Radar radiation source individual identification method based on collaborative learning framework

Abstract

The invention discloses a radar radiation source individual identification method based on a collaborative learning framework, which is used for realizing individual separation and identification of complex radar radiation sources and solves the problem of low identification accuracy caused by small sample quantity and low quality in the prior art. The method comprises the following steps: carrying out preprocessing and bispectrum signal extraction on original intermediate-frequency monopulse data in batches, strengthening the characteristic characterization capability, and generating a sample data set; dividing the sample data set into a training data set, a verification data set and a test data set according to a certain proportion; biLSTM and ResNet18 network models are selected to construct a collaborative learning framework for collaborative training, and a cross entropy loss function is used as a loss function of the model; and inputting the test data set into the neural network model obtained by cooperative training for reasoning, and finally completing classification identification through softmax, and calculating the identification accuracy. The method has a good application prospect in the field of complex radar radiation source individual identification.

Description

Technical Field

[0001] The present invention relates to the field of radar radiation source identification, and in particular to a radar radiation source individual identification method based on a collaborative learning framework. Background Art

[0002] As one of the key technologies in the field of electronic countermeasures, radar emitter individual identification technology can distinguish different individuals of the same radar target and then identify the platform carrying the radar. It is one of the core links in the process of radar emitter intelligence analysis and has extremely important application prospects. With the rapid development of modern science and technology and the popularization of new radar systems such as active phased arrays, the classic method of radar emitter sorting and identification based on traditional pulse modulation parameters has become difficult to work. Therefore, individual identification technology that can distinguish the emitter by using the unintentional features within a single pulse has become increasingly important. However, the extraction of unintentional features of radar emitters has always been one of the difficulties in individual identification.

[0003] The first task of individual recognition is to find the intrinsic representation that can express the individual's uniqueness. The quality of the feature directly determines the recognition effect. However, the subtle features of radar radiation sources are complex in generation mechanism and lack the support of mathematical models. It becomes extremely difficult to quantify such unrobust features.

[0004] In recent years, with the successful application of deep learning methods in many fields, its powerful sample analysis and feature extraction capabilities have attracted widespread attention from scholars in the field of electronic countermeasures. The emergence of deep learning has made it possible to describe and quantify the subtle features of radar radiation sources in high dimensions. Deep learning uses data-driven thinking and does not require manual feature extraction. It can deeply mine high-dimensional features from low-level physical features by training network models, thereby achieving separation and identification of complex radar signals. Summary of the invention

[0005] The purpose of the present invention is to provide a method for identifying individual radar emitters based on a collaborative learning framework, so as to solve the problem that it is difficult to extract characteristic parameters of traditional complex radar signals and cannot effectively perform individual identification.

[0006] In order to achieve the above-mentioned purpose, the present invention provides a radar emitter individual identification method based on a collaborative learning framework, comprising: 1) preprocessing the original radar emitter signal, eliminating the original radar emitter signal with a signal-to-noise ratio lower than 9dB, and performing sample annotation on the preprocessed original radar emitter signal of different targets; 2) performing bispectral analysis on the preprocessed original radar emitter signal to generate a radar emitter individual sample bispectral signal data set; 3) randomly selecting 60% of the data from the radar emitter individual sample bispectral signal data set as a training data set L, selecting 20% ​​of the data as a verification data set U, and the remaining 20% ​​of the data as a test data set; 4) constructing a collaborative learning framework, generating a BiLSTM-ResNet18 network model, and setting the initial training hyperparameters of the BiLSTM-ResNet18 network model; 5) taking the training data set L and the verification data set U as input, using the cross entropy loss function as the BiLSTM-ResNet18 The loss function of the network model is used, and the root mean square transfer algorithm is used as the optimization algorithm of the BiLSTM-ResNet18 network model to train the BiLSTM-ResNet18 network model; 6) When the training accuracy reaches the set threshold and no longer increases, the BiLSTM-ResNet18 network model training is completed, and the optimized BiLSTM-ResNet18 network model is obtained; 7) Taking the test data set as input, the optimized BiLSTM-ResNet18 network model is used to extract the individual subtle features of the radar radiation source signal, and the softmax function is used as the classification function to output the recognition result, and the recognition result is compared and counted with the target sample annotation to calculate the recognition accuracy; 8) After preprocessing and bispectral analysis, the radar radiation source signal to be identified is input into the optimized BiLSTM-ResNet18 network model, and the network model is used to extract the individual subtle features of the radar radiation source signal, and the recognition result and its confidence are output.

[0007] The above-mentioned radar emitter individual identification method based on the collaborative learning framework, wherein the collaborative learning framework selects the residual neural network ResNet18 and the bidirectional long short-term memory network BiLSTM; the input feature map of ResNet18 is a single channel, the output feature map channel is 64, the convolution kernel size is 7×7, the stride is 2, the padding is 3, and the activation function is the ReLu function; the input of BiLSTM is a single channel, the output channel is 224, and the convolution kernel sizes are 2×512, 3×512, and 4×512 respectively.

[0008] The above-mentioned radar emitter individual identification method based on the collaborative learning framework, wherein the step 5) includes: 5-1) dividing the training data set L into two data sets, namely, the training data set L1 and the training data set L2; 5-2) processing the training data set L1 into a matrix of size 224×224, and processing the training data set L2 into a sequence of size 1×50176; 5-3) using the processed training data set L1 as the input of the ResNet18 network model, and using the processed training data set L2 as the input of the BiLSTM network model, using the cross entropy loss function to calculate the difference between the predicted value and the true value of the forward propagation of the network model, if the loss value is too large, then using the root mean square transfer algorithm to update the network model weights, and generating Re after one training round. sNet18 network model and BiLSTM network model; 5-4) Use the trained ResNet18 network model and BiLSTM network model to classify the verification data set U respectively; 5-5) Extract the k data with the highest and lowest classification confidence of the ResNet18 network model for the verification data set U and merge them into the training data set L2; 5-6) Extract the k data with the highest and lowest classification confidence of the BiLSTM network model for the verification data set U and merge them into the training data set L1; 5-7) Remove 2k data in the verification data set U; 5-8) Repeat steps 5-3)-step 5-7) until the verification data set U is empty (that is, until all the data in the verification data set U are removed), at which time the network model training reaches the optimal level.

[0009] Compared with the prior art, the beneficial technical effects of the present invention are:

[0010] The emergence of deep learning has made it possible to describe and quantify the subtle features of radar radiation sources in high dimensions. It uses data-driven thinking to break the barriers of artificial feature extraction and deeply mines high-dimensional features from low-level physical features by training network models, thereby achieving separation and identification of complex radar signals. This method has good robustness and can adapt to data under different signal-to-noise ratio conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The radar emitter individual identification method based on the collaborative learning framework of the present invention is given by the following embodiments and drawings.

[0012] Figure 1 This is a flow chart of a method for identifying individual radar emitters based on a collaborative learning framework according to an embodiment of the present invention.

[0013] Figure 2 Schematic diagram of the time domain signal of the radar radiation source in an embodiment of the present invention.

[0014] Figure 3 Schematic diagram of the dual-spectrum signal of individual samples of radar radiation sources in an embodiment of the present invention.

[0015] Figure 4 This is a training flow chart under the collaborative learning framework in an embodiment of the present invention.

[0016] Figure 5 This is a curve diagram showing how training accuracy changes with the number of training rounds in an embodiment of the present invention.

[0017] Figure 6 This is a curve diagram of the change of training loss with the number of training rounds in an embodiment of the present invention.

[0018] Figure 7 This is a confusion matrix diagram for individual identification of five radars of the same type in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The following will be combined Figure 1 to Figure 7 The radar emitter individual identification method based on the collaborative learning framework of the present invention is further described in detail.

[0020] The idea of ​​the invention is to extract individual subtle features of radar signals by constructing a collaborative learning framework, thereby realizing individual separation and recognition of radar radiation source signals. The collaborative learning framework of the invention can realize the sharing of some features among multiple learning methods, and can achieve good recognition effect even when there are few radar signal samples and the quality is not high.

[0021] Figure 1 Shown is a flow chart of a radar emitter individual identification method based on a collaborative learning framework according to an embodiment of the present invention.

[0022] Combination Figures 1 to 7 The radar emitter individual identification method based on the collaborative learning framework of this embodiment includes:

[0023] 1) Preprocess the original radar radiation source signals, remove the original radar radiation source signals with a signal-to-noise ratio lower than 9dB, and mark the samples of the original radar radiation source signals of different targets after preprocessing;

[0024] The original radar radiation source signal is intermediate frequency single pulse data;

[0025] 2) Perform bispectral analysis on the preprocessed original radar emitter signal to generate a bispectral signal dataset of individual radar emitter samples;

[0026] Bispectral analysis is widely used in the field of signal recognition and can well reflect the unintentional modulation phenomenon of radar radiation source signals. The bispectral signal generated by bispectral analysis is a high-order cumulant that retains the amplitude and phase information of the original signal and can suppress Gaussian noise to a certain extent.

[0027] 2-1) The original intermediate frequency single pulse discrete time sequence sig(n) (n=0,1,2,…,N-1) with a sampling length of N is normalized, and its mathematical expression is:

[0028]

[0029] Among them, sig is a value in the discrete time series; sig max is the maximum value of all values ​​in the discrete time series; sig min is the minimum value of all values ​​in the discrete time series; sig normalization is the normalized value;

[0030] 2-2) The normalized sig(n) is processed into K segments, each with M sampling points, and the discrete Fourier transform coefficients of each segment are calculated respectively. The calculation formula is:

[0031]

[0032] Among them, λ=0,1,L,M / 2; k=1,L,K; y (k) (m) is the kth segment of data;

[0033] 2-3) Calculation step 2-2) Discrete Fourier transform coefficient Y (k) The triple correlation of (λ):

[0034]

[0035] k=1,L,K;0≤λ2≤λ1,λ1+λ2≤f s / 2

[0036] Among them, f s is the signal frequency; △0 = f s / N0, and N0 and L1 should be selected to satisfy M = (2L1 + 1)N0;

[0037] 2-4) Yes Taking the average, we can get the bispectral estimate of K segments:

[0038]

[0039] Among them, ω1=(2πf s / N0)λ1;ω2=(2πf s / N0)λ2; The set is the data set of all samples of radar radiation source individuals after bispectral analysis;

[0040] 3) Randomly select 60% of the data from the radar emitter individual sample bispectral signal data set as the training data set L, select 20% of the data as the verification data set U, and the remaining 20% ​​of the data as the test data set;

[0041] 4) Build a collaborative learning framework, generate a BiLSTM-ResNet18 network model, and set the initial training hyperparameters of the BiLSTM-ResNet18 network model;

[0042] In this embodiment, the collaborative learning framework selects the residual neural network (ResNet18) and the bidirectional long short-term memory network (BiLSTM); the input feature map of ResNet18 is a single channel, the output feature map channel is 64, the convolution kernel size is 7×7, the stride is 2, the padding is 3, and the activation function is the ReLu function; the input of BiLSTM is a single channel, the output channel is 224, and the convolution kernel sizes are 2×512, 3×512, and 4×512 respectively;

[0043] The expression of the ReLu function is:

[0044] relu(x input )=max(0,x input )

[0045] Among them, x input is the input, relu(x input ) is the output of the ReLu function;

[0046] 5) If Figure 4 , taking the training data set L and the validation data set U as input, using the cross entropy loss function as the loss function of the BiLSTM-ResNet18 network model, using the root mean square transfer algorithm as the optimization algorithm of the BiLSTM-ResNet18 network model, and training the BiLSTM-ResNet18 network model;

[0047] 5-1) Divide the training data set L into two data sets, namely, training data set L1 and training data set L2;

[0048] 5-2) Process the training data set L1 into a matrix of size 224×224, and the training data set L2 into a sequence of size 1×50176;

[0049] 5-3) Use the processed training data set L1 as the input of the ResNet18 network model, and use the processed training data set L2 as the input of the BiLSTM network model. Use the cross entropy loss function to calculate the difference between the predicted value and the true value of the forward propagation of the network model. If the loss value is too large, use the root mean square transfer algorithm to update the network model weights. After one training round, generate the ResNet18 network model and the BiLSTM network model;

[0050] 5-4) Use the ResNet18 network model and BiLSTM network model generated by training to classify the verification data set U respectively;

[0051] 5-5) Extract the k data with the highest and lowest classification confidence of the ResNet18 network model for the verification data set U and merge them into the training data set L2;

[0052] 5-6) Extract the k data with the highest and lowest classification confidence of the BiLSTM network model for the verification data set U and merge them into the training data set L1;

[0053] 5-7) Remove 2k data from the verification data set U;

[0054] 5-8) Repeat steps 5-3) to 5-7) until the verification data set U is empty (that is, until all the data in the verification data set U is removed), at which point the network model training reaches the optimal level;

[0055] In order to avoid overfitting, the BiLSTM-ResNet18 network model measures the degree of inconsistency between the predicted value and the true value of the network model. The cross entropy loss function used is expressed as follows:

[0056]

[0057] Among them, y(xi) represents the true probability distribution, Represents the predicted probability distribution. Usually, the output layer of a multi-classification task will be activated by a softmax function, and then the cross entropy is used to characterize the distance between the two probability distributions.

[0058] The RMS transfer algorithm is as follows:

[0059] Take m samples {x1,x2L,x m}, where x i The corresponding true value is y i ;

[0060] Compute the gradient: Among them, L represents the loss function of each sample, f(x i ; θ) represents the input x iThe predicted value of the neural network when

[0061] Cumulative squared gradient:

[0062] r←ρr+(1-ρ)g☉9

[0063] Update parameters:

[0064]

[0065] Among them, ε is the learning rate, which is set to 0.01; θ is the initial parameter, with an initial value of 0.01; r is the initialization accumulator, with an initial value of 0; ρ is the decay rate, which is set to 0.9; δ is the smoothing constant to prevent the divisor from being 0, which is set to 10- 6 ;

[0066] During multiple training sessions, feedback on the training results is provided through the validation dataset, and training hyperparameters are adjusted;

[0067] 6) When the training accuracy reaches the set threshold and no longer increases, the BiLSTM-ResNet18 network model training is completed, and the optimized BiLSTM-ResNet18 network model is obtained;

[0068] 7) Taking the test data set as input, the optimized BiLSTM-ResNet18 network model is used to extract the individual subtle features of the radar radiation source signal, and the softmax function is used as the classification function to output the recognition result. The recognition result is compared and counted with the target sample annotation, and the recognition accuracy is calculated;

[0069] 8) After preprocessing and bispectral analysis, the radar radiation source signal to be identified is input into the optimized BiLSTM-ResNet18 network model, which is used to extract the individual subtle features of the radar radiation source signal and output the identification result and its confidence.

Claims

1. A radar emitter individual identification method based on a collaborative learning framework, characterized in that: include: 1) Preprocess the original radar radiation source signals, remove the original radar radiation source signals with a signal-to-noise ratio lower than 9dB, and mark the samples of the original radar radiation source signals of different targets after preprocessing; 2) Perform bispectral analysis on the preprocessed original radar radiation source signal to generate a bispectral signal data set of individual radar radiation source samples; 3) Randomly select 60% of the data from the radar emitter individual sample bispectral signal data set as the training data set L, select 20% of the data as the verification data set U, and the remaining 20% ​​of the data as the test data set; 4) Build a collaborative learning framework, generate a BiLSTM-ResNet18 network model, and set the initial training hyperparameters of the BiLSTM-ResNet18 network model; 5) Taking the training data set L and the validation data set U as input, using the cross entropy loss function as the loss function of the BiLSTM-ResNet18 network model, using the root mean square transfer algorithm as the optimization algorithm of the BiLSTM-ResNet18 network model, the BiLSTM-ResNet18 network model is trained; 6) When the training accuracy reaches the set threshold and no longer increases, the BiLSTM-ResNet18 network model training is completed, and the optimized BiLSTM-ResNet18 network model is obtained; 7) Taking the test data set as input, the optimized BiLSTM-ResNet18 network model is used to extract the individual subtle features of the radar radiation source signal, and the softmax function is used as the classification function to output the recognition result. The recognition result is compared and counted with the target sample annotation, and the recognition accuracy is calculated; 8) After preprocessing and bispectral analysis, the radar radiation source signal to be identified is input into the optimized BiLSTM-ResNet18 network model, which is used to extract the individual subtle features of the radar radiation source signal and output the identification result and its confidence.

2. The radar emitter individual identification method based on the collaborative learning framework as claimed in claim 1, characterized in that: The step 2) comprises: 2-1) The original intermediate frequency single pulse discrete time series sig(n) with a sampling length of N, n = 0, 1, 2, ..., N-1, is normalized, and its mathematical expression is: Among them, sig is a value in the discrete time series; sig max is the maximum value of all values ​​in the discrete time series; sig min is the minimum value of all values ​​in the discrete time series; sig normalization is the normalized value; 2-2) The normalized sig(n) is processed into K segments, each with M sampling points, and the discrete Fourier transform of each segment is calculated separately, and the coefficients of the calculation formula are: Among them, λ=0,1,L,M / 2; k=1,L,K; y (k) (m) is the kth segment of data; 2-3) Calculation step 2-2) Discrete Fourier transform coefficient Y (k) The triple correlation of (λ): Among them, f s is the signal frequency; △0 = f s / N0, and N0 and L1 should be selected to satisfy M = (2L1 + 1)N0; 2-4) Yes Taking the average, we can get the bispectral estimate of K segments: Among them, ω1=(2πf s / N0)λ1;ω2=(2πf s / N0)λ2; The collection is the data set of all samples of radar radiation source individuals after bispectral analysis.

3. The radar emitter individual identification method based on collaborative learning framework as claimed in claim 1, characterized in that: The collaborative learning framework selects the residual neural network ResNet18 and the bidirectional long short-term memory network BiLSTM; the input feature map of ResNet18 is single-channel, the output feature map channel is 64, the convolution kernel size is 7×7, the stride is 2, the padding is 3, and the activation function is the ReLu function; the input of BiLSTM is single-channel, the output channel is 224, and the convolution kernel sizes are 2×512, 3×512, and 4×512 respectively; The expression of the ReLu function is: relu(x input )=max(0,x input ) Among them, x input is the input, relu(x input ) is the output of the ReLu function.

4. The radar emitter individual identification method based on collaborative learning framework as claimed in claim 1, characterized in that: The step 5) comprises: 5-1) Divide the training data set L into two data sets, namely, training data set L1 and training data set L2; 5-2) Process the training data set L1 into a matrix of size 224×224, and the training data set L2 into a sequence of size 1×50176; 5-3) Use the processed training data set L1 as the input of the ResNet18 network model, and use the processed training data set L2 as the input of the BiLSTM network model. Use the cross entropy loss function to calculate the difference between the predicted value and the true value of the forward propagation of the network model. If the loss value is too large, use the root mean square transfer algorithm to update the network model weights. After one training round, generate the ResNet18 network model and the BiLSTM network model; 5-4) Use the ResNet18 network model and BiLSTM network model generated by training to classify the verification data set U respectively; 5-5) Extract the k data with the highest and lowest classification confidence of the ResNet18 network model for the verification data set U and merge them into the training data set L2; 5-6) Extract the k data with the highest and lowest classification confidence of the BiLSTM network model for the verification data set U and merge them into the training data set L1; 5-7) Remove 2k data from the verification data set U; 5-8) Repeat steps 5-3) to 5-7) until the verification data set U is empty (that is, until all the data in the verification data set U is removed), at which point the network model training reaches the optimal level.

5. The radar emitter individual identification method based on collaborative learning framework as claimed in claim 1, characterized in that: The cross entropy loss function is expressed as follows: Among them, y(x i ) represents the true probability distribution, Represents the predicted probability distribution. Usually, the output layer of a multi-classification task will be activated by a softmax function, and then the cross entropy is used to characterize the distance between the two probability distributions.

6. The radar emitter individual identification method based on collaborative learning framework as claimed in claim 1, characterized in that: The RMS transfer algorithm is as follows: Take m samples {x1,x2L,x m }, where x i The corresponding true value is y i ; Compute the gradient: Among them, L represents the loss function of each sample, f(x i ; θ) represents the input x i The predicted value of the neural network when Cumulative squared gradient: r←ρr+(1-ρ)g☉g Update parameters: Among them, ε is the learning rate, set to 0.01; θ is the initial parameter, the initial value is 0.01; r is the initialization accumulator, the initial value is 0; ρ is the decay rate, set to 0.9; δ is the smoothing constant to prevent the divisor from being 0, set to 10 -6 .

Images