Radiation source identification method based on attention mechanism and multi-domain feature fusion

Abstract

The invention relates to a radiation source identification method based on attention mechanism and multi-domain feature fusion. The invention relates to a radiation source identification method based on attention mechanism and multi-domain feature fusion. The objective of the invention is to solve the problem that the existing method is poor in accuracy and reliability of specific radiation source identification. The method comprises the following steps: step 1, constructing a radiation source data set; step 2, constructing a radiation source identification network; the radiation source identification network comprises a variational mode decomposition (VMD) module, a discrete wavelet transform (DWT) module, a feature extractor, a multi-domain feature fusion module and a linear network; inputting the radiation source data constructed in the step 1 into a radiation source identification network, and outputting an identification result by the radiation source identification network; training the radiation source identification network to obtain a trained radiation source identification network; and step 3, inputting a radiation source to be detected into the trained radiation source identification network, and outputting an identification result by the trained radiation source identification network. The method is applied to the field of radiation source identification.

Description

Technical Field

[0001] The present invention relates to a radiation source recognition method based on attention mechanism and multi-domain feature fusion. Background Art

[0002] Faced with challenges to network security such as illegal device intrusion, data leakage and hacker attacks, radio frequency fingerprinting (RFF) identification technology in the field of specific emitter identification (SEI) has become a hot research topic among many solutions. Radio frequency fingerprinting technology is to compare the unique and unique individual characteristics of the radiation source formed by manufacturing differences and hardware inherent characteristics when the radiation source emits radio frequency signals to fingerprints (also known as radio frequency fingerprints). By analyzing the radio frequency fingerprint features embedded in the transmission signal of a specific radiation source, the specific radiation source object is uniquely identified. This process shows a high degree of stability, reproducibility and uniqueness. Common radiation sources include wireless devices carried by airplanes, ships, drones, etc., as well as radio stations. In terms of IoT security, the uniqueness and difficulty of copying and forging of radio frequency fingerprints can be used to ensure that legally authorized devices can successfully access the network, prevent illegal devices that are copied, counterfeited or have unknown identities from infiltrating, invading or maliciously attacking the system, and provide a line of defense for the security of the IoT system. In life, RF fingerprint recognition technology is used to monitor small aircraft, and it can also track and identify illegal radio stations that illegally occupy radio frequency bands, crack down on illegal broadcasting activities, and protect space security and legal radio communication rights. Therefore, RF fingerprint recognition technology is one of the important key technologies for specific radiation source identification.

[0003] At present, RF fingerprint recognition technology is mainly divided into two categories: traditional RF fingerprint recognition methods and methods based on deep learning. Traditional RF fingerprint recognition methods use expert knowledge, such as wavelet transform, variational mode decomposition, short-time Fourier transform and signal constellation diagram. These methods process the RF signal of the radiation source, obtain its characteristics in the time domain, frequency domain or phase domain, and use it as the unique RFF of the device to identify the radiation source. Traditional methods have advantages such as strong interpretability and no need for training data, but they also have disadvantages such as reliance on the accuracy of statistical models and difficulty in setting detection thresholds. In recent years, RF fingerprint recognition methods based on deep learning are becoming a new direction for identifying individual targets of radiation sources. Deep learning-based methods use neural network models to extract features, directly input the received signals into the network model for training and recognition, and the output end of the network model directly outputs the recognition results without expert knowledge. Deep learning methods use a large amount of data to drive the network to learn and extract features, but end-to-end data-driven methods require a large amount of data to train the network model, and the recognition process is unexplainable. Summary of the invention

[0004] The purpose of the present invention is to solve the problem of poor accuracy and reliability of existing methods for identifying specific radiation sources, and to propose a radiation source identification method based on attention mechanism and multi-domain feature fusion.

[0005] A radiation source recognition method based on attention mechanism and multi-domain feature fusion. The specific process is as follows:

[0006] Step 1: Construct a radiation source dataset;

[0007] Step 2: Build a radiation source identification network;

[0008] The radiation source identification network includes variational mode decomposition VMD module, discrete wavelet transform DWT module, feature extractor Extractor, multi-domain feature fusion module FusionBlock, and linear network;

[0009] The radiation source data constructed in step 1 is input into the radiation source recognition network, and the radiation source recognition network outputs the recognition result;

[0010] Training the radiation source identification network to obtain a trained radiation source identification network;

[0011] Step 3: Input the radiation source to be tested into the trained radiation source identification network, and the trained radiation source identification network outputs the identification result.

[0012] Preferably, in step 1, a radiation source data set is constructed; the specific process is:

[0013] Step 1. Establish a radiation source RF signal collection scenario; the specific process is as follows:

[0014] The radiation source is a wireless walkie-talkie communication radio station, which transmits radio frequency signals and uses the antenna of the universal software radio peripheral B210 to receive the radio frequency signals; the server is connected to the USPR B210 through a USB interface, and the server is used to store the received radio frequency signals transmitted by the communication radio station;

[0015] Step 1 and 2: Set USRPB210 parameters; the specific process is as follows:

[0016] Configure USRPB210 parameters, including center frequency f0, signal bandwidth B, sampling frequency f s ;

[0017] Step 13: Collect and pre-process the RF signal of the radiation source; the specific process is:

[0018] Use the USRPB210 with the parameters configured in steps 1 and 2 to receive RF signals from different communication stations at the same location, and then move the spectrum of the RF signal to the baseband to obtain I and Q signals;

[0019] Step 14: Store and mark the radiation source signal; the specific process is as follows:

[0020] Unify the data size of I and Q signals, and make the length of both signals N;

[0021] The I signal after the uniform size is placed in front and the Q signal is placed in the back to obtain an IQ signal by splicing, and the IQ signal is stored;

[0022] The stored IQ signals are labeled, and the communication stations to which the stored IQ signals belong are labeled to complete the construction of the radiation source data set.

[0023] Preferably, in said step 2, a radiation source identification network is constructed;

[0024] The radiation source identification network includes variational mode decomposition VMD module, discrete wavelet transform DWT module, feature extractor Extractor, multi-domain feature fusion module FusionBlock, and linear network;

[0025] The radiation source data constructed in step 1 is input into the radiation source recognition network, and the radiation source recognition network outputs the recognition result;

[0026] Training the radiation source identification network to obtain a trained radiation source identification network;

[0027] The specific process is:

[0028] Step 21: The IQ signal obtained in step 1 is represented as x(n), and x(n) is represented as a vector x;

[0029] Where n = 1, 2, ..., 2N, N is the data length, and n is the nth sampling point;

[0030] Step 2: Input vector x into the variational mode decomposition (VMD) module, which outputs the instantaneous amplitude feature a. r and the instantaneous phase characteristic Φ r ;

[0031] Step 2: Input the IQ signal x(n) obtained in step 1 into the discrete wavelet transform DWT module. The discrete wavelet transform DWT module outputs the time-frequency characteristics C of the signal x(n). j,m ;

[0032] Step 24: Construct a feature extractor to extract the deep features f of the vector x obtained in step 21. Data Φ obtained in step 2 r and a r The deep features f VMD , the time-frequency features C obtained in steps 2 and 3 j,m The deep features f DWT ;

[0033] Step 25: Extract the deep features f from step 24 Data , deep features f VMD , deep features f DWT Concatenate along the channel dimension into a feature tensor F;

[0034] Step 26: Input the feature tensor F into the multi-domain feature fusion module FusionBlock, and the multi-domain feature fusion module FusionBlock outputs the dynamically weighted fusion feature F Fusion ;

[0035] Step 27: The dynamically weighted fusion feature F output by the multi-domain feature fusion module FusionBlock in step 26 Fusion Input the linear network, and the linear network outputs the recognition result;

[0036] Step 28: Set the loss function;

[0037] Set the optimizer;

[0038] The radiation source recognition network is optimized according to the loss value, and steps 21 to 28 are repeated until the maximum iteration round is reached to obtain a trained radiation source recognition network.

[0039] Preferably, in step 22, the vector x is input into a variational mode decomposition (VMD) module, and the variational mode decomposition (VMD) module outputs an instantaneous amplitude feature a r and the instantaneous phase characteristic Φ r ;

[0040] The specific process is:

[0041] Step 221: Decompose the vector x into a specified number of modal components using variational mode decomposition (VMD), where each modal component includes time, spectrum, and center frequency parameters;

[0042] VMD is expressed as:

[0043]

[0044] Where VMD{·} represents the variational mode decomposition function, V r represents the time component of the rth mode, represents the spectral component of the rth mode, Π r represents the center frequency of the rth mode, and R is the total number of decomposed modes;

[0045] Step 222: Substitute the time component V of the rth mode r After Hilbert transform, the features are obtained

[0046] Step 223: Features The instantaneous amplitude characteristics are obtained by taking the modulus; the expression is:

[0047]

[0048] in, V r The Hilbert transform of r represents the instantaneous amplitude characteristic of the time component of the rth mode, and |·| represents the modulus value;

[0049] Step 224: Features The instantaneous phase characteristics are obtained by processing; the expression is:

[0050]

[0051] Among them, Φ r Represents the instantaneous phase characteristics of the time component of the rth mode, Im(·) means taking the imaginary part, and Re(·) means taking the real part.

[0052] Preferably, in the steps 2 and 3, the IQ signal x(n) obtained in step 1 is input into a discrete wavelet transform DWT module, and the discrete wavelet transform DWT module outputs the time-frequency characteristics C of the signal x(n). j,m ;

[0053] It is expressed as:

[0054]

[0055] Among them, ψ* j,m (n) is the discretized complex Morlet wavelet basis ψ j,m conjugation of (n);

[0056] Discretized complex Morlet wavelet basis ψ j,m The expression of (n) is:

[0057]

[0058] Among them, j is the discrete scale parameter, ranging from 1 to J, and J is the number of decomposition levels;

[0059] m represents the wavelet location parameter, and ψ is the continuous complex Morlet wavelet basis.

[0060] Preferably, in step 24, a feature extractor is constructed to extract the deep features f of the vector x obtained in step 21. Data Φ obtained in step 2 r and a r The deep features f VMD , the time-frequency features C obtained in steps 2 and 3 j,m The deep features f DWT ;

[0061] The specific process is:

[0062] Step 241: Convert the Φ obtained in step 22 r and a r Concatenate to form a feature tensor A of size batchsize×2×R×N;

[0063] The time-frequency features C obtained in steps 2 and 3 are j,m The imaginary part and the real part of are concatenated to form a feature tensor C of size batchsize×2×J×N;

[0064] Concatenate the vectors x obtained in step 2.1 into a tensor X′ of size batchsize×2×N;

[0065] Among them, batchsize is the network training batch size;

[0066] Step 242: Input the feature tensor A into the feature extractor, and the feature extractor outputs Φ r and a r The deep features f VMD ; The expression is:

[0067] f VMD =F VMD (A)

[0068] Among them, F VMD is a feature extractor based on variational mode decomposition;

[0069] Feature Extractor F based on Variational Mode Decomposition VMD It includes in sequence: the first Conv2d, the first BatchNorm layer, the first ReLU layer, the first maximum pooling layer, the second Conv2d, the second BatchNorm layer, the second ReLU layer, the second maximum pooling layer, the third Conv2d, the third BatchNorm layer, the third ReLU layer, the third maximum pooling layer, the fourth Conv2d, the fourth BatchNorm layer, the fourth ReLU layer, the fourth maximum pooling layer, the adaptive average pooling layer, the Flatten layer, and the Linear layer;

[0070] Conv2d is a two-dimensional convolutional layer;

[0071] Step 243: Input the feature tensor C into the feature extractor, and the feature extractor outputs the time-frequency feature C j,m The deep features f DWT ; The expression is:

[0072] f DWT =F DWT (C)

[0073] Among them, F DWT It is a feature extractor based on discrete wavelet transform;

[0074] Feature Extractor F Based on Discrete Wavelet Transform DWT It includes: the first Conv2d, the first BatchNorm layer, the first ReLU layer, the first maximum pooling layer, the second Conv2d, the second BatchNorm layer, the second ReLU layer, the second maximum pooling layer, the third Conv2d, the third BatchNorm layer, the third ReLU layer, the third maximum pooling layer, the fourth Conv2d, the fourth BatchNorm layer, the fourth ReLU layer, the fourth maximum pooling layer, the View layer, and the Linear layer.

[0075] Conv2d is a two-dimensional convolutional layer;

[0076] Step 244: Input the tensor X′ into the feature extractor, and the feature extractor outputs the deep features f of the vector x Data ; The expression is:

[0077] f Data =F Data (X′)

[0078] Among them, F Data is a feature extractor based on IQ signals;

[0079] Feature Extractor F Based on IQ Signal DataIt includes in sequence: the first ComplexConv1d, the first Batch Norm layer, the first ReLU layer, the first maximum pooling layer, the second ComplexConv1d, the second BatchNorm layer, the second ReLU layer, the second maximum pooling layer, the third ComplexConv1d, the third BatchNorm layer, the third ReLU layer, the third maximum pooling layer, the fourth ComplexConv1d, the fourth BatchNorm layer, the fourth ReLU layer, the fourth maximum pooling layer, the fifth ComplexConv1d, the fifth BatchNorm layer, the fifth ReLU layer, the fifth maximum pooling layer, the View layer, and the Linear layer;

[0080] ComplexConv1d is a complex-valued convolutional layer;

[0081] f VMD 、f DWT and f Data The size of each is a feature tensor of batchsize×H×W.

[0082] Preferably, in step 26, the feature tensor F is input into the multi-domain feature fusion module FusionBlock, and the multi-domain feature fusion module FusionBlock outputs the dynamically weighted fusion feature F Fusion ;

[0083] The expression is:

[0084] F Fusion =f FusionBlock (F)

[0085] Among them, f FusionBlock It is a multi-domain feature fusion module;

[0086] F Fusion The size is batchsize×D×H×W, where D is the channel dimension, H is the feature height, and W is the feature width.

[0087] Preferably, the multi-domain feature fusion module f FusionBlock The specific working process is:

[0088] The multi-domain feature fusion module FusionBlock includes the CA mechanism layer and the convolutional network L Conv layer;

[0089] The CA mechanism layer is the channel attention mechanism layer;

[0090] The CA mechanism layer consists of two steps: compression and excitation;

[0091] In the compression step, the global information of each channel is first compressed, and the process is:

[0092] The feature tensor F compresses the information of each channel through global average pooling. The formula is:

[0093]

[0094] Among them, F d,h,w is the feature tensor of the d-th channel of the feature tensor F;

[0095] z d is the global average pooling result of the dth channel of size 1×1, d=1,…,D; D is the total number of channels of the feature tensor F;

[0096] In the excitation step, the channel weight tensor s is obtained as follows:

[0097] By z d Concatenate to form a tensor Z of size batchsize×D×1×1;

[0098] The channel weight tensor s is obtained by s=Sigmoid(L2(ReLU(L1(Z))));

[0099] Among them, L1 and L2 are two different Linear layers, ReLU(·) is the activation function used in CA, and Sigmoid(·) is the activation function used in CA;

[0100] Then the obtained channel weight tensor s is multiplied element by element with the input feature tensor F to obtain the weighted feature tensor F new :

[0101] F new =s⊙F

[0102] Among them, ⊙ represents element-by-element multiplication;

[0103] The weighted feature tensor F new Input the convolutional network layer L in sequence Conv , convolutional network layer L Conv Output multi-domain fusion feature F Fusion =L Conv (F new );

[0104] Convolutional Network L Conv The layers include convolutional layer, BatchNorm layer, ReLU layer, View layer and Dropout layer in sequence.

[0105] Preferably, in step 27, the dynamically weighted fusion feature F output by the multi-domain feature fusion module FusionBlock in step 26 is FusionInput the linear network, and the linear network outputs the recognition result; the specific process is:

[0106] The linear network includes the first Linear layer, the ReLu layer, and the second Linear layer in sequence;

[0107] Dynamically weighted fusion feature F Fusion The first Linear layer, ReLu layer, and second Linear layer are input in sequence, and the second Linear layer outputs the recognition result; it is expressed as:

[0108] Result = f Classifiers (F Fusion )

[0109] Among them, f Classifiers (·) represents the linear network recognition process, and Result represents the recognition result;

[0110] The identification result is the radiation source from which the radiation source signal comes.

[0111] Preferably, the loss function in step 28 is a cross entropy loss function;

[0112] The optimizer uses the Adam optimizer.

[0113] The beneficial effects of the present invention are:

[0114] The present invention combines the feature extraction method based on expert knowledge with the neural network feature extraction method, and uses the channel attention mechanism to weightedly fuse the multi-domain features, so that the signal's multiple transform domain features are fused to form a "fingerprint". By making full use of the interpretability of traditional methods and the automation characteristics of deep learning methods, it can more effectively cope with the challenges of specific radiation source identification in complex environments, improve the accuracy and reliability of specific radiation source identification, and ensure equipment safety and public safety.

[0115] 1. The specific radiation source identification method described in the present invention combines traditional feature extraction and deep learning technology to construct an identification network that uses the attention mechanism method to perform multi-domain feature fusion, which can improve the accuracy and sensitivity of identifying specific radiation sources and ensure equipment safety and public safety.

[0116] 2. The specific radiation source identification method described in the present invention does not rely on a single feature extraction method. It uses variational mode decomposition to extract instantaneous amplitude and phase features, uses discrete wavelet transform to extract time-frequency domain features, and uses deep learning methods to extract deep features. This makes the identification process have both the interpretability of traditional methods and the efficiency and automation characteristics of deep learning methods, which can improve the adaptability and flexibility of the identification system.

[0117] 3. The specific radiation source identification method described in the present invention uses a convolutional neural network to identify specific radiation sources, which can improve the robustness and reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0118] Figure 1 It is a principle block diagram of a radiation source identification method based on attention mechanism and multi-domain feature fusion described in the present invention;

[0119] Figure 2 This is an application scenario model diagram of a radiation source identification method based on attention mechanism and multi-domain feature fusion described in the present invention;

[0120] Figure 3 It is a structure diagram of a feature extractor of a radiation source identification method based on an attention mechanism and multi-domain feature fusion according to the present invention;

[0121] Figure 4 It is a recognition network architecture diagram of a radiation source recognition method based on attention mechanism and multi-domain feature fusion described in the present invention;

[0122] Figure 5 This is a recognition accuracy result diagram of a radiation source recognition method based on attention mechanism and multi-domain feature fusion described in the present invention;

[0123] Figure 6 It is a confusion matrix diagram of the recognition result of the radiation source recognition method based on the attention mechanism and multi-domain feature fusion described in the present invention. DETAILED DESCRIPTION

[0124] Specific implementation method 1: This implementation method is a radiation source identification method based on attention mechanism and multi-domain feature fusion. The specific process is as follows:

[0125] Aiming at the problem of radiation source identification, the present invention combines traditional feature extraction technology with deep learning technology to construct a recognition network that uses the attention mechanism method to fuse multi-domain features, aiming to achieve accurate identification of specific radiation source individuals.

[0126] Step 1: Construct a specific radiation source dataset;

[0127] The specific radiation source is because it is necessary to identify a single individual. There are five devices in total, and see which device is the source of radiation;

[0128] Step 2: Construct a specific radiation source identification network;

[0129] The specific radiation source identification network includes variational mode decomposition VMD module, discrete wavelet transform DWT module, feature extractor Extractor, multi-domain feature fusion module FusionBlock, and linear network;

[0130] The specific radiation source data constructed in step 1 is input into the specific radiation source recognition network, and the specific radiation source recognition network outputs the recognition result;

[0131] Training a specific radiation source recognition network to obtain a trained specific radiation source recognition network;

[0132] Step 3: Input the radiation source to be tested into the trained specific radiation source recognition network, and the trained specific radiation source recognition network outputs the recognition result.

[0133] Specific implementation method 2: This implementation method is different from the specific implementation method 1 in that a specific radiation source data set is constructed in step 1; the specific process is as follows:

[0134] Step 11: Establish a specific radiation source RF signal collection scenario; the specific process is as follows:

[0135] like Figure 2 As shown, the specific radiation source is a wireless walkie-talkie communication station, the communication station transmits a radio frequency signal, and the radio frequency signal is received by using the antenna of a universal software radio peripheral (USRP) B210; the server is connected to the USRP B210 through a USB interface, and the server is used to store the received radio frequency signal transmitted by the communication station;

[0136] Step 1 and 2: Set USRPB210 parameters; the specific process is as follows:

[0137] Configure USRPB210 parameters through software, including center frequency f0, signal bandwidth B, sampling frequency f s ;

[0138] In the experiment, the open source software radio GNU Radio platform was used to set the USRP B210 parameters, center frequency f0 = 409.8MHz, signal bandwidth B = 250kHz, sampling frequency f s =250kHz;

[0139] Step 13: Collect and pre-process the radio frequency signal of a specific radiation source; the specific process is as follows:

[0140] Use the USRPB210 with the parameters configured in steps 1 and 2 to receive the RF signals of different communication stations at the same location, then move the spectrum of the RF signal to the baseband to obtain the I and Q signals to achieve pre-processing of the RF signal;

[0141] Step 14: Store and mark specific radiation source signals; the specific process is as follows:

[0142] Unify the data size of I and Q signals, and make the length of both signals N;

[0143] The I signal after the uniform size is placed in front and the Q signal is placed in the back to obtain an IQ signal by splicing, and the IQ signal is stored;

[0144] The stored IQ signals are labeled, and the communication stations to which the stored IQ signals belong are labeled (the signals of each device are cut into a number of IQ spliced ​​data, and it is marked which device this data belongs to), thus completing the construction of a specific radiation source data set.

[0145] In the experiment, a total of 5 different wireless walkie-talkie communication radio signals with a center frequency of 409.8MHz were collected and marked as 5 categories, with labels 1, 2, 3, 4, and 5 respectively. Each communication radio collected 200 sets of data. The data were I and Q signals with a length of N=2048 sampling points. The front I and rear Q signals were spliced ​​and stored as HDF5 files.

[0146] The other steps and parameters are the same as those in the first embodiment.

[0147] Specific implementation method three: This implementation method is different from specific implementation methods one or two in that a specific radiation source identification network is constructed in step two;

[0148] The specific radiation source identification network includes variational mode decomposition VMD module, discrete wavelet transform DWT module, feature extractor Extractor, multi-domain feature fusion module FusionBlock, and linear network;

[0149] The specific radiation source data constructed in step 1 is input into the specific radiation source recognition network, and the specific radiation source recognition network outputs the recognition result;

[0150] Training a specific radiation source recognition network to obtain a trained specific radiation source recognition network;

[0151] The specific process is:

[0152] Step 21: The IQ signal obtained in step 1 is represented as x(n), and x(n) is represented as a vector x;

[0153] Where n = 1, 2, ..., 2N, N is the data length, and n is the nth sampling point;

[0154] Step 2: Input vector x into the variational mode decomposition (VMD) module, which outputs the instantaneous amplitude feature a. r and the instantaneous phase characteristic Φ r ;

[0155] Step 2: Input the IQ signal x(n) obtained in step 1 into the discrete wavelet transform (DWT) module. The discrete wavelet transform (DWT) module outputs the time-frequency characteristics C of the signal x(n). j,m ;

[0156] Step 24: Construct a feature extractor to extract the deep features f of the vector x obtained in step 21. Data Φ obtained in step 2 r and a r The deep features f VMD , the time-frequency features C obtained in steps 2 and 3 j,m The deep features f DWT ;

[0157] Step 25: Extract the deep features f from step 24 Data , deep features f VMD , deep features f DWT Concatenate along the channel dimension into a feature tensor F of batchsize×D×H×W;

[0158] Where D is the channel dimension, H is the height of the feature, and W is the width of the feature;

[0159] Step 26: Input the feature tensor F into the multi-domain feature fusion module FusionBlock, and the multi-domain feature fusion module FusionBlock outputs the dynamically weighted fusion feature F Fusion ;

[0160] Step 27: The dynamically weighted fusion feature F output by the multi-domain feature fusion module FusionBlock in step 26 Fusion Input the linear network, and the linear network outputs the recognition result;

[0161] Step 28: Set the loss function to calculate the loss between the predicted category and the true category learned by the network;

[0162] Set the optimizer;

[0163] The specific radiation source recognition network is optimized according to the loss value, and steps 21 to 28 are repeated until the maximum number of iterations is reached to obtain a trained specific radiation source recognition network.

[0164] The network architecture of the specific radiation source identification network is as follows Figure 4As shown in the figure, the network consists of VMD function module, DWT function module, feature extractor Extractor, classifier Classifiers, cross-entropy (CE) loss function and optimizer Adam; VMD function module, DWT function module and feature extractor Extractor are responsible for extracting multi-domain features of specific radiation source signals; Classifiers consists of multi-domain feature fusion module FusionBlock module and linear network Linear layers, among which the multi-domain feature fusion module FusionBlock performs feature fusion and the linear network Linearlayers completes recognition; Linear network Linear layers can retain the fused multi-domain fusion features as much as possible as the unique fingerprint feature for identifying a specific radiation source individual. It consists of two Linear layers and a ReLu layer, and the fusion feature F Fusion After passing through the linear layers, the recognition result is obtained;

[0165] The cross entropy loss function is responsible for calculating the loss between the predicted category and the true category learned by the network, while the optimizer Adam optimizes the parameters in the feature extractor Extractor and classifier Classifiers according to the loss value;

[0166] The linear layers architecture configuration table is shown in Table 3;

[0167] In the experiment, the recognition results are divided into 5 categories, namely 1, 2, 3, 4, and 5.

[0168] The other steps and parameters are the same as those in the first or second embodiment.

[0169] Specific implementation method 4: The difference between this implementation method and specific implementation methods 1 to 3 is that in step 22, the vector x is input into the variational mode decomposition VMD module, and the variational mode decomposition VMD module outputs the instantaneous amplitude feature a r and the instantaneous phase characteristic Φ r ; The specific process is:

[0170] Step 221: Decompose the vector x into a specified number of modal components using variational mode decomposition (VMD), where each modal component includes time, spectrum, and center frequency parameters;

[0171] VMD is expressed as:

[0172]

[0173] Where VMD{·} represents the variational mode decomposition function;

[0174] Vr represents the time component of the rth mode;

[0175] represents the spectral component of the rth mode;

[0176] Π r represents the center frequency of the rth mode;

[0177] R is the total number of decomposed modes;

[0178] Step 222: Substitute the time component V of the rth mode r After Hilbert transform, the features are obtained

[0179] Step 223: Features The instantaneous amplitude characteristics are obtained by taking the modulus; the expression is:

[0180]

[0181] in, V r The Hilbert transform of

[0182] a r Represents the instantaneous amplitude characteristics of the time component of the rth mode;

[0183] |·| represents the modulus value;

[0184] Step 224: Features The instantaneous phase characteristics are obtained by processing; the expression is:

[0185]

[0186] Among them, Φ r represents the instantaneous phase characteristic of the time component of the rth mode, Im(·) means taking the imaginary part, and Re(·) means taking the real part;

[0187] In the experiment, R=4.

[0188] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0189] Specific implementation mode 5: This implementation mode is different from any one of the specific implementation modes 1 to 4 in that in the steps 2 and 3, the IQ signal x(n) obtained in step 1 is input into a discrete wavelet transform DWT (Discrete Wavelet Transform) module, and the discrete wavelet transform DWT module outputs the time-frequency characteristics C of the signal x(n). j,m ;

[0190] It is expressed as:

[0191]

[0192] Among them, ψ * j,m (n) is the discretized complex Morlet wavelet basis ψ j,m conjugation of (n);

[0193] Discretized complex Morlet wavelet basis ψ j,m The expression of (n) is:

[0194]

[0195] Among them, j is the discrete scale parameter, ranging from 1 to J, and J is the number of decomposition levels;

[0196] m represents the wavelet location parameter, ψ is the continuous complex Morlet wavelet basis;

[0197] In the experiment, J=256, and m is automatically calculated by the software.

[0198] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0199] Specific implementation method 6: This implementation method is different from specific implementation methods 1 to 5 in that a feature extractor is constructed in step 24 to extract the deep features f of the vector x obtained in step 21. Data Φ obtained in step 2 r and a r The deep features f VMD , the time-frequency features C obtained in steps 2 and 3 j,m The deep features f DWT ;

[0200] The specific process is:

[0201] Step 241: Convert the Φ obtained in step 22 r and a r The concatenation size is batchsize×2×R×N feature tensor A;

[0202] The time-frequency features C obtained in steps 2 and 3 are j,m The imaginary part and the real part of are concatenated to form a feature tensor C of size batchsize×2×J×N;

[0203] Concatenate the vectors x obtained in step 2.1 into a tensor X′ of size batchsize×2×N;

[0204] Among them, batchsize is the network training batch size; for example, there are 100 iterations, and the batch size of each round is 10;

[0205] Step 242: Input the feature tensor A into the feature extractor, and the feature extractor outputs Φ r and a r The deep features f VMD ; The expression is:

[0206] f VMD =F VMD (A)

[0207] Among them, F VMD is a feature extractor based on variational mode decomposition;

[0208] Feature Extractor F based on Variational Mode Decomposition VMD It includes in sequence: the first convolution layer Conv2d, the first BatchNorm layer, the first ReLU layer, the first maximum pooling (Maxpooling) layer, the second convolution layer Conv2d, the second Batch Norm layer, the second ReLU layer, the second maximum pooling (Maxpooling) layer, the third convolution layer Conv2d, the third Batch Norm layer, the third ReLU layer, the third maximum pooling (Maxpooling) layer, the fourth convolution layer Conv2d, the fourth Batch Norm layer, the fourth ReLU layer, the fourth maximum pooling (Maxpooling) layer, the adaptive average pooling (adaptive pool) layer, the Flatten layer, and the Linear layer;

[0209] Conv2d is a two-dimensional convolutional layer;

[0210] Step 243: Input the feature tensor C into the feature extractor, and the feature extractor outputs the time-frequency feature C j,m The deep features f DWT ; The expression is:

[0211] f DWT =F DWT (C)

[0212] Among them, F DWT It is a feature extractor based on discrete wavelet transform;

[0213] Feature Extractor F Based on Discrete Wavelet Transform DWTIt includes in sequence: the first convolution layer Conv2d, the first BatchNorm layer, the first ReLU layer, the first maximum pooling (Maxpooling) layer, the second convolution layer Conv2d, the second Batch Norm layer, the second ReLU layer, the second maximum pooling (Maxpooling) layer, the third convolution layer Conv2d, the third Batch Norm layer, the third ReLU layer, the third maximum pooling (Maxpooling) layer, the fourth convolution layer Conv2d, the fourth Batch Norm layer, the fourth ReLU layer, the fourth maximum pooling (Maxpooling) layer, the View layer, and the Linear layer;

[0214] Conv2d is a two-dimensional convolutional layer;

[0215] Step 244: Input the tensor X′ into the feature extractor, and the feature extractor outputs the deep features f of the vector x Data ; The expression is:

[0216] f Data =F Data (X′)

[0217] Among them, F Data is a feature extractor based on IQ signals;

[0218] Feature Extractor F Based on IQ Signal Data It includes in sequence: the first convolution layer ComplexConv1d, the first BatchNorm layer, the first ReLU layer, the first maximum pooling (Maxpooling) layer, the second convolution layer ComplexConv1d, the second Batch Norm layer, the second ReLU layer, the second maximum pooling (Max pooling) layer, the third convolution layer ComplexConv1d, the third BatchNorm layer, the third ReLU layer, the third maximum pooling (Maxpooling) layer, the fourth convolution layer ComplexConv1d, the fourth BatchNorm layer, the fourth ReLU layer, the fourth maximum pooling (Maxpooling) layer, the fifth convolution layer ComplexConv1d, the fifth BatchNorm layer, the fifth ReLU layer, the fifth maximum pooling (Max pooling) layer, the View layer, and the Linear layer;

[0219] f VMD 、f DWT and f Data The size of each is a feature tensor of batchsize×H×W.

[0220] ComplexConv1d is a complex-valued convolutional layer;

[0221] like Figure 3 As shown in Figure 2, the feature extractor consists of three parallel deep learning network modules, namely, the feature learning network F based on variational mode decomposition VMD , Feature learning network based on discrete wavelet transform F DWT and the feature learning network F based on IQ signals Data ;

[0222] The convolution layer extracts features, but these features are easily disturbed by noise and tend to learn irrelevant features in the data set, which reduces robustness; normalizing the data through the Batch Norm layer can reduce the interference of data noise and outliers, making the feature extraction network training faster and more stable; introducing nonlinearity through the ReLu layer enhances the network learning ability and prevents overfitting; the Max pooling layer can suppress local features and retain the main features; the Dropout layer "discards" the data tensor in proportion to avoid overfitting and improve the generalization ability of the network; the adaptive pooling layer can dynamically adjust the output size according to the size of the input feature map, so that the output size remains consistent under different input sizes. The data is compressed from high dimensions to one dimension through Flatten or View operations, and finally after passing through the Linear layer, the output size is a deep feature tensor of batchsize*H*W;

[0223] The specific network architecture configuration of the feature extractor is shown in Table 1;

[0224] In the experiment, batchsize=4, H=256, W=1.

[0225] Table 1 Network configuration table of feature extractor

[0226]

[0227] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0228] Specific implementation method 7: This implementation method is different from any one of the specific implementation methods 1 to 6 in that in step 26, the feature tensor F is input into the multi-domain feature fusion module FusionBlock, and the multi-domain feature fusion module FusionBlock outputs the dynamically weighted fusion feature F Fusion ;

[0229] The expression is:

[0230] F Fusion =f FusionBlock (F)

[0231] Among them, f FusionBlock It is a multi-domain feature fusion module;

[0232] F Fusion The size is batchsize×D×H×W, where D is the channel dimension, H is the feature height, and W is the feature width.

[0233] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0234] Specific implementation eight: This implementation differs from specific implementations one to seven in that the multi-domain feature fusion module f FusionBlock The specific working process is:

[0235] The multi-domain feature fusion module FusionBlock includes the CA mechanism layer and the convolutional network L Conv layer;

[0236] The CA mechanism layer is the channel attention (CA) mechanism layer;

[0237] The CA mechanism improves the representation ability of the network by explicitly modeling the interdependencies between features. The CA mechanism layer consists of two steps: compression and excitation.

[0238] In the compression step, the global information of each channel is first compressed, and the process is:

[0239] The feature tensor F compresses the information of each channel through global average pooling. The formula is:

[0240]

[0241] Among them, F d,h,w is the feature tensor of the d-th channel of the feature tensor F;

[0242] z d is the global average pooling result of the dth channel of size 1×1, d=1,…,D; D is the total number of channels of the feature tensor F;

[0243] In the excitation step, the channel weight tensor s is obtained as follows:

[0244] By z d Concatenate to form a tensor Z of size batchsize×D×1×1;

[0245] The channel weight tensor s is obtained by s=Sigmoid(L2(ReLU(L1(Z))));

[0246] Among them, L1 and L2 are two different Linear layers, ReLU(·) is the activation function used in CA, and Sigmoid(·) is the activation function used in CA;

[0247] Then the obtained channel weight tensor s is multiplied element by element with the input feature tensor F to obtain the weighted feature tensor F new :

[0248] F new =s⊙F

[0249] Among them, ⊙ represents element-by-element multiplication;

[0250] Through compression and excitation, CA can help the network automatically learn which features are more important for the current task and enhance the model's ability to express key features;

[0251] In CA, we dynamically weight multi-domain features to obtain the feature tensor F new After that, the weighted feature tensor F new Input the convolutional network layer L in sequence Conv , convolutional network layer L Conv Output multi-domain fusion feature F Fusion =L Conv (F new );

[0252] Convolutional Network L Conv The layers include convolutional layer (Conv1d), BatchNorm layer, ReLU layer, View layer and Dropout layer in sequence;

[0253] The network configuration of the multi-domain feature fusion module FusionBlock is shown in Table 2;

[0254] In the experiment, D=3.

[0255] Table 2 Network configuration table of multi-domain feature fusion module FusionBlock

[0256]

[0257] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0258] Specific implementation method 9: This implementation method is different from any one of specific implementation methods 1 to 8 in that in step 27, the dynamically weighted fusion feature F output by the multi-domain feature fusion module FusionBlock in step 26 is Fusion Input the linear network, and the linear network outputs the recognition result; the specific process is:

[0259] The linear network includes the first Linear layer, the ReLu layer, and the second Linear layer in sequence;

[0260] Dynamically weighted fusion feature F Fusion The first Linear layer, ReLu layer, and second Linear layer are input in sequence, and the second Linear layer outputs the recognition result; it is expressed as:

[0261] Result = f Classifiers (F Fusion )

[0262] Among them, f Classifiers (·) represents the linear layers recognition process, and Result represents the recognition result;

[0263] The identification result is which radiation source the radiation source signal comes from (which device emits the radiation source).

[0264] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0265] Specific implementation ten: This implementation is different from specific implementations one to nine in that the loss function in step 28 is a cross entropy loss function, which is responsible for calculating the loss between the predicted category and the true category learned by the network;

[0266] The optimizer uses the Adam optimizer.

[0267] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0268] The network architecture of the specific radiation source identification network is as follows Figure 4 As shown in the figure, the network consists of VMD function module, DWT function module, feature extractor Extractor, classifier Classifiers, cross-entropy (CE) loss function and optimizer Adam; VMD function module, DWT function module and feature extractor Extractor are responsible for extracting multi-domain features of specific radiation source signals; Classifiers Classifiers consists of multi-domain feature fusion module FusionBlock module and linear network Linear layers, in which the multi-domain feature fusion module FusionBlock performs feature fusion, and the linear network Linearlayers completes recognition;

[0269] The cross entropy loss function is responsible for calculating the loss between the predicted category and the true category learned by the network, while the optimizer Adam optimizes the parameters in the feature extractor Extractor and classifier Classifiers according to the loss value;

[0270] The linear layers architecture configuration table is shown in Table 3;

[0271] In the experiment, the recognition results are divided into 5 categories, namely 1, 2, 3, 4, and 5;

[0272] Table 3 Linear layers architecture configuration table

[0273]

[0274] Training a specific emitter identification network

[0275] The specific radiation source data set is divided into a training set and a test set according to a certain ratio; the training set is used for network training, the validation set is used for evaluating the learning effect during the network model training process, and the test set is used to evaluate the performance of the network model; in one training, the training set data is input into the radiation source recognition network, multi-domain feature extraction is performed through the VMD function module, the DWT function module and the feature extractor Extractor, multi-domain feature fusion and recognition are performed through the classifiers Classifiers, the network model is learned and updated through the cross entropy loss function and the optimizer Adam, and the learning and convergence degree of the network model is tested using the validation set; after several iterative trainings, the trained network model is saved after the network model converges; finally, the recognition verification is performed on this model using the test set to obtain the recognition result;

[0276] In the experiment, the training set consists of 5 types of radio walkie-talkie communication stations, 160 groups of data are randomly selected from 200 groups of data for each station, totaling 800 groups of data; the test set consists of 40 groups of data randomly selected from each station, totaling 200 groups of data; the network model is iterated and trained 100 times; after 100 training times, the network model has converged and the trained network model is saved; finally, the test set is used to perform recognition verification on this model for 5 types of wireless walkie-talkie communication stations to obtain the recognition results.

[0277] The method of the present invention is used to identify five types of wireless intercom communication stations. The recognition accuracy of the training set and the verification set is as follows: Figure 5 As shown. Figure 5 It can be seen that the recognition accuracy of the training set and the verification set are finally stabilized at more than 95.0%, indicating that the recognition network constructed by the method of the present invention has good generalization and feature learning effects. Figure 6 This is the confusion matrix of the recognition results obtained by the method of the present invention using the test set. Figure 6 It can be seen that the method of the present invention has a good recognition effect on the five types of wireless walkie-talkie communication stations, with an accuracy rate of more than 97.44% and an average accuracy rate of up to 99.0%.

[0278] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A radiation source identification method based on attention mechanism and multi-domain feature fusion, characterized by: The specific process of the method is: Step 1: Construct a radiation source dataset; Step 2: Build a radiation source identification network; The radiation source identification network includes variational mode decomposition VMD module, discrete wavelet transform DWT module, feature extractor Extractor, multi-domain feature fusion module FusionBlock, and linear network; The radiation source data constructed in step 1 is input into the radiation source recognition network, and the radiation source recognition network outputs the recognition result; Training the radiation source identification network to obtain a trained radiation source identification network; Step 3: Input the radiation source to be tested into the trained radiation source identification network, and the trained radiation source identification network outputs the identification result.

2. According to claim 1, a radiation source identification method based on attention mechanism and multi-domain feature fusion is characterized in that: In step 1, a radiation source data set is constructed; the specific process is as follows: Step 1. Establish a radiation source RF signal collection scenario; the specific process is as follows: The radiation source is a wireless walkie-talkie communication radio station, which transmits radio frequency signals and uses the antenna of the universal software radio peripheral B210 to receive the radio frequency signals; the server is connected to the USPR B210 through a USB interface, and the server is used to store the received radio frequency signals transmitted by the communication radio station; Step 1 and 2: Set USRPB210 parameters; the specific process is as follows: Configure USRPB210 parameters, including center frequency f0, signal bandwidth B, sampling frequency f s ; Step 13: Collect and pre-process the RF signal of the radiation source; the specific process is: Use the USRPB210 with the parameters configured in steps 1 and 2 to receive RF signals from different communication stations at the same location, and then move the spectrum of the RF signal to the baseband to obtain I and Q signals; Step 14: Store and mark the radiation source signal; the specific process is as follows: Unify the data size of I and Q signals, and make the length of both signals N; The I signal after the uniform size is placed in front and the Q signal is placed in the back to obtain an IQ signal by splicing, and the IQ signal is stored; The stored IQ signals are labeled, and the communication stations to which the stored IQ signals belong are labeled to complete the construction of the radiation source data set.

3. According to claim 2, a radiation source identification method based on attention mechanism and multi-domain feature fusion is characterized in that: In the step 2, a radiation source identification network is constructed; The radiation source identification network includes variational mode decomposition VMD module, discrete wavelet transform DWT module, feature extractor Extractor, multi-domain feature fusion module FusionBlock, and linear network; The radiation source data constructed in step 1 is input into the radiation source recognition network, and the radiation source recognition network outputs the recognition result; Training the radiation source identification network to obtain a trained radiation source identification network; The specific process is: Step 21: The IQ signal obtained in step 1 is represented as x(n), and x(n) is represented as a vector x; Where n = 1, 2, ..., 2N, N is the data length, and n is the nth sampling point; Step 2: Input vector x into the variational mode decomposition (VMD) module, which outputs the instantaneous amplitude feature a. r and the instantaneous phase characteristic Φ r ; Step 2: Input the IQ signal x(n) obtained in step 1 into the discrete wavelet transform DWT module. The discrete wavelet transform DWT module outputs the time-frequency characteristics C of the signal x(n). j,m ; Step 24: Construct a feature extractor to extract the deep features f of the vector x obtained in step 21. Data Φ obtained in step 2 r and a r The deep features f VMD , the time-frequency features C obtained in steps 2 and 3 j,m The deep features f DWT ; Step 25: Extract the deep features f from step 24 Data , deep features f VMD , deep features f DWT Concatenate along the channel dimension into a feature tensor F; Step 26: Input the feature tensor F into the multi-domain feature fusion module FusionBlock, and the multi-domain feature fusion module FusionBlock outputs the dynamically weighted fusion feature F Fusion ; Step 27: The dynamically weighted fusion feature F output by the multi-domain feature fusion module FusionBlock in step 26 Fusion Input the linear network, and the linear network outputs the recognition result; Step 28: Set the loss function; Set the optimizer; The radiation source recognition network is optimized according to the loss value, and steps 21 to 28 are repeated until the maximum iteration round is reached to obtain a trained radiation source recognition network.

4. According to claim 3, a radiation source identification method based on attention mechanism and multi-domain feature fusion is characterized in that: In step 22, the vector x is input into the variational mode decomposition (VMD) module, and the variational mode decomposition (VMD) module outputs the instantaneous amplitude feature a r and the instantaneous phase characteristic Φ r ; The specific process is: Step 221: Decompose the vector x into a specified number of modal components using variational mode decomposition (VMD), where each modal component includes time, spectrum, and center frequency parameters; VMD is expressed as: Where VMD{·} represents the variational mode decomposition function, V r represents the time component of the rth mode, represents the spectral component of the rth mode, Π r represents the center frequency of the rth mode, and R is the total number of decomposed modes; Step 222: Substitute the time component V of the rth mode r After Hilbert transform, the features are obtained ; Step 223: Features The instantaneous amplitude characteristics are obtained by taking the modulus; the expression is: in, V r The Hilbert transform of r represents the instantaneous amplitude characteristic of the time component of the rth mode, and |·| represents the modulus value; Step 224: Features The instantaneous phase characteristics are obtained by processing; the expression is: Among them, Φ r Represents the instantaneous phase characteristics of the time component of the rth mode, Im(·) means taking the imaginary part, and Re(·) means taking the real part.

5. According to claim 4, a radiation source identification method based on attention mechanism and multi-domain feature fusion is characterized in that: In the steps 2 and 3, the IQ signal x(n) obtained in step 1 is input into a discrete wavelet transform DWT module, and the discrete wavelet transform DWT module outputs the time-frequency characteristics C of the signal x(n) j,m ; It is expressed as: Among them, ψ * j,m (n) is the discretized complex Morlet wavelet basis ψ j,m conjugation of (n); Discretized complex Morlet wavelet basis ψ j,m The expression of (n) is: Among them, j is the discrete scale parameter, ranging from 1 to J, and J is the number of decomposition levels; m represents the wavelet location parameter, and ψ is the continuous complex Morlet wavelet basis.

6. The radiation source identification method based on attention mechanism and multi-domain feature fusion according to claim 5 is characterized in that: In step 24, a feature extractor is constructed to extract the deep features f of the vector x obtained in step 21. Data Φ obtained in step 2 r and a r The deep features f VMD , the time-frequency features C obtained in steps 2 and 3 j,m The deep features f DWT ; The specific process is: Step 241: Convert the Φ obtained in step 22 r and a r Concatenate to form a feature tensor A of size batchsize×2×R×N; The time-frequency features C obtained in steps 2 and 3 are j,m The imaginary part and the real part of are concatenated to form a feature tensor C of size batchsize×2×J×N; Concatenate the vectors x obtained in step 2.1 into a tensor X′ of size batchsize×2×N; Among them, batchsize is the network training batch size; Step 242: Input the feature tensor A into the feature extractor, and the feature extractor outputs Φ r and a r The deep features f VMD ; The expression is: f VMD =F VMD (A) Among them, F VMD is a feature extractor based on variational mode decomposition; Feature Extractor F based on Variational Mode Decomposition VMD It includes in sequence: the first Conv2d, the first BatchNorm layer, the first ReLU layer, the first maximum pooling layer, the second Conv2d, the second BatchNorm layer, the second ReLU layer, the second maximum pooling layer, the third Conv2d, the third BatchNorm layer, the third ReLU layer, the third maximum pooling layer, the fourth Conv2d, the fourth BatchNorm layer, the fourth ReLU layer, the fourth maximum pooling layer, the adaptive average pooling layer, the Flatten layer, and the Linear layer; Conv2d is a two-dimensional convolutional layer; Step 243: Input the feature tensor C into the feature extractor, and the feature extractor outputs the time-frequency feature C j,m The deep features f DWT ; The expression is: f DWT =F DWT (C) Among them, F DWT It is a feature extractor based on discrete wavelet transform; Feature Extractor F Based on Discrete Wavelet Transform DWT It includes: the first Conv2d, the first BatchNorm layer, the first ReLU layer, the first maximum pooling layer, the second Conv2d, the second BatchNorm layer, the second ReLU layer, the second maximum pooling layer, the third Conv2d, the third BatchNorm layer, the third ReLU layer, the third maximum pooling layer, the fourth Conv2d, the fourth BatchNorm layer, the fourth ReLU layer, the fourth maximum pooling layer, the View layer, and the Linear layer. Conv2d is a two-dimensional convolutional layer; Step 244: Input the tensor X′ into the feature extractor, and the feature extractor outputs the deep features f of the vector x Data ; The expression is: f Data =F Data (X′) Among them, F Data is a feature extractor based on IQ signals; Feature Extractor F Based on IQ Signal Data It includes in sequence: the first ComplexConv1d, the first BatchNorm layer, the first ReLU layer, the first maximum pooling layer, the second ComplexConv1d, the second BatchNorm layer, the second ReLU layer, the second maximum pooling layer, the third ComplexConv1d, the third BatchNorm layer, the third ReLU layer, the third maximum pooling layer, the fourth ComplexConv1d, the fourth BatchNorm layer, the fourth ReLU layer, the fourth maximum pooling layer, the fifth ComplexConv1d, the fifth BatchNorm layer, the fifth ReLU layer, the fifth maximum pooling layer, the View layer, and the Linear layer; ComplexConv1d is a complex-valued convolutional layer; f VMD 、f DWT and f Data The size of each is a feature tensor of batchsize×H×W.

7. The radiation source identification method based on attention mechanism and multi-domain feature fusion according to claim 6 is characterized in that: In step 26, the feature tensor F is input into the multi-domain feature fusion module FusionBlock, and the multi-domain feature fusion module FusionBlock outputs the dynamically weighted fusion feature F Fusion ; The expression is: F Fusion =f FusionBlock (F) Among them, f FusionBlock It is a multi-domain feature fusion module; F Fusion The size is batchsize×D×H×W, where D is the channel dimension, H is the feature height, and W is the feature width.

8. The radiation source identification method based on attention mechanism and multi-domain feature fusion according to claim 7 is characterized in that: The multi-domain feature fusion module f FusionBlock The specific working process is: The multi-domain feature fusion module FusionBlock includes the CA mechanism layer and the convolutional network L Conv layer; The CA mechanism layer is the channel attention mechanism layer; The CA mechanism layer consists of two steps: compression and excitation; In the compression step, the global information of each channel is first compressed, and the process is: The feature tensor F compresses the information of each channel through global average pooling. The formula is: Among them, F d,h,w is the feature tensor of the d-th channel of the feature tensor F; z d is the global average pooling result of the dth channel of size 1×1, d=1,...,D; D is the total number of channels of the feature tensor F; In the excitation step, the channel weight tensor s is obtained as follows: By z d Concatenate to form a tensor Z of size batchsize×D×1×1; The channel weight tensor s is obtained by s=Sigmoid(L2(ReLU(L1(Z)))); Among them, L1 and L2 are two different Linear layers, ReLU(·) is the activation function used in CA, and Sigmoid(·) is the activation function used in CA; Then the obtained channel weight tensor s is multiplied element by element with the input feature tensor F to obtain the weighted feature tensor F new : F new =s⊙F Among them, ⊙ represents element-by-element multiplication; The weighted feature tensor F new Input the convolutional network layer L in sequence Conv , convolutional network layer L Conv Output multi-domain fusion feature F Fusion =L Conv (F new ); Convolutional Network L Conv The layers include convolutional layer, BatchNorm layer, ReLU layer, View layer and Dropout layer in sequence.

9. The radiation source identification method based on attention mechanism and multi-domain feature fusion according to claim 8 is characterized in that: In step 27, the dynamically weighted fusion feature F output by the multi-domain feature fusion module FusionBlock in step 26 is Fusion Input the linear network, and the linear network outputs the recognition result; the specific process is: The linear network includes the first Linear layer, the ReLu layer, and the second Linear layer in sequence; Dynamically weighted fusion feature F Fusion The first Linear layer, ReLu layer, and second Linear layer are input in sequence, and the second Linear layer outputs the recognition result; it is expressed as: Result=f Classifiers (F Fusion ) Among them, f Classifiers (·) represents the linear network recognition process, and Result represents the recognition result; The identification result is the radiation source from which the radiation source signal comes.

10. The radiation source identification method based on attention mechanism and multi-domain feature fusion according to claim 9 is characterized in that: The loss function in step 28 is a cross entropy loss function; The optimizer uses the Adam optimizer.

Images