A method and system for identifying radiation source individuals based on transfer learning

Abstract

The present invention discloses a method and system for identifying radiation source individuals based on transfer learning, belonging to the technical field of radiation source individual identification. First, bispectrum transformation is used for feature extraction to extract fingerprint information in the signal; then, a convolutional neural network is used to learn the individual data of the source and target domains to obtain a network model; finally, the source domain model is migrated through a domain adaptation transfer learning method based on model parameter migration, maximum mean discrepancy, and hyperbolic space, realizing the intelligent migration of radiation source individuals. The present invention can have good recognition accuracy at a low signal-to-noise ratio and still has good recognition ability under different channels, solves the problems of network cold start and slow convergence, and improves the accuracy of radiation source individual identification at the same time.

Description

Technical Field

[0001] The present invention belongs to the technical field of radiation source individual recognition, and particularly relates to a radiation source individual recognition method and system based on transfer learning. Background Art

[0002] Radiation source individual recognition refers to identifying a target individual by extracting one or more modulation features exhibited by the received signal. As the number of air target individuals increases day by day, it is necessary to judge whether the airborne target radiation source is working properly and the enemy / foe attribute of abnormal airborne targets. All these judgments require the individual recognition of airborne radiation sources as a prerequisite, and how to make a quick and accurate judgment is an important means to achieve national security.

[0003] Techniques for identifying individual radiation sources based on fingerprint information began in the last century. Sa K, Lang D et al. proposed a method that uses a constellation diagram to extract fingerprint information and inputs it into a convolutional neural network for identification. However, this method has relatively strict requirements for time synchronization information (Sa K, Lang D, Wang C, et al. Specific Emitter Identification Techniques for the Internet of Things[J]. IEEE Access, 2020, 8:1644-1652.). Song C, Xu J et al. proposed a method for analyzing non-stationary signals using empirical mode decomposition, but this method has the problem of mode mixing (Song C, Xu J, Zhan Y. A method for specific emitter identification based on empirical mode decomposition[C] / / IEEE International Conference on Wireless Communications. IEEE, 2010.). Gok G, Alp Y K et al. proposed a method based on variational mode decomposition (VMD) that uses the envelope and instantaneous frequency of the received signal as a set of models to identify different radiation source signals. However, this method has a high complexity and is not conducive to practical operation (Gok G, Alp Y K, Arikan O. A New Method for Specific Emitter Identification With Results on Real Radar Measurements[J]. IEEE Transactions on Information Forensics and Security, 2020, PP(99):1-1.). Shieh C S et al. proposed using statistics such as the direction of arrival (DOA), pulse width (PW), pulse repetition frequency (PRF), and radar frequency (RF) of the radiation source's conventional parameters as the basis for classification and identification. However, this method has poor identification accuracy at low signal-to-noise ratios (Shieh C S, Lin C T. A vector neural network for emitter identification[J]. IEEE Transactions on Antennas and Propagation, 2002, 50(8):1120-1127.).Dan Xu, Bo Yang et al. proposed a kernel principal component analysis (KPCA) prediction learning method, which solved the problem of processing different data clusters with complex non-linear distributions. However, this method is difficult to be applied to other communication scenarios under non-Gaussian channels (Dan Xu, Bo Yang, Wenli Jiang,. An improved SVDU-IKPCA algorithm for Specific Emitter Identification[C] / / International Conference on Information &Automation. IEEE, 2008.). Yue Chen et al. proposed the individual identification of communication radiation sources based on IQ diagram features. However, the features extracted by this method are not obvious, and its performance is very poor under low signal-to-noise ratio (Yue Chen, Yingke Lei, Xin Li, Ling Ye, Fan Mei. Individual Identification of Communication Radiation Sources Based on IQ Diagram Features[J]. Signal Processing). Peng Chen et al. proposed a method using the Frechet distance to calculate the distance between signals, pulse envelopes or instantaneous frequencies, so as to realize the individual identification of radiation sources. This method also has poor identification performance under low signal-to-noise ratio (P. Chen, G. Li, K. Xu and J. Wan, "Applying the frechet distance to the specific emitter identification," 2016 IEEE 13th International Conference on Signal Processing(ICSP), Chengdu, 2016, pp.1027-1030.). D'Agostino S proposed an algorithm for feature extraction of signals using the short-time discrete Fourier transform (STDFT), and finally used the clustering algorithm for classification and identification. However, this method is highly dependent on the selection of sliding windows, and the identification accuracy of the clustering algorithm is relatively low (D'Agostino S. Specific emitter identification based on amplitude features[C] / / IEEE International Conference on Signal &Image Processing Applications. IEEE, 2015.).

[0004] The above method solves the problem of individual identification of radiation sources to a certain extent, but the identification performance is poor in a low signal-to-noise ratio environment, and the generalization ability in different channel environments is insufficient; in addition, traditional training strategies require a large amount of high-quality data as training data, but such idealized data cannot be obtained in the current environment, and the network has problems of cold start and slow convergence, so the traditional training method is not applicable to the actual environment.

[0005] The difficulties in solving the above problems and defects are as follows: due to the influence of the current complex electromagnetic environment and channels, it is difficult for common airborne radiation source individual signals to reach a high signal-to-noise ratio. It is normal for the signal-to-noise ratio of the signal to be low and the interference to be strong. It is difficult to find a feature extraction method with strong anti-interference ability and good identification ability for most channels; data needs to be saved and archived, and data storage has a large demand for memory space. At the same time, high-quality data is scarce and difficult to obtain, which is not conducive to improving the training speed and maximizing the utilization of resources. Summary of the Invention

[0006] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a method and system for individual identification of radiation sources based on transfer learning, which can accurately identify individual radiation sources in a complex electromagnetic environment with low signal-to-noise ratio and strong interference, and use the previous training results to quickly update the new individual data identification system, so as to improve the electromagnetic situation awareness ability.

[0007] In order to achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0008] A method for individual identification of radiation sources based on transfer learning uses bispectrum transformation for feature extraction to extract fingerprint information in the signal; learns the individual data of the source target domain through a convolutional neural network to obtain a source network model; migrates the source network model through a domain adaptation transfer learning method based on model parameter migration, maximum mean discrepancy, and hyperbolic space.

[0009] The method for individual identification of radiation sources based on transfer learning includes the following steps:

[0010] Step 1: First, collect training data and perform screening and noise reduction preprocessing on it;

[0011] Step 2: Use bispectrum transformation to extract features from the preprocessed signal, perform downsampling and cropping operations on the obtained feature map to obtain an input of n*1*256*256, where n represents the total amount of data;

[0012] Step 3: Use the individual data of the source target domain to train the convolutional neural network, add a hyperbolic space network layer before the Soft-Max layer of the convolutional neural network to obtain the classification loss of the source network, and retain the parameter information of the source network model;

[0013] Step 4: Migrate the source network model to the target network and freeze the first five convolutional layers;

[0014] Step 5: Add an adaptation error layer and a hyperbolic space layer to the target network, and train the target network by combining a small part of the source data and the target data. Use the classification loss and the MMD loss as the new loss functions to update the network parameters and perform optimization to obtain the final network model after iteration and optimization.

[0015] The processing of the signal using the bispectrum transform described in Step 2 includes:

[0016] For the received signal r(n), first divide it into Γ segments of signals, and the third-order cyclic cumulant is:

[0017]

[0018] where χ γ (τ1,τ2) is the third-order cyclic cumulant of each segment of the signal, τ1,τ2 are two different time delays, then the bispectrum estimation is expressed as:

[0019]

[0020] where δ < Δ - 1, ω(τ1,τ2) is the hexagonal window function, then the dimension reduction of the bispectrum is expressed as:

[0021]

[0022] where {U1,...,U M} is the compressed bispectrum, {V1,...V M} is the compressed projection space of the original selected bispectrum; s and d are the bispectrum weight matrices of signals from different or the same transmitters. When s ≠ d, s ij = 0; when s = d, ε is a positive real number, and the above formula is transformed into:

[0023]

[0024] where, o ii = ∑ j s ij , f ii = ∑ j d ij , is matrix multiplication;

[0025] Let If δ is a non-zero constant, the Lagrangian equation is expressed as:

[0026]

[0027] To make minimum, let we get:

[0028]

[0029] Calculate W = [ω1,..., w d , then the compressed bispectrum is calculated by the following formula:

[0030] U i = V i W.

[0031] The convolutional neural network learning described in step three includes:

[0032] First, perform bispectrum transform feature extraction on the signal, and then the convolutional layer learns the input data; the convolutional layer contains multiple convolutional kernels inside, and each element that makes up the convolutional kernel corresponds to a weight coefficient and a bias. Each neuron in the convolutional layer is connected to multiple neurons in a region close to its position in the previous layer, and the size of the region depends on the size of the convolutional kernel. The calculation formula is:

[0033]

[0034]

[0035] where b is the bias, ω is the weight value, x and y represent the convolutional kernel performing convolutional processing on the entire feature map, Z l and Z l+1 represent the convolutional input and output of the (l + 1)-th layer, also known as the feature map, L l+1 is the size of Z l+1 , assuming the length and width of the feature map are equal; Z(i, j) corresponds to the pixel of the feature map, K is the number of channels of the feature map, f is the size of the convolutional kernel, s0 is the convolutional stride, and p is the padding size; the convolutional layer contains an activation function to assist in expressing complex features. The rectified linear unit RELU is used to make the neurons in the convolutional neural network have sparse activation. Its representation form is as follows:

[0036]

[0037] Secondly, after feature extraction in the convolutional layer, the output features will be passed to the pooling layer for feature selection and information filtering; after the convolutional layer, a pooling layer operation is added. The pooling layer selects the pooling region in the same way as the convolutional kernel scans the feature map, which is controlled by the pooling size, stride, and padding. The expression is as follows:

[0038]

[0039] In the formula, A k represents the input feature map, and other parameters are the same as those of the convolutional layer.

[0040] The transfer learning method of the maximum mean discrepancy (MMD) described in step five includes:

[0041] First, define a feature transformation T to adjust the joint distribution so that the source domain and the target domain are jointly mapped into the feature space;

[0042]

[0043] Among them, Q(y|x) is the conditional probability distribution. Then, kernel principal component analysis (PCA) is adopted to reduce the dimension and reconstruct the error, that is, to find an orthogonal transformation matrix A:

[0044]

[0045] Among them, I is the identity matrix, X is the input feature matrix, where tr(·) represents the trace of the matrix, and this formula is optimized as the eigenvalue decomposition of XHX T A = A T The effective solution of φ, where φ is the k largest eigenvalues;

[0046] Adopt the maximum mean discrepancy to measure the distance between domains, which belongs to non-parametric measurement and is estimated as follows:

[0047]

[0048] Among them, X s is the source domain sample set, X t is the target domain sample set, φ(x) maps each sample to the Hilbert space H associated with the kernel k(x i , x j ) = φ(x i ), T φ(x j ), n s , n t represent the sizes of the source domain and the target domain sample sets respectively; substituting into PCA, we get:

[0049]

[0050] It is simplified to:

[0051] MMD(X s , X t ) = tr(A T XM0XT A);

[0052] where tr(·) is the trace of a matrix, and M0 is defined as follows:

[0053]

[0054] Minimize the loss function, i.e.:

[0055]

[0056] where represents the classification loss of the model on the true label y and the input data X L MMD(X S , X T ) represents the distance between the source data X S and the target data X T , and λ is a hyperparameter that determines the strength of the confusion domain.

[0057] The domain adaptation in the hyperbolic space described in step five includes:[[]]

[0058] The Poincaré ball model space of the hyperbolic space is defined by the manifold , where τ is a hyperparameter representing the curvature of the Poincaré ball; in the Poincaré ball model, the induced distance between any two points z i , is expressed as:[[]]

[0059]

[0060] Add a hyperbolic network layer at the end of the original deep learning network, and finally map the Euclidean space R n to the hyperbolic manifold which is expressed as:[[]]

[0061]

[0062] A radiation source individual recognition system based on transfer learning, comprising:[[]]

[0063] Feature extraction module: First, collect the training data, and perform screening and noise reduction preprocessing on it; perform feature extraction on the preprocessed signal using bispectral transformation, and perform downsampling and cropping operations on the obtained feature map to obtain an input of n*1*256*256, where n represents the total amount of data;

[0064] Convolutional neural network learning module: Use the individual data of the source and target domains to train the convolutional neural network. Add a hyperbolic space network layer before the Soft-Max layer of the convolutional neural network to obtain the classification loss of the source network, and retain the source network model parameter information;

[0065] Target domain migration module: Migrate the source network model to the target network and freeze the first five convolutional layers; Add an adaptation error layer and a hyperbolic space layer to the target network, and jointly train the target network with a small part of the source data and target data. Use the classification loss and the MMD loss as the new loss functions to update the network parameters and perform optimization to obtain the final network model through iteration and optimization.

[0066] An information data processing terminal is used to implement the radiation source individual recognition system based on transfer learning described above.

[0067] Combining all the above technical solutions, the advantages and positive effects of the present invention are as follows: The present invention solves the problems of network cold start and slow convergence, has the advantages of good recognition accuracy at low signal-to-noise ratios and good recognition performance under different channels, and realizes the intelligent migration of radiation source individual recognition. Description of the Drawings

[0068] Figure 1 is the flowchart of the radiation source individual recognition method based on transfer learning in the embodiment of the present invention.

[0069] Figure 2 is the training accuracy graph of transfer learning in the embodiment of the present invention.

[0070] Figure 3 is the schematic diagram of the recognition accuracy of the radiation source individual recognition based on transfer learning in the embodiment of the present invention under different Doppler frequency shifts.

[0071] Figure 4 is the schematic diagram of the recognition performance of 8 types of radiation source individuals under different signal-to-noise ratios in the embodiment of the present invention. Detailed Embodiment

[0072] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0073] As Figure 1 shown, a radiation source individual recognition method based on transfer learning includes:

[0074] S101, Use bispectral transformation for feature extraction to extract the fingerprint information in the signal;

[0075] S102, Learn the individual data of the source and target domains through a convolutional neural network to obtain a network model;

[0076] S103. Migrate the source domain model through a domain adaptation transfer learning method based on model-based parameter transfer, maximum mean discrepancy, and hyperbolic space to achieve intelligent transfer of radiation source individuals.

[0077] A radiation source individual recognition system based on transfer learning, comprising:

[0078] Feature extraction module: First, collect training data, and perform screening and noise reduction preprocessing on it; perform feature extraction on the preprocessed signal using bispectrum transformation, and perform downsampling and cropping operations on the obtained feature map to obtain an input of n*1*256*256, where n represents the total amount of data;

[0079] Convolutional neural network learning module: Use the individual data of the source and target domains to train the convolutional neural network. Add a hyperbolic space network layer before the Soft-Max layer of the convolutional neural network to obtain the classification loss of the source network, and retain the parameter information of the source network model;

[0080] Target domain migration module: Migrate the source network model to the target network and freeze the first five convolutional layers; add an adaptation error layer and a hyperbolic space layer to the target network, and jointly train the target network with a small part of the source data and target data. Use the classification loss and MMD loss as the new loss function to update the network parameters and perform optimization solving. After iteration and optimization, obtain the final network model.

[0081] The technical solution of the present invention will be further described below in conjunction with embodiments.

[0082] The described radiation source individual recognition method based on transfer learning includes the following steps:

[0083] S101. Use bispectrum transformation for feature extraction to extract fingerprint information in the signal;

[0084] First, collect training data, and perform a series of preprocessing such as screening and noise reduction on it; perform feature extraction on the preprocessed signal using bispectrum transformation, and perform downsampling and cropping operations on the obtained feature map to obtain an input of n*1*256*256, where n represents the total amount of data;

[0085] For the received signal r(n), first divide it into Γ segments of signals, and the third-order cyclic cumulant is:

[0086]

[0087] where χ γ (τ1,τ2) is the third-order cyclic cumulant of each segment of signal, τ1,τ2 are two different time delays, then the bispectrum estimation is expressed as:

[0088]

[0089] Among them, δ < Δ - 1, ω(τ1, τ2) is the hexagonal window function, then the dimensionality reduction of the bispectrum can be expressed as:

[0090]

[0091] Among them, {U1,..., U M} is the compressed bispectrum, and {V1,... V M} is the compressed projection space of the original selected bispectrum. s and d are the bispectrum weight matrices of signals from different or the same transmitters. When s ≠ d, s ij = 0; when s = d, ε is a positive real number, and the above formula can be transformed into:

[0092]

[0093] Among them, o ii = ∑ j s ij , f ii = ∑ j d ij , is matrix multiplication;

[0094] Let δ be a non - zero constant, then the Lagrangian equation can be expressed as:

[0095]

[0096] To minimize Let We get:

[0097]

[0098] Calculate W = [ω1,..., w d , then the compressed bispectrum can be calculated by the following formula:

[0099] U i = V i W;

[0100] S102. Learn the individual data of the source - target domain through a convolutional neural network to obtain a network model;

[0101] First, perform bispectrum transform feature extraction on the signal, and then let the convolutional layer learn the input data. The convolutional layer contains multiple convolutional kernels inside. Each element that makes up the convolutional kernel corresponds to a weight coefficient and a bias. Each neuron in the convolutional layer is connected to multiple neurons in a region with a close position in the previous layer. The size of the region depends on the size of the convolutional kernel, and the calculation formula is:

[0102]

[0103]

[0104] Among them, b is the bias, ω is the weight value, x and y represent that the convolutional kernel performs convolutional processing on the entire feature map, Z l and Z l+1 represent the convolutional input and output of the (l + 1)-th layer, and are also called feature maps. L l+1 is the size of Z l+1 . Assume that the length and width of the feature map are equal; Z(i, j) corresponds to the pixel of the feature map, K is the number of channels of the feature map, f is the size of the convolutional kernel, s0 is the convolutional stride, and p is the padding size. The convolutional layer contains an activation function to assist in expressing complex features. Use the rectified linear unit RELU to make the neurons in the convolutional neural network have sparse activation. Its representation form is as follows:

[0105]

[0106] Secondly, after feature extraction in the convolutional layer, the output features will be passed to the pooling layer for feature selection and information filtering. After adding a pooling layer operation after the convolutional layer, the pooling layer selects the pooling area in the same step as the convolutional kernel scans the feature map, which is controlled by the pooling size, stride, and padding. The expression is as follows:

[0107]

[0108] In the formula, A k represents the input feature map, and other parameters are the same as those in the convolutional layer;

[0109] S103. Through the domain adaptation transfer learning method based on model parameter transfer, maximum mean discrepancy, and hyperbolic space, transfer the source domain model to achieve intelligent transfer of radiation source individuals.

[0110] First, define the feature transformation T to adjust the joint distribution so that the source domain and the target domain are jointly mapped into the feature space;

[0111]

[0112] Among them, Q(y|x) is the conditional probability distribution. Then, the present invention adopts kernel principal component analysis (PCA) to perform dimensionality reduction and reconstruction on the error, that is, to find an orthogonal transformation matrix A:

[0113]

[0114] Among them, I is the identity matrix, X is the input feature matrix, where tr(·) represents the trace of the matrix. This formula can be optimized to the eigenvalue decomposition of XHX T A = A T The effective solution of φ, where φ is the k largest eigenvalues;

[0115] The distribution difference between domains still needs to be further optimized. The maximum mean discrepancy is adopted to measure the distance between domains, which belongs to non-parametric measurement and is estimated as follows:

[0116]

[0117] Among them, X s is the source domain sample set, X t is the target domain sample set, and φ(x) maps each sample to the Hilbert space associated with the kernel k(x i , x j ) = φ(x i ). T φ(x j ) In the Hilbert space, n s , n t respectively represent the sizes of the source domain and target domain sample sets. Substituting into PCA, we can get:

[0118]

[0119] It is simplified to:

[0120] MMD(X s , X t ) = tr(A T XM0X T A);

[0121] Among them, tr(·) is the trace of the matrix, and M0 is defined as follows:

[0122]

[0123] In addition to MMD, it is also hoped that the paradigm of transfer learning can contribute to the representation of the classifier. In this way, it is necessary to minimize the loss function, that is:

[0124]

[0125] Among them, The classification loss of the representative model on the true label y and the input data X L , MMD(X S , X T ) represents the distance between the source data X S and the target data X T . λ is a hyperparameter that determines the strength of the confusion domain.

[0126] Domain adaptation in hyperbolic space includes:

[0127] The Poincaré ball model space in hyperbolic space, which is defined by a manifold , where τ is a hyperparameter representing the curvature of the Poincaré ball; in the Poincaré ball model, the induced distance between any two points can be expressed as:

[0128]

[0129] Adding a hyperbolic network layer at the end of the original deep learning network, and finally mapping the Euclidean space to the hyperbolic manifold is expressed as:

[0130]

[0131] The technical effects of the present invention will be described in detail below in combination with simulation experiments.

[0132] To evaluate the performance of the present invention, simulation verification is carried out. Consider a secondary radar transceiver signal MARK, and the simulation parameters are set as follows: the center frequency f0 of the MARK-X interrogation signal is 50 MHz, the carrier frequency f c is 20 MHz, the sampling frequency f s of the MARK-XII interrogation signal is 20 MHz, and the carrier frequency f c is 5 MHz. The MARX-X response signal consists of 16-bit information code bits, and the numbers of each bit are in turn: F1, C1, A1, C2, A2, C4, A4, X, B1, D1, B2, D2, B4, D4, F2, and SPI. Each information code is a pulse, and there are only two levels of "0" and "1". The MARK-XII response signal consists of one code word for every three pulses, with a pulse group interval of 4 μs and a total of 16 groups. The sampling frequencies f s of the MARK-X response signal and the MARK-XII response signal are both 20 MHz, and the carrier frequency f c is 5 MHz.

[0133] Figure 2Shows the training accuracy of the radiation source individual recognition method based on transfer learning. The simulation results of the method proposed in the present invention under different Doppler frequency shifts are as follows Figure 3 shown, from Figure 3 it can be seen that after the Doppler frequency shift is mixed in, the recognition accuracy is affected to a certain extent, but the individual recognition accuracy still remains above 90%, indicating that after transfer learning, the robustness of the network against the Doppler frequency shift has been improved. Figure 4 Gives the recognition accuracy change of different signals with the signal-to-noise ratio. From Figure 4 it can be seen that except for the MARK-X response signal, the recognition accuracy of the 8 types of radiation source individuals changes significantly with the signal-to-noise ratio, and the changes in the recognition performance of the other 7 types of radiation source individuals are not significant, indicating that the method proposed in the present invention has a certain noise suppression ability.

[0134] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented in whole or in part in the form of a computer program product, the computer program product includes one or more computer instructions. When the computer program instructions are loaded or executed on a computer, the processes or functions described in the embodiments are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wire (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that the computer can access or a data storage device such as a server or data center that includes one or more integrated available media. The available medium can be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)).

[0135] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention shall be covered by the protection scope of the present invention.

Claims

1. A method for identifying radiation source individuals based on transfer learning, characterized in that: Feature extraction is carried out using bispectrum transformation to extract fingerprint information in the signal; the individual data of the source target domain is learned through a convolutional neural network to obtain a source network model; the source network model is migrated through a domain adaptation transfer learning method based on model parameter transfer, maximum mean discrepancy (MMD), and hyperbolic space; The transfer learning method of the maximum mean discrepancy (MMD) includes: First, a feature transformation T is defined to adjust the joint distribution so that the source domain and the target domain are jointly mapped into the feature space; Among them, Q(y|x) is the conditional probability distribution. Then, kernel principal component analysis (PCA) is used to reduce the dimension and reconstruct the error, that is, to find an orthogonal transformation matrix A: where I is the identity matrix, X is the input feature matrix, where tr(·) represents the trace of a matrix, and this formula is optimized as the eigenvalue decomposition of XHX T A = A T the valid solution of φ, where φ are the k largest eigenvalues; The maximum mean discrepancy is used to measure the distance between domains, which belongs to non-parametric measurement and is estimated as follows: Among them, X s is the source domain sample set, and X t is the target domain sample set. φ(x) maps each sample to the Hilbert space associated with the kernel k(x i , x j ) = φ(x i ). T In the Hilbert space associated with φ(x j ), n , and n s , n t represent the sizes of the source domain and target domain sample sets respectively; substituting into PCA gives: It is simplified to the following through the kernel function: MMD(X s ,X t ) = tr(A T XM0X T A); Among them, tr(·) is the trace of the matrix, and M0 is defined as follows: Reduce the loss function, that is: Among them, represents the classification loss of the model on the true label y and the input data X L MMD(X S , X T ) represents the distance between the source data X S and the target data X T ; λ is a hyperparameter that determines the strength of the confusion domain. The domain adaptation transfer learning method in hyperbolic space includes: The Poincaré ball model of hyperbolic space, through the manifold is defined, where τ is a hyperparameter representing the curvature of the Poincaré ball; in the Poincaré ball model, for any two points the induced distance is expressed as: Add a hyperbolic network layer at the end of the original deep learning network, and finally map the Euclidean space to the hyperbolic manifold which is expressed as:

2. The method according to claim 1, characterized in that, It includes the following steps: Step 1: First, collect the training data and perform screening and noise reduction preprocessing on it; Step 2: Use bispectrum transformation to extract features from the preprocessed signal, perform downsampling and cropping operations on the obtained feature map to obtain an input of n*1*256*256, where n represents the total amount of data; Step 3: Use the individual data of the source target domain to train the convolutional neural network. A hyperbolic space network layer is added before the Soft-Max layer of the convolutional neural network to obtain the classification loss of the source network and retain the parameter information of the source network model; Step 4: Migrate the source network model to the target network and freeze the first five convolutional layers; Step 5: Add an adaptation error layer and a hyperbolic space layer to the target network, jointly train the target network with a small part of the source data and target data, use the classification loss and MMD loss as the new loss function to update the network parameters and perform optimization solution. After iteration and optimization, the final network model is obtained.

3. The method according to claim 2, wherein The processing of the signal using bispectrum transformation in Step 2 includes: For the received signal r(n), it is first divided into Γ segments of signals, and the third-order cyclic cumulant is as follows: where X γ (τ1,τ2) is the third-order cyclic cumulant of each signal segment, τ1 and τ2 are two different time delays, then the bispectrum estimation is expressed as: Among them, δ < Δ - 1, ω(τ1,τ2) is the hexagonal window function, then the dimension reduction of the bispectrum is expressed as: Among them, {U1,...,U M} is the compressed bispectrum, and {V1,...V M} is the compressed projection space of the originally selected bispectrum; s and d are the bispectrum weight matrices of signals from different or the same transmitters. When s≠d, s ij = 0; when s = d, ε is a positive real number, and the above formula is transformed into: wherein, o ii = Σ j s ij ,f ii = Σ j d ij , is matrix multiplication; Let δ be a non-zero constant, then the Lagrange equation is expressed as: To minimize Let We obtain: Calculate \(W = [\omega_1,...,\omega d \), then the compressed bispectrum is calculated by the following formula: U i = V i W.

4. The method according to claim 2, characterized in that The convolutional neural network learning in Step 3 includes: First, bispectrum transformation feature extraction is performed on the signal, and then the convolutional layer learns the input data; the convolutional layer contains multiple convolutional kernels inside, and each element that makes up the convolutional kernel corresponds to a weight coefficient and a bias. Each neuron in the convolutional layer is connected to multiple neurons in the area close to the position in the previous layer, and the size of the area depends on the size of the convolutional kernel. The calculation formula is: Among them, b is the deviation, ω is the weight value, x and y represent the convolution kernel performing convolution processing on the entire feature map, and Z l and Z l+1 represent the convolution input and output of the (l + 1)-th layer, which are also called feature maps. L l+1 is the size of Z l+1 Assume that the length and width of the feature map are equal; Z(i, j) corresponds to the pixel of the feature map, K is the number of channels of the feature map, f is the size of the convolution kernel, s0 is the convolution stride, and p is the padding size; the convolution layer contains an activation function to assist in expressing complex features. The rectified linear unit RELU is used to make the neurons in the convolutional neural network have sparse activation. Its representation form is as follows: Secondly, after feature extraction in the convolutional layer, the output features will be passed to the pooling layer for feature selection and information filtering; a pooling layer operation is added after the convolutional layer. The pooling layer selects the pooling area in the same step as the convolutional kernel scans the feature map, which is controlled by the pooling size, stride, and padding. The expression is as follows: where A k represents the input feature map, and other parameters are the same as those of the convolutional layer.

5. A radiation source individual recognition system based on transfer learning for implementing the method according to claim 1, characterized in that, It includes: Feature extraction module: First, collect the training data and perform screening and noise reduction preprocessing on it; Feature extraction is performed on the preprocessed signal using bispectrum transformation. Downsampling and cropping operations are performed on the obtained feature map to obtain an input of n*1*256*256, where n represents the total amount of data. Convolutional neural network learning module: The convolutional neural network is trained using the individual data of the source and target domains. A hyperbolic space network layer is added before the Soft-Max layer of the convolutional neural network to obtain the classification loss of the source network, and the parameter information of the source network model is retained. Target domain migration module: The source network model is migrated to the target network, and the first five convolutional layers are frozen. An adaptation error layer and a hyperbolic space layer are added to the target network. The target network is trained using a small part of the source data and the target data. The classification loss and the MMD loss are used as the new loss functions to update the network parameters and perform optimization. After iteration and optimization, the final network model is obtained.

6. An information data processing terminal, characterized in that It is used to implement the radiation source individual recognition system based on transfer learning described in claim 5.

Images