency of the subsequent network input dimensions. The details are as follows: , , Where: Indicates The radiation source at time The received signal at represents the signal length of the sampling point, represents an imaginary unit; right Perform absolute value processing and normalization, that is, , and the length is The input features are sample features; Furthermore, the data set in step 1 ,in represents the sample characteristics, Indicates the label to which the sample belongs.
[0028] Furthermore, FFT pooling is used in step 3, and the specific process is as follows: Assume that the input feature before pooling is ,right The 2D Discrete Fourier Transform (DTF) is: , Where: is the spectrum component after DFT transformation, and They are the input features height and width, and , , ; The input feature before FFT pooling is , the dimension is , after FFT transformation, the characteristic is , the size remains unchanged, only the original time domain representation is converted into frequency domain representation; Furthermore, after the FFT pooling in step 3, if and only if and hour, , It is essentially the expression of global average pooling GAP, that is, all input features are The mean is obtained by summing and averaging; Furthermore, wavelet pooling is used in step 3, and the specific process is as follows: Assume that the input feature before pooling is , using Haar wavelet function Perform a 2D wavelet transform and obtain its approximate coefficients, horizontal detail coefficients, vertical detail coefficients, and diagonal detail coefficients: , , , , Where: represents the number and scale of decomposition, and is the position coordinate of the transformed feature, and denote the scale function and detail function respectively; when and hour, It is expressed as: , The input feature before wavelet pooling is , the dimension is , the characteristics after wavelet transformation are , the size becomes , the original two-dimensional signal is divided into wavelet coefficients of different scales and directions.
[0029] Furthermore, the step 4 is specifically as follows: The frequency component adopts the low-frequency frequency component criterion or the highest recognition accuracy criterion. After FFT transformation and wavelet transformation, the main information of the feature map is concentrated in the low-frequency area, and the low-frequency frequency component criterion is adopted: Assume that after transformation It means that the frequency component is selected by using the low-frequency component criterion. , For FFT transform or wavelet change dimension: ; The schematic diagram of the transformation before and after using the low-frequency component criterion is as follows Figure 6 As shown, assuming The dimension is In the figure, LF and HF represent low-frequency components and high-frequency components respectively. The indexes of rows and columns are 0~7. The index number selected by GAP pooling is the frequency component with the value (0,0), as marked by the red box in the left figure. The remaining 48 frequency components are discarded. After adopting the low-frequency frequency component criterion, the index number selected is the part marked by the triangle box, and the lower right part is discarded. right frequency components for recognition accuracy test, select the first K corresponding frequency components with the highest recognition accuracy, and use the highest recognition accuracy criterion. P = [ p 1 , p 2 , ⋯ p K ] = max K ( p 1 , p 2 ⋯ p MN ) , Where: The representative is from Select the K frequency components with the highest recognition accuracy from the largest recognition accuracy, and record them as [ p 1 , p 2 , ⋯ p K ] , discard the remaining frequency components; Directly [ p 1 , p 2 , ⋯ p K ] Identify and obtain the radar radiation source identification result. Select the frequency index corresponding to the first K recognition accuracies and The frequency components are discarded. The schematic diagram of the transformation before and after using the highest recognition accuracy criterion is as follows Figure 6 Assume The dimension is In the figure, LF and HF represent low-frequency components and high-frequency components respectively. The indexes of rows and columns are 0~7. The frequency component with index number (0,0) selected by GAP pooling is marked with a red box in the left figure, and the remaining 48 frequency components are discarded.
[0030] top-X represents the ranking of the recognition accuracy of each frequency component. The first K frequency components are selected according to the recognition accuracy, and the remaining frequency components are discarded.
[0031] Example 4. This example is an experiment of the method described in Example 3, and is as follows: A. Dataset Setting The radar individual signals collected in this application are from the X-band sea navigation radar, and the parameters are shown in Table 1.
[0032] Table 1 Radar model parameters , The receiver uses a portable ultra-wideband real-time spectrum analyzer, and the receiver instantaneous receiving bandwidth is 80MHz. At the same time, 8 radars of the same model are turned on to collect signals. The distances from the collection points to each radar are shown in Table 2 below.
[0033] Table 2 Experimental radar deployment table , In order to verify the performance of the algorithm, a sample sufficient condition data set was constructed, including cases where the number of radiation source individuals was 3, 4, 5, 6, 7, and 8; B. Experimental Conditions The deep learning model is built based on the pytorch2.0.1 framework. The computer configuration used is Intel(R)Xeon(R): W-2295 CPU: 128G memory, graphics card NVIDIA RTX A6000, operating system is Windows 10, no special instructions, network training times epoch=100, batch size=64, learning rate is set to The signal-to-noise ratio of the collected signal is 3dB~18dB, the number of radiation sources of each type is 13,000, and the spectrum neural network is built based on ResNet34, and the low-frequency component criterion is adopted.
[0034] C. Experimental results analysis Algorithm recognition performance analysis (1) FFT pooling recognition performance analysis To verify the recognition performance of FFT pooling, the FFT pooling recognition accuracy is as follows Figure 7 As shown, when the number of radiation sources The change curve of loss cost function is as follows: Figure 8 As shown. Figure 7 It can be seen that when the number of radiation sources When , the FFT pooling recognition accuracy is 99.66%, 99.35%, 99.52%, 98.61%, 97.75%, 97.81% respectively. However, compared with the global average pooling method (GAP), the recognition performance of the method proposed in this application is improved by about 5% under the same conditions. This is mainly because FFT pooling retains more subtle information and improves the recognition accuracy of the subsequent network. Figure 8 As we know, when epoch=80, the cost loss function tends to be stable, indicating that the network model converges.
[0035] (2) Wavelet transform pooling recognition performance analysis To verify the recognition performance of wavelet transform pooling, the algorithm recognition accuracy is as follows Fig. 9 As shown, when the number of radiation sources The change curve of loss cost function is as follows: Fig.10 As shown. Fig. 9 It can be seen that when the number of radiation sources When , the wavelet pooling recognition accuracy is 99.22%, 98.42%, 99.264%, 97.78%, 97.64% and 97.62% respectively. However, compared with the global average pooling method (GAP), the recognition performance of the algorithm proposed in this application is improved by about 5% under the same conditions. This is mainly because wavelet pooling retains more subtle information. The individual identification of radar emitters pays more attention to the subtle features of the signal. Therefore, after wavelet pooling, the neural network pays attention to more subtle features, which improves the recognition accuracy to a certain extent. Fig.10 As we know, when epoch=80, the cost loss function tends to be stable, and the value of the cost loss function tends to 0, indicating that the training neural network model has converged.
[0036] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the defined scope.
Claims
1. A spectral neural network model, characterized in that: For radar radiation source individual identification, it includes a neural network, a pooling layer, and a fully connected layer, wherein the pooling layer is connected to the back of the neural network, and the fully connected layer is connected to the back of the pooling layer; The neural network is used to extract data features; The pooling layer pools the data output by the neural network; The fully connected layer is used to realize individual identification of radar emitters.
2. The spectrum neural network model for radar emitter individual identification according to claim 1 is characterized in that: The neural network is resnet34 or resnet50 or resnet101; The pooling layer is a GAP pooling layer, an FFT pooling layer or a wavelet pooling layer, and the pooling layer is selected according to different needs; The fully connected layer uses a low-frequency component criterion or a maximum recognition accuracy criterion to achieve individual recognition of radar emitters.
3. A spectrum neural network model for radar emitter individual identification according to claim 1 or 2, characterized in that: The GAP pooling layer performs maximum average pooling processing, and the feature dimension before pooling is , for each channel feature Perform global average pooling. After pooling, each channel feature is averaged to a number, and the pooling dimension is ; The FFT pooling layer performs FFT transformation, and uses FFT transformation to decompose the features into zero-frequency frequency components, low-frequency frequency components and high-frequency frequency components. The FFT transformation does not change the feature dimension, but only changes the feature from the time domain dimension to the frequency domain dimension; The wavelet pooling layer performs wavelet t transform, and uses wavelet transform to decompose the features into approximation coefficients, horizontal detail coefficients, vertical detail coefficients and diagonal detail coefficients. Wavelet transform will change the feature dimension, and the transformation dimension is becomes .
4. A radar radiation source individual identification system composed of the neural network model described in any one of claims 1 to 3, characterized in that: It includes a signal acquisition module for acquiring radar radiation source signals; A signal receiving module, used to receive the signal collected by the signal collecting module; A signal preprocessing module, used to receive the signal from the signal receiving module and perform Fourier transform processing; The spectral neural network model is used to perform pooling and recognition on the Fourier transformed signal.
5. A radar radiation source individual identification system according to claim 4, characterized in that: The signal acquisition module is a radar, and 8 radars of the same model are set to start up and collect signals at the same time; The signal receiving module is a receiver, which adopts a portable ultra-wideband real-time spectrum analyzer, and the receiver's instantaneous receiving bandwidth is 80 MHz.
6. A method for identifying a radar radiation source individual using the radar radiation source individual identification system according to claim 4 or 5, characterized in that: The following steps are involved: Step 1: Perform FFT preprocessing on the received data, i.e., the I / Q data stream, to obtain a data set D; Step 2: Build a neural network, provide resnet34 or resnet50 or resnet101, and select different network structures for extracting data features according to the characteristics of the data set D described in step 1; Step 3: Pool the output of the network structure in step 2; Step 4: Select frequency components for the data after pooling in step 3, classify the data, and realize individual identification of radar emitters.
7. A method for identifying individual radar radiation sources according to claim 6, characterized in that: In step 1: assuming that the received radar radiation source signal has been sorted into individual radiation source signals, considering the nonlinear system response of the power amplifier, the high-frequency signal is input It is expressed as: , Where: represents the radar baseband modulated signal, Indicates the signal carrier frequency, Indicates the signal sampling frequency; Taylor polynomial amplifier output Describe it as: , Where: are Taylor polynomial coefficients, is the order of the polynomial; The sampled signal R from the i-th radiation source intercepted by the receiver (i) (nT) is expressed as: , Where: is the channel fading coefficient from the transmitter to the receiver, is the channel additive noise; The Fourier transform in step 1 is specifically as follows: , , Where: Indicates The radiation source at time The received signal at represents the signal length of the sampling point, represents an imaginary unit; right Perform absolute value processing and normalization, that is, , and the length is The input features are sample features; The data set in step 1 ,in represents the sample characteristics, Indicates the label to which the sample belongs.
8. The method for identifying individual radar radiation sources according to claim 6, characterized in that: FFT pooling is used in step 3, and the specific process is as follows: Assume that the input feature before pooling is ,right The 2D discrete Fourier transform DTF is: , Where: is the spectrum component after DFT transformation, and They are the input features height and width, and , , ; The input feature before FFT pooling is , the dimension is , after FFT transformation, the characteristic is , the size remains unchanged, only the original time domain representation is converted into frequency domain representation; After the FFT pooling in step 3, if and only if and hour, , It is essentially the expression of global average pooling GAP, that is, all input features are The sum and then average gives the mean.
9. The method for identifying individual radar radiation sources according to claim 6, characterized in that: Wavelet pooling is used in step 3, and the specific process is as follows: Assume that the input feature before pooling is , using Haar wavelet function Perform a 2D wavelet transform and obtain its approximate coefficients, horizontal detail coefficients, vertical detail coefficients, and diagonal detail coefficients: ， ， ， ， Where: represents the number and scale of decomposition, and is the position coordinate of the transformed feature, and denote the scale function and detail function respectively; when and hour, It is expressed as: , The input feature before wavelet pooling is , the dimension is , the characteristics after wavelet transformation are , the size becomes , the original two-dimensional signal is divided into wavelet coefficients of different scales and directions.
10. The method for identifying individual radar radiation sources according to claim 6, characterized in that: The step 4 is specifically as follows: The frequency component adopts the low-frequency frequency component criterion or the highest recognition accuracy criterion. After FFT transformation and wavelet transformation, the main information of the feature map is concentrated in the low-frequency area, and the low-frequency frequency component criterion is adopted: Assume that after transformation It means that the frequency component is selected by using the low-frequency component criterion. , For FFT transform or wavelet change dimension: ; right frequency components for recognition accuracy test, select the first K corresponding frequency components with the highest recognition accuracy, and use the highest recognition accuracy criterion: , Where: The representative is from Select the K frequency components with the highest recognition accuracy from the largest recognition accuracy, and record them as , discard the remaining frequency components; Directly Identify and obtain the radar radiation source identification result.