Multi-domain feature fusion radiation source individual identification method

Abstract

The invention provides a radiation source individual identification method based on multi-domain feature fusion, which mainly realizes radiation source individual identification through signal acquisition, signal preprocessing, signal pre-classification, feature extraction and model pre-training. Compared with a traditional method, the method has the advantages that the fused multi-domain features can better reflect subtle differences of radiation sources, and the recognition effect is obviously improved; by respectively extracting the features of each domain, the dimension of data can be effectively reduced, so that the complexity of calculation and storage is reduced; meanwhile, pertinent processing and fusion are allowed to be carried out on the features of different domains, and the classification precision is improved; the features of different domains can be selected and weighted, the influence of important features is enhanced, and the model is more flexible and accurate; common computer hardware is used for configuration, practicability is high, and economical efficiency is good.

Description

Technical Field

[0001] The invention belongs to the technical field of signal recognition, and in particular relates to a method for identifying individual radiation sources. Background Art

[0002] Communication radiation source individual identification technology is a technology that identifies specific targets by identifying the unique subtle features of the signal sent by the radiation source. These subtle features are inadvertently modulated on the signal when the radiation source is in a non-ideal state, and they are mainly manifested as special distortions in the signal time domain, frequency domain and other aspects. These features are unique, stable and measurable, and can be regarded as the unique "fingerprint" of the radiation source, which is used to identify and distinguish different individual radiation sources.

[0003] The fingerprint features of radiation sources can be divided into transient features and steady-state features. Although the pulse signal in the transient process contains rich individual features, its duration is extremely short, similar to noise, and it is necessary to determine the starting point of the transient process, so it is difficult to use. The main research direction at this stage is based on steady-state features. A large number of steady-state signals can extract effective recognition features, avoid the limitations of transient features, and have better applicability in practical applications. Existing research based on steady-state features is mainly carried out from the time domain, frequency domain, and time-frequency domain of steady-state signals. The method of using time domain I / Q signals as network input is simple to implement and has strong practicality, but the recognition rate is not high in the case of strong noise, and it is too dependent on the design of neural networks. Frequency domain analysis is mainly used to reveal the frequency characteristics of signals, while high-order spectrum analysis focuses on describing the high-order statistical characteristics of signals, which is suitable for processing complex signals and non-ideal signal analysis. Cao Jiakun et al. proposed a method for extracting and identifying subtle features of orthogonal modulation of communication radiation sources based on local bispectrum and Longformer network, "A method for extracting and identifying subtle features of orthogonal modulation of communication radiation sources" Journal of Radio Science 1-7. Specifically, the integrated bispectrum is used to extract high-level statistical features of the signal, distinguish the characteristic signals of different transmitters, and improve the accuracy and robustness of transmitter identification by analyzing the nonlinear characteristics and phase coupling relationship of the transmitted signal. However, this method is highly dependent on the quality and quantity of training data, and the lack of time domain features will affect the accuracy and robustness of the recognition results. At present, methods based on transform domain features are more popular. By performing time-frequency analysis, the information of the signal in the joint time-frequency domain can be obtained, thereby revealing more subtle differences between different signals, but this is generally effective for the analysis of nonlinear signals.

[0004] In complex communication environments, especially in non-cooperative communication, the electromagnetic environment is very complex and the signal conditions are diverse. Relying solely on the characteristics of a single domain signal often cannot fully and accurately characterize the slight differences between individual radiators, thus limiting the accuracy of individual identification. In addition, in order to forcibly improve the recognition accuracy, more complex neural network models are often used in research, which places a greater demand on the number of signal samples. However, in actual communication environments, it is difficult to obtain enough signal samples. Summary of the invention

[0005] In order to overcome the problems existing in the prior art, the present invention proposes a method for individual identification of radiation sources by fusion of multi-domain features. The basic idea is: using a combination of high-order powers of time domain signals and frequency domain spectrum data FD-SP (Frequency domain spectrum data) as the model input signal, and characterizing the differences of different signals by time domain features and frequency domain features, which can provide more comprehensive signal features for radiation source identification, and convert the problem of radiation source identification into decision area division in a high-dimensional feature space to improve the accuracy and robustness of identification.

[0006] The I / Q data of the signal, as the time domain characteristics of the radiation source, is the mainstream method of signal acquisition and storage, and contains all the information of the signal in the time domain. However, I / Q data is difficult to highlight subtle feature differences and is not easy to extract fingerprint features.

[0007] The square of I / Q data, SI / Q, can effectively characterize the amplitude characteristics of the signal and highlight the originally blurred or confused signal spectrum. In a low signal-to-noise ratio environment, it can relatively gently amplify the signal and noise, which is suitable for situations where the signal characteristics are more obvious but need to be moderately enhanced. In addition, the nonlinear distortion introduced by SI / Q is small and the computational complexity is low, which is particularly suitable for real-time processing scenarios.

[0008] The fourth power of I / Q data, FI / Q, can effectively highlight the difference between complex radiation source signals and background signals. Since complex radiation source signals are usually composed of multiple simple signals and their amplitudes vary dramatically, the fourth power processing can more significantly amplify the characteristics of these signals, thereby distinguishing them from the background signals. However, this processing method will also introduce higher nonlinear effects and computational complexity, so it is suitable for occasions that require strong feature extraction.

[0009] High-order power processing of signals can enhance the nonlinear characteristics of signals, help improve the signal's ability to resist multipath fading and other nonlinear distortions, and more reliably distinguish different signal states. By pre-classifying the signal, the appropriate signal processing method can be effectively selected, such as square or fourth power processing, to improve the accuracy and efficiency of recognition.

[0010] In order to achieve the above object, the present invention provides the following technical solutions:

[0011] A method for identifying individual radiation sources by fusion of multi-domain features, comprising:

[0012] Step S1: signal acquisition. Use a signal acquisition device to collect time domain signals from multiple communication radiation sources. The individual communication radiation sources are radiation source devices of the same type, and there are radiation source fingerprint differences between the radiation source devices of the same type. The signal emitted by the communication radiation source is collected by the signal acquisition device, and its signal is stored in the form of time domain data.

[0013] Step S2: Signal preprocessing. The collected signal is preprocessed to improve the quality and consistency of the signal. The preprocessing includes signal segmentation, normalization and standardization.

[0014] Step S2-1: Signal segmentation. The continuously collected long-term signal is segmented into multiple short-term signal segments for subsequent processing and analysis. The sliding window technology is used to move a fixed step size each time to segment the signal into signal segments to improve the time resolution.

[0015] Step S2-2: Normalization processing: Use the minimum and maximum normalization method to linearly transform the amplitude value of the signal into the range of [0, 1], eliminate the amplitude difference between different signals, and make the signals compared and analyzed on the same scale.

[0016]

[0017] Where x is the amplitude value of the signal segment, x min and x max are the minimum and maximum values ​​of the signal, respectively, x norm is the normalized signal amplitude value.

[0018] Step S2-3: Standardization: Perform mean and variance standardization on the signal to make the mean of the signal data 0 and the variance 1, so as to improve the stability of subsequent feature extraction and classification recognition.

[0019]

[0020] Among them, x norm is the normalized signal amplitude value, μ is the mean value of the signal, σ is the standard deviation of the signal, and xns is the normalized signal amplitude value.

[0021] Step S3: Signal pre-classification: Select a suitable high-order power processing method based on the modulation type of the signal.

[0022] Step S3-1: Modulation mode identification: A modulation mode identification method based on a deep residual shrinkage network (DSRN) is used, and the pre-processed signal is used as the input of the deep residual shrinkage network model to identify the modulation mode.

[0023] Step S3-1-1: Shallow feature extraction. Let the signal x ns After the first convolutional layer in the deep residual contraction network, the shallow feature y of the modulation mode is obtained 0 .

[0024] y 0 =ReLU(W0*x sn +b0)

[0025] Among them, * represents the convolution operation, W0 is the convolution kernel weight, b0 is the bias, and ReLU is the activation function.

[0026] Step S3-1-2: Modulation mode shallow feature y 0 The deep feature y is extracted through the i-th residual shrinkage block (Residual Shrikage Building Unit, RSBU) i .

[0027] First, the feature h is extracted through the first two convolutional layers in the RSBU unit. i :

[0028] h i =ReLU(W i *y 0 +b i )

[0029] Among them, W i is the convolution kernel weight, b i is the bias and ReLU is the activation function.

[0030] The extracted feature h i Perform global average pooling to obtain the feature mean A i :

[0031] A i =GAP(|h i |)

[0032] Use the attention mechanism module in the RSBU unit to generate the feature hi The scaling parameter α for each channel in i :

[0033] α i =σ(W att ·A i +b att )

[0034] Where σ(·) is the sigmoid activation function, W att is the weight of the attention mechanism, b att is the bias of the attention mechanism. The adaptive threshold τ is calculated through the attention mechanism i :

[0035] τ i =α i ×A i

[0036] Apply soft thresholding to feature h i Perform shrinkage to obtain the shrinkage feature h i ′:

[0037] h i ′=sign(h i )·max(|h i |-τ i ,0)

[0038] Among them, sign() is the sign function and max() takes the larger value.

[0039] The shrink feature h i ′ and input feature y i-1 Perform residual connection to obtain deep feature y i :

[0040] y i =h i ′+y i-1

[0041] As above, the deep feature y i After 13 RSBU units are stacked and extracted, the final feature y is obtained. 13 .

[0042] Step S3-1-3: Classifier output. The final features are passed through the fully connected layer of the deep residual shrinkage network and the softmax activation function to obtain the classification output and determine the modulation mode z:

[0043] z=softmax(W out ·y 13 +b out )

[0044] Among them, Wout The weight of the fully connected layer, b out is the bias of the fully connected layer, and z is the predicted probability vector for each modulation type.

[0045] Step S3-2: High-order power processing selection. For low-order modulation methods, that is, the order is less than or equal to 4, and the amount of information carried by each symbol is less than or equal to 2 bits, such as AM, FM, PM, 2-FSK, 4-FSK, 4-PAM, BPSK, QPSK and 4-QAM, the time domain signal is processed by quadratic power. For high-order modulation methods, that is, the order is greater than 4, and the amount of information carried by each symbol is greater than 2 bits, such as 8-FSK, 16-FSK, 8-PAM, 16-PAM, 8PSK, 16PSK, 8QAM and 16QAM, the time domain signal is processed by quartic power.

[0046] Step S4: Extract features. Perform high-order power processing and fast Fourier transform processing on the preprocessed signal to obtain time domain features and frequency domain features.

[0047] Step S4-1: high-order power processing in the time domain, i.e., square processing or quartic processing. Perform a square operation on each value in the preprocessed signal segment to obtain SI / Q, which describes the amplitude characteristics of the signal and highlights the originally blurred or confused signal spectrum.

[0048]

[0049] Each value in the signal segment is raised to the fourth power to obtain FI / Q, which highlights the difference between the complex radiation source signal and other background signals.

[0050]

[0051] Step S4-2: Fast Fourier Transform. Perform fast Fourier transform on the signal segment to obtain the signal spectrum information X k , including the amplitude and phase of the frequency component, namely FD-SP.

[0052]

[0053] Among them, x j represents the value of the time domain signal at the jth sampling point, n is the total length of the signal, that is, the number of sampling points, k is the frequency index, i is the imaginary unit, and X k is the frequency domain representation of the signal at frequency k, which is a complex number.

[0054] Step S5: Model pre-training. Pre-train a multi-domain feature fusion recognition model, which includes two feature extraction networks and one feature fusion recognition network. Back propagation based on the gradient descent method is used in the training process, and the loss function uses a weighted cross entropy loss function.

[0055] Step S5-1: Feature vector extraction. The signal combination SI / Q+FD-SP or FI / Q+FD-SP is used as the input of the multi-domain feature fusion recognition model, and features are extracted by two feature extraction networks, namely the convolutional neural network, to obtain a feature vector F containing the time domain. t and contains the frequency domain eigenvector F s , linearly concatenate the two eigenvectors into a multi-domain eigenvector F in terms of size m .

[0056] Step S5-2: Feature vector fusion. The multi-domain feature vector F m As the input of the feature fusion recognition network, the fully connected layer and ReLU activation layer in the feature fusion recognition network are transformed into the multi-domain feature vector F m Weighted and nonlinear transformations are performed to further refine and fuse features and enhance the expressiveness of features. The use of multiple fully connected layers and ReLU activation functions can gradually extract more complex and advanced features. Multi-domain feature vector F m The formula for multiple fully connected layers and activation layers is as follows:

[0057] F1 output =ReLU(W1·F m +b1)]

[0058] F2 output =ReLU(W2·F1 output +b2)

[0059] F3 output =ReLU(W3·F2 output +b3)

[0060] Where W1, W2, and W3 are weight matrices of the fully connected layer, b1, b2, and b3 are bias vectors, and the ReLU activation function is f(z)=max(0,z), where z is the input vector from the previous neural network layer.

[0061] Step S5-3: Calculate the category probability. The vector F3 that successfully integrates the time domain features and frequency domain features in step S5-2 is output Send it to the softmax function classifier. When there are D categories in total, it outputs a D-dimensional vector, which is the probability of each category to which the sample belongs:

[0062]

[0063] in represents the probability that the sample belongs to the i-th category, z i is the i-th element in the input vector z, zd is the d-th element in the input vector z.

[0064] Step S5-4: Calculate the classification loss. Use the cross entropy loss function L to measure the difference between the predicted probability distribution and the true distribution:

[0065]

[0066] where y=[y1,y2,...,y D ] is the true label vector, where y i Indicates whether the sample belongs to the i-th category. If the sample belongs to the i-th category, then y i =1, otherwise y i =0.

[0067] Step S5-5: Back propagation. Adjust the optimization model hyperparameters according to the Adam optimization algorithm until the cross entropy loss function converges. The Adam optimization algorithm combines the advantages of momentum and RMSProp (Root Mean Square Propagation), adaptively adjusts the learning rate, and makes the training process more stable and efficient.

[0068] Initialize the hyperparameters of the Adam algorithm: learning rate η, exponential decay rate β1 of the first-order moment estimate, exponential decay rate β2 of the second-order moment estimate, and bias correction term ε, with typical values ​​of η = 0.001, β1 = 0.9, β2 = 0.999, ε = 10 -8 , then initialize the first-order moment vector m t and the second-order moment vector v t is the zero vector.

[0069] At each iteration, the loss function is calculated with respect to the model parameters θ t The gradient g t :

[0070]

[0071] Update the first-order moment estimate m t and the second-order moment estimate v t :

[0072] m t =β1m t-1 +(1-β1)g t

[0073] v t =β2v t-1 +(1-β2)g t 2

[0074] For m t and vt Correction for deviation and

[0075]

[0076] Update the model parameters θ using the corrected moment estimates t :

[0077]

[0078] Where η is the learning rate, It is β1 to the power of t, which represents the bias correction factor of the first-order moment estimate in the tth iteration. It is β2 to the power of t, which represents the bias correction factor of the second-order moment estimate in the tth iteration. t is the gradient, L(·) cross entropy is the loss function, θ t-1 is the model parameter of the previous iteration, m t-1 and v t-1 They represent the first-order moment estimate and the second-order moment estimate of the previous iteration respectively.

[0079] Step S6: Individual identification of radiation sources. The newly received signal is segmented, normalized and standardized, and multi-domain features are extracted according to the above steps. The extracted features are fused to form a feature vector, which is input into the trained classification model to obtain the identification probability of each radiation source.

[0080] The present invention also provides a radiation source individual identification device with multi-domain feature fusion, comprising:

[0081] The signal acquisition module is used to collect time domain signals, collect the signals emitted by the communication radiation source, and store the signals in the form of time domain data;

[0082] The signal preprocessing module is used to preprocess the collected signals, including operations such as signal segmentation, normalization and standardization, so as to improve the quality and consistency of the signals;

[0083] The signal pre-classification module selects the appropriate high-order power processing method based on the modulation type of the signal;

[0084] The signal processing module performs high-order power processing and fast Fourier transform processing on the signal to obtain time domain features and frequency domain features;

[0085] Model pre-training module, used to pre-train multi-domain feature fusion recognition model;

[0086] The radiation source individual identification module is used to segment, normalize and standardize the newly received signals, extract multi-domain features according to the above steps, fuse the extracted features to form a feature vector, input the trained classification model, and obtain the identification probability of each radiation source;

[0087] The signal receiver in the signal acquisition module is a USRP-2954R general-purpose software radio device with a center frequency of 375MHz. The sampling rate is set to 100MHz to perform I / Q orthogonal dual-channel sampling on the transmitted signal. The communication signals sent by four simulated radiation sources are collected. The total number of samples is 20,000. The sample order is shuffled and divided into training set, test set and validation set, with ratios of 0.6, 0.2, and 0.2 respectively; the signal preprocessing module and signal processing module are implemented in the CPU; the signal pre-classification module, model pre-training module, and radiation source individual identification module are implemented in the GPU and CPU. The CPU uses the 13th Gen Intel(R)Core(TM)i5-13400F and the GPU uses the NVIDIA Geforce RTX 4060Ti.

[0088] Compared with the prior art, the beneficial effects of the present invention are: the fused multi-domain features can better reflect the subtle differences in radiation sources, and the recognition effect is significantly improved; by extracting the features of each domain separately, the dimension of the data can be effectively reduced, thereby reducing the complexity of calculation and storage; at the same time, it allows targeted processing and fusion of features from different domains to improve classification accuracy; features from different domains can be selected and weighted to enhance the influence of important features and make the model more flexible and accurate; it uses ordinary computer hardware configuration, and is highly practical and economical. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 is a flow chart of the steps of the method of the present invention;

[0090] Figure 2 It is a flow chart of the deep residual shrinkage network identification method in the method of the present invention;

[0091] Figure 3 It is a structural diagram of a convolutional neural network in the method of the present invention;

[0092] Figure 4 It is a feature fusion recognition network structure diagram in the method of the present invention;

[0093] Figure 5 Identify the effect diagram for a single signal;

[0094] Figure 6 It is a diagram showing the recognition effects of different signal combinations in the method of the present invention;

[0095] Figure 7This is a performance comparison diagram of the method of the present invention and other methods;

[0096] Figure 8 This is a structural diagram of the radiation source individual identification device in the present invention. DETAILED DESCRIPTION

[0097] In order to make the objectives, technical methods and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0098] Attached Figure 1 The figure is a flow chart of the steps of the method of the present invention.

[0099] In a specific embodiment of the present invention.

[0100] The experimental data set uses matlab simulation to generate four radiation source signals with a signal carrier frequency of 1800 Hz and a modulation mode of 8PSK.

[0101] Step S1: signal acquisition. The sampling frequency is 40000 Hz, the number of signal samples for each radiation source is 10000, and a total of 40000 samples are stored in the form of time domain data. The sample signal is passed through the Rayleigh fading channel to simulate the multipath effect encountered in wireless propagation. The ratios of the training set, test set, and validation set are 0.6, 0.2, and 0.2, respectively.

[0102] Step S2: Signal preprocessing. The signal is preprocessed, including signal segmentation, normalization and standardization, to improve the quality and consistency of the signal. Then the signal is processed by square, fourth power and fast Fourier transform to provide time domain features and frequency domain features.

[0103] Step S2-1: Use the sliding window technique, move a fixed step size each time, and divide the signal into signal segments to improve the time resolution. Each sample contains 2048 sampling points.

[0104] Step S2-2: Eliminate the amplitude differences between different signals so that the signals can be compared and analyzed at the same scale. Use the minimum and maximum normalization method to linearly transform the amplitude values ​​of the signals into the range of [0,1].

[0105] Step S2-3: perform mean and variance normalization on the signal so that the mean of the signal is 0 and the variance is 1, so as to improve the stability of subsequent feature extraction and classification recognition.

[0106] Step S3: Signal pre-classification: Select a suitable high-order power processing method based on the modulation type of the signal.

[0107] Step S3-1: Modulation mode identification. The modulation mode of the signal is identified by the modulation mode identification method based on the deep residual shrinkage network DRSN, and it can be known that the modulation mode is 8PSK. The deep residual shrinkage network identification method, such as Figure 2 shown.

[0108] Step S3-2: High-order power processing mode selection: 8PSK is a high-order modulation mode, and the time domain signal is processed by the fourth power.

[0109] Step S4: extract features, perform high-order power processing and fast Fourier transform processing on the signal to provide time domain features and frequency domain features.

[0110] Step S4-1: Perform a fourth power operation on each value in the signal segment to highlight the difference between the complex radiation source signal and other background signals.

[0111] Step S4-2: Perform fast Fourier transform on the signal segment to obtain the spectrum information of the signal, including the amplitude and phase of the frequency component.

[0112] Step S5: Model pre-training. The multi-domain feature fusion recognition model is pre-trained, including two feature extraction networks and one feature fusion recognition network. The training process uses back propagation based on the gradient descent method, and the loss function uses the weighted cross entropy loss function.

[0113] Step S5-1: Feature extraction. The signal combination FI / Q+FD-SP is used as the input of the model and is processed by two methods such as Figure 3 The convolutional neural network is used to extract features. The convolutional neural network contains 8 convolutional layers with a step size of 2 and a convolution kernel size of 1×3. Each convolutional layer is connected to a batch normalization layer and a ReLU activation layer, and finally ends with an average pooling layer. The time domain feature vector F with a size of 1×8×512 is obtained through two convolutional neural networks. t and the frequency domain feature vector F with size 1×8×512 s , linearly concatenate the two feature vectors into a multi-domain feature vector F with a size of 1×16×512 m .

[0114] Step S5-2: Feature fusion. Multi-domain feature vector F m As Figure 4 The input of the feature fusion recognition network, the fully connected layer and the ReLU activation layer in the network further refine and fuse the features by weighting and nonlinearly transforming the features, thereby enhancing the expressiveness of the features.

[0115] The feature fusion recognition network contains three fully connected layers, with the number of nodes being 512, 128, and 128 respectively. The number of fully connected nodes in the last layer is set to the number of categories of the radiation source. Each neuron in the fully connected layer is connected to all neurons in the previous layer to form a dense neural network structure. The fully connected layer performs weighted summation of the input features through linear transformation, efficiently learns the relationship between features, and realizes the linear combination of features. The weight coefficients in the weight matrix reflect the importance of each feature, and then nonlinearity is introduced through the activation function, so that the fully connected layer can capture the complex relationship and high-order interaction information between features. The fully connected layer achieves feature dimensionality reduction by reducing the number of neurons layer by layer, and maps high-dimensional features to low-dimensional space. This not only reduces the computational complexity, but also extracts more concise and abstract features.

[0116] Step S5-3: Calculate the category probability. The vector F3 that successfully integrates the time domain features and frequency domain features in step S5-2 is output Send it to the softmax function classifier and output a 4-dimensional vector.

[0117] Softmax is used as the activation function before the model output. It is used to predict the probability of the output. The larger the output value of the neuron, the more likely the neuron corresponds to the true category. The recognition probability of each radiation source is:

[0118]

[0119] Where D is the total number of communication radiation source devices, and the output values ​​of the multi-classification are converted into a probability distribution ranging from [0, 1] and 1.

[0120] Step S5-4: Calculate the classification loss. The cross entropy loss function is used to measure the difference between the predicted probability distribution and the true distribution.

[0121] Step S5-5: Back propagation, adjust the optimization model hyperparameters according to the optimization method Adam until the loss function converges. The Adam optimization method combines the advantages of momentum and RMSProp, adaptively adjusts the learning rate, and makes the training process more stable and efficient.

[0122] Step S6: Individual identification of radiation sources. The newly received signal is segmented, normalized and standardized, and multi-domain features are extracted according to the above steps. The extracted features are fused to form a feature vector, which is input into the trained classification model to obtain the identification probability of each radiation source.

[0123] In order to explore the recognition effect of multi-domain feature fusion, the recognition effect of four single signals was first tested, using CNN as the trunk network of this design. Figure 5. FI / Q has the best recognition effect in single signal input. When the signal-to-noise ratio is -5dB, the recognition accuracy can reach 59.72%. When the signal-to-noise ratio is 10dB to 30dB, the recognition accuracy is above 80%, and the accuracy reaches 90.45% at 30dB. The recognition effect of TD-I / Q is the worst, with an accuracy of only 65.22% at 30dB. SI / Q has better recognition effect than FD-SP at 5dB to 30dB, with a recognition accuracy of 85.67% at 30dB. The recognition effects of the three time domain signals are greatly affected by noise, and the recognition accuracy curve is steeper. FD-SP has stronger noise resistance, a relatively flat recognition accuracy curve, and an overall recognition effect is better than TD-I / Q.

[0124] Single-domain features cannot fully characterize the fingerprint features of communication radiation sources. In order to avoid the limitations of single-domain features, the features of time domain and frequency domain are extracted and fused for identification. Figure 6 As shown in the figure, TD-I / Q and FD-SP are signals that have been repeatedly used in previous explorations. Therefore, the combination of TD-I / Q+FD-SP is used as the basic control group. The time domain features and frequency domain features are integrated. The accuracy curve is shown in the square label curve. The accuracy of -5dB~10dB rises rapidly from 68.78% to 87.2.%, and the accuracy of 15dB~30dB remains above 91%. Compared with the recognition effect of TD-I / Q and FD-SP input separately, the recognition effect is significantly improved, which preliminarily verifies the effectiveness of the method proposed in the present invention.

[0125] exist Figure 5 It can be observed that due to the low recognition rate of TD-I / Q, the time domain features contained in SI / Q and FI / Q are more significant. Therefore, SI / Q and FI / Q are combined with FD-SP respectively. The experimental results show that the recognition accuracy of the two combinations is significantly improved compared with the combination of TD-I / Q+FD-SP. The accuracy curve of the combination SI / Q+FD-SP is about 6% higher than that of the combination TD-I / Q+FD-SP. The recognition effect of the combination FI / Q+FD-SP is the best, with an accuracy of 85.65% at -5dB and an accuracy of more than 99% above 15dB. It maintains the high recognition rate of FI / Q at high signal-to-noise ratio and combines the good noise resistance of FD-SP at low signal-to-noise ratio.

[0126] In order to eliminate the influence of SI / Q and FI / Q single domain features on the experiment, two combination experiments, TD-I / Q+SI / Q and TD-I / Q+FI / Q, which do not contain frequency domain features, were set up to demonstrate the advantages of multi-domain feature fusion. The recognition results of the two combinations improved rapidly in the accuracy curve of -5dB to 15dB, indicating the sensitivity of time domain features to noise and the limitations of single domain features. The recognition effect of the combination TD-I / Q+SI / Q is completely inferior to the combination SI / Q+FD-SP. The combination TD-I / Q+FI / Q also has no advantage over the combination FI / Q+FD-SP, and the combination containing frequency domain features performs relatively smoothly at low signal-to-noise ratios.

[0127] The recognition performance of the invention is compared with that of other radiation source signal feature extraction methods. In order to ensure the rigor of the experimental data and the effectiveness of the comparison, the control variable method is used to complete the relevant experiments. The original data set and feature extraction network used in the experiment are consistent with Example 1. Cao et al. proposed a method for extracting and identifying subtle features of orthogonal modulation of communication radiation sources based on local bispectrum. Cun et al. converted the original signal into multi-domain signals such as I / Q eye diagram, vector diagram and Hilbert-VMD time-frequency spectrum as network input. Wang et al. extracted the mean, variance, kurtosis, skewness and information entropy of HHT marginal spectrum and bispectrum for splicing and fusion to achieve classification. Figure 7 The recognition performance of different methods is compared and it is found that the combination of SI / Q+FD-SP and the combination of FI / Q+FD-SP in this embodiment has the best recognition effect, and its performance at low signal-to-noise ratio is also better than other methods, and the computational complexity is reduced.

[0128] In another specific embodiment of the present invention, the radiation source identification signal data set is generated by a radiation source signal simulation platform of an AD9910 direct digital frequency synthesizer and a USRP-2954R software radio device, and the modulation method uniformly adopts QPSK.

[0129] The radiation source signal simulation platform consists of three parts: signal transmitter, signal receiver and signal processor. The general computer in the signal processor is used to uniformly schedule and complete the experimental data acquisition. The signal transmitter uses the STM32f401ccu6 microcontroller as the main controller to receive the signal generation command from the host computer software, and then controls the AD9910 direct digital frequency synthesizer through the SPI serial data bus to generate the corresponding modulated radar pulse signal. The simulation of different radar radiation sources is achieved by switching different power amplifier channels.

[0130] Step S1: Signal acquisition: Time domain signals are collected from multiple radiation sources using the same receiving device.

[0131] The signal receiver is a USRP-2954R general software radio device with a center frequency of 375MHz and a sampling rate of 100MHz. The transmitted signal is sampled in I / Q quadrature dual channels, and the communication signals sent by four simulated radiation sources are collected. The total number of samples is 20,000. At the same time, the order of the samples is disrupted and divided into training set, test set and validation set, with the ratios of 0.6, 0.2 and 0.2 respectively. The signal collection location is indoors, simulating a relatively ideal communication environment.

[0132] Step S2: Preprocess the signal, including signal segmentation, normalization and standardization, to improve the quality and consistency of the signal.

[0133] Step S3: Signal pre-classification: Select a suitable high-order power processing method based on the modulation type of the signal.

[0134] Step S3-1: Modulation mode identification: The modulation mode of the signal is identified by a modulation mode identification method based on a deep residual shrinkage network DRSN, and it can be known that the modulation mode is QPSK.

[0135] Step S3-2: High-order power processing method selection: QPSK is a simple modulation method, and the time domain signal is processed by quadratic power.

[0136] Step S4: extract features, perform high-order power processing and fast Fourier transform processing on the signal to provide time domain features and frequency domain features.

[0137] Step S4-1: Perform a square operation on each value in the signal segment to characterize the amplitude characteristics of the signal and highlight the signal spectrum lines that were originally blurred or confused.

[0138] Step S4-2: Perform fast Fourier transform on the signal segment to obtain the spectrum information of the signal, including the amplitude and phase of the frequency component.

[0139] Step S5: Model pre-training. The signal combination SI / Q+FD-SP is used as the input of the model to pre-train the multi-domain feature fusion recognition model. The back propagation based on the gradient descent method is used in the training process, and the loss function adopts the weighted cross entropy loss function.

[0140] Step S6: Individual identification of radiation sources. The newly received signal is segmented, normalized and standardized, and multi-domain features are extracted according to the above steps. The extracted features are fused to form a feature vector, which is input into the trained classification model to obtain the identification probability of each radiation source.

[0141] The recognition results of four single signal inputs are shown in Table 1. Since the data collected by the receiver indoors can be regarded as an ideal channel environment, the recognition accuracy of TD-I / Q, SI / Q and FD-SP is close to the experimental results based on the simulation data set at 30dB. After the QPSK modulated signal is processed by square and fourth power, the constellation diagrams of the two are transformed from four constellation points to two constellation points, but the positions in the phase space are different, resulting in the recognition results of S-IQ and F-IQ being relatively close, 86.78% and 86.25% respectively.

[0142] Table 1 Recognition effects of different single signals

[0143]

[0144] The recognition results of different signal form combination inputs are shown in Table 2. The recognition effects of the five combination inputs are all above 95%, which is higher than the recognition accuracy of a single signal input. The combination TD-I / Q+FD-SP, which integrates time domain and frequency domain features, has the lowest recognition accuracy of 95.21%. The combination TD-I / Q+SI / Q and the combination TD-I / Q+FI / Q contain richer time domain features and ignore the influence of noise. The recognition results are better than the combination TD-I / Q+SI / Q, which are 96.72% and 96.08% respectively. The combination of SI / Q and FI / Q with significant time domain features and FD-SP has the best effect of more than 98.6.

[0145] Table 2 Recognition effects of different signal combinations

[0146]

[0147] The verification of the measured signal further illustrates the effectiveness of the multi-domain feature fusion method. The feature fusion learning neural network fully exploits the diversity and complementarity between multi-domain features, and the recognition accuracy is significantly better than other single signal input methods. The TD-I / Q signal is processed by high-order power, which can map the signal to a high-dimensional space, increase the distance between adjacent constellation points in the constellation diagram, and help improve the signal's anti-interference ability against multipath fading and other nonlinear distortions, thereby helping to fully exploit the signal's time domain characteristics and improve the discrimination of radiation sources.

[0148] The present invention also provides a radiation source individual identification device with multi-domain feature fusion, such as Figure 8 As shown, including:

[0149] The signal acquisition module is used to collect time domain signals, collect the signals emitted by the communication radiation source, and store the signals in the form of time domain data;

[0150] The signal preprocessing module is used to preprocess the signal collected by the signal acquisition module, including operations such as signal segmentation, normalization and standardization, so as to improve the quality and consistency of the signal;

[0151] The signal pre-classification module identifies the modulation category of the signal output by the signal pre-processing module and selects an appropriate high-order power processing method;

[0152] The signal processing module performs high-order power processing and fast Fourier transform processing on the signal output by the signal preprocessing module to obtain time domain features and frequency domain features;

[0153] A model pre-training module uses the signal output by the signal pre-processing module to pre-train a multi-domain feature fusion recognition model;

[0154] The radiation source individual identification module is used to segment, normalize and standardize the newly received signal, extract multi-domain features according to the above steps, fuse the extracted features to form a feature vector, input the trained multi-domain feature fusion recognition model, and obtain the recognition probability of each radiation source;

[0155] The signal acquisition module uses USRP-2954R general software radio equipment with a center frequency of 375MHz and a sampling rate of 100MHz to perform I / Q orthogonal dual-channel sampling on the transmitted signal, and collects communication signals sent by 4 simulated radiation source individuals. The total number of samples is 20,000, and the sample order is shuffled and divided into training set, test set and verification set, with proportions of 0.6, 0.2, and 0.2 respectively. The signal preprocessing module and the signal processing module are implemented using the CPU, and the signal pre-classification module, the model pre-training module and the radiation source individual identification module are implemented using the CPU and the GPU. The CPU is a 13th Gen Intel(R)Core(TM)i5-13400F, and the GPU is an NVIDIA Geforce RTX 4060Ti, both of which are ordinary computer hardware configurations. It can be seen that the present invention is very practical and economical.

[0156] Aiming at the limitations of the current mainstream single-domain feature radiation source identification method, the present invention proposes a communication radiation source identification method based on multi-domain feature fusion from the perspective of fully characterizing fingerprint features. The original time domain I / Q signal is processed by high-order power and fast Fourier transform. Different signal forms can reflect different fingerprint features, forming a combined input multi-domain feature fusion recognition model. The features are extracted by two sub-networks respectively and fused and recognized by the fully connected layer to make up for the shortcomings of single domain features. The experimental results of the simulation data set and the measured data set show that compared with the method based on single domain feature recognition, the multi-domain features fused by the present invention can better reflect the subtle differences of the radiation source, and the recognition effect is significantly improved.

[0157] The advantages of this multi-domain feature fusion model include reduced dimensionality and high flexibility. By extracting the features of each domain separately, the dimension of the data can be effectively reduced, thereby reducing the complexity of calculation and storage. This method allows targeted processing and fusion of features from different domains to improve classification accuracy; it can also select and weight features from different domains to enhance the influence of important features, making the model more flexible and accurate.

[0158] The method of the present invention can be applied to scenarios such as intelligent communication confrontation, electromagnetic spectrum monitoring, public safety monitoring and equipment fault diagnosis, effectively improving the intelligence and automation level of various systems and having strong practicality.

Claims

1. A method for identifying individual radiation sources by fusion of multi-domain features. include: Step S1: signal acquisition, i.e., acquiring time domain signals from multiple communication radiation sources using a signal acquisition device; Step S2: signal preprocessing, i.e. preprocessing the collected signal, including signal segmentation, normalization and standardization; Step S3: signal pre-classification, that is, starting from the modulation type of the pre-processed signal, selecting an appropriate high-order power processing method; Step S4: extracting features, i.e. performing high-order power processing and fast Fourier transform processing on the preprocessed signal to obtain time domain features and frequency domain features; Step S5: model pre-training, i.e. pre-training a multi-domain feature fusion recognition model; Step S6: Individual identification of radiation sources, that is, segmenting, normalizing and standardizing the newly received signals, extracting multi-domain features, fusing the extracted features to form feature vectors, inputting the trained classification model, and obtaining the identification probability of each radiation source.

2. A radiation source individual identification method based on multi-domain feature fusion as claimed in claim 1, Features: In step S1, the signal emitted by the communication radiation source is collected by a signal collection device, and the signal is stored in the form of time domain data; Step S2 includes: Step S2-1: signal segmentation, that is, segmenting the continuously collected long-time signal into multiple short-time signal segments, using a sliding window technique, moving a fixed step size each time, to segment the signal into signal segments; Step S2-2: Normalization processing, that is, using the minimum and maximum normalization method to linearly transform the amplitude value of the signal to the range of [0,1]; Where x is the amplitude value of the signal segment, x min and x max are the minimum and maximum values ​​of the signal, respectively, x norm is the normalized signal amplitude value; Step S2-3: Standardization processing, that is, performing mean and variance standardization processing on the signal so that the mean of the signal data is 0 and the variance is 1; Among them, x norm is the normalized signal amplitude value, μ is the mean value of the signal, σ is the standard deviation of the signal, and x ns is the normalized signal amplitude value.

3. A radiation source individual identification method based on multi-domain feature fusion as claimed in claim 2, Features: Step S3 includes: Step S3-1: modulation mode recognition, that is, using a modulation mode recognition method based on a deep residual shrinkage network, taking the preprocessed signal as the input of the deep residual shrinkage network model to recognize the modulation mode; Step S3-2: High-power processing selection.

4. The method for identifying individual radiation sources by fusion of multi-domain features as claimed in claim 3, Features: Step S3-1 includes: Step S3-1-1: Shallow feature extraction, that is, signal x ns After the first convolutional layer in the deep residual contraction network, the shallow feature y of the modulation mode is obtained 0 ; y 0 =ReLU(W 0 *x sn +b 0 ) Among them, * represents the convolution operation, W 0 is the convolution kernel weight, b 0 is the bias, ReLU is the activation function; Step S3-1-2: Modulation mode shallow feature y 0 The deep feature yi is extracted through the i-th residual shrinkage block (Residual Shrikage Building Unit, RSBU); First, the feature h is extracted through the first two convolutional layers in the RSBU unit. i : h i =ReLU(W i *y 0 +b i ) Among them, W i is the convolution kernel weight, b i is the bias, ReLU is the activation function; The extracted feature h i Perform global average pooling to obtain the feature mean A i : A i =GAP(|h i |) Use the attention mechanism module in the RSBU unit to generate the feature h i The scaling parameter α for each channel in i : a i =σ(W att ·A i +b att ) Where σ(·) is the sigmoid activation function, W att is the weight of the attention mechanism, b att is the bias of the attention mechanism; Calculate the adaptive threshold τ through the attention mechanism i : t i =a i ×A i Apply soft thresholding to feature h i Perform shrinkage to obtain the shrinkage feature h′ i : h′ i =sign(h i )·max(|h i |-τ i ,0) Among them, sign() is the sign function, and max() takes the largest value; The shrink feature h′ i With the input feature y i-1 Perform residual connection to obtain deep feature y i : y i =h′ i +y i-1 As above, the deep feature y i After 13 RSBU units are stacked and extracted, the final feature y is obtained. 13 ; Step S3-1-3: Classifier output, that is, the final feature is passed through the fully connected layer of the deep residual shrinkage network and the softmax activation function to obtain the classification output and determine the modulation mode z: z =softmax(W out ·y 13 +b out ) Among them, W out The weight of the fully connected layer, b out is the bias of the fully connected layer, z is the predicted probability vector for each modulation type; In step S3-2, for low-order modulation methods, that is, the order is less than or equal to 4, the amount of information carried by each symbol is less than or equal to 2 bits, and the time domain signal is processed quadratically; for high-order modulation methods, that is, the order is greater than 4, the amount of information carried by each symbol is greater than 2 bits, and the time domain signal is processed quartically.

5. The method for identifying individual radiation sources by fusion of multi-domain features as claimed in claim 4, Features: In step S3-2, low-order modulation modes include: AM, FM, PM, 2-FSK, 4-FSK, 4-PAM, BPSK, QPSK and 4-QAM; high-order modulation modes include: 8-FSK, 16-FSK, 8-PAM, 16-PAM, 8PSK, 16PSK, 8QAM and 16QAM.

6. The method for identifying individual radiation sources by fusion of multi-domain features as claimed in claim 1, Features: Step S4 includes: Step S4-1: high-order power processing in the time domain, i.e., square processing or quartic processing, performs a square operation on each value in the preprocessed signal segment to obtain SI / Q; Take each value in the signal segment and raise it to the fourth power to get FI / Q; Step S4-2: Fast Fourier transform, that is, fast Fourier transform is performed on the signal segment to obtain the signal spectrum information X k , including the amplitude and phase of the frequency component, i.e., FD-SP; Among them, x j represents the value of the time domain signal at the jth sampling point, n is the total length of the signal, that is, the number of sampling points, k is the frequency index, i is the imaginary unit, X k is the frequency domain representation of the signal at frequency k, which is a complex number.

7. The method for identifying individual radiation sources by fusion of multi-domain features as claimed in claim 1, Features: Step S5 includes: Step S5-1: feature vector extraction, that is, the signal combination SI / Q+FD-SP or FI / Q+FD-SP is used as the input of the multi-domain feature fusion recognition model, and features are extracted by two feature extraction networks, namely the convolutional neural network, to obtain a feature vector F containing the time domain t and contains the eigenvector F in the frequency domain s , linearly concatenate the two eigenvectors into a multi-domain eigenvector F in terms of size m ; Step S5-2: Feature vector fusion, that is, multi-domain feature vector F m As the input of the feature fusion recognition network, the fully connected layer and ReLU activation layer in the feature fusion recognition network are transformed into the multi-domain feature vector F m Perform weighted and nonlinear transformation, multi-domain feature vector F m The formula after 3 fully connected layers and activation layers is as follows: F1 output =ReLU(W1·F m +b1)] <h2 style=";text-align:left;direction:ltr">F2<h2 style=";text-align:left;direction:ltr"> output <h2 style=";text-align:left;direction:ltr"> =ReLU(W2·F1<h2 style=";text-align:left;direction:ltr"> output <h2 style=";text-align:left;direction:ltr"> +b2) <h2 style=";text-align:left;direction:ltr">F3<h2 style=";text-align:left;direction:ltr"> output <h2 style=";text-align:left;direction:ltr"> =ReLU(W3·F2<h2 style=";text-align:left;direction:ltr"> output <h2 style=";text-align:left;direction:ltr"> +b3) Where W1, W2, W3 are weight matrices of the fully connected layer, b1, b2, b3 are bias vectors, the ReLU activation function is f(z)=max(0, z), and z is the input vector from the previous neural network layer; Step S5-3: Calculate the category probability, that is, the vector F3 that successfully integrates the time domain features and frequency domain features in step S5-2 output Send it to the softmax function classifier. When there are D categories in total, it outputs a D-dimensional vector, which is the probability of each category to which the sample belongs: in represents the probability that the sample belongs to the i-th category, z i is the i-th element in the input vector z, z d is the d-th element in the input vector z; Step S5-4: Calculate the classification loss, that is, use the cross entropy loss function L to measure the difference between the predicted probability distribution and the true distribution: where y = [y 1 ,y 2 , ..., y D ] is the true label vector, where y i Indicates whether the sample belongs to the i-th category. If the sample belongs to the i-th category, then y i =1, otherwise y i =0; Step S5-5: back propagation, i.e. adjusting the optimization model hyperparameters according to the Adam optimization algorithm until the cross entropy loss function converges; Initialize the hyperparameter learning rate η of the Adam algorithm and the exponential decay rate β of the first-order moment estimate 1 , the exponential decay rate of the second-order moment estimate β 2 and the bias correction term ε, with a typical value of η = 0.001, β 1 =0.9,β 2 =0.999,ε=10 -8 , then initialize the first-order moment vector m t and the second-order moment vector v t is the zero vector; At each iteration, the loss function is calculated with respect to the model parameters θ t The gradient g t : Update the first-order moment estimate m t and the second moment estimate v t : m t =b 1 m t-1 +(1-β 1 )g t v t =b 2 v t-1 +(1-β 2 )g t 2 For m t and v t Correction for deviation and Update the model parameters θ using the corrected moment estimates t : Where η is the learning rate, Is β 1 The t-th power represents the bias correction factor of the first-order moment estimate in the t-th iteration; Is β 2 The t-th power represents the bias correction factor of the second-order moment estimate in the t-th iteration; g t is the gradient, L(·) cross entropy is the loss function, θ t-1 is the model parameter of the previous iteration, m t-1 and v t-1 They represent the first-order moment estimate and the second-order moment estimate of the previous iteration respectively.

8. A radiation source individual identification device based on multi-domain feature fusion, include: The signal acquisition module is used to collect time domain signals, collect the signals emitted by the communication radiation source, and store the signals in the form of time domain data; The signal preprocessing module is used to preprocess the collected signals, including signal segmentation, normalization and standardization, so as to improve the quality and consistency of the signals; The signal pre-classification module selects the appropriate high-order power processing method based on the modulation type of the signal; The signal processing module performs high-order power processing and fast Fourier transform processing on the signal to obtain time domain features and frequency domain features; Model pre-training module, used to pre-train multi-domain feature fusion recognition model; The radiation source individual identification module is used to segment, normalize and standardize the newly received signals, extract multi-domain features according to the above steps, fuse the extracted features to form a feature vector, input the trained classification model, and obtain the identification probability of each radiation source.

9. The radiation source individual identification device using multi-domain feature fusion as claimed in claim 8, Features: The signal receiver in the signal acquisition module is a USRP-2954R general-purpose software radio device with a center frequency of 375MHz. The sampling rate is set to 100MHz to perform I / Q orthogonal dual-channel sampling on the transmitted signal. The communication signals sent by four simulated radiation sources are collected. The total number of samples is 20,000. The sample order is shuffled and divided into training set, test set and validation set, with ratios of 0.6, 0.2, and 0.2 respectively; the signal preprocessing module and signal processing module are implemented in the CPU; the signal pre-classification module, model pre-training module, and radiation source individual identification module are implemented in the GPU and CPU.

10. The radiation source individual identification device of claim 9 using multi-domain feature fusion, Features: The CPU uses the 13th Gen Intel(R)Core(TM)i5-13400F, and the GPU uses the NVIDIA Geforce RTX4060Ti.

Images