Method for identifying individual radar radiation sources across receivers based on deep learning

Abstract

The present invention discloses a method for individual recognition of radar radiation sources based on deep learning, which mainly solves the problem of the decline in recognition accuracy caused by receiver distortion pollution in the individual recognition of radar radiation sources across receivers in the prior art. The implementation scheme is as follows: processing and dividing data from different receivers; constructing a dual-branch deep learning-based radar radiation source individual recognition model composed of a feature extractor and two recognition classifiers; using the backpropagation method to iteratively train the model, and making the feature extractor complete the difference compensation of the same radiation source signal under different receiving environments through the maximization of domain labels; using the trained model for individual recognition of radar radiation source signals. The present invention enhances the robustness to different non-ideal pollution factors, improves the accuracy of individual recognition of radar radiation sources in the cross-receiver application scenario, and can be used for radar individual information reconnaissance across receivers in complex electromagnetic environments.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal processing, and particularly relates to a method for identifying radar radiation sources, which can be applied to the reconnaissance of radar individual information across receivers in a complex electromagnetic environment. Background Art

[0002] A radar radiation source is an electromagnetic wave transmitting device used to generate detectable electromagnetic wave signals within a certain range so that a radar receiver can detect a target. A radar radiation source includes components such as an antenna, a radio frequency transmitter, a modulator, etc., and usually determines the target position and characteristics through the emitted electromagnetic wave and the signal reflected therefrom. Common radar radiation sources include civilian radars and military radars, which are widely used for communication, navigation, monitoring, and reconnaissance.

[0003] Radar radiation source individual identification is a technology that matches a signal with a unique transmitter based on the unintentional modulation characteristics attached to the signal. Since the unintentional modulation characteristics originate from the characteristics of the hardware circuit of the radiation source and are unique to each transmitter, it is also known as "fingerprint" identification. Identifying the individual radar radiation sources existing in space can obtain the most intuitive radar distribution and the action strategy of the radar-mounted equipment, which is the basis for subsequent situation analysis and decision-making.

[0004] However, with the increasing maturity of electronic information technology, the emergence of new radar systems, the wide popularization of multifunctional radars, the application of new waveforms, and the improvement of the performance of electronic devices, it has caused great difficulties in constructing the identification feature library of radar radiation sources. At the same time, the radar signal coverage frequency band has been broadened, the working parameter domains of different radar individuals in the same batch overlap highly, and the signal forms of multifunctional radars are agile, making the traditional sorting and identification technologies based on pulse description word parameters such as carrier frequency, pulse width, amplitude, and pulse repetition frequency difficult to adapt to the current severe and complex electromagnetic spectrum environment. The identification of radar individuals relies on the radiation source individual identification technology that uses the "fingerprint" characteristics of the transmitter for identification. In the current application scenarios of radar radiation source individual identification, since the known signal and the signal to be identified come from different receivers, the different receiver distortions contained in the signals will cause feature drift and model mismatch, thus leading to a decrease in identification accuracy.

[0005] To solve the above problems, it is necessary to improve the design idea of the radar radiation source individual identification algorithm, enhance the robustness to different receiver distortions, and improve the versatility in different receiving scenarios.

[0006] The patent document with the application number CN202010494523.1 discloses a "Radar Radiation Source Individual Recognition Method under Different Signal-to-Noise Ratios". By splicing multiple features such as the rising edge time, pulse width, and bispectrum waveform, it effectively extracts the fingerprint features of radiation sources in different dimensions and has a high recognition accuracy in the simulation environment with different signal-to-noise ratios. However, since this method only stacks the number of features to improve the recognition accuracy in an environment with noise differences, it does not design relevant algorithms to weaken pollution differences, nor does it consider pollutions with a high degree of coupling such as receiver distortion. Therefore, the individual recognition accuracy in the cross-receiver application scenario is low. Summary of the Invention

[0007] The purpose of the present invention is to overcome the defects of the above-mentioned existing technologies and propose a cross-receiver radar radiation source individual recognition method based on deep learning to improve the recognition accuracy of radar radiation source individuals in cross-receiver application scenarios.

[0008] To achieve the above purpose, the technical solutions adopted by the present invention are as follows:

[0009] (1) Obtain pulse signal data from different receivers, define the domain labels of each pulse signal, and delimit the source domain data and the target domain data; divide more than half of the source domain data and the target domain data into the training data set Q respectively, and the remaining into the test data set T;

[0010] (2) Perform bispectrum transformation on each pulse signal to obtain the reduced-dimensional one-dimensional feature domain SIB;

[0011] (3) Construct a radar radiation source individual recognition model O based on deep learning:

[0012] (3a) Establish a feature extractor G composed of a first-level one-dimensional convolutional layer, a max pooling layer 1, a second-level one-dimensional convolutional layer, a max pooling layer 2, and a flattening layer cascaded in sequence f ;

[0013] (3b) Establish an individual recognition classifier G composed of a first-level fully connected layer, a second-level fully connected layer, and a Softmax classifier cascaded in sequence y ;

[0014] (3c) Establish a domain recognition classifier G composed of a gradient reversal layer R, a first-level fully connected layer, a second-level fully connected layer, and a Softmax classifier cascaded in sequence d , where the gradient reversal layer R changes the optimization direction of domain classification to blur the differences between different domains and suppress the feature drift pollution caused by different receivers;

[0015] (3d) Connect G f and G yThe cascade forms an individual recognition branch, and then G f is cascaded with G d to form a domain recognition branch;

[0016] (4) Iteratively train the network recognition model O to obtain the optimized model O * :

[0017] (4a) Initialize the iteration number i, set the maximum iteration number to I, I≥50, set the loss function for network optimization to the cross-entropy function, and the learning rate to μ p , and initialize the network parameters of O to θ;

[0018] (4b) Randomly select a batch of data from the source domain training data and input it into the network model O. Use the network prediction probability that the nth sample in the source domain of the individual recognition branch belongs to the cth radiation source individual to obtain the individual classification prediction label Use the network prediction probability that the nth sample in the source domain of the domain recognition branch comes from the kth receiver , to obtain the domain prediction label , and calculate the loss L1 between the individual classification prediction label and the actual classification label Lf1 and the loss L2 between the domain prediction label and the actual domain label Ld1 respectively;

[0019] (4c) Randomly select a batch of data from the target domain training data and input it into the network model O. Use the domain recognition branch to calculate the network prediction probability that the lth sample in the target domain comes from the kth receiver , to obtain the domain prediction label Calculate the loss L3 between this domain prediction label and the actual domain label Ld2;

[0020] (4d) Sum all the losses to obtain the total loss L = L1 + L2 + L3 of the network, and perform backpropagation accordingly. Use the gradient descent method to optimize the network parameters to complete one iteration;

[0021] (4e) Repeat steps (4b)-(4d) until i≥I is satisfied, stop the iteration, and obtain the optimized network model O * ;

[0022] (5) Input the test sample T into the trained network model O * , and use the individual recognition branch to calculate the recognition result L p .

[0023] Compared with the prior art, the present invention has the following advantages:

[0024] First, since the present invention uses the bispectrum contour integral as the feature domain, it can reduce the feature dimension without losing feature information as much as possible. Compared with the existing technology that performs recognition through the two-dimensional bispectrum domain, its computational complexity is smaller.

[0025] Second, since the present invention introduces a domain recognition and classification branch containing a gradient reversal layer to compensate for the drift of the individual fingerprint features of radar radiation sources caused by different pollution factors, it has a good inhibitory effect on both the feature drift caused by different receiver distortions and the feature measurement errors caused by noise.

[0026] Third, since the present invention constructs a radar radiation source individual recognition model O based on deep learning, the deep learning technology it uses has a higher-dimensional feature processing ability and a more excellent deep feature extraction ability compared with the existing support vector machine technology, improving the recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is the overall flowchart for the implementation of the present invention;

[0028] Figure 2 is the application example scenario diagram of the present invention;

[0029] Figure 3 is the network model structure diagram in the present invention;

[0030] Figure 4 is the sub-flowchart for training the network model in the present invention;

[0031] Figure 5 is the comparison diagram of the feature domains of the signals of three radiation sources passing through two receivers at 40 dB in the present invention;

[0032] Figure 6 is the comparison diagram of the feature domains of the signals of three radiation sources passing through two receivers at 20 dB in the present invention;

[0033] Figure 7 is the curve comparison diagram of the classification accuracy between the present invention and the prior art embodiments. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] The embodiments and effects of the present invention will be further described in detail below with reference to the accompanying drawings.

[0035] Referring to Figure 2 , the usage scenario of this example includes three radar radiation sources and two signal receivers. The three radar radiation sources emit signals, which are respectively received by two different receivers. Among them, the signal of receiver 1 has a classification and recognition label. The hardware parameters of the three radar radiation sources are different, and the signal waveforms and pulse parameters they emit are the same. The mixing parameters of the two receivers are the same, but the hardware parameters are different.

[0036] In this scenario, the present invention identifies some signals of receiver 1 and all signals of receiver 2.

[0037] Referring to Figure 1 , the implementation steps of this example are as follows:

[0038] Step 1: Obtain the pulse signal data of the two receivers, and process and divide it.

[0039] (1.1) Define the domain label of the pulse according to the receiver source of the pulse signal. The domain label Ld1 of the pulse signal received by receiver 1 is 1, and the domain label Ld2 of the pulse signal received by receiver 2 is 2;

[0040] (1.2) Designate the signals received by receiver 1 as source domain data, and the signals received by receiver 2 as target domain data;

[0041] (1.3) Divide more than half of the source domain data and target domain data into training data sets respectively, and the rest are test data sets.

[0042] Step 2: Perform bispectrum transformation on each pulse signal to obtain the reduced-dimensional one-dimensional feature domain SIB.

[0043] (2.1) Perform zero-mean normalization on the pulse signal with a sampling frequency of f s to obtain the processed pulse sampling sequence as x(0), x(1),... x(N - 1), and divide it into K segments, each segment containing M sampling points. The data of the k-th segment is x (k) (0), x (k) (1),... x (k) (M - 1);

[0044] (2.2) Calculate the discrete Fourier transform DFT coefficients of each segment of data:

[0045]

[0046] where λ is the independent variable of the Fourier transform;

[0047] (2.3) Calculate the triple correlation of the DFT coefficients to obtain the bispectrum estimate value of this segment of data

[0048]

[0049] In the formula, Δ0 = f s / N0, L1 is the total sliding window length of the correlation operation, N0 is the normalization factor of the correlation operation, N0 and L1 satisfy M = (2L1 + 1)N0, and λ1, λ2 are the sliding independent variables of two different dimensions of the correlation operation;

[0050] (2.4) Take the mean of the bispectrum estimates of all K segments of signals to obtain the bispectrum estimate of the entire segment of signal

[0051]

[0052] where are the two frequency dimensions of the bispectrum;

[0053] (2.5) Perform contour integration on the bispectrum estimate value to obtain the reduced-dimensional bispectrum feature domain SIB(ω)

[0054]

[0055] Step 3, construct a radar emitter individual recognition model O based on deep learning.

[0056] Refer to Figure 3 , and the specific implementation of this step is as follows:

[0057] (3.1) Establish a feature extractor G composed of a first-level one-dimensional convolutional layer, a first max-pooling layer, a second-level one-dimensional convolutional layer, a second max-pooling layer, and a flattening layer cascaded in sequence f , where:

[0058] The number of channels of the first convolutional layer is 16, the convolutional kernel size is 1x32, and the activation function is the ReLU function;

[0059] The number of channels of the second convolutional layer is 16, the convolutional kernel size is 1x25, and the activation function is the ReLU function;

[0060] The pooling kernel sizes of the two max-pooling layers are both 1x2;

[0061] The number of nodes of the flattening layer is 192;

[0062] (3.2) Establish an individual recognition classifier G composed of a first-level fully connected layer, a second-level fully connected layer, and a Softmax classifier cascaded in sequence y , where:

[0063] The number of nodes of the first-level fully connected layer is 192, and the activation function is the Sigmoid function;

[0064] The number of nodes of the second-level fully connected layer is C, the activation function is the Sigmoid function, and the number of nodes C is the number of radar emitter individuals included in the dataset;

[0065] (3.3) Establish a domain recognition classifier G composed of a gradient reversal layer R, a first-level fully connected layer, a second-level fully connected layer, and a Softmax classifier cascaded in sequence d , where:

[0066] The number of nodes in the first fully connected layer is 192, and the activation function is the Sigmoid function;

[0067] The number of nodes in the second fully connected layer is N, and the activation function is the Sigmoid function, where N is the number of receivers for receiving data;

[0068] Gradient Reversal Layer R, during the forward calculation of the network, its output is equal to the input x, that is, R(x)=x; during the backpropagation optimization of the network, R takes the negative of the loss gradient of the input, that is , where p is the ratio of the current iteration number i to the total iteration number I, representing the iteration process of the network, Loss is the network loss during backpropagation, and γ = 10 is a hyperparameter;

[0069] (3.4) Concatenate G f and G y to form an individual recognition branch, and then concatenate G f and G d to form a domain recognition branch.

[0070] Step 4, perform iterative training on the network recognition model O to obtain the optimized model O * .

[0071] (4.1) Initialize the iteration number i = 1, set the maximum iteration number to I = 50, set the loss function for network optimization to the cross-entropy function, and the learning rate to μ p , initialize the network parameters of O, where the learning rate μ p is calculated as follows:

[0072]

[0073] where μ p has an initial value of μ0 = 0.01, the multiplication hyperparameter α = 10, and β = 0.75 is an exponential hyperparameter;

[0074] (4.2) Input the training set data into the network and update the network parameters in a backpropagation manner:

[0075] Refer to Figure 4 , the implementation of this step is as follows:

[0076] (4.2.1) Randomly select 32 pulse feature domain data from the source domain training data and input them into the network model O, and use the individual recognition branch to calculate the network prediction probability that the nth sample of this batch of source domain data belongs to the cth radiation source individual The specific calculation steps are as follows:

[0077] First, the input feature domain SIB passes through the feature extractor G with parameters θ1f , the output feature vector f:

[0078] f = G f (SIB; θ1);

[0079] Next, input the feature vector f into the individual recognition classifier G with parameter θ2 y for classification prediction to obtain a classification prediction vector of length C

[0080]

[0081] where C is the number of radar emitter individuals, represents the possibility that the nth source domain input sample belongs to each radar emitter individual;

[0082] Finally, let the cth point of be z c , according to the classification prediction vector calculate the probability that the nth source domain input sample belongs to each classification label

[0083]

[0084] p n c is the network prediction probability that the nth sample in the source domain belongs to the cth label, F is the total number of labels, where (4.2.2) Calculate the loss L1 between the individual classification prediction result and the true classification label Lf1:

[0085]

[0086] where U is the number of radar emitter individuals, N is the total number of samples, is the network prediction probability that the nth sample in the source domain belongs to the cth label, is the true probability that the nth sample in the source domain belongs to the cth label, belonging to the cth emitter is 1, not belonging to the cth emitter is 0;

[0087] (4.2.3) Use the domain recognition branch to calculate the network prediction probability that the mth sample in this batch of source domain data comes from the kth receiver The calculation method is the same as (4.2.1);

[0088] (4.2.4) Calculate the loss L2 between the domain classification prediction result and the true domain classification label Ld1;

[0089]

[0090] Among them, H is the number of receiver individuals, and M is the total number of samples. is the network prediction probability that the nth sample in the source domain comes from the kth receiver. is the true probability that the nth sample in the source domain belongs to the kth receiver. If it belongs to the kth receiver, it is 1; if it does not belong to the kth receiver, it is 0.

[0091] (4.2.5) Randomly select 32 pulse feature domain data from the target domain training data and input them into the network model O. Use the domain recognition branch to calculate the network prediction probability that the lth sample in this batch of target domain data comes from the kth receiver The calculation method is the same as (4.2.1).

[0092] (4.2.6) Calculate the loss L3 between the domain classification prediction result and the true domain classification label:

[0093]

[0094] Among them, H is the number of receiver individuals, and L is the total number of samples. is the network prediction probability that the lth sample in the target domain comes from the kth receiver. is the true probability that the lth sample belongs to the kth receiver. If it belongs to the kth receiver, it is 1; if it does not belong to the kth receiver, it is 0.

[0095] (4.3) Sum all the losses to obtain the total loss L of the network, L = L1 + L2 + L3, and perform backpropagation accordingly. Use the gradient descent method to optimize the network parameters to complete one iteration.

[0096] (4.4) Repeat steps (4.2)-(4.3) until i ≥ I is satisfied, stop the iteration, and obtain the optimized network model O * .

[0097] Step 5, input the test sample T into the trained network model O * , and obtain the recognition result L through the individual recognition branch p , and complete the recognition of the radar emitter individuals across receivers.

[0098] The technical effects of the present invention will be further described below in combination with simulation experiments:

[0099] 1. Experimental conditions:

[0100] Hardware environment: The CPU is Inter(R)core(TM)i7-10700F with a main frequency of 2.90GHz, the memory is 16.0GB, and the operating system is 64-bit; the graphics card is NVIDIA RTX2060; Software environment: Microsoft windows 10 Professional Edition, Python3.8, SystemVue10.0.

[0101] Three radar radiation sources and two signal receivers are set. The signals emitted by the radar radiation sources are all single-frequency pulses with the same amplitude, a pulse width of 2 μs, and a carrier frequency of 250 MHz. The hardware performance parameters of different radar radiation sources and receivers are different. The mixing parameters of the devices are shown in Table 1, the hardware performance parameter settings of the transmitter are shown in Table 2, and the hardware performance parameter settings of the receiver are shown in Table 3:

[0102] Table 1 Mixing parameter settings

[0103] <![CDATA[Transmitter intermediate frequency f EI > 50 MHz <![CDATA[Radio frequency f R > 200 MHz <![CDATA[Intermediate frequency of the receiver, f RI > 120 MHz Receiver Bandwidth 50 MHz

[0104] Table 2 Hardware performance parameter settings of the transmitter

[0105]

[0106]

[0107] Table 3 Hardware performance parameter settings of the receiver

[0108] Receiver 1 Receiver 2 Amplifier 1 Third-Order Intercept Point / dBm 45 35 1 dB Compression Point of Amplifier 1 / dBm 30 20 Mixer Second-Order Intercept Point / dBm 40 30 Mixer Third-Order Intercept Point / dBm 30 20 Amplifier 2 Third-Order Intercept Point / dBm 45 35 1 dB Compression Point of Amplifier 2 / dBm 30 20

[0109] Referring to the dataset partitioning method of previous studies, 200 samples are generated for each radiation source, of which 120 samples are used as training data and 80 samples are used as test data. Two receivers are set, and the data of receiver 1 is used as the source domain data, and the data of receiver 2 is used as the target domain data.

[0110] The recognition effect is evaluated by the recognition accuracy rate. The calculation formula for the recognition accuracy rate is:

[0111]

[0112] II. Simulation content

[0113] Simulation 1: Under the condition that the signal-to-noise ratio is 40 dB, the signals generated by radar radiation sources 1, 2, and 3 set in Table 2 are respectively passed through receiver 1 and receiver 2 to obtain received signals. The double-spectrum contour integral transform is performed on all received signals using the present invention to obtain the pulse feature domain, and the feature differences between different transmitters and the influence of the receiver on the radiation source features are simulated. The results are as Figure 5 shown, where Figure 5(a) The feature domains of the signals generated by radar radiation sources 1, 2, and 3 passing through receiver 1 from left to right in sequence; Figure 5 (b) The feature domains of the signals generated by radar radiation sources 1, 2, and 3 passing through receiver 2 from left to right in sequence.

[0114] Simulation 2: Under the condition that the signal-to-noise ratio is 20 dB, the signals generated by radar radiation sources 1, 2, and 3 set in Table 2 pass through receiver 1 and receiver 2 respectively to obtain received signals. The present invention is used to perform bispectrum contour integral transformation on all received signals to obtain the pulse feature domain, and the influence of simulation noise and the receiver on the radiation source characteristics is simulated. The results are as Figure 6 shown, Figure 6 (a) is the feature domain of the signals generated by radar radiation sources 1, 2, and 3 passing through receiver 1, Figure 6 (b) is the feature domain of the signals generated by radar radiation sources 1, 2, and 3 passing through receiver 2.

[0115] Simulation 3: Under the conditions that the signal-to-noise ratios are 5 dB, 10 dB, 20 dB, 30 dB, and 40 dB respectively, the signals generated by radar radiation sources 1, 2, and 3 set in Table 2 pass through receiver 1 and receiver 2 respectively to obtain received signals. The present invention and the existing radar radiation source individual recognition technology based on one-dimensional convolutional neural network are used to perform simulation recognition on them respectively, and the recognition accuracies of the two technologies under different signal-to-noise ratios are obtained. The results are as Figure 7 shown. In the figure, 1D-DANN is the present invention, and 1D-CNN is the existing technology.

[0116] 3. Analysis of simulation results:

[0117] From Figure 4 and 5 it can be seen that under the condition of relatively high signal-to-noise ratio, after the signals of the same radiation source pass through different receivers, different degrees of distortion occur in the feature domain. As the hardware performance of the receiver decreases, the degree of distortion will also increase. When there is noise and the noise power is greater than the unintentional pollution power generated by the receiver, the noise becomes the main factor polluting the feature domain.

[0118] From Figure 6 it can be seen that when the signal-to-noise ratio is higher than 20 dB, the recognition accuracy of the existing technology is between 80% and 90%, and the recognition accuracy of the present invention is higher than 95%; when the signal-to-noise ratio is 5 dB, the recognition accuracy of the existing technology drops to 50%, while the present invention still has a recognition accuracy of 79%. The simulation results show that: under different signal-to-noise ratio conditions, the recognition effect of the present invention is more than 8% higher than that of the existing algorithm, verifying that the present invention has a good optimization effect on both the hardware characteristics and the feature domain drift caused by noise, and has a good recognition effect on radar radiation source signals across receivers.

Claims

1. A method for identifying individual radar radiation sources across receivers based on deep learning, characterized in that, The following are included: (1) Obtain pulse signal data from different receivers, define the domain labels for each pulse signal, and delimit the source domain data and the target domain data; respectively divide more than half of the source domain data and the target domain data into the training data set Q, and the remaining into the test data set T; (2) Perform bispectrum transformation processing on each pulse signal to obtain the reduced-dimensional one-dimensional feature domain SIB; (3) Construct a radar emitter individual recognition model O based on deep learning: (3a)Construct a feature extractor G by cascading a first-level one-dimensional convolutional layer, a max pooling layer 1, a second-level one-dimensional convolutional layer, a max pooling layer 2, and a flattening layer in sequence f ; (3b) Establish an individual recognition classifier G composed of a cascaded first-level fully connected layer, a second-level fully connected layer, and a Softmax classifier in sequence y ; (3c) Establish a domain recognition classifier G composed of a gradient reversal layer R, a first-level fully connected layer, a second-level fully connected layer, and a Softmax classifier cascaded in sequence d ; (3d) Cascade G f with G y to form an individual recognition branch, and then cascade G f with G d to form a domain recognition branch; (4) Iteratively train the network recognition model O to obtain the optimized model O * : (4a) Initialize the iteration number i, set the maximum iteration number as I, where I ≥ 50, set the loss function for network optimization as the cross-entropy function, and the learning rate as μ p , and initialize the network parameters of O as θ; (4b) Randomly select a batch of data from the source domain training data and input it into the network model O, and use the network prediction probability that the nth sample in the source domain of the individual recognition branch belongs to the cth radiation source individual Obtain the individual classification prediction label Lf1 * , and use the network prediction probability that the nth sample in the source domain of the domain recognition branch comes from the kth receiver Obtain the domain prediction label Calculate the loss L1 between the individual classification prediction label Lf1 * and the actual classification label Lf1, and the loss L2 between the domain prediction label and the actual domain label Ld1; (4c) Randomly select a batch of data from the target domain training data and input it into the network model O. Use the domain recognition branch to calculate the network prediction probability that the l-th sample in the target domain comes from the k-th receiver. Obtain the domain prediction label Calculate this domain prediction label And the loss L3 with the actual domain label Ld2; (4d) Sum all the losses to obtain the total loss L of the network, L = L1 + L2 + L3, and perform backpropagation accordingly, and use the gradient descent method to optimize the network parameters to complete one iteration; (4e) Repeat steps (4b)-(4d) until i≥I is satisfied, stop the iteration, and obtain the optimized network model O * ; (5) Input the test sample T into the trained network model O * , and use the individual recognition branch to calculate the recognition result L p .

2. The method according to claim 1, wherein In the above (1), defining the domain labels for each pulse signal and delimitating the source domain data and the target domain data are realized as follows: (1a) Define the domain label of the pulse according to the receiver source of the pulse signal, that is, assign the same domain label to the pulse signals received from the same receiver, and assign different domain labels to the signals from different receivers; (1b) Delimit the source domain data and the target domain data according to whether the pulse signal has an individual recognition label: If the pulse has an individual recognition label, it is delimitated as the source domain data; Conversely, it is delimitated as the target domain data to be recognized.

3. The method according to claim 1, wherein In the above (2), performing bispectrum transformation processing on each pulse signal received by different receivers is realized as follows: (2a) Perform zero-mean normalization on the pulse signal with a sampling frequency of f s to obtain the processed pulse sampling sequence as x(0), x(1),... x(N-1), and divide it into K segments, each segment containing M sampling points. The data of the k-th segment is x (k) (0), x (k) (1),... x (k) (M-1); (2b) Calculate the discrete Fourier transform DFT coefficients of each segment of data: (2c) Calculate the triple correlation of the DFT coefficients to obtain the bispectrum estimate of this data segment where Δ0 = f s / N0, N0 and L1 satisfy M = (2L1 + 1)N0, and λ1, λ2 are sliding independent variables in two different dimensions of related operations; (2d) Take the mean of the bispectrum estimates of all K-segment signals to obtain the bispectrum estimate of the entire signal segment wherein (2e)Perform contour integration on the bispectrum estimate value to obtain the reduced-dimension bispectrum feature domain SIB(ω) 4. The method according to claim 1, wherein The feature extractor G constructed in step (3a) f , and its parameters for each layer are as follows: The number of channels of the first convolutional layer is 16, the convolutional kernel size is 1x32, and the activation function is the ReLU function; The number of channels of the second convolutional layer is 16, the convolutional kernel size is 1x25, and the activation function is the ReLU function; The pooling kernel sizes of the two max pooling layers are both 1x2; The number of nodes in the flattening layer is 192.

5. The method according to claim 1, wherein The individual recognition classifier G constructed in step (3b) y has the following parameters for each layer: The number of nodes in the first-level fully connected layer is 192, and the activation function is the Sigmoid function; The number of nodes in the second-level fully connected layer is C, and the activation function is the Sigmoid function, and the number of nodes C is the number of radar emitter individuals included in the data set.

6. The method according to claim 1, wherein The domain recognition classifier G constructed in step (3c) d has the following structure and parameters for each layer: The number of nodes in the first-level fully connected layer is 192, and the activation function is the Sigmoid function; The number of nodes in the second-level fully connected layer is N, and the activation function is the Sigmoid function, where N is the number of receivers for receiving data; When the gradient reversal layer R performs forward calculation of the network, its output is equal to the input x, that is, R(x) = x, When optimizing the network by backpropagation, R takes the negative of the loss gradient of the input: In the formula, p is the ratio of the current iteration number i to the total iteration number I, representing the iteration process of the network, Loss is the network loss during backpropagation, and γ is the constant 10.

7. The method according to claim 1, characterized in that, The learning rate μ in step (4a) p , and the calculation formula is as follows: where μ0 is the initial value of μ p , α is the multiplication hyperparameter, and β is the exponential hyperparameter.

8. The method according to claim 1, characterized in that Calculate the individual classification prediction label Lf1 in step (4b) * The loss L1 with the actual classification label Lf1 is as follows: where U is the number of radar radiation source individuals, and N is the total number of samples. is the network prediction probability that the n-th sample in the source domain belongs to the c-th label. is the true probability that the n-th sample in the source domain belongs to the c-th label. If it belongs to the c-th radiation source, is 1; if it does not belong to the c-th radiation source, is 0.

9. The method according to claim 1, characterized in that, Calculate the loss L2 between the domain classification prediction label in step (4b) and the actual classification label Ld1, with the formula as follows: where \(H\) is the number of receivers and \(M\) is the total number of samples, is the network prediction probability that the \(n\)th sample in the source domain comes from the \(k\)th receiver, is the true probability that the \(n\)th sample in the source domain belongs to the \(k\)th receiver. If it belongs to the \(k\)th receiver, it is 1; if it does not belong to the \(k\)th receiver, it is 0.

10. The method according to claim 1, characterized in that, Calculate the domain prediction label in step (4c) The loss L3 with the actual domain label, and the formula is as follows where H is the number of receivers and L is the total number of samples, is the network prediction probability that the l-th sample in the target domain comes from the k-th receiver, r l k is the true probability that the l-th sample belongs to the k-th receiver. If it belongs to the k-th receiver, r l k is 1; if it does not belong to the k-th receiver, r l k is 0.

11. The method according to claim 8, wherein The network prediction probability in the L1 formula is calculated as follows: First, the input feature domain SIB passes through the feature extractor G with parameter θ1 f , and outputs the feature vector f: f = G f (SIB; θ1); Next, input the feature vector f into the individual recognition classifier G with parameter θ2 y for classification prediction to obtain a classification prediction vector of length C Among them, C is the number of individual radar radiation sources, indicating the likelihood that the nth source domain input sample belongs to each individual radar radiation source; Finally, let the c-th point of c be z and calculate the probability that the n-th source domain input sample belongs to each classification label according to the classification prediction vector is the network prediction probability that the nth sample in the source domain belongs to the cth label, and F is the total number of labels, where

Images