An intelligent identification method for individual communication radiation sources in cross-period scenarios

Abstract

The present invention discloses a method for intelligently identifying individual communication radiation sources in cross-period scenarios, relating to the technical field of radiation source individual identification. To address the problem of the degradation of individual distinguishability of radiation sources over time due to the inherent characteristics of communication signal radiation sources and the influence of complex electromagnetic environmental factors, the present invention, based on a generative adversarial network, constructs a generator using residual blocks to establish a mapping relationship between the feature distributions of data from two periods. By adding noise to the discriminator and discarding neurons with a certain probability, the model training process is stabilized and the risk of overfitting is reduced. Through adversarial training, a robust classifier with strong generalization capabilities is constructed. Through the above technical solutions, the present invention can achieve stable and effective individual identification of communication radiation sources in cross-period scenarios.

Description

Technical Field

[0001] The present invention relates to the field of radiation source individual identification, and more specifically, to a method for communication radiation source individual identification based on a generative adversarial network in a cross-period scenario. Background Art

[0002] Emitter identification is a technology that intercepts and analyzes the signals emitted by emitters to identify specific emitters. It provides decision-making support for subsequent positioning and tracking, and is a crucial research topic in electronic countermeasures, communications reconnaissance, and spectrum management.

[0003] Traditional communication radiation source identification methods extract signal features based on prior information and establish an individual feature library. Subsequent identification is achieved by querying the library and comparing features. Classic features include instantaneous frequency and instantaneous phase. This method relies too much on prior information, and its applicability has been greatly reduced with the development of modern electronic technology. Subsequently, researchers have adopted several feature extraction methods based on modern signal processing. Typical examples include short-time Fourier transform, Hilbert-Huang transform, wavelet transform, empirical mode decomposition, and variational mode decomposition. After extracting transform domain features, machine learning algorithms such as support vector machines, nearest neighbor methods, and decision trees are used for classification and identification. These methods have achieved good recognition results when the degree of differentiation between individuals is relatively high.

[0004] In recent years, with the continuous development of machine learning, artificial intelligence technologies, represented by deep learning, have emerged and are gradually being used to identify individual radiation sources. Neural networks can implement complex nonlinear function mappings, effectively and simply solving some difficult nonlinear mapping problems. They can also extract abstract features that are difficult to define and extract using traditional methods, enabling end-to-end feature extraction and classification. This has achieved relatively good results in the field of identifying individual radiation sources.

[0005] Both of the above methods only identify individual emitters from the same period. However, in real-world scenarios, individual emitters often come from different periods, creating a cross-period scenario. Identifying individual emitters in this cross-period scenario involves using a recognition model trained with signal data from individual emitters from a previous period to identify the same individual emitter over time. The main difference between this and conventional identification is that the training and test data come from two different periods, and the feature distributions of the data from these two periods may be inconsistent. This can lead to varying degrees of decline in the recognition accuracy of previously trained models on later data over time.

[0006] Finding effective methods for extracting individual features and achieving effective identification across time periods is of great research significance. Researchers studying the emission mechanisms of radiation sources have discovered that variations in the circuits and components of the transmitters themselves can lead to subtle differences in the transmitted signals. Analyzing and extracting these subtle differences as individual features can enable identification of individual radiation sources. However, traditional feature extraction methods struggle to define, capture, and extract these subtle differences in the transmitted signals. Furthermore, over time, as the transmitter's operating environment changes and components age to varying degrees, these subtle differences in the transmitted signals also vary. This presents new challenges and difficulties for the stable and effective identification of individual radiation sources across time periods. Summary of the Invention

[0007] The purpose of the present invention is to provide an intelligent identification method for individual communication radiation sources in cross-period scenarios in response to the defects of the existing technology and the problems existing in the identification of individual radiation sources in cross-period scenarios, so as to achieve stable and effective identification of individual radiation sources in cross-period scenarios.

[0008] To achieve the above object, the present invention adopts the following technical solutions:

[0009] A method for intelligently identifying individual communication radiation sources in a cross-period scenario, comprising the steps of:

[0010] Step 1. Acquire (receive) periodic transmission signals of the same individual communication radiation source at least three months apart, and obtain a first periodic transmission signal (a periodic transmission signal of the previous period) and a second periodic transmission signal (a periodic transmission signal of the next period) of each individual communication radiation source;

[0011] Step 2. Perform signal preprocessing on the first and second periodic transmission signals of each individual communication radiation source, including pulse detection, pulse segmentation, and normalization, to obtain standardized single pulse signal data. The standardized single pulse signal of the first periodic transmission signal is used as the first pulse data, and the standardized single pulse signal of the second periodic transmission signal is used as the second pulse data.

[0012] Setting a training data set and a test data set based on the first and second pulse data, wherein the first pulse data is labeled data and is provided with a radiation source label, and the second pulse data is unlabeled data;

[0013] Step 3. Set the model structure of the neural network, which includes a generator, a discriminator, and a classifier. The generator is used to learn the mapping relationship between the first pulse data and the second pulse data, and output the generated data of the first pulse data (also called false data); the discriminator is used to learn the similarity between the generated data output by the generator and the second pulse data, and the classifier is used to learn and output the classification results of the first pulse data and the generated data;

[0014] Step 4. Since the discriminator is used to implement a binary classification task, the binary cross entropy is selected as the discriminator's loss function; since the classifier is used for (radiant source) identification, the cross entropy is selected as the classifier's loss function; and the adaptive moment estimator Adam is selected as the optimizer. The network parameters of the generator, discriminator, and classifier are updated alternately to train the neural network.

[0015] Step 5. Select recognition accuracy as the performance metric of the classifier, test the classifier and output the recognition results.

[0016] Furthermore, in step 1, the transmission signal of the individual communication radiation source is received through the antenna, the received transmission signal is digitally filtered, amplified and mixed to obtain an intermediate frequency signal, and then the intermediate frequency signal is A / D (analog / digital) sampled, digitally filtered and digitally down-converted to obtain a zero intermediate frequency signal, and the first periodic transmission signal and the second periodic transmission signal of the individual communication radiation source are obtained based on the zero intermediate frequency signals of the previous and next two periods.

[0017] Furthermore, step 2 specifically includes first decoding the signal according to the communication protocol to extract the radiation source individual information in the frame header as a label;

[0018] Furthermore, among the multiple single pulse signals obtained after pulse detection and pulse segmentation of the first periodic transmission signal, only the long pulse signals therein are standardized to obtain the first pulse data; among the multiple single pulse signals obtained after pulse detection and pulse segmentation of the second periodic transmission signal, only the long pulse signals therein are standardized to obtain the second pulse data.

[0019] Furthermore, the normalization process is as follows: the mean and standard deviation of each pulse signal are calculated, the mean of the pulse signal is subtracted from the pulse signal, and then the pulse signal is divided by the standard deviation to obtain the normalized pulse signal, namely, the z-score normalization process.

[0020] Furthermore, in step 4, the neural network set for training is specifically:

[0021] (1) sampling noise from a uniform distribution and inputting it into a generator together with the first pulse data, and the generator outputs generated data (false data);

[0022] (2) The generated data and the second pulse data are fed into the discriminator successively, and the discriminator is trained to judge the generated data as false and the second pulse data as true;

[0023] (3) The first pulse data and the generated data are successively fed into the classifier, and the classifier is trained to classify and identify the radiation source of the current input data, so that the classifier has good classification performance on both the first pulse data and the generated data;

[0024] (4) Send the generated data to the discriminator and classifier in turn to train the discriminator and classifier;

[0025] (5) Repeat the above four steps until the preset training end condition (loss function convergence) is met. When the discriminator converges, the generator can generate data from the next period that is indistinguishable from the fake data. At this time, the classifier has good classification performance on both the fake data and the data from the previous period. In this way, the purpose of transferring the data features from the previous period to train a classifier with good performance on the data from the next period is achieved.

[0026] Furthermore, in step 3, the network structures of the generator, discriminator, and classifier are specifically as follows:

[0027] The generator includes, in sequence: a fully connected layer, a first convolution module, a residual network consisting of a plurality of residual blocks stacked, and a second convolution module, wherein the first convolution module includes, in sequence: a one-dimensional convolution layer, a batch normalization layer, and a PReLu activation function layer; the residual block includes, from front to back (in the direction of forward propagation), a one-dimensional convolution layer 1, a batch normalization layer 1, a PReLu activation function layer, a one-dimensional convolution layer 2, and a batch normalization layer 2, and there is a short connection between the input and output of the residual block (that is, the input feature map of the residual block is added to the feature map output by the batch normalization layer 2 to obtain the output feature map of the residual block); the second convolution module includes, in sequence: a one-dimensional convolution layer, a batch normalization layer, and a PReLu activation function layer;

[0028] The fully connected layer of the generator inputs noise data (N) sampled from the uniform distribution initialization Uniform (-1, 1) num ,N c ,N len ), where N num Indicates the number of samples, N c represents the noise channel, N len Represents the noise length. The output of the fully connected layer has the same dimension as the input, but the noise length changes. The output of the fully connected layer and the first pulse data are then concatenated (according to the channel dimension) and input into the first convolution module.

[0029] The discriminator includes: a convolutional network composed of several stacked third convolutional modules and a fully connected layer; the third convolutional module includes: a one-dimensional convolutional layer, a batch normalization layer and a LeakyReLU activation function layer; the fully connected layer outputs two-dimensional data, namely the number of samples and the probability of the sample being true or false;

[0030] The classifier includes, in sequence: a convolutional network consisting of several stacked fourth convolutional modules and several fully connected layers; wherein the fourth convolutional module includes, in sequence: a one-dimensional convolutional layer, a batch normalization layer, and a LeakyReLU activation function layer; for any non-last fully connected layer, a batch normalization layer and a ReLU activation function layer are set in sequence thereafter, and the last fully connected layer has a Softmax activation function, which is used to output the classification probability of each individual communication radiation source, that is, each individual communication radiation source is regarded as a classification category.

[0031] Furthermore, the convolution kernel size of the one-dimensional convolution layer of the first convolution module of the generator is 1×3, the convolution kernel size of the one-dimensional convolution layer 1 and the one-dimensional convolution layer 2 of the residual block is the same, both are 1×3, and the convolution kernel size of the one-dimensional convolution layer of the second convolution module is 1×1; the convolution kernel size of the one-dimensional convolution layer of the third convolution module of the discriminator is 1×3, and the third convolution module also includes a noise insertion layer or the third convolution module also includes a noise insertion layer (for enhancing the robustness of the network) and an anti-overfitting layer in sequence, for discarding some neurons with a specified probability (preferably 90%), that is, a random dropout layer (dropout) can also be set after the noise insertion; the convolution kernel size of the one-dimensional convolution layer of the fourth convolution module of the classifier is 1×5.

[0032] Furthermore, the generator includes 8 residual blocks; the discriminator includes 8 third convolution modules, and the number of convolution kernel channels included in the one-dimensional convolution layer of each third convolution module increases layer by layer; the classifier includes 3 fourth convolution modules, including 4 fully connected layers.

[0033] Furthermore, step 5 is specifically as follows: counting the recognition results of all test samples on the trained classifier, calculating the recognition accuracy, and visualizing all recognition results through the confusion matrix.

[0034] The technical solution provided by the present invention brings at least the following beneficial effects:

[0035] In this invention, although the characteristic distributions of the signal data from individual radiators in the two periods are inconsistent, they still have similarities. The generator, through joint training with the discriminator and classifier, adds noise to the data from the previous period to generate data sampled from the distribution of the data from the next period. At the same time, the classifier achieves excellent classification performance on both the generated data and the data from the previous period. Through this technical solution, the present invention can achieve stable and effective identification of individual radiators in cross-period scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A schematic diagram of the processing process of a method for intelligently identifying individual communication radiation sources in a cross-period scenario provided by an embodiment of the present invention;

[0037] Figure 2 is the time domain characteristic of the signal sample used in the embodiment of the present invention;

[0038] Figure 3 is a single pulse signal sample after normalization used in an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of the neural network model structure used in an embodiment of the present invention;

[0040] Figure 5 In the embodiment of the present invention, the classifier trained with data from the previous period is tested on data from the next period, and the confusion matrix and classification accuracy of the individual identification of five radiation sources are shown.

[0041] Figure 6 In the embodiment of the present invention, the classifier trained by the recognition method provided by the embodiment of the present invention is tested on the data of the latter period, and the confusion matrix and classification accuracy of the individual recognition of 5 radiation sources are shown. DETAILED DESCRIPTION

[0042] To help those skilled in the art better understand the objectives, technical solutions, and advantages of the embodiments of the present invention, the implementation of the technical solutions will be further described in detail below with reference to the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by those skilled in the art based on the embodiments of the present invention without creative effort are also within the scope of protection of the present invention.

[0043] like Figure 1 As shown, this embodiment provides a method for intelligently identifying individual communication radiation sources in a cross-period scenario, which includes the following steps:

[0044] Step 1: Receive 1090M ADS-B signals from five transmitters via antennas in two periods, five months apart. Digitally filter, amplify, and mix the received signals to obtain an intermediate frequency (IF) signal. A / D sample, digitally filter, and digitally down-convert the IF signal to obtain a zero IF signal. Extract the transmitter's individual information from the frame header and save each frame of data as a data file named according to the individual information. Signal data from the two periods is obtained. The sampling rate is 186.667MHz, the sampling bandwidth is 6M, and the acquisition environment and location remain the same.

[0045] Step 2: Preprocess the signal data collected before and after the acquisition.

[0046] (1) Use MATLAB to read the collected signal data sample file and extract the data value. Its time domain characteristics are as follows: Figure 2 As shown, first extract the I / Q signals, combine them into a complex signal, and calculate the signal envelope. Observe the envelope time domain characteristics to determine the relevant parameters. Take the noise threshold NoiseAp = 200, the rising edge amplitude threshold PeakAp = 100, and the smoothing window length L win =3, cutting accuracy C precision =1 (a cutting signal contains one pulse), long and short pulse length threshold L threshold ;

[0047] (2) According to the above parameters, the peaks and valleys are detected by sliding, and the long pulse signal samples with a dimension of 1×456 and the short pulse signal samples with a dimension of 1×256 are cut out in the two periods before and after. Figure 3 As shown; that is, 456 and 256 are the length thresholds of long and end pulses respectively;

[0048] (3) In order to eliminate the influence of the differences in the dimensions and value ranges between the indicators, the obtained single pulse signal data is standardized and the data is scaled to a specific area according to a certain ratio. Specifically, the mean value of the single pulse signal x is first calculated. With the standard deviation σ, then by the formula: Get standardized pulse data with a mean of 0 and a standard deviation of 1;

[0049] Step 3: Organize and create the data set.

[0050] The number of long and short pulse signals of each individual radiation source obtained by cutting is counted, and the long pulses are selected as experimental data, that is, the long pulses are used as pulse samples; among them, the pulse samples of the previous period are labeled, and the pulse samples of the next period are unlabeled. The label can be based on the communication protocol decoding of the signal and extracting the individual radiation source information in the frame header of the data frame as the radiation source label. Then, the training set and test set are selected according to the specified ratio. In this embodiment, the pulse samples of the previous period (with labels) and 80% of the pulse samples of the next period (without labels) are used as the training set, and the remaining 20% of the pulse samples of the next period (without labels) are used as the test set.

[0051] Step 4: Build and train the neural network model, such as Figure 4 As shown, the neural network model consists of three parts: generator, discriminator and classifier, which are specifically:

[0052] (1) The first part is the generator, which consists of 11 layers:

[0053] The first layer is the fully connected layer (Linear);

[0054] The second layer is a CNN module. The CNN module stacks a one-dimensional convolution layer with 64 weight matrices of size (3, 1). The activation function of the convolution layer is PReLu, and batch normalization is performed before activation.

[0055] The 3rd to 10th layers are residual blocks, which are: 1 one-dimensional convolution layer with 64 weight matrices, weight matrix size (3, 1), batch normalization layer, PReLu activation function layer, one-dimensional convolution layer, batch normalization layer; a short connection (also called jump connection) is set between the input of each residual block (the input of the first one-dimensional convolution layer) and the output of the second batch normalization layer;

[0056] The 11th layer is a CNN module. The CNN module stacks a one-dimensional convolutional layer, contains a weight matrix, the weight matrix size is (1,1), the activation function is PReLu, and batch normalization is performed before activation.

[0057] Among them, the first layer of the fully connected layer inputs the noise (N) sampled from Uniform (-1,1) num ,N c ,N len ); including three dimensions, the number of samples, the noise channel and the noise length. In this embodiment, the sampled noise is (N batch_size ,1,10), where N batch_sizeRepresents the batch size, that is, the number of samples; the output has the same dimension as the input, but the noise length changes; the output of the first fully connected layer and the spike data of the previous period are concatenated to obtain a dimension of (N batch_size ,2,456) matrix is input to the CNN module. The input contains three dimensions, namely the number of samples, the number of sample channels, and the sample length. The output is the same as the input dimension. The weight matrix of all convolutional layers contains 4 dimensions, namely the input channel, the output channel, the convolution kernel length, and the convolution kernel width.

[0058] (2) The second part is the discriminator, which consists of 9 layers:

[0059] The first layer is a CNN module, which stacks a one-dimensional convolutional layer with 64 weight matrices of size (3, 1). The activation function of the convolutional layer is LeakyReLU, and batch normalization is performed before activation.

[0060] The second layer is a CNN module, which stacks a one-dimensional convolutional layer with 128 weight matrices of size (3, 1). The activation function is LeakyReLU, and batch normalization is performed before activation. To enhance the robustness of the network, a noise insertion layer is added after activation. To prevent overfitting, some neurons are discarded with a probability of 90%.

[0061] The third layer is a CNN module, which stacks a one-dimensional convolutional layer with 256 weight matrices of size (3,1). The activation function is LeakyReLU, and batch normalization is performed before activation. To enhance the robustness of the network, a noise insertion layer is added after activation. To prevent overfitting, some neurons are discarded with a probability of 90%.

[0062] The fourth layer is a CNN module, which stacks a one-dimensional convolutional layer with 512 weight matrices of size (3, 1). The activation function is LeakyReLU, and batch normalization is performed before activation. To enhance the robustness of the network, a noise insertion layer is added after activation. To prevent overfitting, some neurons are discarded with a probability of 90%.

[0063] The fifth layer is a CNN module, which stacks a one-dimensional convolutional layer with 1024 weight matrices of size (3, 1). The activation function is LeakyReLU, and batch normalization is performed before activation. To enhance the robustness of the network, a noise insertion layer is added after activation. To prevent overfitting, some neurons are discarded with a probability of 90%.

[0064] The sixth layer is a CNN module, which stacks a one-dimensional convolutional layer with 2048 weight matrices of size (3, 1). The activation function is LeakyReLU, and batch normalization is performed before activation. To enhance the robustness of the network, a noise insertion layer is added after activation. To prevent overfitting, some neurons are discarded with a probability of 90%.

[0065] The 7th layer is a CNN module. The CNN module stacks a one-dimensional convolutional layer with 4096 weight matrices of size (3,1). The activation function is LeakyReLU, and batch normalization is performed before activation. To enhance the robustness of the network, a noise insertion layer is added after activation. To prevent overfitting, some neurons are discarded with a probability of 90%.

[0066] The eighth layer is a CNN module, which stacks a one-dimensional convolutional layer with 8092 weight matrices of size (3, 1). The activation function is LeakyReLU, and batch normalization is performed before activation. To enhance the robustness of the network, a noise insertion layer is added after activation. To prevent overfitting, some neurons are discarded with a probability of 90%.

[0067] The 9th layer is a fully connected layer.

[0068] The discriminator inputs the data output by the generator or the individual signal data of the next period (the pulse samples of the next period), whose dimensions are the number of samples, the number of sample channels, and the sample length. The fully connected layer finally outputs data of two dimensions: the number of samples and the probability of the sample being true or false. All CNN module inputs contain three dimensions: the number of samples, the number of sample channels, and the sample length. The output has the same number of dimensions as the input, but the sample length varies. The weight matrices of all convolutional layers contain four dimensions: the input channels, the output channels, the convolution kernel length, and the convolution kernel width.

[0069] (3) The third part is the classifier, which consists of 7 layers:

[0070] The first layer is a CNN module, which stacks a one-dimensional convolutional layer with 6 weight matrices of size (5, 1). After batch normalization, the activation function is PReLU.

[0071] The second layer is a CNN module, which stacks a one-dimensional convolutional layer with 16 weight matrices of size (5, 1). After batch normalization, the activation function is PReLU.

[0072] The third layer is a CNN module, which stacks a one-dimensional convolutional layer with 32 weight matrices of size (5, 1). After batch normalization, the activation function is PReLU.

[0073] The 4th, 5th, 6th, and 7th layers are fully connected layers, and the activation function of the last fully connected layer is Softmax.

[0074] Among them, the classifier input is the data output by the generator or the individual signal data of the previous period as described in step 1. Its dimensions are the number of samples, the number of sample channels, and the sample length. The last layer of full connection outputs two dimensions of data, namely the number of samples and the probability of the sample corresponding to each category; all CNN module inputs contain three dimensions, namely the number of samples, the number of sample channels, and the sample length. The output is the same as the input dimension, and the sample length changes; the weight matrix of all convolutional layers contains 4 dimensions, namely the input channel, the output channel, the convolution kernel length, and the convolution kernel width;

[0075] During training, the optimization objective function of the network model can be set as:

[0076]

[0077] The overall loss is composed of the adversarial loss l between the generator and the discriminator d (D,G), cross entropy loss l of the classifier c (G, C) consists of two parts, α represents the adversarial loss weight, β represents the cross entropy loss weight, D represents the discriminator, and the internal parameters are determined by θ D Characterization; G represents the generator, and the internal parameters are represented by θ G Characterization, C represents the classifier, and the internal parameters are represented by θ C representation;

[0078] in:

[0079]

[0080]

[0081] Among them, x s Represents the data of the previous period (pulse sample), y s Indicates the data label of the previous period, x t represents the data of the next period (pulse sample), z represents noise, E represents output expectation, p data () represents the data distribution, that is, x t ~p data (x t ) represents the data x in the next period t The data distribution of x s ~p data (x s ) represents the data x of the previous period s Data distribution, z~p Z(z) represents the prior Gaussian distribution of noise z. D() represents the output of the discriminator, G() represents the output of the generator, that is, the generated data, and C() represents the output of the classifier.

[0082] In this embodiment, the specific training process of the generator, discriminator, and classifier is as follows:

[0083] (1) First train the generator D:

[0084] (1-1) x t (The data of the next period) is input into the discriminator D, and the loss of the current output of the discriminator D is calculated based on the binary cross entropy loss function, which is recorded as

[0085] (1-2) x s (previous period data) and noise are fed into the generator to obtain the generated data x f (false data), x f Input the discriminator D, and calculate the loss of the current output of the discriminator D based on the binary cross entropy, which is recorded as

[0086] (1-3) Minimize The network parameters of the generator D are updated through back propagation. It should be noted that when training the generator D, the discriminator G participates in the calculation, but the network parameters of G are not updated.

[0087] (2) Retrain classifier C:

[0088] (2-1) x s Send it to classifier C, and calculate the classification loss of the current output of classifier C based on the cross entropy loss function, which is recorded as

[0089] (2-2) When training generator D, the data x output from generator D f Send it to classifier C, and calculate the classification loss of the current output of classifier C based on the cross entropy loss function, which is recorded as

[0090] (2-3) Minimize The network parameters of the classifier C are updated through back propagation. It should be noted that when training the classifier C, although the generator G participates in the calculation, the network parameters of G are not updated.

[0091] (3) Finally, train the generator G:

[0092] (3-1) x s and noise are fed into the generator G to obtain the generated data x f , and x fSend it to the discriminator D, and calculate the loss of the current output of the discriminator D based on the two-category cross entropy loss function, which is recorded as

[0093] (3-2) Then x f Send it to classifier C, and calculate the classification loss of the current output of classifier C based on the cross entropy loss function, which is recorded as

[0094] (3-3) Minimize Update the network parameters of the generator G through backpropagation.

[0095] Repeat the above training process ((1) to (3)) until the preset training convergence conditions are met, such as the number of training times reaches the preset maximum value, or the loss convergence.

[0096] Step 5: Obtain recognition results based on the classifier of the trained neural network model.

[0097] When the training is completed, the data of the next period is identified and processed based on the trained classifier. That is, in the present invention, the neural network model is trained based on the historical signal data and the current data to be tested (without labels), and then the individual communication radiation sources are intelligently identified and processed based on the trained classifier.

[0098] As a possible implementation, in this embodiment, the specific settings for training the neural network model are:

[0099] Optimizer: Adam, L2 weight decay: 1e-5, betas: (0.5, 0.999), generator and discriminator learning rates: 2e-4; discriminator learning rate: 1e-3; weight α is set to 0.4, weight β is set to 0.6; spike samples from the previous epoch (with labels) and 80% of spike samples from the next epoch (without labels) are used as the training set, and the remaining 20% of spike samples from the next epoch (without labels) are used as the test set. The training and test sets are randomly shuffled, with 256 samples per iteration, and training is carried out for 50 epochs. After the network converges, the classifier trained only on the data from the previous epoch is tested on the data from the next epoch. The recognition results are shown in the attached figure. Figure 5 The classifier trained using the above steps is tested on the data from the next period. The recognition results are shown in the attached figure. Figure 6 .

[0100] This embodiment comprehensively considers the characteristics of individual radiation sources that change over time, and can use noise to establish a mapping relationship between data from the previous period and data with different distributions in the next period, thereby achieving stable and effective individual radiation source identification in cross-period scenarios.

[0101] Note that the above preferred embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the present invention. Those skilled in the art should understand that any modifications, equivalent replacements, and improvements made within the scope of the present invention are intended to fall within the scope of protection of the present invention.

Claims

1. A method for intelligently identifying individual communication radiation sources in a cross-period scenario, the method comprising the following steps: Step 1: Obtain periodic transmission signals of the same individual communication radiation source at least three months apart, and obtain a first periodic transmission signal and a second periodic transmission signal of each individual communication radiation source; Step 2: performing signal preprocessing on the first and second periodic transmission signals of each individual communication radiation source, including pulse detection, pulse segmentation, and normalization, to obtain a standardized single pulse signal, using the standardized single signal of the first periodic transmission signal as the first pulse data, and using the standardized single pulse signal of the second periodic transmission signal as the second pulse data; Setting a training data set and a test data set based on the first and second pulse data, wherein the first pulse data is set with a radiation source label; Step 3, setting a model structure of a neural network, wherein the neural network includes a generator, a discriminator, and a classifier, wherein the generator is used to learn a mapping relationship between the first pulse data and the second pulse data, and output the generated data of the first pulse data; the discriminator is used to learn the similarity between the generated data output by the generator and the second pulse data, and the classifier is used to learn and output the classification results of the first pulse data and the generated data; Step 4: Use binary cross entropy as the loss function of the discriminator, cross entropy as the loss function of the classifier, and adaptive moment estimation Adam as the optimizer to alternately update the network parameters of the generator, discriminator, and classifier to train the set neural network; Step 5: Select recognition accuracy as the performance metric of the classifier, test the classifier and output the recognition results.

2. The method according to claim 1, characterized in that In step 1, the transmission signal of the individual communication radiation source is received through the antenna, the received transmission signal is digitally filtered, amplified and mixed to obtain an intermediate frequency signal, and then the intermediate frequency signal is A / D sampled, digitally filtered and digitally down-converted to obtain a zero intermediate frequency signal. Based on the zero intermediate frequency signals of the previous and next two periods, the first periodic transmission signal and the second periodic transmission signal of the individual communication radiation source are obtained.

3. The method according to claim 1, characterized in that In step 2, the radiation source label of the first pulse data is: the radiation source individual information in the frame header of the data frame is extracted according to the communication protocol decoding as the radiation source label.

4. The method according to claim 1, wherein Pulse detection and pulse segmentation are specifically as follows: obtaining the signal envelope and determining: noise threshold, rising edge amplitude threshold, smoothing window length, cutting accuracy, long and short pulse length threshold; recording the start and end positions of each pulse, and cutting based on the start and end positions of the pulse and the pulse length threshold to obtain a single pulse signal.

5. The method according to claim 1, wherein In step 2, after pulse detection and pulse segmentation of the first periodically transmitted signal, multiple single pulse signals are obtained, and only the long pulse signal therein is normalized to obtain first pulse data; Among the multiple single pulse signals obtained after pulse detection and pulse segmentation of the second periodic transmission signal, only the long pulse signal is standardized to obtain second pulse data.

6. The method according to claim 1, characterized in that In step 2, the standardization process is as follows: calculating the mean and standard deviation of each pulse signal, subtracting the mean from the pulse signal and then dividing it by the standard deviation to obtain a standardized pulse signal.

7. The method according to claim 1, wherein In step 4, the neural network set for training is specifically: (1) sampling noise from a uniform distribution and inputting the noise together with the first pulse data into a generator, and the generator outputs generated data; (2) The generated data and the second pulse data are fed into the discriminator successively, and the discriminator is trained to judge the generated data as false and the second pulse data as true; (3) The first pulse data and the generated data are successively fed into the classifier, and the classifier is trained to classify and identify the radiation source of the current input data, so that the classifier has good classification performance on both the first pulse data and the generated data; (4) Send the generated data to the discriminator and classifier in turn to train the discriminator and classifier; (5) Repeat the above four steps until the preset training end conditions are met.

8. The method according to claim 1, characterized in that In step 3, the network structures of the generator, discriminator and classifier are specifically as follows: The generator includes, in sequence: a fully connected layer, a first convolution module, a residual network consisting of a plurality of residual blocks stacked, and a second convolution module, wherein the first convolution module includes, in sequence: a one-dimensional convolution layer, a batch normalization layer, and a PReLu activation function layer; the residual block includes, from front to back: a one-dimensional convolution layer 1, a batch normalization layer 1, a PReLu activation function layer, a one-dimensional convolution layer 2, and a batch normalization layer 2, and there is a short connection between the input and output of the residual block; the second convolution module includes, in sequence: a one-dimensional convolution layer, a batch normalization layer, and a PReLu activation function layer; The fully connected layer of the generator inputs noise data (N) sampled from a uniform distribution initialized from (-1,1) num ,N c ,N len ), where N num Indicates the number of samples, N c represents the noise channel, N len Represents the noise length. The output of the fully connected layer has the same dimension as the input, but the noise length changes. The output of the fully connected layer and the first pulse data are then concatenated and input into the first convolution module. The discriminator includes: a convolutional network composed of several stacked third convolutional modules and a fully connected layer; the third convolutional module includes: a one-dimensional convolutional layer, a batch normalization layer and a LeakyReLU activation function layer; the fully connected layer outputs two-dimensional data, namely the number of samples and the probability of the sample being true or false; The classifier includes, in sequence: a convolutional network consisting of several stacked fourth convolutional modules and several fully connected layers; wherein the fourth convolutional module includes, in sequence: a one-dimensional convolutional layer, a batch normalization layer, and a LeakyReLU activation function layer; for any non-last fully connected layer, a batch normalization layer and a ReLU activation function layer are set in sequence thereafter, and the last fully connected layer has a Softmax activation function.

9. The method according to claim 3, characterized in that The convolution kernel size of the one-dimensional convolution layer of the first convolution module of the generator is 1×3, the convolution kernel size of the one-dimensional convolution layer 1 and the one-dimensional convolution layer 2 of the residual block are the same, both 1×3, and the convolution kernel size of the one-dimensional convolution layer of the second convolution module is 1×1; The convolution kernel size of the one-dimensional convolution layer of the third convolution module of the discriminator is 1×3, and the third convolution module further includes a noise insertion layer or the third convolution module further includes a noise insertion layer and a random dropout layer in sequence; The convolution kernel size of the one-dimensional convolution layer of the fourth convolution module of the classifier is 1×5.

10. The method according to claim 9, wherein The generator includes 8 residual blocks; the discriminator includes 8 third convolution modules, and the number of convolution kernel channels included in the one-dimensional convolution layer included in each third convolution module increases layer by layer; the classifier includes 3 fourth convolution modules, including 4 fully connected layers.

Images