Radar emitter recognition method based on feature fusion and ensemble learning

Abstract

The present invention proposes a radar emitter recognition method based on feature fusion and ensemble learning, and the implementation steps are as follows: (1) Obtain a training sample set and a test sample set; (2) Construct an ensemble learning model; (3) Perform iterative training on the ensemble learning model; (4) Obtain the emitter recognition result. The present invention constructs an ensemble learning model including three base classifiers, and uses a stacked autoencoder to perform feature fusion on the pulse description word features and time-frequency distribution features of the radar emitter signal, solves the problems of incomplete feature utilization and low utilization rate in the prior art, and improves the recognition robustness and accuracy of radar emitter recognition in complex electromagnetic environments.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal processing, and relates to a method for identifying radar radiation sources, specifically to a method for identifying radar radiation sources based on feature fusion and ensemble learning, which can be applied to the field of target recognition in complex electromagnetic environments. Background Art

[0002] A radar radiation source is an electromagnetic wave emitting device used to generate detectable electromagnetic wave signals within a certain range so that a radar receiver can detect targets. A radar radiation source includes components such as an antenna, a radio frequency transmitter, and a modulator, and usually determines the target position and characteristics through the emitted electromagnetic waves and the signals reflected therefrom. Common radar radiation sources include civilian radars, military radars, etc., and are widely used in fields such as communication, navigation, monitoring, and reconnaissance.

[0003] Radar radiation source identification refers to the process of using radar technology to identify targets and determine information such as the type, category, quantity, and position of the targets. Different radar radiation sources have different characteristics. By analyzing radar signals, the characteristics of radar radiation sources can be identified, and then the types of radar radiation sources can be determined, which helps to effectively track and monitor targets and provides a reliable basis for subsequent identification of radiation source working modes, situation analysis, and decision-making.

[0004] Radar radiation source identification collects and analyzes signals in the electromagnetic space, extracts features from the signals, and classifies the received radar pulses from different radiation sources and identifies the radar models of each group of separated pulses by using the correlation of characteristic parameters of the same radiation source and the difference of characteristic parameters of signals from different radiation sources, which is a key link in completing target identification.

[0005] However, with the rapid development of electronic communication technologies and equipment, a large number of various electronic information devices are continuously put into use. In a narrow electromagnetic space range, there may be dozens of radiation sources simultaneously performing various tasks such as communication, navigation, tracking, and ranging under the interference of environmental clutter noise, and the receiver needs to simultaneously receive and process pulse signals from a large number of radiation sources. Moreover, due to the increasing demand for radar anti-interception and anti-jamming capabilities in modern times, a large number of advanced radar systems are put into use. These radars perform complex inter-pulse modulation or intra-pulse modulation on the emitted signals, such as parameter hopping, phase coding, etc. The complex modulation of signals makes the parameter domain aliasing of pulses more serious. Traditional radar radiation source identification technologies are difficult to adapt to this complex and intertwined electromagnetic environment, and there are problems such as too low sorting accuracy and robustness when facing a large number of radiation sources.

[0006] To solve the above problems, it is necessary to optimize the radar emitter recognition method, improve the adaptability of the algorithm to complex environments, and enhance the anti-interference ability of the algorithm to achieve accurate and stable radar emitter recognition in various environments. For example, a patent application with the publication number CN114358050A and the title "An Intelligent Radar Emitter Recognition Method Based on Bicubic Interpolation and WVD Feature Matrix" in 2022 discloses an intelligent radar emitter recognition method based on bicubic interpolation and WVD feature matrix. This method processes radar emitter signals through WVD transformation and bicubic interpolation method, constructs a WVD feature matrix, and uses a CNN model to train and identify the constructed matrix, realizing the recognition of 6 different intra-pulse modulation radar emitters, improving the accuracy and adaptability of radar emitter recognition and classification to a certain extent. However, this method only considers the time-frequency distribution information of the signal and cannot perform accurate recognition in a complex environment with a large number of emitters. Moreover, due to the use of a single neural network, when the network training falls into problems such as gradient disappearance and local optimum, it is easy to cause the decline of recognition accuracy and robustness. Summary of the Invention

[0007] The object of the present invention is to overcome the defects of the above-mentioned existing technologies, and propose a radar emitter recognition method based on feature fusion and ensemble learning to solve the technical problems of low recognition robustness and accuracy in the existing technologies when facing a complex electromagnetic environment with a large number of emitters and serious parameter domain aliasing.

[0008] To achieve the above object, the technical solutions adopted by the present invention include the following steps:

[0009] (1) Obtain a training sample set and a test sample set:

[0010] (1a) Obtain N pulse signals S = {s n |1≤n≤N} including C radar emitter categories and their corresponding pulse description words and labels, and preprocess each pulse signal to obtain time-frequency distribution features TF = {TF n |1≤n≤N}. At the same time, normalize the pulse description words corresponding to each pulse signal to obtain pulse description word features PDW = {PDW n |1≤n≤N}, where C≥30, N≥9000, PDW n and TF n respectively represent the pulse description word features and time-frequency distribution features corresponding to the nth pulse signal;

[0011] (1b) Select a total of R time-frequency distribution features {TF r |1≤r≤R} and pulse description word features {PDW r|1 ≤ r ≤ R} and the label {L r |1 ≤ r ≤ R} form the training sample set B train , and the remaining E time-frequency distribution features {TF e |1 ≤ e ≤ E} and the pulse descriptor words features {PDW e |1 ≤ e ≤ E} form the test sample set B test , where R > 0.5N, E = N - R, L r represents the label corresponding to the r-th pulse signal;

[0012] (2) Construct an ensemble learning model:

[0013] Construct an ensemble learning model O including a first base classifier O1, a second base classifier O2, and a third base classifier O3 arranged in parallel, and a decision maker O4 cascaded with the output ends of O1, O2, and O3; where O1 includes a first K-nearest neighbor classification model connected in sequence a first label mapping layer RLab1 and a second K-nearest neighbor classification model where the input end of which is also connected to the input end of; O2 includes a first convolutional neural network connected in sequence a second label mapping layer RLab2 and a second convolutional neural network where the input end of which is also connected to the input end of; O3 includes a stacked autoencoder DAE and a third K-nearest neighbor classification model KNN connected in sequence CFS , the output end of which is also connected to the input end of DAE; where and both include a first convolutional layer, an average pooling layer, a second convolutional layer, a max pooling layer, a flattening layer, and multiple fully connected layers connected in sequence; DAE includes multiple hidden layers stacked in sequence;

[0014] (3) Iteratively train the ensemble learning model O:

[0015] (3a) Initialize the iteration number as i, the maximum iteration number as I, I ≥ 400, and the network parameters of are θ T1 , θ T2 , the network parameter of DAE is θ D , and let i = 0;

[0016] (3b) Use the training sample set B train as the input of the ensemble learning model O for forward propagation:

[0017] (3b1) The first K-nearest neighbor classification model in the first base classifier O1 Classify the pulse description word feature PDW in each training sample r ; The first label mapping layer RLab1 is based on The class fuzzy vector obtained by classification Map the label L in each training sample r ; The second K-nearest neighbor classification model Based on the mapped label obtained by RLab1 Classify the pulse description word feature PDW in each training sample r to obtain a class fuzzy vector

[0018] (3b2) The first convolutional neural network in the second base classifier O2 Extract features and classify the time-frequency distribution feature TF in each training sample r separately to obtain the output feature Vet r and the class fuzzy vector The second label mapping layer RLab2 is based on Map the label L in each training sample r to obtain the mapped label The second convolutional neural network Based on the mapped label obtained by RLab2 Classify the time-frequency distribution feature TF in each training sample r to obtain a class fuzzy vector

[0019] (3b3) The stacked autoencoder DAE in the third base classifier O3 performs feature fusion on the pulse description word feature PDW in each training sample r and the Vet output by the first convolutional neural network r ; The third K-nearest neighbor classification model KNN CFS Classify the fused feature vector CFS obtained by DAE feature fusion r to obtain a class fuzzy vector

[0020] (3c) Adopt the cross-entropy loss function and calculate the loss value of the first convolutional neural network through r and L ; Calculate the loss value of the second convolutional neural network through and ; Calculate the loss value of the second convolutional neural network through through and L r Calculate the loss value of the stacked auto - encoder DAE Then, adopt the back - propagation method to pass through Calculate respectively the gradients of the network parameters of the DAE, and then adopt the gradient - descent algorithm to pass through the gradients of the network parameters of the DAE to respectively the network parameter θ T1 、 the network parameter θ T2 、the network parameter θ of the DAE D perform an update to obtain the ensemble learning model O of this iteration i ;

[0021] (3d) Judge whether i≥I holds. If so, obtain the trained ensemble learning model O * , otherwise, let i = i + 1, O i =O, and execute step (3b);

[0022] (4) Obtain the radiation source recognition result:

[0023] (4a) Use the test sample set B test as the input of the trained ensemble learning model O * . The first base classifier O1 classifies the pulse - description - word feature PDW in each test sample e to obtain the class - fuzzy vector The second base classifier O2 classifies the time - frequency distribution feature TF in each test sample e to obtain the output feature Vet e and the class - fuzzy vector The third base classifier O3 classifies the pulse - description - word feature PDW in each test sample e after fusing with the Vet output by the second base classifier O2 e to obtain the class - fuzzy vector

[0024] (4b) The discriminator O4 makes a decision on the class - fuzzy vectors in each test sample output by O1, O2, and O3 to obtain the radar radiation source recognition result Res = {Res e |1≤e≤E}.

[0025] Compared with the prior art, the present invention has the following advantages:

[0026] (1) The stacked autoencoder is included in the integrated learning model constructed by the present invention. During the process of training the model and obtaining the radar emitter recognition result, the stacked autoencoder realizes the effective fusion of pulse description word features and time-frequency distribution features, removes redundant features, fully excavates the discriminative features and potential connections for the target task in each feature, improves the utilization rate of features, and can effectively improve the accuracy of radar emitter recognition in a complex electromagnetic environment with a large number of emitters and severely overlapping parameter domains compared with the prior art.

[0027] (2) The integrated learning model constructed by the present invention includes three base classifiers and a decision maker. During the process of training the model and obtaining the radar emitter recognition result, the three base classifiers respectively complete the recognition of radar emitters based on pulse description word features, time-frequency distribution features and fusion features, and the decision maker jointly determines the recognition results of the three base classifiers to obtain the final radar emitter recognition result. Compared with the prior art, the integrated learning model realizes the multiple reuse of features, combines the advantages of different classification methods, and improves the accuracy and robustness of radar emitter recognition. Description of the Drawings

[0028] Figure 1 is the implementation flowchart of the present invention.

[0029] Figure 2 is the structural schematic diagram of the integrated learning model of the present invention.

[0030] Figure 3 is the structural schematic diagram of the convolutional neural network of the present invention.

[0031] Figure 4 is the radar emitter recognition result diagram of the embodiment of the present invention at a signal-to-noise ratio of -6 dB.

[0032] Figure 5 is the curve comparison diagram of the recognition accuracy between the present invention and the prior art embodiment. Detailed Embodiment

[0033] The present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0034] Refer to Figure 1 , the present invention includes the following steps:

[0035] Step 1) Obtain a training sample set and a test sample set:

[0036] (1a) Obtain N pulse signals S = {s including C radar emitter categories n{|1 ≤ n ≤ N} and its corresponding pulse description words and labels, and preprocess each pulse signal to obtain the time-frequency distribution feature TF = {TF n |1 ≤ n ≤ N}, and normalize the pulse description words corresponding to each pulse signal to obtain the pulse description word feature PDW = {PDW n |1 ≤ n ≤ N}, where C ≥ 30, N ≥ 9000, PDW n and TF n respectively represent the pulse description word feature and the time-frequency distribution feature corresponding to the nth pulse signal.

[0037] The implementation steps for preprocessing each pulse signal are as follows:

[0038] (1a1) Use the CWD time-frequency distribution to perform time-frequency transformation on each pulse signal s n to obtain the matrix CWD n , and reduce the matrix of CWD n and then perform grayscale processing to obtain the grayscale matrix G n ;

[0039] (1a2) Calculate the mean μ n of the elements in the grayscale matrix G B that are greater than the preset binarization threshold T More and the mean μ B of the elements that are less than T Less , and obtain the binarization threshold T B ' = (μ More + μ Less ) / 2 for this iteration;

[0040] (1a3) Determine whether |T B ' - T B | ≥ 0.001 holds. If so, let T B = T B ', and execute step (1a2); otherwise, perform binarization on the grayscale matrix G B according to T n to obtain the binary matrix Q n ; and perform an opening operation on Q n to obtain the opening matrix OP n , and then transform all the elements in the grayscale matrix G n according to OP n and Q n to obtain the time-frequency distribution feature TF = {TF n |1 ≤ n ≤ N}, where the element tf n in the x-th row and y-th column of the nth time-frequency distribution feature TF xy is calculated by the formula:

[0041]

[0042] Among them, T B is the binarization threshold, and g xy , q xy , op xy are the elements in the x-th row and y-th column of matrices G n , Q n , OP n respectively.

[0043] The above preprocessing has completed the denoising of the time-frequency diagram, retained the fine edge features of the time-frequency diagram, and weakened the adverse effects of noise on performance.

[0044] (1b) Select a total of R time-frequency distribution features {TF r |1 ≤ r ≤ R}, pulse description word features {PDW r |1 ≤ r ≤ R}, and labels {L r |1 ≤ r ≤ R} corresponding to more than half of the pulse signals in each radar emitter category to form a training sample set B train , and form a test sample set B e with the remaining E time-frequency distribution features {TF e |1 ≤ e ≤ E} and pulse description word features {PDW test |1 ≤ e ≤ E}, where R > 0.5N, E = N - R, and L r represents the label corresponding to the r-th pulse signal.

[0045] Step 2) Construct an ensemble learning model, whose structure is as Figure 2 shown:

[0046] Construct an ensemble learning model O including a first base classifier O1, a second base classifier O2, and a third base classifier O3 arranged in parallel, and a decision maker O4 cascaded with the output ends of O1, O2, and O3; where O1 includes a first K-nearest neighbor classification model connected in sequence, a first label mapping layer RLab1, and a second K-nearest neighbor classification model where the input end of is also connected to the input end of ; O2 includes a first convolutional neural network connected in sequence, a second label mapping layer RLab2, and a second convolutional neural network where the input end of CFS is also connected to the input end of DAE; the output end of is also connected to the input end of DAE; where and both include a first convolutional layer, an average pooling layer, a second convolutional layer, a max pooling layer, a flattening layer, and multiple fully connected layers in a sequential cascade; the DAE includes multiple hidden layers stacked in sequence;

[0047] wherein the first convolutional neural network and the second convolutional neural network are structured as Figure 3 shown, both including a first convolutional layer, an average pooling layer, a second convolutional layer, a max pooling layer, a flattening layer, and 4 fully connected layers in a sequential cascade. The specific structural parameters are as follows: the number of channels of the first convolutional layer and the second convolutional layer are 30 and 60 respectively, the kernel sizes of the convolutional kernels are 3*3 and 5*5 respectively, and the activation functions are both sigmoid; the kernel sizes of the max pooling layer and the average pooling layer are both 2*2, and the activation function is sigmoid; the number of nodes in the flattening layer is 61440, and the activation function is ReLU; the number of nodes in the 4 fully connected layers are 500, 5, 100, and 100 in sequence, and the activation functions are linear, ReLU, linear, and Softmax in sequence; the stacked autoencoder DAE includes 3 hidden layers, and the number of nodes in the hidden layers are 512, 4096, and 512 in sequence, and the activation functions are all sigmoid.

[0048] Step 3) Iteratively train the ensemble learning model O:

[0049] (3a) Initialize the iteration number as i, the maximum iteration number as I, I≥400, and the network parameters of T1 and T2 are θ D , respectively, the network parameter of the DAE is θ

[0050] (3b) Use the training sample set B train as the input of the ensemble learning model O for forward propagation:

[0051] (3b1) The first base classifier O1 classifies the pulse description word feature PDW r in each training sample:

[0052] (3b11) The first K-nearest neighbor classification model classifies the pulse description word feature PDW r in each training sample;

[0053] (3b12) The first label mapping layer RLab1 constructs a confusion matrix according to the class fuzzy vector obtained by classification and the label L r in each training sample and according to Construct the dominant classification set Set_KNN that meets the following conditions of advantage classification PDW ={type PDW (p)|1 ≤ p ≤ P}, and then according to the mapping relationship map the label L in each training sample r to obtain the mapped label

[0054]

[0055] where represents the probability that the a-th class in the classifier is judged as the b-th class, and Set_KNN PDW is composed of P disjoint class sets, and type PDW (p) represents the p-th class set among them, and σ PDW is the confusion threshold of RLab1.

[0056] The above processing of the class fuzzy vector obtained by the classification of and the completion of label mapping result in new class labels that can be accurately distinguished in the pulse descriptor feature space, making the classification result of the first base classifier more reliable and conducive to improving the recognition accuracy of the method.

[0057] (3b13) Second K-nearest neighbor classification model Classify the pulse descriptor feature PDW in each training sample according to the mapped label obtained by mapping RLab1 r to obtain the class fuzzy vector

[0058] (3b2) The second base classifier O2 classifies the time-frequency distribution feature TF in each training sample r as follows:

[0059] (3b21) First convolutional neural network Extract features and classify the time-frequency distribution feature TF in each training sample r separately, and obtain the output feature Vet r through the second fully connected layer and obtain the class fuzzy vector

[0060] through the 4th fully connected layer. The feature extraction, vectorization, and dimensionality reduction of the time-frequency distribution feature matrix are completed, providing conditions for the feature fusion of the third base classifier.

[0061] (3b22) The second label mapping layer RLab2 is based on and Lr Construct a confusion matrix and according to construct a dominant classification set Set_CNN that satisfies the following conditions TF ={type TF (t)|1≤t≤T}, and then according to the mapping relationship map the label L r in each training sample to obtain the mapped label

[0062]

[0063] where represents the probability that the a-th class in the classifier is judged as the b-th class, and Set_CNN TF is composed of T non-overlapping class sets, and type TF (t) represents the t-th class set among them, and σ TF is the confusion threshold of RLab2.

[0064] The above processing of the class fuzzy vector obtained by the classification of and the completion of label mapping result in new class labels that can be accurately distinguished in the time-frequency distribution feature space, making the classification result of the second base classifier more reliable and conducive to improving the recognition accuracy of the method.

[0065] (3b23) The second convolutional neural network classifies the time-frequency distribution feature TF in each training sample according to the mapped label obtained by the RLab2 mapping r to obtain the class fuzzy vector

[0066] (3b3) The stacked autoencoder DAE in the third base classifier O3 fuses the pulse descriptor feature PDW r in each training sample with the Vet output by the first convolutional neural network r ; The third K-nearest neighbor classification model KNN CFS classifies the fused feature vector CFS r obtained by the DAE feature fusion to obtain the class fuzzy vector

[0067] The above effectively fuses the pulse descriptor feature and the time-frequency distribution feature through the stacked autoencoder DAE, removes redundant features, fully excavates the discriminative features and potential connections for the task in both, and improves the utilization rate of features.

[0068] (3c) The cross-entropy loss function is adopted, and the loss value of the first convolutional neural network is calculated through and L r the first convolutional neural network ; Through and the loss value of the second convolutional neural network is calculated ; Through and L r the loss value of the stacked autoencoder DAE is calculated Then, the backpropagation method is adopted. Through calculate respectively the network parameter gradients of the DAE, and then adopt the gradient descent algorithm. Through the network parameter gradients of the DAE are used to the network parameter θ T1 , the network parameter θ T2 , the network parameter θ of the DAE D are updated to obtain the ensemble learning model O of this iteration i ;

[0069] (3d) Determine whether i≥I holds. If so, obtain the trained ensemble learning model O * , otherwise, set i = i + 1, O i = O, and execute step (3b);

[0070] Step 4) Obtain the radiation source recognition result:

[0071] (4a) Use the test sample set B test as the input of the trained ensemble learning model O * . The first base classifier O1 classifies the pulse description word feature PDW in each test sample e to obtain the class fuzzy vector The second base classifier O2 classifies the time-frequency distribution feature TF in each test sample e to obtain the output feature Vet e and the class fuzzy vector The third base classifier O3 classifies the pulse description word feature PDW in each test sample e after fusing the features with the Vet output by the second base classifier O2 e to obtain the class fuzzy vector

[0072] (4b) The discriminator O4 judges the class fuzzy vectors in each test sample output by O1, O2, and O3 Make a judgment, and the judgment steps are as follows:

[0073] (4b1) Calculate the subscripts idx corresponding to the maximum values in the category fuzzy vectors respectively, and obtain the predicted labels corresponding to idx PDW , idx TF , idx CFS by looking up the index table; here, P PDW , idx TF , idx CFS ∈ [1, C]; P e CFS ∈ [1, C];

[0074] (4b2) Judge whether or holds. If so, the radar emitter recognition result Res e = P e CFS . Otherwise, calculate the joint prediction label according to type PDW , type TF and execute step (4b3); and execute step (4b3);

[0075] (4b3) Judge whether and has only one element. If so, the radar emitter recognition result Otherwise, execute step (4b4);

[0076] (4b4) Judge whether holds. If so, the radar emitter recognition result Otherwise, execute step (4b5);

[0077] (4b5) Judge whether holds. If so, calculate the radar emitter prediction candidate vector and obtain the radar emitter recognition result through ; Otherwise, calculate the radar emitter prediction candidate vector and obtain the radar emitter recognition result through ;

[0078] where σ acc is the accuracy threshold of the discriminator O4, and type TF is a subset of the dominant classification set in the second label mapping layer RLab2.

[0079] The above joint determination of the results of the three base classifiers in the ensemble learning model enables the model to combine the advantages of different classifiers, adapt to different electromagnetic environments, and improve the robustness of the method.

[0080] The following further illustrates the technical effects of the present invention in combination with simulation experiments:

[0081] 1. Experimental conditions:

[0082] Hardware environment: The CPU is Inter(R)Xeon(R)CPU E3-1231 v3 with a main frequency of 3.40 GHz, 32.0 GB of memory, and a 64-bit operating system. Software environment: Microsoft windows 10 Professional Edition, MATLAB 2021 simulation software, Python 3.8.

[0083] 2. Simulation content:

[0084] Set 30 different modulated radiation sources, and the modulation parameters of the radiation sources are shown in the following table:

[0085] Frequency / MHz Pulse width / us Angle of arrival / ° Intrapulse modulation 1 Fixed 690 Jitter 2.8 - 3.2 Dither 63 - 66 BPSK 2 Linear 692 - 698 Jitter 3.8 - 4.2 Linear 63 - 66 Frank 3 Dither 700 - 710 Jitter 3.2 - 3.6 Dither 65 - 68 T2 4 Dither 700 - 710 Jitter 3.2 - 3.6 Dither 65 - 68 Frank 5 Fixed 690 Jitter 2.8 - 3.2 Dither 63 - 66 P2 6 Dither 700 - 720 Jitter 2.25 - 2.75 Dither 63 - 65 T4 7 Jitter 675 - 695 Jitter 4.15 - 4.65 Linear 61 - 64 P3 8 Linear 692 - 698 Jitter 3.8 - 4.2 Linear 63 - 66 T3 9 Jitter 682 - 688 Dither 3.3 - 3.7 Dither 63 - 66 T1 10 Dither 700 - 710 Jitter 3.2 - 3.6 Dither 65 - 68 T3 11 Jitter 675 - 695 Jitter 4.15 - 4.65 Linear 61 - 64 T2 12 Linear 692 - 698 Jitter 3.8 - 4.2 Linear 63 - 66 Costas 13 Jitter 712 - 718 Jump 2.65 - 3.15 Linear 65 - 68 T3 14 Jitter 722 - 728 Dither 3.2 - 3.6 Linear 65 - 68 T2 15 Jitter 722 - 728 Dither 3.2 - 3.6 Linear 65 - 68 T1 16 Dither 700 - 720 Jitter 2.25 - 2.75 Dither 63 - 65 Costas 17 Jitter 712 - 718 Jump 2.65 - 3.15 Linear 65 - 68 P3 18 Linear 692 - 698 Jitter 3.8 - 4.2 Linear 63 - 66 P4 19 Fixed 690 Jitter 2.8 - 3.2 Dither 63 - 66 T2 20 Linear 692 - 698 Jitter 3.8 - 4.2 Linear 63 - 66 P1 21 Jitter 712 - 718 Jump 2.65 - 3.15 Linear 65 - 68 T1 22 Linear 692 - 698 Jitter 3.8 - 4.2 Linear 63 - 66 BPSK 23 Dither 700 - 720 Jitter 2.25 - 2.75 Dither 63 - 65 P4 24 Jitter 682 - 688 Dither 3.3 - 3.7 Dither 63 - 66 T4 25 Jitter 722 - 728 Dither 3.2 - 3.6 Linear 65 - 68 P4 26 Jitter 722 - 728 Dither 3.2 - 3.6 Linear 65 - 68 P3 27 Jitter 675 - 695 Jitter 4.15 - 4.65 Linear 61 - 64 P1 28 Jitter 712 - 718 Jump 2.65 - 3.15 Linear 65 - 68 BPSK 29 Jitter 712 - 718 Jump 2.65 - 3.15 Linear 65 - 68 T4 30 Jitter 682 - 688 Dither 3.3 - 3.7 Dither 63 - 66 Frank

[0086] Generate more than 300 samples for each radiation source, and use 200 samples in each radiation source sample as the training set and the remaining 100 as the test set.

[0087] When the radiation source identification result in the simulation experiment of the present invention is the same as the label of the radiation source in the test set of the dataset, it is considered that the radiation source identification result is correct; when the radiation source identification result in the simulation experiment of the present invention is different from the label of the radiation source in the test set of the dataset, it is considered that the radiation source identification result is incorrect.

[0088] Simulation 1: Simulate the present invention at a signal-to-noise ratio of -6 dB, and the results are as Figure 4 shown.

[0089] Simulation 2: Simulation comparison of the present invention and the prior art in different signal-to-noise ratio environments, and the results are as Figure 5 shown.

[0090] 3. Analysis of simulation results:

[0091] Referring to Figure 4 , at a signal-to-noise ratio of -6 dB, the overall recognition accuracy of the present invention is 97.6%, and the accuracy of the recognition results of each radiation source exceeds 90%. The results show that the recognition accuracy of the present invention is high and it can effectively adapt to the complex electromagnetic environment with a large number of radiation sources.

[0092] Referring to Figure 5, with the horizontal axis being the environmental signal-to-noise ratio and the vertical axis being the recognition accuracy rate. Compared with the prior art, the present invention has a high accuracy rate in the electromagnetic environment with a signal-to-noise ratio ranging from -20 dB to 10 dB. In the electromagnetic environment with a signal-to-noise ratio above -4 dB, the recognition accuracy rate is stably above 99%. This shows that compared with the prior art, the present invention deeply excavates the information volume of multi-dimensional features, improves the accuracy of radiation source recognition, has better adaptability to complex environments, and has high robustness.

Claims

1. A radar emitter recognition method based on feature fusion and ensemble learning, characterized in that It includes the following steps: (1) Obtain a training sample set and a test sample set: (1a) Obtain N pulse signals S = {s n | 1 ≤ n ≤ N} including C categories of radar radiation sources, and their corresponding pulse description words and labels, and preprocess each pulse signal to obtain time-frequency distribution features TF = {TF n | 1 ≤ n ≤ N}, and at the same time normalize the pulse description words corresponding to each pulse signal to obtain pulse description word features PDW = {PDW n | 1 ≤ n ≤ N}, where C ≥ 30, N ≥ 9000, PDW n , TF n respectively represent the pulse description word features and time-frequency distribution features corresponding to the nth pulse signal; (1b) Select a total of R time-frequency distribution features {TF r | 1 ≤ r ≤ R}, pulse description word features {PDW r | 1 ≤ r ≤ R}, and labels {L r | 1 ≤ r ≤ R} corresponding to more than half of the pulse signals in each radar radiation source category to form a training sample set B train , and form a test sample set B test with the remaining E time-frequency distribution features {TF e | 1 ≤ e ≤ E} and pulse description word features {PDW e | 1 ≤ e ≤ E}, where R > 0.5N, E = N - R, and L r represents the label corresponding to the r-th pulse signal; (2) Construct an ensemble learning model: Construct an ensemble learning model O that includes a first base classifier O1, a second base classifier O2, and a third base classifier O3 arranged in parallel, and a decision maker O4 cascaded with the output ends of O1, O2, and O3; where O1 includes a first K-nearest neighbor classification model connected in sequence a first label mapping layer RLab1 and a second K-nearest neighbor classification model where the input end of which is also connected to the input end of; O2 includes a first convolutional neural network connected in sequence a second label mapping layer RLab2 and a second convolutional neural network where the input end of which is also connected to the input end of; O3 includes a stacked autoencoder DAE and a third K-nearest neighbor classification model KNN connected in sequence CFS , the output end of which is also connected to the input end of DAE; where and both include a first convolutional layer, an average pooling layer, a second convolutional layer, a max pooling layer, a flattening layer, and multiple fully connected layers connected in sequence; DAE includes multiple hidden layers stacked in sequence; (3) Iteratively train the ensemble learning model O: (3a) Initialize the iteration number as i and the maximum iteration number as I, where I ≥ 400, and the network parameters of T1 and T2 are θ D respectively, the network parameter of DAE is θ D , and let i = 0; (3b) Use the training sample set B train as the input of the ensemble learning model O for forward propagation: (3b1) The first K-nearest neighbor classification model in the first base classifier O1 Classify the pulse descriptor word feature PDW in each training sample r ; The first label mapping layer RLab1 is based on The class fuzzy vector obtained by classification Map the label L in each training sample r ; The second K-nearest neighbor classification model According to the mapped label obtained by RLab1 Classify the pulse descriptor word feature PDW in each training sample r To obtain a class fuzzy vector (3b2) The first convolutional neural network in the second base classifier O2 For the time-frequency distribution feature TF in each training sample r Feature extraction and classification are respectively performed to obtain the output feature Vet r And the class fuzzy vector The second label mapping layer RLab2 is based on For the label L in each training sample r Perform mapping to obtain the mapped label The second convolutional neural network According to the mapped label obtained by the RLab2 mapping For the time-frequency distribution feature TF in each training sample r Perform classification to obtain the class fuzzy vector (3b3) The stacked autoencoder DAE in the third base classifier O3 performs feature fusion on the pulse descriptor word features PDW in each training sample r with the first convolutional neural network output Vet r ; The third K-nearest neighbor classification model KNN CFS classifies the fused feature vector CFS obtained by DAE feature fusion r to obtain a class fuzzy vector (3c) The cross-entropy loss function is adopted, and the loss value of the first convolutional neural network and L r is calculated ; the loss value of the second convolutional neural network is calculated and ; the loss value of the stacked autoencoder DAE is calculated ; then, the backpropagation method is adopted, and the network parameter gradients of DAE are calculated and r respectively. Then, the gradient descent algorithm is used to update the network parameters θ of respectively, and the network parameters θ of and the network parameters θ of T1 , of T2 and the network parameters θ D of DAE to obtain the ensemble learning model O i for this iteration; (3d) Determine whether i≥I holds. If so, obtain the trained ensemble learning model O * , otherwise, set i = i + 1, O i = O, and execute step (3b); (4) Obtain the radiation source recognition result: (4a) Use the test sample set B test as the input of the trained ensemble learning model O * The first base classifier O1 classifies the pulse descriptor word feature PDW in each test sample e to obtain a class fuzzy vector The second base classifier O2 classifies the time-frequency distribution feature TF in each test sample e to obtain the output feature Vet e and a class fuzzy vector The third base classifier O3 classifies the pulse descriptor word feature PDW in each test sample e after fusing features with the Vet output by the second base classifier O2 e to obtain a class fuzzy vector (4b) The discriminator O4 makes a decision on the class fuzzy vectors in each test sample output by O1, O2, and O3, and obtains the radar emitter recognition result Res = {Res |1 ≤ e ≤ E}. e ​ 2. The radar emitter recognition method based on feature fusion and ensemble learning according to claim 1, wherein The preprocessing of each pulse signal described in step (1a) is implemented as follows: (1a1) When using the CWD time-frequency distribution, perform time-frequency transformation on each pulse signal s n to obtain the matrix CWD n , and after performing matrix scaling on CWD n and grayscaling it, obtain the grayscale matrix G n ; (1a2) Calculate the grayscale matrix G n The mean value μ of the elements greater than a pre-set binarization threshold T B and the mean value μ of the elements less than T More are obtained, and the binarization threshold T B ' of this iteration is obtained as T Less '=(μ B +μ More +μ Less ) / 2; (1a3) Determine |T B '-T B |≥0.001 holds. If so, let T B =T B ', and execute step (1a2); otherwise, according to T B perform binarization on the grayscale matrix G n to obtain the binary matrix Q n ; and perform an opening operation on Q n to obtain the opening matrix OP n , and then according to OP n and Q n perform transformation on all elements in the grayscale matrix G n to obtain the time-frequency distribution feature TF = {TF n |1 ≤ n ≤ N}, where the element tf n in the x-th row and y-th column of the n-th time-frequency distribution feature TF xy is calculated by the formula: Among them, T B is the binarization threshold, and g xy , q xy , op xy are the elements in the x-th row and y-th column of matrices G n , Q n , OP n respectively.

3. The radar emitter recognition method based on feature fusion and ensemble learning according to claim 1, wherein The ensemble learning model O described in step (2), where: The first convolutional neural network and the second convolutional neural network contain 4 fully connected layers; and The structural parameters are as follows: the number of channels of the first convolutional layer and the second convolutional layer are 30 and 60 respectively, the kernel sizes of the convolutional kernels are 3*3 and 5*5 respectively, and the activation functions are both sigmoid; the kernel sizes of the max pooling layer and the average pooling layer are both 2*2, and the activation function is sigmoid; the number of nodes in the flattening layer is 61440, and the activation function is ReLU; the number of nodes in the 4 fully connected layers are 500, 5, 100, and 100 in sequence, and the activation functions are linear, ReLU, linear, and Softmax in sequence; the number of hidden layers contained in the stacked autoencoder DAE is 3, and the number of nodes in the hidden layers are 512, 4096, and 512 in sequence, and the activation functions are all sigmoid.

4. The radar emitter recognition method based on feature fusion and ensemble learning according to claim 1, characterized in that The first label mapping layer RLab1 described in step (3b1) is based on the class fuzzy vector obtained by classification to map the label L in each training sample r The specific steps are as follows: The first label mapping layer RLab1 is based on and L r to construct a confusion matrix and based on to construct a dominant classification set Set_KNN that satisfies the following conditions PDW ={type PDW (p)|1 ≤ p ≤ P}, and then according to the mapping relationship map the label L in each training sample r to obtain the mapped label Among them denotes the probability that the a-th class in the classifier is judged as the b-th class, Set_KNN PDW is composed of P disjoint class sets, type PDW (p) represents the p-th class set among them, σ PDW is the confusion threshold of RLab1.

5. The radar emitter recognition method based on feature fusion and ensemble learning according to claim 1, characterized in that The first convolutional neural network described in step (3b2) extracts and classifies the time-frequency distribution features TF in each training sample r respectively, and the specific steps are as follows: The first convolutional layer performs convolution on each time-frequency distribution feature TF r ; the average pooling layer performs average pooling on the matrix r after TF convolution; the second convolutional layer performs convolution on the matrix after average pooling; the max pooling layer performs max pooling on the matrix after the second convolutional layer convolution; the flattening layer vectorizes the matrix after max pooling; the first fully connected layer maps the vector V vectorized by the flattening layer; the second fully connected layer maps the vector r after the first fully connected layer mapping to obtain the output feature Vet ; the third fully connected layer maps the output feature Vet r of the second fully connected layer; the fourth fully connected layer performs fuzzy probability on the class vector r after the third fully connected layer mapping to obtain the class fuzzy vector ; 6. The radar emitter recognition method based on feature fusion and ensemble learning according to claim 1, wherein The discriminator O4 in step (4b) makes a judgment on the class fuzzy vectors in each test sample output by O1, O2, and O3. The specific steps are as follows: (4b1) Calculate the category fuzzy vectors separately The subscript idx corresponding to the maximum value in PDW , idx TF , idx CFS , and by looking up the index table, obtain idx PDW , idx TF , idx CFS The corresponding predicted labels (4b2) Judgment Or Is established. If so, the radar radiation source recognition result Res e = P e CFS . Otherwise, according to type PDW , type TF Calculate the joint prediction label And execute step (4b3); (4b3) Judgment And Determine whether there is only one element. If so, the radar radiation source recognition result Otherwise, execute step (4b4); (4b4) Judge Whether it holds. If so, the radar emitter recognition result Otherwise, execute step (4b5); (4b5) Judge whether it holds. If so, calculate the predicted candidate vector of the radar radiation source and obtain the radar radiation source recognition result through Otherwise, calculate the predicted candidate vector of the radar radiation source and obtain the radar radiation source recognition result through ​​ Among them, σ acc is the accuracy threshold of the discriminator O4, and type TF is the subset of the dominant classification set in the second label mapping layer RLab2.

Images